Latent Space: The AI Engineer Podcast

From rewriting Google’s search stack in the early 2000s to reviving sparse trillion-parameter models and co-designing TPUs with frontier ML research, Jeff Dean has quietly shaped nearly every layer of the modern AI stack. As Chief AI Scientist at Google and a driving force behind Gemini, Jeff has lived through multiple scaling revolutions from CPUs and sharded indices to multimodal models that reason across text, video, and code.
Jeff joins us to unpack what it really means to “own the Pareto frontier,” why distillation is the engine behind every Flash model breakthrough, how energy (in picojoules) not FLOPs is becoming the true bottleneck, what it was like leading the charge to unify all of Google’s AI teams, and why the next leap won’t come from bigger context windows alone, but from systems that give the illusion of attending to trillions of tokens.
We discuss:
* Jeff’s early neural net thesis in 1990: parallel training before it was cool, why he believed scaling would win decades early, and the “bigger model, more data, better results” mantra that held for 15 years
* The evolution of Google Search: sharding, moving the entire index into memory in 2001, softening query semantics pre-LLMs, and why retrieval pipelines already resemble modern LLM systems
* Pareto frontier strategy: why you need both frontier “Pro” models and low-latency “Flash” models, and how distillation lets smaller models surpass prior generations
* Distillation deep dive: ensembles → compression → logits as soft supervision, and why you need the biggest model to make the smallest one good
* Latency as a first-class objective: why 10–50x lower latency changes UX entirely, and how future reasoning workloads will demand 10,000 tokens/sec
* Energy-based thinking: picojoules per bit, why moving data costs 1000x more than a multiply, batching through the lens of energy, and speculative decoding as amortization
* TPU co-design: predicting ML workloads 2–6 years out, speculative hardware features, precision reduction, sparsity, and the constant feedback loop between model architecture and silicon
* Sparse models and “outrageously large” networks: trillions of parameters with 1–5% activation, and why sparsity was always the right abstraction
* Unified vs. specialized models: abandoning symbolic systems, why general multimodal models tend to dominate vertical silos, and when vertical fine-tuning still makes sense
* Long context and the illusion of scale: beyond needle-in-a-haystack benchmarks toward systems that narrow trillions of tokens to 117 relevant documents
* Personalized AI: attending to your emails, photos, and documents (with permission), and why retrieval + reasoning will unlock deeply personal assistants
* Coding agents: 50 AI interns, crisp specifications as a new core skill, and how ultra-low latency will reshape human–agent collaboration
* Why ideas still matter: transformers, sparsity, RL, hardware, systems — scaling wasn’t blind; the pieces had to multiply together
Show Notes:
* Gemma 3 Paper
* Gemma 3
* Gemini 2.5 Report
* Jeff Dean’s “Software Engineering Advice from
Building Large-Scale Distributed Systems” Presentation (with Back of the Envelope Calculations)
* Latency Numbers Every Programmer Should Know by Jeff Dean
* The Jeff Dean Facts
* Jeff Dean Google Bio
* Jeff Dean on “Important AI Trends” @Stanford AI Club
* Jeff Dean & Noam Shazeer — 25 years at Google (Dwarkesh)
—
Jeff Dean
* LinkedIn: https://www.linkedin.com/in/jeff-dean-8b212555
* X: https://x.com/jeffdean
Google
* https://google.com
* https://deepmind.google

Full Video Episode
Timestamps
00:00:04 — Introduction: Alessio & Swyx welcome Jeff Dean, chief AI scientist at Google, to the Latent Space podcast00:00:30 — Owning the Pareto Frontier & balancing frontier vs low-latency models00:01:31 — Frontier models vs Flash models + role of distillation00:03:52 — History of distillation and its original motivation00:05:09 — Distillation’s role in modern model scaling00:07:02 — Model hierarchy (Flash, Pro, Ultra) and distillation sources00:07:46 — Flash model economics & wide deployment00:08:10 — Latency importance for complex tasks00:09:19 — Saturation of some tasks and future frontier tasks00:11:26 — On benchmarks, public vs internal00:12:53 — Example long-context benchmarks & limitations00:15:01 — Long-context goals: attending to trillions of tokens00:16:26 — Realistic use cases beyond pure language00:18:04 — Multimodal reasoning and non-text modalities00:19:05 — Importance of vision & motion modalities00:20:11 — Video understanding example (extracting structured info)00:20:47 — Search ranking analogy for LLM retrieval00:23:08 — LLM representations vs keyword search00:24:06 — Early Google search evolution & in-memory index00:26:47 — Design principles for scalable systems00:28:55 — Real-time index updates & recrawl strategies00:30:06 — Classic “Latency numbers every programmer should know”00:32:09 — Cost of memory vs compute and energy emphasis00:34:33 — TPUs & hardware trade-offs for serving models00:35:57 — TPU design decisions & co-design with ML00:38:06 — Adapting model architecture to hardware00:39:50 — Alternatives: energy-based models, speculative decoding00:42:21 — Open research directions: complex workflows, RL00:44:56 — Non-verifiable RL domains & model evaluation00:46:13 — Transition away from symbolic systems toward unified LLMs00:47:59 — Unified models vs specialized ones00:50:38 — Knowledge vs reasoning & retrieval + reasoning00:52:24 — Vertical model specialization & modules00:55:21 — Token count considerations for vertical domains00:56:09 — Low resource languages & contextual learning00:59:22 — Origins: Dean’s early neural network work01:10:07 — AI for coding & human–model interaction styles01:15:52 — Importance of crisp specification for coding agents01:19:23 — Prediction: personalized models & state retrieval01:22:36 — Token-per-second targets (10k+) and reasoning throughput01:23:20 — Episode conclusion and thanks
Transcript
Alessio Fanelli [00:00:04]: Hey everyone, welcome to the Latent Space podcast. This is Alessio, founder of Kernel Labs, and I’m joined by Swyx, editor of Latent Space.
Shawn Wang [00:00:11]: Hello, hello. We’re here in the studio with Jeff Dean, chief AI scientist at Google. Welcome. Thanks for having me. It’s a bit surreal to have you in the studio. I’ve watched so many of your talks, and obviously your career has been super legendary. So, I mean, congrats. I think the first thing must be said, congrats on owning the Pareto Frontier.
Jeff Dean [00:00:30]: Thank you, thank you. Pareto Frontiers are good. It’s good to be out there.
Shawn Wang [00:00:34]: Yeah, I mean, I think it’s a combination of both. You have to own the Pareto Frontier. You have to have like frontier capability, but also efficiency, and then offer that range of models that people like to use. And, you know, some part of this was started because of your hardware work. Some part of that is your model work, and I’m sure there’s lots of secret sauce that you guys have worked on cumulatively. But, like, it’s really impressive to see it all come together in, like, this slittily advanced.
Jeff Dean [00:01:04]: Yeah, yeah. I mean, I think, as you say, it’s not just one thing. It’s like a whole bunch of things up and down the stack. And, you know, all of those really combine to help make UNOS able to make highly capable large models, as well as, you know, software techniques to get those large model capabilities into much smaller, lighter weight models that are, you know, much more cost effective and lower latency, but still, you know, quite capable for their size. Yeah.
Alessio Fanelli [00:01:31]: How much pressure do you have on, like, having the lower bound of the Pareto Frontier, too? I think, like, the new labs are always trying to push the top performance frontier because they need to raise more money and all of that. And you guys have billions of users. And I think initially when you worked on the CPU, you were thinking about, you know, if everybody that used Google, we use the voice model for, like, three minutes a day, they were like, you need to double your CPU number. Like, what’s that discussion today at Google? Like, how do you prioritize frontier versus, like, we have to do this? How do we actually need to deploy it if we build it?
Jeff Dean [00:02:03]: Yeah, I mean, I think we always want to have models that are at the frontier or pushing the frontier because I think that’s where you see what capabilities now exist that didn’t exist at the sort of slightly less capable last year’s version or last six months ago version. At the same time, you know, we know those are going to be really useful for a bunch of use cases, but they’re going to be a bit slower and a bit more expensive than people might like for a bunch of other broader models. So I think what we want to do is always have kind of a highly capable sort of affordable model that enables a whole bunch of, you know, lower latency use cases. People can use them for agentic coding much more readily and then have the high-end, you know, frontier model that is really useful for, you know, deep reasoning, you know, solving really complicated math problems, those kinds of things. And it’s not that. One or the other is useful. They’re both useful. So I think we’d like to do both. And also, you know, through distillation, which is a key technique for making the smaller models more capable, you know, you have to have the frontier model in order to then distill it into your smaller model. So it’s not like an either or choice. You sort of need that in order to actually get a highly capable, more modest size model. Yeah.
Alessio Fanelli [00:03:24]: I mean, you and Jeffrey came up with the solution in 2014.
Jeff Dean [00:03:28]: Don’t forget, L’Oreal Vinyls as well. Yeah, yeah.
Alessio Fanelli [00:03:30]: A long time ago. But like, I’m curious how you think about the cycle of these ideas, even like, you know, sparse models and, you know, how do you reevaluate them? How do you think about in the next generation of model, what is worth revisiting? Like, yeah, they’re just kind of like, you know, you worked on so many ideas that end up being influential, but like in the moment, they might not feel that way necessarily. Yeah.
Jeff Dean [00:03:52]: I mean, I think distillation was originally motivated because we were seeing that we had a very large image data set at the time, you know, 300 million images that we could train on. And we were seeing that if you create specialists for different subsets of those image categories, you know, this one’s going to be really good at sort of mammals, and this one’s going to be really good at sort of indoor room scenes or whatever, and you can cluster those categories and train on an enriched stream of data after you do pre-training on a much broader set of images. You get much better performance. If you then treat that whole set of maybe 50 models you’ve trained as a large ensemble, but that’s not a very practical thing to serve, right? So distillation really came about from the idea of, okay, what if we want to actually serve that and train all these independent sort of expert models and then squish it into something that actually fits in a form factor that you can actually serve? And that’s, you know, not that different from what we’re doing today. You know, often today we’re instead of having an ensemble of 50 models. We’re having a much larger scale model that we then distill into a much smaller scale model.
Shawn Wang [00:05:09]: Yeah. A part of me also wonders if distillation also has a story with the RL revolution. So let me maybe try to articulate what I mean by that, which is you can, RL basically spikes models in a certain part of the distribution. And then you have to sort of, well, you can spike models, but usually sometimes... It might be lossy in other areas and it’s kind of like an uneven technique, but you can probably distill it back and you can, I think that the sort of general dream is to be able to advance capabilities without regressing on anything else. And I think like that, that whole capability merging without loss, I feel like it’s like, you know, some part of that should be a distillation process, but I can’t quite articulate it. I haven’t seen much papers about it.
Jeff Dean [00:06:01]: Yeah, I mean, I tend to think of one of the key advantages of distillation is that you can have a much smaller model and you can have a very large, you know, training data set and you can get utility out of making many passes over that data set because you’re now getting the logits from the much larger model in order to sort of coax the right behavior out of the smaller model that you wouldn’t otherwise get with just the hard labels. And so, you know, I think that’s what we’ve observed. Is you can get, you know, very close to your largest model performance with distillation approaches. And that seems to be, you know, a nice sweet spot for a lot of people because it enables us to kind of, for multiple Gemini generations now, we’ve been able to make the sort of flash version of the next generation as good or even substantially better than the previous generations pro. And I think we’re going to keep trying to do that because that seems like a good trend to follow.
Shawn Wang [00:07:02]: So, Dara asked, so it was the original map was Flash Pro and Ultra. Are you just sitting on Ultra and distilling from that? Is that like the mother load?
Jeff Dean [00:07:12]: I mean, we have a lot of different kinds of models. Some are internal ones that are not necessarily meant to be released or served. Some are, you know, our pro scale model and we can distill from that as well into our Flash scale model. So I think, you know, it’s an important set of capabilities to have and also inference time scaling. It can also be a useful thing to improve the capabilities of the model.
Shawn Wang [00:07:35]: And yeah, yeah, cool. Yeah. And obviously, I think the economy of Flash is what led to the total dominance. I think the latest number is like 50 trillion tokens. I don’t know. I mean, obviously, it’s changing every day.
Jeff Dean [00:07:46]: Yeah, yeah. But, you know, by market share, hopefully up.
Shawn Wang [00:07:50]: No, I mean, there’s no I mean, there’s just the economics wise, like because Flash is so economical, like you can use it for everything. Like it’s in Gmail now. It’s in YouTube. Like it’s yeah. It’s in everything.
Jeff Dean [00:08:02]: We’re using it more in our search products of various AI mode reviews.
Shawn Wang [00:08:05]: Oh, my God. Flash past the AI mode. Oh, my God. Yeah, that’s yeah, I didn’t even think about that.
Jeff Dean [00:08:10]: I mean, I think one of the things that is quite nice about the Flash model is not only is it more affordable, it’s also a lower latency. And I think latency is actually a pretty important characteristic for these models because we’re going to want models to do much more complicated things that are going to involve, you know, generating many more tokens from when you ask the model to do so. So, you know, if you’re going to ask the model to do something until it actually finishes what you ask it to do, because you’re going to ask now, not just write me a for loop, but like write me a whole software package to do X or Y or Z. And so having low latency systems that can do that seems really important. And Flash is one direction, one way of doing that. You know, obviously our hardware platforms enable a bunch of interesting aspects of our, you know, serving stack as well, like TPUs, the interconnect between. Chips on the TPUs is actually quite, quite high performance and quite amenable to, for example, long context kind of attention operations, you know, having sparse models with lots of experts. These kinds of things really, really matter a lot in terms of how do you make them servable at scale.
Alessio Fanelli [00:09:19]: Yeah. Does it feel like there’s some breaking point for like the proto Flash distillation, kind of like one generation delayed? I almost think about almost like the capability as a. In certain tasks, like the pro model today is a saturated, some sort of task. So next generation, that same task will be saturated at the Flash price point. And I think for most of the things that people use models for at some point, the Flash model in two generation will be able to do basically everything. And how do you make it economical to like keep pushing the pro frontier when a lot of the population will be okay with the Flash model? I’m curious how you think about that.
Jeff Dean [00:09:59]: I mean, I think that’s true. If your distribution of what people are asking people, the models to do is stationary, right? But I think what often happens is as the models become more capable, people ask them to do more, right? So, I mean, I think this happens in my own usage. Like I used to try our models a year ago for some sort of coding task, and it was okay at some simpler things, but wouldn’t do work very well for more complicated things. And since then, we’ve improved dramatically on the more complicated coding tasks. And now I’ll ask it to do much more complicated things. And I think that’s true, not just of coding, but of, you know, now, you know, can you analyze all the, you know, renewable energy deployments in the world and give me a report on solar panel deployment or whatever. That’s a very complicated, you know, more complicated task than people would have asked a year ago. And so you are going to want more capable models to push the frontier in the absence of what people ask the models to do. And that also then gives us. Insight into, okay, where does the, where do things break down? How can we improve the model in these, these particular areas, uh, in order to sort of, um, make the next generation even better.
Alessio Fanelli [00:11:11]: Yeah. Are there any benchmarks or like test sets they use internally? Because it’s almost like the same benchmarks get reported every time. And it’s like, all right, it’s like 99 instead of 97. Like, how do you have to keep pushing the team internally to it? Or like, this is what we’re building towards. Yeah.
Jeff Dean [00:11:26]: I mean, I think. Benchmarks, particularly external ones that are publicly available. Have their utility, but they often kind of have a lifespan of utility where they’re introduced and maybe they’re quite hard for current models. You know, I, I like to think of the best kinds of benchmarks are ones where the initial scores are like 10 to 20 or 30%, maybe, but not higher. And then you can sort of work on improving that capability for, uh, whatever it is, the benchmark is trying to assess and get it up to like 80, 90%, whatever. I, I think once it hits kind of 95% or something, you get very diminishing returns from really focusing on that benchmark, cuz it’s sort of, it’s either the case that you’ve now achieved that capability, or there’s also the issue of leakage in public data or very related kind of data being, being in your training data. Um, so we have a bunch of held out internal benchmarks that we really look at where we know that wasn’t represented in the training data at all. There are capabilities that we want the model to have. Um, yeah. Yeah. Um, that it doesn’t have now, and then we can work on, you know, assessing, you know, how do we make the model better at these kinds of things? Is it, we need different kind of data to train on that’s more specialized for this particular kind of task. Do we need, um, you know, a bunch of, uh, you know, architectural improvements or some sort of, uh, model capability improvements, you know, what would help make that better?
Shawn Wang [00:12:53]: Is there, is there such an example that you, uh, a benchmark inspired in architectural improvement? Like, uh, I’m just kind of. Jumping on that because you just.
Jeff Dean [00:13:02]: Uh, I mean, I think some of the long context capability of the, of the Gemini models that came, I guess, first in 1.5 really were about looking at, okay, we want to have, um, you know,
Shawn Wang [00:13:15]: immediately everyone jumped to like completely green charts of like, everyone had, I was like, how did everyone crack this at the same time? Right. Yeah. Yeah.
Jeff Dean [00:13:23]: I mean, I think, um, and once you’re set, I mean, as you say that needed single needle and a half. Hey, stack benchmark is really saturated for at least context links up to 1, 2 and K or something. Don’t actually have, you know, much larger than 1, 2 and 8 K these days or two or something. We’re trying to push the frontier of 1 million or 2 million context, which is good because I think there are a lot of use cases where. Yeah. You know, putting a thousand pages of text or putting, you know, multiple hour long videos and the context and then actually being able to make use of that as useful. Try to, to explore the über graduation are fairly large. But the single needle in a haystack benchmark is sort of saturated. So you really want more complicated, sort of multi-needle or more realistic, take all this content and produce this kind of answer from a long context that sort of better assesses what it is people really want to do with long context. Which is not just, you know, can you tell me the product number for this particular thing?
Shawn Wang [00:14:31]: Yeah, it’s retrieval. It’s retrieval within machine learning. It’s interesting because I think the more meta level I’m trying to operate at here is you have a benchmark. You’re like, okay, I see the architectural thing I need to do in order to go fix that. But should you do it? Because sometimes that’s an inductive bias, basically. It’s what Jason Wei, who used to work at Google, would say. Exactly the kind of thing. Yeah, you’re going to win. Short term. Longer term, I don’t know if that’s going to scale. You might have to undo that.
Jeff Dean [00:15:01]: I mean, I like to sort of not focus on exactly what solution we’re going to derive, but what capability would you want? And I think we’re very convinced that, you know, long context is useful, but it’s way too short today. Right? Like, I think what you would really want is, can I attend to the internet while I answer my question? Right? But that’s not going to happen. I think that’s going to be solved by purely scaling the existing solutions, which are quadratic. So a million tokens kind of pushes what you can do. You’re not going to do that to a trillion tokens, let alone, you know, a billion tokens, let alone a trillion. But I think if you could give the illusion that you can attend to trillions of tokens, that would be amazing. You’d find all kinds of uses for that. You would have attend to the internet. You could attend to the pixels of YouTube and the sort of deeper representations that we can find. You could attend to the form for a single video, but across many videos, you know, on a personal Gemini level, you could attend to all of your personal state with your permission. So like your emails, your photos, your docs, your plane tickets you have. I think that would be really, really useful. And the question is, how do you get algorithmic improvements and system level improvements that get you to something where you actually can attend to trillions of tokens? Right. In a meaningful way. Yeah.
Shawn Wang [00:16:26]: But by the way, I think I did some math and it’s like, if you spoke all day, every day for eight hours a day, you only generate a maximum of like a hundred K tokens, which like very comfortably fits.
Jeff Dean [00:16:38]: Right. But if you then say, okay, I want to be able to understand everything people are putting on videos.
Shawn Wang [00:16:46]: Well, also, I think that the classic example is you start going beyond language into like proteins and whatever else is extremely information dense. Yeah. Yeah.
Jeff Dean [00:16:55]: I mean, I think one of the things about Gemini’s multimodal aspects is we’ve always wanted it to be multimodal from the start. And so, you know, that sometimes to people means text and images and video sort of human-like and audio, audio, human-like modalities. But I think it’s also really useful to have Gemini know about non-human modalities. Yeah. Like LIDAR sensor data from. Yes. Say, Waymo vehicles or. Like robots or, you know, various kinds of health modalities, x-rays and MRIs and imaging and genomics information. And I think there’s probably hundreds of modalities of data where you’d like the model to be able to at least be exposed to the fact that this is an interesting modality and has certain meaning in the world. Where even if you haven’t trained on all the LIDAR data or MRI data, you could have, because maybe that’s not, you know, it doesn’t make sense in terms of trade-offs of. You know, what you include in your main pre-training data mix, at least including a little bit of it is actually quite useful. Yeah. Because it sort of tempts the model that this is a thing.
Shawn Wang [00:18:04]: Yeah. Do you believe, I mean, since we’re on this topic and something I just get to ask you all the questions I always wanted to ask, which is fantastic. Like, are there some king modalities, like modalities that supersede all the other modalities? So a simple example was Vision can, on a pixel level, encode text. And DeepSeq had this DeepSeq CR paper that did that. Vision. And Vision has also been shown to maybe incorporate audio because you can do audio spectrograms and that’s, that’s also like a Vision capable thing. Like, so, so maybe Vision is just the king modality and like. Yeah.
Jeff Dean [00:18:36]: I mean, Vision and Motion are quite important things, right? Motion. Well, like video as opposed to static images, because I mean, there’s a reason evolution has evolved eyes like 23 independent ways, because it’s such a useful capability for sensing the world around you, which is really what we want these models to be. So I think the only thing that we can be able to do is interpret the things we’re seeing or the things we’re paying attention to and then help us in using that information to do things. Yeah.
Shawn Wang [00:19:05]: I think motion, you know, I still want to shout out, I think Gemini, still the only native video understanding model that’s out there. So I use it for YouTube all the time. Nice.
Jeff Dean [00:19:15]: Yeah. Yeah. I mean, it’s actually, I think people kind of are not necessarily aware of what the Gemini models can actually do. Yeah. Like I have an example I’ve used in one of my talks. It had like, it was like a YouTube highlight video of 18 memorable sports moments across the last 20 years or something. So it has like Michael Jordan hitting some jump shot at the end of the finals and, you know, some soccer goals and things like that. And you can literally just give it the video and say, can you please make me a table of what all these different events are? What when the date is when they happened? And a short description. And so you get like now an 18 row table of that information extracted from the video, which is, you know, not something most people think of as like a turn video into sequel like table.
Alessio Fanelli [00:20:11]: Has there been any discussion inside of Google of like, you mentioned tending to the whole internet, right? Google, it’s almost built because a human cannot tend to the whole internet and you need some sort of ranking to find what you need. Yep. That ranking is like much different for an LLM because you can expect a person to look at maybe the first five, six links in a Google search versus for an LLM. Should you expect to have 20 links that are highly relevant? Like how do you internally figure out, you know, how do we build the AI mode that is like maybe like much broader search and span versus like the more human one? Yeah.
Jeff Dean [00:20:47]: I mean, I think even pre-language model based work, you know, our ranking systems would be built to start. I mean, I think even pre-language model based work, you know, our ranking systems would be built to start. With a giant number of web pages in our index, many of them are not relevant. So you identify a subset of them that are relevant with very lightweight kinds of methods. You know, you’re down to like 30,000 documents or something. And then you gradually refine that to apply more and more sophisticated algorithms and more and more sophisticated sort of signals of various kinds in order to get down to ultimately what you show, which is, you know, the final 10 results or, you know, 10 results plus. Other kinds of information. And I think an LLM based system is not going to be that dissimilar, right? You’re going to attend to trillions of tokens, but you’re going to want to identify, you know, what are the 30,000 ish documents that are with the, you know, maybe 30 million interesting tokens. And then how do you go from that into what are the 117 documents I really should be paying attention to in order to carry out the tasks that the user has asked? And I think, you know, you can imagine systems where you have, you know, a lot of highly parallel processing to identify those initial 30,000 candidates, maybe with very lightweight kinds of models. Then you have some system that sort of helps you narrow down from 30,000 to the 117 with maybe a little bit more sophisticated model or set of models. And then maybe the final model is the thing that looks. So the 117 things that might be your most capable model. So I think it has to, it’s going to be some system like that, that is really enables you to give the illusion of attending to trillions of tokens. Sort of the way Google search gives you, you know, not the illusion, but you are searching the internet, but you’re finding, you know, a very small subset of things that are, that are relevant.
Shawn Wang [00:22:47]: Yeah. I often tell a lot of people that are not steeped in like Google search history that, well, you know, like Bert was. Like he was like basically immediately inside of Google search and that improves results a lot, right? Like I don’t, I don’t have any numbers off the top of my head, but like, I’m sure you guys, that’s obviously the most important numbers to Google. Yeah.
Jeff Dean [00:23:08]: I mean, I think going to an LLM based representation of text and words and so on enables you to get out of the explicit hard notion of, of particular words having to be on the page, but really getting at the notion of this topic of this page or this page. Paragraph is highly relevant to this query. Yeah.
Shawn Wang [00:23:28]: I don’t think people understand how much LLMs have taken over all these very high traffic system, very high traffic. Yeah. Like it’s Google, it’s YouTube. YouTube has this like semantics ID thing where it’s just like every token or every item in the vocab is a YouTube video or something that predicts the video using a code book, which is absurd to me for YouTube size.
Jeff Dean [00:23:50]: And then most recently GROK also for, for XAI, which is like, yeah. I mean, I’ll call out even before LLMs were used extensively in search, we put a lot of emphasis on softening the notion of what the user actually entered into the query.
Shawn Wang [00:24:06]: So do you have like a history of like, what’s the progression? Oh yeah.
Jeff Dean [00:24:09]: I mean, I actually gave a talk in, uh, I guess, uh, web search and data mining conference in 2009, uh, where we never actually published any papers about the origins of Google search, uh, sort of, but we went through sort of four or five or six. generations, four or five or six generations of, uh, redesigning of the search and retrieval system, uh, from about 1999 through 2004 or five. And that talk is really about that evolution. And one of the things that really happened in 2001 was we were sort of working to scale the system in multiple dimensions. So one is we wanted to make our index bigger, so we could retrieve from a larger index, which always helps your quality in general. Uh, because if you don’t have the page in your index, you’re going to not do well. Um, and then we also needed to scale our capacity because we were, our traffic was growing quite extensively. Um, and so we had, you know, a sharded system where you have more and more shards as the index grows, you have like 30 shards. And then if you want to double the index size, you make 60 shards so that you can bound the latency by which you respond for any particular user query. Um, and then as traffic grows, you add, you add more and more replicas of each of those. And so we eventually did the math that realized that in a data center where we had say 60 shards and, um, you know, 20 copies of each shard, we now had 1200 machines, uh, with disks. And we did the math and we’re like, Hey, one copy of that index would actually fit in memory across 1200 machines. So in 2001, we introduced, uh, we put our entire index in memory and what that enabled from a quality perspective was amazing. Um, and so we had more and more replicas of each of those. Before you had to be really careful about, you know, how many different terms you looked at for a query, because every one of them would involve a disk seek on every one of the 60 shards. And so you, as you make your index bigger, that becomes even more inefficient. But once you have the whole index in memory, it’s totally fine to have 50 terms you throw into the query from the user’s original three or four word query, because now you can add synonyms like restaurant and restaurants and cafe and, uh, you know, things like that. Uh, bistro and all these things. And you can suddenly start, uh, sort of really, uh, getting at the meaning of the word as opposed to the exact semantic form the user typed in. And that was, you know, 2001, very much pre LLM, but really it was about softening the, the strict definition of what the user typed in order to get at the meaning.
Alessio Fanelli [00:26:47]: What are like principles that you use to like design the systems, especially when you have, I mean, in 2001, the internet is like. Doubling, tripling every year in size is not like, uh, you know, and I think today you kind of see that with LLMs too, where like every year the jumps in size and like capabilities are just so big. Are there just any, you know, principles that you use to like, think about this? Yeah.
Jeff Dean [00:27:08]: I mean, I think, uh, you know, first, whenever you’re designing a system, you want to understand what are the sort of design parameters that are going to be most important in designing that, you know? So, you know, how many queries per second do you need to handle? How big is the internet? How big is the index you need to handle? How much data do you need to keep for every document in the index? How are you going to look at it when you retrieve things? Um, what happens if traffic were to double or triple, you know, will that system work well? And I think a good design principle is you’re going to want to design a system so that the most important characteristics could scale by like factors of five or 10, but probably not beyond that because often what happens is if you design a system for X. And something suddenly becomes a hundred X, that would enable a very different point in the design space that would not make sense at X. But all of a sudden at a hundred X makes total sense. So like going from a disk space index to a in memory index makes a lot of sense once you have enough traffic, because now you have enough replicas of the sort of state on disk that those machines now actually can hold, uh, you know, a full copy of the, uh, index and memory. Yeah. And that all of a sudden enabled. A completely different design that wouldn’t have been practical before. Yeah. Um, so I’m, I’m a big fan of thinking through designs in your head, just kind of playing with the design space a little before you actually do a lot of writing of code. But, you know, as you said, in the early days of Google, we were growing the index, uh, quite extensively. We were growing the update rate of the index. So the update rate actually is the parameter that changed the most. Surprising. So it used to be once a month.
Shawn Wang [00:28:55]: Yeah.
Jeff Dean [00:28:56]: And then we went to a system that could update any particular page in like sub one minute. Okay.
Shawn Wang [00:29:02]: Yeah. Because this is a competitive advantage, right?
Jeff Dean [00:29:04]: Because all of a sudden news related queries, you know, if you’re, if you’ve got last month’s news index, it’s not actually that useful for.
Shawn Wang [00:29:11]: News is a special beast. Was there any, like you could have split it onto a separate system.
Jeff Dean [00:29:15]: Well, we did. We launched a Google news product, but you also want news related queries that people type into the main index to also be sort of updated.
Shawn Wang [00:29:23]: So, yeah, it’s interesting. And then you have to like classify whether the page is, you have to decide which pages should be updated and what frequency. Oh yeah.
Jeff Dean [00:29:30]: There’s a whole like, uh, system behind the scenes that’s trying to decide update rates and importance of the pages. So even if the update rate seems low, you might still want to recrawl important pages quite often because, uh, the likelihood they change might be low, but the value of having updated is high.
Shawn Wang [00:29:50]: Yeah, yeah, yeah, yeah. Uh, well, you know, yeah. This, uh, you know, mention of latency and, and saving things to this reminds me of one of your classics, which I have to bring up, which is latency numbers. Every programmer should know, uh, was there a, was it just a, just a general story behind that? Did you like just write it down?
Jeff Dean [00:30:06]: I mean, this has like sort of eight or 10 different kinds of metrics that are like, how long does a cache mistake? How long does branch mispredict take? How long does a reference domain memory take? How long does it take to send, you know, a packet from the U S to the Netherlands or something? Um,
Shawn Wang [00:30:21]: why Netherlands, by the way, or is it, is that because of Chrome?
Jeff Dean [00:30:25]: Uh, we had a data center in the Netherlands, um, so, I mean, I think this gets to the point of being able to do the back of the envelope calculations. So these are sort of the raw ingredients of those, and you can use them to say, okay, well, if I need to design a system to do image search and thumb nailing or something of the result page, you know, how, what I do that I could pre-compute the image thumbnails. I could like. Try to thumbnail them on the fly from the larger images. What would that do? How much dis bandwidth than I need? How many des seeks would I do? Um, and you can sort of actually do thought experiments in, you know, 30 seconds or a minute with the sort of, uh, basic, uh, basic numbers at your fingertips. Uh, and then as you sort of build software using higher level libraries, you kind of want to develop the same intuitions for how long does it take to, you know, look up something in this particular kind of.
Shawn Wang [00:31:21]: I’ll see you next time.
Shawn Wang [00:31:51]: Which is a simple byte conversion. That’s nothing interesting. I wonder if you have any, if you were to update your...
Jeff Dean [00:31:58]: I mean, I think it’s really good to think about calculations you’re doing in a model, either for training or inference.
Jeff Dean [00:32:09]: Often a good way to view that is how much state will you need to bring in from memory, either like on-chip SRAM or HBM from the accelerator. Attached memory or DRAM or over the network. And then how expensive is that data motion relative to the cost of, say, an actual multiply in the matrix multiply unit? And that cost is actually really, really low, right? Because it’s order, depending on your precision, I think it’s like sub one picodule.
Shawn Wang [00:32:50]: Oh, okay. You measure it by energy. Yeah. Yeah.
Jeff Dean [00:32:52]: Yeah. I mean, it’s all going to be about energy and how do you make the most energy efficient system. And then moving data from the SRAM on the other side of the chip, not even off the off chip, but on the other side of the same chip can be, you know, a thousand picodules. Oh, yeah. And so all of a sudden, this is why your accelerators require batching. Because if you move, like, say, the parameter of a model from SRAM on the, on the chip into the multiplier unit, that’s going to cost you a thousand picodules. So you better make use of that, that thing that you moved many, many times with. So that’s where the batch dimension comes in. Because all of a sudden, you know, if you have a batch of 256 or something, that’s not so bad. But if you have a batch of one, that’s really not good.
Shawn Wang [00:33:40]: Yeah. Yeah. Right.
Jeff Dean [00:33:41]: Because then you paid a thousand picodules in order to do your one picodule multiply.
Shawn Wang [00:33:46]: I have never heard an energy-based analysis of batching.
Jeff Dean [00:33:50]: Yeah. I mean, that’s why people batch. Yeah. Ideally, you’d like to use batch size one because the latency would be great.
Shawn Wang [00:33:56]: The best latency.
Jeff Dean [00:33:56]: But the energy cost and the compute cost inefficiency that you get is quite large. So, yeah.
Shawn Wang [00:34:04]: Is there a similar trick like, like, like you did with, you know, putting everything in memory? Like, you know, I think obviously NVIDIA has caused a lot of waves with betting very hard on SRAM with Grok. I wonder if, like, that’s something that you already saw with, with the TPUs, right? Like that, that you had to. Uh, to serve at your scale, uh, you probably sort of saw that coming. Like what, what, what hardware, uh, innovations or insights were formed because of what you’re seeing there?
Jeff Dean [00:34:33]: Yeah. I mean, I think, you know, TPUs have this nice, uh, sort of regular structure of 2D or 3D meshes with a bunch of chips connected. Yeah. And each one of those has HBM attached. Um, I think for serving some kinds of models, uh, you know, you, you pay a lot higher cost. Uh, and time latency, um, bringing things in from HBM than you do bringing them in from, uh, SRAM on the chip. So if you have a small enough model, you can actually do model parallelism, spread it out over lots of chips and you actually get quite good throughput improvements and latency improvements from doing that. And so you’re now sort of striping your smallish scale model over say 16 or 64 chips. Uh, but as if you do that and it all fits in. In SRAM, uh, that can be a big win. So yeah, that’s not a surprise, but it is a good technique.
Alessio Fanelli [00:35:27]: Yeah. What about the TPU design? Like how much do you decide where the improvements have to go? So like, this is like a good example of like, is there a way to bring the thousand picojoules down to 50? Like, is it worth designing a new chip to do that? The extreme is like when people say, oh, you should burn the model on the ASIC and that’s kind of like the most extreme thing. How much of it? Is it worth doing an hardware when things change so quickly? Like what was the internal discussion? Yeah.
Jeff Dean [00:35:57]: I mean, we, we have a lot of interaction between say the TPU chip design architecture team and the sort of higher level modeling, uh, experts, because you really want to take advantage of being able to co-design what should future TPUs look like based on where we think the sort of ML research puck is going, uh, in some sense, because, uh, you know, as a hardware designer for ML and in particular, you’re trying to design a chip starting today and that design might take two years before it even lands in a data center. And then it has to sort of be a reasonable lifetime of the chip to take you three, four or five years. So you’re trying to predict two to six years out where, what ML computations will people want to run two to six years out in a very fast changing field. And so having people with interest. Interesting ML research ideas of things we think will start to work in that timeframe or will be more important in that timeframe, uh, really enables us to then get, you know, interesting hardware features put into, you know, TPU N plus two, where TPU N is what we have today.
Shawn Wang [00:37:10]: Oh, the cycle time is plus two.
Jeff Dean [00:37:12]: Roughly. Wow. Because, uh, I mean, sometimes you can squeeze some changes into N plus one, but, you know, bigger changes are going to require the chip. Yeah. Design be earlier in its lifetime design process. Um, so whenever we can do that, it’s generally good. And sometimes you can put in speculative features that maybe won’t cost you much chip area, but if it works out, it would make something, you know, 10 times as fast. And if it doesn’t work out, well, you burned a little bit of tiny amount of your chip area on that thing, but it’s not that big a deal. Uh, sometimes it’s a very big change and we want to be pretty sure this is going to work out. So we’ll do like lots of carefulness. Uh, ML experimentation to show us, uh, this is actually the, the way we want to go. Yeah.
Alessio Fanelli [00:37:58]: Is there a reverse of like, we already committed to this chip design so we can not take the model architecture that way because it doesn’t quite fit?
Jeff Dean [00:38:06]: Yeah. I mean, you, you definitely have things where you’re going to adapt what the model architecture looks like so that they’re efficient on the chips that you’re going to have for both training and inference of that, of that, uh, generation of model. So I think it kind of goes both ways. Um, you know, sometimes you can take advantage of, you know, lower precision things that are coming in a future generation. So you can, might train it at that lower precision, even if the current generation doesn’t quite do that. Mm.
Shawn Wang [00:38:40]: Yeah. How low can we go in precision?
Jeff Dean [00:38:43]: Because people are saying like ternary is like, uh, yeah, I mean, I’m a big fan of very low precision because I think that gets, that saves you a tremendous amount of time. Right. Because it’s picojoules per bit that you’re transferring and reducing the number of bits is a really good way to, to reduce that. Um, you know, I think people have gotten a lot of luck, uh, mileage out of having very low bit precision things, but then having scaling factors that apply to a whole bunch of, uh, those, those weights. Scaling. How does it, how does it, okay.
Shawn Wang [00:39:15]: Interesting. You, so low, low precision, but scaled up weights. Yeah. Huh. Yeah. Never considered that. Yeah. Interesting. Uh, w w while we’re on this topic, you know, I think there’s a lot of, um, uh, this, the concept of precision at all is weird when we’re sampling, you know, uh, we just, at the end of this, we’re going to have all these like chips that I’ll do like very good math. And then we’re just going to throw a random number generator at the start. So, I mean, there’s a movement towards, uh, energy based, uh, models and processors. I’m just curious if you’ve, obviously you’ve thought about it, but like, what’s your commentary?
Jeff Dean [00:39:50]: Yeah. I mean, I think. There’s a bunch of interesting trends though. Energy based models is one, you know, diffusion based models, which don’t sort of sequentially decode tokens is another, um, you know, speculative decoding is a way that you can get sort of an equivalent, very small.
Shawn Wang [00:40:06]: Draft.
Jeff Dean [00:40:07]: Batch factor, uh, for like you predict eight tokens out and that enables you to sort of increase the effective batch size of what you’re doing by a factor of eight, even, and then you maybe accept five or six of those tokens. So you get. A five, a five X improvement in the amortization of moving weights, uh, into the multipliers to do the prediction for the, the tokens. So these are all really good techniques and I think it’s really good to look at them from the lens of, uh, energy, real energy, not energy based models, um, and, and also latency and throughput, right? If you look at things from that lens, that sort of guides you to. Two solutions that are gonna be, uh, you know, better from, uh, you know, being able to serve larger models or, you know, equivalent size models more cheaply and with lower latency.
Shawn Wang [00:41:03]: Yeah. Well, I think, I think I, um, it’s appealing intellectually, uh, haven’t seen it like really hit the mainstream, but, um, I do think that, uh, there’s some poetry in the sense that, uh, you know, we don’t have to do, uh, a lot of shenanigans if like we fundamentally. Design it into the hardware. Yeah, yeah.
Jeff Dean [00:41:23]: I mean, I think there’s still a, there’s also sort of the more exotic things like analog based, uh, uh, computing substrates as opposed to digital ones. Uh, I’m, you know, I think those are super interesting cause they can be potentially low power. Uh, but I think you often end up wanting to interface that with digital systems and you end up losing a lot of the power advantages in the digital to analog and analog to digital conversions. You end up doing, uh, at the sort of boundaries. And periphery of that system. Um, I still think there’s a tremendous distance we can go from where we are today in terms of energy efficiency with sort of, uh, much better and specialized hardware for the models we care about.
Shawn Wang [00:42:05]: Yeah.
Alessio Fanelli [00:42:06]: Um, any other interesting research ideas that you’ve seen, or like maybe things that you cannot pursue a Google that you would be interested in seeing researchers take a step at, I guess you have a lot of researchers. Yeah, I guess you have enough, but our, our research.
Jeff Dean [00:42:21]: Our research portfolio is pretty broad. I would say, um, I mean, I think, uh, in terms of research directions, there’s a whole bunch of, uh, you know, open problems and how do you make these models reliable and able to do much longer, kind of, uh, more complex tasks that have lots of subtasks. How do you orchestrate, you know, maybe one model that’s using other models as tools in order to sort of build, uh, things that can accomplish, uh, you know, much more. Yeah. Significant pieces of work, uh, collectively, then you would ask a single model to do. Um, so that’s super interesting. How do you get more verifiable, uh, you know, how do you get RL to work for non-verifiable domains? I think it’s a pretty interesting open problem because I think that would broaden out the capabilities of the models, the improvements that you’re seeing in both math and coding. Uh, if we could apply those to other less verifiable domains, because we’ve come up with RL techniques that actually enable us to do that. Uh, effectively, that would, that would really make the models improve quite a lot. I think.
Alessio Fanelli [00:43:26]: I’m curious, like when we had Noam Brown on the podcast, he said, um, they already proved you can do it with deep research. Um, you kind of have it with AI mode in a way it’s not verifiable. I’m curious if there’s any thread that you think is interesting there. Like what is it? Both are like information retrieval of JSON. So I wonder if it’s like the retrieval is like the verifiable part. That you can score or what are like, yeah, yeah. How, how would you model that, that problem?
Jeff Dean [00:43:55]: Yeah. I mean, I think there are ways of having other models that can evaluate the results of what a first model did, maybe even retrieving. Can you have another model that says, is this things, are these things you retrieved relevant? Or can you rate these 2000 things you retrieved to assess which ones are the 50 most relevant or something? Um, I think those kinds of techniques are actually quite effective. Sometimes I can even be the same model, just prompted differently to be a, you know, a critic as opposed to a, uh, actual retrieval system. Yeah.
Shawn Wang [00:44:28]: Um, I do think like there, there is that, that weird cliff where like, it feels like we’ve done the easy stuff and then now it’s, but it always feels like that every year. It’s like, oh, like we know, we know, and the next part is super hard and nobody’s figured it out. And, uh, exactly with this RLVR thing where like everyone’s talking about, well, okay, how do we. the next stage of the non-verifiable stuff. And everyone’s like, I don’t know, you know, Ellen judge.
Jeff Dean [00:44:56]: I mean, I feel like the nice thing about this field is there’s lots and lots of smart people thinking about creative solutions to some of the problems that we all see. Uh, because I think everyone sort of sees that the models, you know, are great at some things and they fall down around the edges of those things and, and are not as capable as we’d like in those areas. And then coming up with good techniques and trying those. And seeing which ones actually make a difference is sort of what the whole research aspect of this field is, is pushing forward. And I think that’s why it’s super interesting. You know, if you think about two years ago, we were struggling with GSM, eight K problems, right? Like, you know, Fred has two rabbits. He gets three more rabbits. How many rabbits does he have? That’s a pretty far cry from the kinds of mathematics that the models can, and now you’re doing IMO and Erdos problems in pure language. Yeah. Yeah. Pure language. So that is a really, really amazing jump in capabilities in, you know, in a year and a half or something. And I think, um, for other areas, it’d be great if we could make that kind of leap. Uh, and you know, we don’t exactly see how to do it for some, some areas, but we do see it for some other areas and we’re going to work hard on making that better. Yeah.
Shawn Wang [00:46:13]: Yeah.
Alessio Fanelli [00:46:14]: Like YouTube thumbnail generation. That would be very helpful. We need that. That would be AGI. We need that.
Shawn Wang [00:46:20]: That would be. As far as content creators go.
Jeff Dean [00:46:22]: I guess I’m not a YouTube creator, so I don’t care that much about that problem, but I guess, uh, many people do.
Shawn Wang [00:46:27]: It does. Yeah. It doesn’t, it doesn’t matter. People do judge books by their covers as it turns out. Um, uh, just to draw a bit on the IMO goal. Um, I’m still not over the fact that a year ago we had alpha proof and alpha geometry and all those things. And then this year we were like, screw that we’ll just chuck it into Gemini. Yeah. What’s your reflection? Like, I think this, this question about. Like the merger of like symbolic systems and like, and, and LMS, uh, was a very much core belief. And then somewhere along the line, people would just said, Nope, we’ll just all do it in the LLM.
Jeff Dean [00:47:02]: Yeah. I mean, I think it makes a lot of sense to me because, you know, humans manipulate symbols, but we probably don’t have like a symbolic representation in our heads. Right. We have some distributed representation that is neural net, like in some way of lots of different neurons. And activation patterns firing when we see certain things and that enables us to reason and plan and, you know, do chains of thought and, you know, roll them back now that, that approach for solving the problem doesn’t seem like it’s going to work. I’m going to try this one. And, you know, in a lot of ways we’re emulating what we intuitively think, uh, is happening inside real brains in neural net based models. So it never made sense to me to have like completely separate. Uh, discrete, uh, symbolic things, and then a completely different way of, of, uh, you know, thinking about those things.
Shawn Wang [00:47:59]: Interesting. Yeah. Uh, I mean, it’s maybe seems obvious to you, but it wasn’t obvious to me a year ago. Yeah.
Jeff Dean [00:48:06]: I mean, I do think like that IMO with, you know, translating to lean and using lean and then the next year and also a specialized geometry model. And then this year switching to a single unified model. That is roughly the production model with a little bit more inference budget, uh, is actually, you know, quite good because it shows you that the capabilities of that general model have improved dramatically and, and now you don’t need the specialized model. This is actually sort of very similar to the 2013 to 16 era of machine learning, right? Like it used to be, people would train separate models for lots of different, each different problem, right? I have, I want to recognize street signs and something. So I train a street sign. Recognition recognition model, or I want to, you know, decode speech recognition. I have a speech model, right? I think now the era of unified models that do everything is really upon us. And the question is how well do those models generalize to new things they’ve never been asked to do and they’re getting better and better.
Shawn Wang [00:49:10]: And you don’t need domain experts. Like one of my, uh, so I interviewed ETA who was on, who was on that team. Uh, and he was like, yeah, I, I don’t know how they work. I don’t know where the IMO competition was held. I don’t know the rules of it. I just trained the models, the training models. Yeah. Yeah. And it’s kind of interesting that like people with these, this like universal skill set of just like machine learning, you just give them data and give them enough compute and they can kind of tackle any task, which is the bitter lesson, I guess. I don’t know. Yeah.
Jeff Dean [00:49:39]: I mean, I think, uh, general models, uh, will win out over specialized ones in most cases.
Shawn Wang [00:49:45]: Uh, so I want to push there a bit. I think there’s one hole here, which is like, uh. There’s this concept of like, uh, maybe capacity of a model, like abstractly a model can only contain the number of bits that it has. And, uh, and so it, you know, God knows like Gemini pro is like one to 10 trillion parameters. We don’t know, but, uh, the Gemma models, for example, right? Like a lot of people want like the open source local models that are like that, that, that, and, and, uh, they have some knowledge, which is not necessary, right? Like they can’t know everything like, like you have the. The luxury of you have the big model and big model should be able to capable of everything. But like when, when you’re distilling and you’re going down to the small models, you know, you’re actually memorizing things that are not useful. Yeah. And so like, how do we, I guess, do we want to extract that? Can we, can we divorce knowledge from reasoning, you know?
Jeff Dean [00:50:38]: Yeah. I mean, I think you do want the model to be most effective at reasoning if it can retrieve things, right? Because having the model devote precious parameter space. To remembering obscure facts that could be looked up is actually not the best use of that parameter space, right? Like you might prefer something that is more generally useful in more settings than this obscure fact that it has. Um, so I think that’s always attention at the same time. You also don’t want your model to be kind of completely detached from, you know, knowing stuff about the world, right? Like it’s probably useful to know how long the golden gate be. Bridges just as a general sense of like how long are bridges, right? And, uh, it should have that kind of knowledge. It maybe doesn’t need to know how long some teeny little bridge in some other more obscure part of the world is, but, uh, it does help it to have a fair bit of world knowledge and the bigger your model is, the more you can have. Uh, but I do think combining retrieval with sort of reasoning and making the model really good at doing multiple stages of retrieval. Yeah.
Shawn Wang [00:51:49]: And reasoning through the intermediate retrieval results is going to be a, a pretty effective way of making the model seem much more capable, because if you think about, say, a personal Gemini, yeah, right?
Jeff Dean [00:52:01]: Like we’re not going to train Gemini on my email. Probably we’d rather have a single model that, uh, we can then use and use being able to retrieve from my email as a tool and have the model reason about it and retrieve from my photos or whatever, uh, and then make use of that and have multiple. Um, you know, uh, stages of interaction. that makes sense.
Alessio Fanelli [00:52:24]: Do you think the vertical models are like, uh, interesting pursuit? Like when people are like, oh, we’re building the best healthcare LLM, we’re building the best law LLM, are those kind of like short-term stopgaps or?
Jeff Dean [00:52:37]: No, I mean, I think, I think vertical models are interesting. Like you want them to start from a pretty good base model, but then you can sort of, uh, sort of viewing them, view them as enriching the data. Data distribution for that particular vertical domain for healthcare, say, um, we’re probably not going to train or for say robotics. We’re probably not going to train Gemini on all possible robotics data. We, you could train it on because we want it to have a balanced set of capabilities. Um, so we’ll expose it to some robotics data, but if you’re trying to build a really, really good robotics model, you’re going to want to start with that and then train it on more robotics data. And then maybe that would. It’s multilingual translation capability, but improve its robotics capabilities. And we’re always making these kind of, uh, you know, trade-offs in the data mix that we train the base Gemini models on. You know, we’d love to include data from 200 more languages and as much data as we have for those languages, but that’s going to displace some other capabilities of the model. It won’t be as good at, um, you know, Pearl programming, you know, it’ll still be good at Python programming. Cause we’ll include it. Enough. Of that, but there’s other long tail computer languages or coding capabilities that it may suffer on or multi, uh, multimodal reasoning capabilities may suffer. Cause we didn’t get to expose it to as much data there, but it’s really good at multilingual things. So I, I think some combination of specialized models, maybe more modular models. So it’d be nice to have the capability to have those 200 languages, plus this awesome robotics model, plus this awesome healthcare, uh, module that all can be knitted together to work in concert and called upon in different circumstances. Right? Like if I have a health related thing, then it should enable using this health module in conjunction with the main base model to be even better at those kinds of things. Yeah.
Shawn Wang [00:54:36]: Installable knowledge. Yeah.
Jeff Dean [00:54:37]: Right.
Shawn Wang [00:54:38]: Just download as a, as a package.
Jeff Dean [00:54:39]: And some of that installable stuff can come from retrieval, but some of it probably should come from preloaded training on, you know, uh, a hundred billion tokens or a trillion tokens of health data. Yeah.
Shawn Wang [00:54:51]: And for listeners, I think, uh, I will highlight the Gemma three end paper where they, there was a little bit of that, I think. Yeah.
Alessio Fanelli [00:54:56]: Yeah. I guess the question is like, how many billions of tokens do you need to outpace the frontier model improvements? You know, it’s like, if I have to make this model better healthcare and the main. Gemini model is still improving. Do I need 50 billion tokens? Can I do it with a hundred, if I need a trillion healthcare tokens, it’s like, they’re probably not out there that you don’t have, you know, I think that’s really like the.
Jeff Dean [00:55:21]: Well, I mean, I think healthcare is a particularly challenging domain, so there’s a lot of healthcare data that, you know, we don’t have access to appropriately, but there’s a lot of, you know, uh, healthcare organizations that want to train models on their own data. That is not public healthcare data, uh, not public health. But public healthcare data. Um, so I think there are opportunities there to say, partner with a large healthcare organization and train models for their use that are going to be, you know, more bespoke, but probably, uh, might be better than a general model trained on say, public data. Yeah.
Shawn Wang [00:55:58]: Yeah. I, I believe, uh, by the way, also this is like somewhat related to the language conversation. Uh, I think one of your, your favorite examples was you can put a low resource language in the context and it just learns. Yeah.
Jeff Dean [00:56:09]: Oh, yeah, I think the example we used was Calamon, which is truly low resource because it’s only spoken by, I think 120 people in the world and there’s no written text.
Shawn Wang [00:56:20]: So, yeah. So you can just do it that way. Just put it in the context. Yeah. Yeah. But I think your whole data set in the context, right.
Jeff Dean [00:56:27]: If you, if you take a language like, uh, you know, Somali or something, there is a fair bit of Somali text in the world that, uh, or Ethiopian Amharic or something, um, you know, we probably. Yeah. Are not putting all the data from those languages into the Gemini based training. We put some of it, but if you put more of it, you’ll improve the capabilities of those models.
Shawn Wang [00:56:49]: Yeah.
Jeff Dean [00:56:49]: So, or of those languages.
Shawn Wang [00:56:52]: Uh, yeah, cool. Uh, it’s, uh, I have a side interest in linguistics. I, I, I did, uh, uh, a few classes back in college and like, uh, part of me, like if I was a linguist and I could have access to all these models, I would just be asking really fundamental questions about language itself. Yeah. Like, uh, one is th there’s one very obvious one, which is Sapir-Whorf, like how much does like the language that you speak affect your thinking, but then also there’s some languages where there’s just concepts that are not represented in other languages, but some others, many others that are just duplicates, right. Where, uh, there’s also another paper that people love called the platonic representation where, you know, like the, the, an image of a cup is, uh, if you say learn a model on that and you, you, you have a lot of texts with the word cup eventually maps to it, like roughly the same place. And so like that should apply to languages except where it doesn’t. And that’s actually like very interesting differences in what humanity has discovered as concepts that maybe English doesn’t have.
Shawn Wang [00:57:54]: I don’t know. It’s just like my, my rant on languages. Yeah.
Jeff Dean [00:57:58]: I mean, I, I did some work on a early model that fused together a language based model with you have, you know, nice word based representations and then an image model where you have. Trained it on image net like things. Yes. And then you fuse together the top layers of, uh, no, this is devise, uh, uh, the, you do a little bit more training to fuse together those representations. And what you found was that if you give a novel image that is not in any of the categories in the image model, it was trained on the model can often assigns kind of the right cat, the right label to that image. Um, so for example, um, I think, uh, telescope and, uh, binoculars were both in the training, uh, categories for the image model, but, um, microscope was not. Hmm. And so if you’re given an image of a microscope, it actually can come up with something that’s, uh, got the word microscope as the label that it assigns, even though it’s never actually seen an image labeled that.
Shawn Wang [00:59:01]: Oh, that’s nice. That’s kind of cool. Yeah.
Jeff Dean [00:59:04]: Um, so yeah.
Shawn Wang [00:59:07]: Useful. Uh, cool. Uh, I think. There, there’s more general, like broad questions, but like, I guess what, what do you, uh, wish you were asked more in, in, in general, like, you know, like you, you have such a broad scope. We’ve covered the hardware, we’ve covered the, the, the models research. Yeah.
Jeff Dean [00:59:22]: I mean, I think, uh, one thing that’s kind of interesting is, you know, I, I did a undergrad thesis on neural network, uh, training, uh, uh, parallel neural network training, uh, back in 1990 when I got exposed to neural nets and I always felt kind of, they were the right abstraction. Uh, but we just needed way more compute than we had then. Mm-hmm. So like the 32 processors in the department parallel computer, you know, could get you a, a little bit more interesting, uh, model, but not, not enough to solve real problems. And so starting in 2008 or nine, you know, the world started to have enough computing power through Moore’s law and, you know, larger, interesting data sets to train on to actually, you know, start training neural nets that could tackle real problems that people cared about. Yeah. Speech recognition. Vision, and eventually, uh, language. Um, and so, um, when I started working on neural nets at Google in, in late 2011, um, you know, I really just felt like we should scale up the size of neural networks we can train using, you know, large amounts of parallel computation. And so, uh, I actually, uh, revived some ideas from my undergrad thesis where I’d done both model parallel and data parallel, uh, training and I compared them. Uh, I, I called them. I’ve been doing this since I was eight. It was something different. There was like pattern partitioned and, you know, model partitioned or something.
Shawn Wang [01:00:43]: Well, I have to, is it, is it public? And we can go dig it up?
Jeff Dean [01:00:45]: Yeah, it’s on, it’s on the web. Okay, nice. Um, but, uh, you know, I think combining a lot of those techniques and really just trying to push on scaling things up over the last, you know, 15 years has been, you know, really important. And that means, you know, improvements in the hardware. So, you know, pushing on building specialized hardware like TPUs. Uh, it also means, you know, pushing on software, abstraction layers to let people express their ideas to the computer. Thank you for having me.
Jeff Dean [01:01:40]: Thank you for having me.
Shawn Wang [01:07:10]: If that’s something you would agree with at the time, or is there a different post-mortem?
Jeff Dean [01:07:15]: The brain marketplace for compute quotas.
Shawn Wang [01:07:18]: Compute quotas, where basically he was like, okay, David worked at OpenAI as VP Engine and then he worked at Google. He was like, fundamentally, OpenAI was willing to go all in, like, bet the farm on one thing, whereas Google was more democratic. Everyone had a quota. And I was like, okay, if you believe in scaling as an important thing, that’s an important organizational-wide decision to do.
Jeff Dean [01:07:41]: Yeah. Yeah, I mean, I think I would somewhat agree with that. I mean, I think I actually wrote a one-page memo saying we were being stupid by fragmenting our resources. So in particular, at the time, we had efforts within Google Research. And in the brain team in particular, on large language models. We also had efforts on multimodal models in other parts of brain and Google Research. And then Legacy DeepMind had efforts like Chinchilla models and Flamingo models. And so really, we were fragmenting not only our compute across those separate efforts, but also our best people and our best. And so I said, this is just stupid. Why don’t we combine things and have one effort to train an awesome single unified model that is multimodal from the start, that’s good at everything. And that was the origin of the Gemini effort.
Shawn Wang [01:08:52]: And my one-page memo worked, which is good. Did you have the name? Because also for those who don’t know, you named Gemini.
Jeff Dean [01:08:58]: I did. There was another name proposed. And I said, you know what? You know, it’s sort of like these two organizations really are like twins in some sense coming together. So I kind of like that. And then there’s also the NASA interpretation of the early Gemini project being an important thing on your way to the Apollo project. So it seemed like a good name. Twins coming together. Right.
Alessio Fanelli [01:09:27]: Yeah. Nice. I know we’re already running out of time, but I’m curious how you use AI. Today to code. So, I mean, you’re probably one of the most prolific engineers in the history of computer science. Um, I was reading on through the article about you and Sanjay’s friendship and how you work together. And you have one quote about, you need to find someone that you’re going to pair program with who’s compatible with your way of thinking so that the two of you together are a complimentary force. And I was thinking about how you think about coding agents and this, like, how do you shape a coding agents to be compatible with your way of thinking? Like. How would you rate the tools today? Like, where should things go? Yeah.
Jeff Dean [01:10:07]: I mean, first, I think the coding tools are, you know, getting vastly better compared to where they were a year or two, two years ago. So now you can actually rely on them to do more complex things that you as a, as a software engineer want to accomplish. And you can sort of delegate, you know, pretty complex things to these tools. And I think one really nice aspect about the, uh, interaction between, uh, uh, human, uh, software engineer and, uh, uh, coding model that they’re working with is your way of talking to that, uh, coding model actually sort of, uh, dictates how it interacts with you, right? Like you could ask it, please write a bunch of good tests for this. You could ask it, please help me brainstorm. Performance ideas and your way of doing that is going to shape how the model responds, what kinds of problems it tackles, you know, how much do you want the model to go off and do things that are larger and more independent versus interact with it, uh, more to make sure that you’re shaping the right kinds of, of things. And I think it’s not the case that any one style is the right thing for everything, right? Like some kinds of problems you actually want, uh, maybe a more frequent interaction style with a model. And other ones, you’re just like, yeah, please just go write this because I, I know I need this thing. I can specify it well enough, um, and go off and do it and come back when you’re done. And so I do think there’s going to be more of a style of having lots of independent, uh, software agents off doing things on your behalf and figuring out the right sort of human computer interaction model and UI and so on for when should it interrupt you and say, Hey, I need a little more guidance here, or I’ve done this thing. Now what, now what should I do? Um, I think we, we’re not at the end all answer to that question. And as the models get better, that, uh, set of decisions you put into how the interaction should happen may, may change, right? Like if you, if you have a team of 50 interns, how would you manage that if they were people? And I think it’s not, do you want 50 interns? You might, if they’re really good, right?
Shawn Wang [01:12:23]: It’s a lot of management. But it’s a lot of, uh.
Jeff Dean [01:12:25]: Uh, yeah. I mean, I think that is probably within the realm of possibilities that lots of people could have 50 interns. Yeah. And so how would you actually deal with that as a person, right? Like you would probably want them to form small sub teams, so you don’t have to interact with 50 of them. You can interact with five, five of those teams and they’re off doing things on your behalf, but I don’t know exactly what the, how this is going to unfold.
Alessio Fanelli [01:12:52]: Hmm. Yeah. How do you think about bringing people? Like the pair programming is always helpful to like get net new ideas in the distribution, so to speak. It feels as we have more of these coding agents, write the code, it’s hard to bring other people into the problem. So you go to like, you know, you have 50 interns, right? And then you want to go to Noam Shazier be like, Hey, no, I’m, I want to like pair on this thing. But now there’s like this huge amount of work that has been done in parallel that you need to catch him up on. Right. And I’m curious, like if people are going to be in a way more isolated in their teams, where it’s. It’s like, okay, there’s so much context in these 50 interns that it’s just hard for me to like relay everything back to you.
Jeff Dean [01:13:33]: Maybe. I mean, on the other hand, like imagine a classical software organization without any AI assisted tools, right. You would have, you know, 50 people doing stuff and their interaction style is going to be naturally very hierarchical because, um, you know, these 50 people are going to be working on this part of the system and not. Not interact that much with these other people over here. But if you have, you know, five people each managing 50 virtual agents, you know, they might be able to actually have much higher bandwidth communication among the five people, uh, then you would have among five people who are also trying to coordinate, you know, a 50 person software team. Each.
Alessio Fanelli [01:14:15]: So how, how do you, I’m curious how you change your just working rhythm, you know, like you spend more time ahead with people going through SPACs and design. Goals. Like,
Jeff Dean [01:14:26]: um, I mean, I do think it’s interesting that, you know, whenever people were taught how to write software, they were taught that it’s really important to write specifications super clearly, but no one really believed that. Like it was like, yeah, whatever. I don’t need to do that. I’m going to really, I don’t know. I mean, writing the English language specification was never kind of an artifact that was really paid a lot of attention to. I mean, it was important, but it wasn’t sort of the thing. That drove the actual creative process quite as much as if you specify what software you want the agent to write for you, you’d better be pretty darn careful of and how you specify that because that’s going to dictate the quality of the output, right? Like if you, if you don’t cover that it needs to handle this kind of thing, or that this is a super important corner case, or that, you know, you really care about the performance of this part of it, you know, it may, uh, not do what you want. Yeah. And the better you get at interacting with these models. And I think one of the ways people will get better is they will get really good at crisply specifying things rather than leaving things to ambiguity. And that is actually probably not a bad thing. It’s not a bad skill to have, regardless of whether you’re a software engineer or a, you know, trying to do some other kind of, uh, task, you know, being able to crisply specify what it is you want. It’s going to be really important. Yeah.
Shawn Wang [01:15:52]: My, my joke is, um, you know, good. Yeah. I think one thing is in, uh, indistinguishable from sufficiently advanced executive communication, like it’s like writing an internal memo, like weigh your words very carefully and also I think very important to be multimodal, right? I think, uh, one thing that, uh, anti-gravity from, from Google also did was like, just come out the gate to very, very strong multimodal, including videos, and that’s the highest bandwidth communication prompt that you can give to the model, which is fantastic. Yeah.
Alessio Fanelli [01:16:20]: How do you collect things that you often you will have in your mind? So you have this amazing, like performance sense thing that you’ve heard about how to look for performance improvements. And is there a lot more value in like people writing these like generic things down so that they can then put them back as like potential retrieval artifacts for the model? Like, or do I have like the edge cases is like a good example, right? It’s like, if you’re building systems, you already have in your mind, specific edge cases, depending on it. But now you have to like, every time repeat it. Like, are you having people spend a lot more time writing? Are you finding out more generic things to bring back?
Jeff Dean [01:16:56]: Or, um, I mean, I do think well-written guides of, of how to do good software engineering are going to be useful because they can be used as input to models or, you know, read by other developers so that their prompts are, you know, more clear about what the, the underlying software system should, should be doing. Um, you know, I think it may not be that you need to create a custom one. For every situation, if you have general guides and put those into, you know, the context of a coding agent, that, that can be helpful. Like in, you can imagine one for distributed systems, you could say, okay, think about failures of these kinds of things. And these are some techniques you can deal with failures. You know, you can have, uh, you know, Paxos like replication, or, you know, you can, uh, send the request to two places and tolerate failure because you only need one of them to come back. You know, a little. Description of 20 techniques like that in building distributed systems, probably would go a long way to having a coding agent be able to sort of cobble up more reliable and robust distributed systems.
Shawn Wang [01:18:07]: Yeah. Yeah. I wonder when Gemini will be able to build Spanner, right?
Alessio Fanelli [01:18:12]: Probably already has the code inside, you know?
Alessio Fanelli [01:18:16]: Yeah. That, I mean, that’s a good example, right? When you have like, you know, the cap theorem and it’s like, well, this is like truth and you cannot break that. And then you build something that broke it.
Shawn Wang [01:18:26]: Like, I’m curious, like models in a way are like, would he say he broke it? Did you, would you say you broke cap theorem? Really? Yeah. Okay. All right. I mean, under local assumptions. Yeah. Under some, some, yeah. And they’re like, you know, good clocks. Yeah. Yeah.
Alessio Fanelli [01:18:41]: It’s like some, sometimes you don’t have to like always follow what is known to be true. Right. And I, I think models in a way, like if you tell them something, they’re like really buy into that, you know? Um, yeah. So yeah, just more. Thinking than any answer on how to fix it.
Jeff Dean [01:18:57]: Yeah, my, my, uh, you know, it’s just on this, like, like big prompting and, and, uh, iteration, you know, I think that coming back to your latency point, um, I always, I always try to one, one AB test or experiment or benchmark or research I would like is what is the performance difference between, let’s say three dumb fast model calls with human alignment because the human will correct human alignment, being human looks at the first one and produces a new prompt.
Shawn Wang [01:19:23]: For the second one. Correct. Okay. As opposed to like, you spec it out, you know, it’s been a long time writing as a pro a big, big fat prompt, and then you have a very smart model. Do it right. Right. You know, cause, uh, really is, is, uh, our lacks in performance, uh, an issue of like, well, you just haven’t specified well enough. There’s no universe in which I can produce what you want because you just haven’t told me. Right.
Jeff Dean [01:19:44]: It’s underspecified. So I could produce 10 different things and only one of them is the thing you wanted. Yeah.
Shawn Wang [01:19:49]: And the multi-turn taking with a flash model is enough. Yeah.
Jeff Dean [01:19:54]: Yeah, I’m, I’m a big believer in pushing on latency because I think being able to have really low latency interactions with a system you’re using is just much more delightful than something that is, you know, 10 times as slow or 20 times as slow. And I think, you know, in the future we’ll see models that are, and, and underlying software and hardware systems that are 20X lower latency than what we have today, 50X lower latency. And that’s going to be really, really important for systems. That need to do a lot of stuff, uh, between your interactions.
Shawn Wang [01:20:27]: Yeah. Yeah. There, there’s two extremes, right? And then meanwhile, you also have DeepThink, which is all the way on the other side. Right.
Jeff Dean [01:20:33]: But you would use DeepThink all the time if it weren’t for cost and latency, right? If, if you could have that capability in a model because the latency improvement was 20X, uh, in the underlying hardware and system and costs, you know, there’s no reason you wouldn’t want that.
Shawn Wang [01:20:50]: Yeah.
Jeff Dean [01:20:52]: But at the same time, then you’d probably have a model. That is even better. That would take you 20X longer, even on that new hardware. Yeah.
Shawn Wang [01:21:00]: Uh, you know, there, there’s, uh, the Pareto curve keeps climbing. Um, yeah, onward and outward, onward and outward. Yeah. Should we ask him for predictions to, to go? I don’t know if you have any predictions that you, that you like to keep, you know, like, uh, one, one way to do this is you have your tests whenever a new model comes out that you run, uh, what’s something that you’re, you’re not quite happy with yet. That you think we’ll get done soon.
Jeff Dean [01:21:29]: Um, let me make two predictions that are not quite in that vein. Yeah. So I think a personalized model that knows you and knows all your state and is able to retrieve over all state you have access to, that you opt into is going to be incredibly useful compared to a more generic model that doesn’t have access to that. So like, can something attend to everything I’ve ever seen? Yeah. Every email, every photo, every. Yeah. Video I’ve watched, that’s going to be really useful. Uh, I think, uh, more and more specialized hardware is going to enable much lower latency models and much more capable models for affordable prices, uh, than say the current, current status quo. Uh, that’s going to be also quite important. Yeah.
Shawn Wang [01:22:16]: When you say much lower latency, uh, people usually talk in tokens per second. Is that a term that is okay? Okay. Uh, you know, we’re at, let’s say a hundred. Now we can go to a thousand. Is it meaningful to go 10,000? Yes. Really? Okay. Absolutely. Right. Yeah. Because of chain of thought and chain of thought reasoning.
Jeff Dean [01:22:36]: I mean, you could think, you know, uh, many more tokens, you could do many more parallel rollouts. You could generate way more code, uh, and check that the code is cracked with a chain of thought reasoning. So I think, you know, being able to do that at 10,000 tokens per second would be awesome. Yeah.
Shawn Wang [01:22:52]: At 10,000 tokens per second, you are no longer reading code. Yeah. Like you will just generate it. You’ll, I’m not reading it.
Jeff Dean [01:22:58]: Well, remember, it may not, it may not end up with 10,000 tokens of code. Yeah. It may be a thousand tokens of code that with 9,000 tokens of reasoning behind it, which would actually be probably much better code to read. Yeah.
Alessio Fanelli [01:23:11]: Yeah. If I had more time, I would have written a shorter letter. Yeah. Yeah. Yeah. Um, awesome. Jeff, this was amazing. Thanks for taking the time. Thank you.
Jeff Dean [01:23:20]: It’s been fun. Thanks for having me.


Get full access to Latent.Space at www.latent.space/subscribe

This podcast features Gabriele Corso and Jeremy Wohlwend, co-founders of Boltz and authors of the Boltz Manifesto, discussing the rapid evolution of structural biology models from AlphaFold to their own open-source suite, Boltz-1 and Boltz-2. The central thesis is that while single-chain protein structure prediction is largely “solved” through evolutionary hints, the next frontier lies in modeling complex interactions (protein-ligand, protein-protein) and generative protein design, which Boltz aims to democratize via open-source foundations and scalable infrastructure.

Full Video Pod
On YouTube!

Timestamps
* 00:00 Introduction to Benchmarking and the “Solved” Protein Problem
* 06:48 Evolutionary Hints and Co-evolution in Structure Prediction
* 10:00 The Importance of Protein Function and Disease States
* 15:31 Transitioning from AlphaFold 2 to AlphaFold 3 Capabilities
* 19:48 Generative Modeling vs. Regression in Structural Biology
* 25:00 The “Bitter Lesson” and Specialized AI Architectures
* 29:14 Development Anecdotes: Training Boltz-1 on a Budget
* 32:00 Validation Strategies and the Protein Data Bank (PDB)
* 37:26 The Mission of Boltz: Democratizing Access and Open Source
* 41:43 Building a Self-Sustaining Research Community
* 44:40 Boltz-2 Advancements: Affinity Prediction and Design
* 51:03 BoltzGen: Merging Structure and Sequence Prediction
* 55:18 Large-Scale Wet Lab Validation Results
* 01:02:44 Boltz Lab Product Launch: Agents and Infrastructure
* 01:13:06 Future Directions: Developpability and the “Virtual Cell”
* 01:17:35 Interacting with Skeptical Medicinal Chemists
Key Summary
Evolution of Structure Prediction & Evolutionary Hints
* Co-evolutionary Landscapes: The speakers explain that breakthrough progress in single-chain protein prediction relied on decoding evolutionary correlations where mutations in one position necessitate mutations in another to conserve 3D structure.
* Structure vs. Folding: They differentiate between structure prediction (getting the final answer) and folding (the kinetic process of reaching that state), noting that the field is still quite poor at modeling the latter.
* Physics vs. Statistics: RJ posits that while models use evolutionary statistics to find the right “valley” in the energy landscape, they likely possess a “light understanding” of physics to refine the local minimum.
The Shift to Generative Architectures
* Generative Modeling: A key leap in AlphaFold 3 and Boltz-1 was moving from regression (predicting one static coordinate) to a generative diffusion approach that samples from a posterior distribution.
* Handling Uncertainty: This shift allows models to represent multiple conformational states and avoid the “averaging” effect seen in regression models when the ground truth is ambiguous.
* Specialized Architectures: Despite the “bitter lesson” of general-purpose transformers, the speakers argue that equivariant architectures remain vastly superior for biological data due to the inherent 3D geometric constraints of molecules.
Boltz-2 and Generative Protein Design
* Unified Encoding: Boltz-2 (and BoltzGen) treats structure and sequence prediction as a single task by encoding amino acid identities into the atomic composition of the predicted structure.
* Design Specifics: Instead of a sequence, users feed the model blank tokens and a high-level “spec” (e.g., an antibody framework), and the model decodes both the 3D structure and the corresponding amino acids.
* Affinity Prediction: While model confidence is a common metric, Boltz-2 focuses on affinity prediction—quantifying exactly how tightly a designed binder will stick to its target.
Real-World Validation and Productization
* Generalized Validation: To prove the model isn’t just “regurgitating” known data, Boltz tested its designs on 9 targets with zero known interactions in the PDB, achieving nanomolar binders for two-thirds of them.
* Boltz Lab Infrastructure: The newly launched Boltz Lab platform provides “agents” for protein and small molecule design, optimized to run 10x faster than open-source versions through proprietary GPU kernels.
* Human-in-the-Loop: The platform is designed to convert skeptical medicinal chemists by allowing them to run parallel screens and use their intuition to filter model outputs.
Transcript
RJ [00:05:35]: But the goal remains to, like, you know, really challenge the models, like, how well do these models generalize? And, you know, we’ve seen in some of the latest CASP competitions, like, while we’ve become really, really good at proteins, especially monomeric proteins, you know, other modalities still remain pretty difficult. So it’s really essential, you know, in the field that there are, like, these efforts to gather, you know, benchmarks that are challenging. So it keeps us in line, you know, about what the models can do or not.
Gabriel [00:06:26]: Yeah, it’s interesting you say that, like, in some sense, CASP, you know, at CASP 14, a problem was solved and, like, pretty comprehensively, right? But at the same time, it was really only the beginning. So you can say, like, what was the specific problem you would argue was solved? And then, like, you know, what is remaining, which is probably quite open.
RJ [00:06:48]: I think we’ll steer away from the term solved, because we have many friends in the community who get pretty upset at that word. And I think, you know, fairly so. But the problem that was, you know, that a lot of progress was made on was the ability to predict the structure of single chain proteins. So proteins can, like, be composed of many chains. And single chain proteins are, you know, just a single sequence of amino acids. And one of the reasons that we’ve been able to make such progress is also because we take a lot of hints from evolution. So the way the models work is that, you know, they sort of decode a lot of hints. That comes from evolutionary landscapes. So if you have, like, you know, some protein in an animal, and you go find the similar protein across, like, you know, different organisms, you might find different mutations in them. And as it turns out, if you take a lot of the sequences together, and you analyze them, you see that some positions in the sequence tend to evolve at the same time as other positions in the sequence, sort of this, like, correlation between different positions. And it turns out that that is typically a hint that these two positions are close in three dimension. So part of the, you know, part of the breakthrough has been, like, our ability to also decode that very, very effectively. But what it implies also is that in absence of that co-evolutionary landscape, the models don’t quite perform as well. And so, you know, I think when that information is available, maybe one could say, you know, the problem is, like, somewhat solved. From the perspective of structure prediction, when it isn’t, it’s much more challenging. And I think it’s also worth also differentiating the, sometimes we confound a little bit, structure prediction and folding. Folding is the more complex process of actually understanding, like, how it goes from, like, this disordered state into, like, a structured, like, state. And that I don’t think we’ve made that much progress on. But the idea of, like, yeah, going straight to the answer, we’ve become pretty good at.
Brandon [00:08:49]: So there’s this protein that is, like, just a long chain and it folds up. Yeah. And so we’re good at getting from that long chain in whatever form it was originally to the thing. But we don’t know how it necessarily gets to that state. And there might be intermediate states that it’s in sometimes that we’re not aware of.
RJ [00:09:10]: That’s right. And that relates also to, like, you know, our general ability to model, like, the different, you know, proteins are not static. They move, they take different shapes based on their energy states. And I think we are, also not that good at understanding the different states that the protein can be in and at what frequency, what probability. So I think the two problems are quite related in some ways. Still a lot to solve. But I think it was very surprising at the time, you know, that even with these evolutionary hints that we were able to, you know, to make such dramatic progress.
Brandon [00:09:45]: So I want to ask, why does the intermediate states matter? But first, I kind of want to understand, why do we care? What proteins are shaped like?
Gabriel [00:09:54]: Yeah, I mean, the proteins are kind of the machines of our body. You know, the way that all the processes that we have in our cells, you know, work is typically through proteins, sometimes other molecules, sort of intermediate interactions. And through that interactions, we have all sorts of cell functions. And so when we try to understand, you know, a lot of biology, how our body works, how disease work. So we often try to boil it down to, okay, what is going right in case of, you know, our normal biological function and what is going wrong in case of the disease state. And we boil it down to kind of, you know, proteins and kind of other molecules and their interaction. And so when we try predicting the structure of proteins, it’s critical to, you know, have an understanding of kind of those interactions. It’s a bit like seeing the difference between... Having kind of a list of parts that you would put it in a car and seeing kind of the car in its final form, you know, seeing the car really helps you understand what it does. On the other hand, kind of going to your question of, you know, why do we care about, you know, how the protein falls or, you know, how the car is made to some extent is that, you know, sometimes when something goes wrong, you know, there are, you know, cases of, you know, proteins misfolding. In some diseases and so on, if we don’t understand this folding process, we don’t really know how to intervene.
RJ [00:11:30]: There’s this nice line in the, I think it’s in the Alpha Fold 2 manuscript, where they sort of discuss also like why we even hopeful that we can target the problem in the first place. And then there’s this notion that like, well, four proteins that fold. The folding process is almost instantaneous, which is a strong, like, you know, signal that like, yeah, like we should, we might be... able to predict that this very like constrained thing that, that the protein does so quickly. And of course that’s not the case for, you know, for, for all proteins. And there’s a lot of like really interesting mechanisms in the cells, but yeah, I remember reading that and thought, yeah, that’s somewhat of an insightful point.
Gabriel [00:12:10]: I think one of the interesting things about the protein folding problem is that it used to be actually studied. And part of the reason why people thought it was impossible, it used to be studied as kind of like a classical example. Of like an MP problem. Uh, like there are so many different, you know, type of, you know, shapes that, you know, this amino acid could take. And so, this grows combinatorially with the size of the sequence. And so there used to be kind of a lot of actually kind of more theoretical computer science thinking about and studying protein folding as an MP problem. And so it was very surprising also from that perspective, kind of seeing. Machine learning so clear, there is some, you know, signal in those sequences, through evolution, but also through kind of other things that, you know, us as humans, we’re probably not really able to, uh, to understand, but that is, models I’ve, I’ve learned.
Brandon [00:13:07]: And so Andrew White, we were talking to him a few weeks ago and he said that he was following the development of this and that there were actually ASICs that were developed just to solve this problem. So, again, that there were. There were many, many, many millions of computational hours spent trying to solve this problem before AlphaFold. And just to be clear, one thing that you mentioned was that there’s this kind of co-evolution of mutations and that you see this again and again in different species. So explain why does that give us a good hint that they’re close by to each other? Yeah.
RJ [00:13:41]: Um, like think of it this way that, you know, if I have, you know, some amino acid that mutates, it’s going to impact everything around it. Right. In three dimensions. And so it’s almost like the protein through several, probably random mutations and evolution, like, you know, ends up sort of figuring out that this other amino acid needs to change as well for the structure to be conserved. Uh, so this whole principle is that the structure is probably largely conserved, you know, because there’s this function associated with it. And so it’s really sort of like different positions compensating for, for each other. I see.
Brandon [00:14:17]: Those hints in aggregate give us a lot. Yeah. So you can start to look at what kinds of information about what is close to each other, and then you can start to look at what kinds of folds are possible given the structure and then what is the end state.
RJ [00:14:30]: And therefore you can make a lot of inferences about what the actual total shape is. Yeah, that’s right. It’s almost like, you know, you have this big, like three dimensional Valley, you know, where you’re sort of trying to find like these like low energy states and there’s so much to search through. That’s almost overwhelming. But these hints, they sort of maybe put you in. An area of the space that’s already like, kind of close to the solution, maybe not quite there yet. And, and there’s always this question of like, how much physics are these models learning, you know, versus like, just pure like statistics. And like, I think one of the thing, at least I believe is that once you’re in that sort of approximate area of the solution space, then the models have like some understanding, you know, of how to get you to like, you know, the lower energy, uh, low energy state. And so maybe you have some, some light understanding. Of physics, but maybe not quite enough, you know, to know how to like navigate the whole space. Right. Okay.
Brandon [00:15:25]: So we need to give it these hints to kind of get into the right Valley and then it finds the, the minimum or something. Yeah.
Gabriel [00:15:31]: One interesting explanation about our awful free works that I think it’s quite insightful, of course, doesn’t cover kind of the entirety of, of what awful does that is, um, they’re going to borrow from, uh, Sergio Chinico for MIT. So he sees kind of awful. Then the interesting thing about awful is God. This very peculiar architecture that we have seen, you know, used, and this architecture operates on this, you know, pairwise context between amino acids. And so the idea is that probably the MSA gives you this first hint about what potential amino acids are close to each other. MSA is most multiple sequence alignment. Exactly. Yeah. Exactly. This evolutionary information. Yeah. And, you know, from this evolutionary information about potential contacts, then is almost as if the model is. of running some kind of, you know, diastro algorithm where it’s sort of decoding, okay, these have to be closed. Okay. Then if these are closed and this is connected to this, then this has to be somewhat closed. And so you decode this, that becomes basically a pairwise kind of distance matrix. And then from this rough pairwise distance matrix, you decode kind of the
Brandon [00:16:42]: actual potential structure. Interesting. So there’s kind of two different things going on in the kind of coarse grain and then the fine grain optimizations. Interesting. Yeah. Very cool.
Gabriel [00:16:53]: Yeah. You mentioned AlphaFold3. So maybe we have a good time to move on to that. So yeah, AlphaFold2 came out and it was like, I think fairly groundbreaking for this field. Everyone got very excited. A few years later, AlphaFold3 came out and maybe for some more history, like what were the advancements in AlphaFold3? And then I think maybe we’ll, after that, we’ll talk a bit about the sort of how it connects to Bolt. But anyway. Yeah. So after AlphaFold2 came out, you know, Jeremy and I got into the field and with many others, you know, the clear problem that, you know, was, you know, obvious after that was, okay, now we can do individual chains. Can we do interactions, interaction, different proteins, proteins with small molecules, proteins with other molecules. And so. So why are interactions important? Interactions are important because to some extent that’s kind of the way that, you know, these machines, you know, these proteins have a function, you know, the function comes by the way that they interact with other proteins and other molecules. Actually, in the first place, you know, the individual machines are often, as Jeremy was mentioning, not made of a single chain, but they’re made of the multiple chains. And then these multiple chains interact with other molecules to give the function to those. And on the other hand, you know, when we try to intervene of these interactions, think about like a disease, think about like a, a biosensor or many other ways we are trying to design the molecules or proteins that interact in a particular way with what we would call a target protein or target. You know, this problem after AlphaVol2, you know, became clear, kind of one of the biggest problems in the field to, to solve many groups, including kind of ours and others, you know, started making some kind of contributions to this problem of trying to model these interactions. And AlphaVol3 was, you know, was a significant advancement on the problem of modeling interactions. And one of the interesting thing that they were able to do while, you know, some of the rest of the field that really tried to try to model different interactions separately, you know, how protein interacts with small molecules, how protein interacts with other proteins, how RNA or DNA have their structure, they put everything together and, you know, train very large models with a lot of advances, including kind of changing kind of systems. Some of the key architectural choices and managed to get a single model that was able to set this new state-of-the-art performance across all of these different kind of modalities, whether that was protein, small molecules is critical to developing kind of new drugs, protein, protein, understanding, you know, interactions of, you know, proteins with RNA and DNAs and so on.
Brandon [00:19:39]: Just to satisfy the AI engineers in the audience, what were some of the key architectural and data, data changes that made that possible?
Gabriel [00:19:48]: Yeah, so one critical one that was not necessarily just unique to AlphaFold3, but there were actually a few other teams, including ours in the field that proposed this, was moving from, you know, modeling structure prediction as a regression problem. So where there is a single answer and you’re trying to shoot for that answer to a generative modeling problem where you have a posterior distribution of possible structures and you’re trying to sample this distribution. And this achieves two things. One is it starts to allow us to try to model more dynamic systems. As we said, you know, some of these structures can actually take multiple structures. And so, you know, you can now model that, you know, through kind of modeling the entire distribution. But on the second hand, from more kind of core modeling questions, when you move from a regression problem to a generative modeling problem, you are really tackling the way that you think about uncertainty in the model in a different way. So if you think about, you know, I’m undecided between different answers, what’s going to happen in a regression model is that, you know, I’m going to try to make an average of those different kind of answers that I had in mind. When you have a generative model, what you’re going to do is, you know, sample all these different answers and then maybe use separate models to analyze those different answers and pick out the best. So that was kind of one of the critical improvement. The other improvement is that they significantly simplified, to some extent, the architecture, especially of the final model that takes kind of those pairwise representations and turns them into an actual structure. And that now looks a lot more like a more traditional transformer than, you know, like a very specialized equivariant architecture that it was in AlphaFold3.
Brandon [00:21:41]: So this is a bitter lesson, a little bit.
Gabriel [00:21:45]: There is some aspect of a bitter lesson, but the interesting thing is that it’s very far from, you know, being like a simple transformer. This field is one of the, I argue, very few fields in applied machine learning where we still have kind of architecture that are very specialized. And, you know, there are many people that have tried to replace these architectures with, you know, simple transformers. And, you know, there is a lot of debate in the field, but I think kind of that most of the consensus is that, you know, the performance... that we get from the specialized architecture is vastly superior than what we get through a single transformer. Another interesting thing that I think on the staying on the modeling machine learning side, which I think it’s somewhat counterintuitive seeing some of the other kind of fields and applications is that scaling hasn’t really worked kind of the same in this field. Now, you know, models like AlphaFold2 and AlphaFold3 are, you know, still very large models.
RJ [00:29:14]: in a place, I think, where we had, you know, some experience working in, you know, with the data and working with this type of models. And I think that put us already in like a good place to, you know, to produce it quickly. And, you know, and I would even say, like, I think we could have done it quicker. The problem was like, for a while, we didn’t really have the compute. And so we couldn’t really train the model. And actually, we only trained the big model once. That’s how much compute we had. We could only train it once. And so like, while the model was training, we were like, finding bugs left and right. A lot of them that I wrote. And like, I remember like, I was like, sort of like, you know, doing like, surgery in the middle, like stopping the run, making the fix, like relaunching. And yeah, we never actually went back to the start. We just like kept training it with like the bug fixes along the way, which was impossible to reproduce now. Yeah, yeah, no, that model is like, has gone through such a curriculum that, you know, learned some weird stuff. But yeah, somehow by miracle, it worked out.
Gabriel [00:30:13]: The other funny thing is that the way that we were training, most of that model was through a cluster from the Department of Energy. But that’s sort of like a shared cluster that many groups use. And so we were basically training the model for two days, and then it would go back to the queue and stay a week in the queue. Oh, yeah. And so it was pretty painful. And so we actually kind of towards the end with Evan, the CEO of Genesis, and basically, you know, I was telling him a bit about the project and, you know, kind of telling him about this frustration with the compute. And so luckily, you know, he offered to kind of help. And so we, we got the help from Genesis to, you know, finish up the model. Otherwise, it probably would have taken a couple of extra weeks.
Brandon [00:30:57]: Yeah, yeah.
Brandon [00:31:02]: And then, and then there’s some progression from there.
Gabriel [00:31:06]: Yeah, so I would say kind of that, both one, but also kind of these other kind of set of models that came around the same time, were kind of approaching were a big leap from, you know, kind of the previous kind of open source models, and, you know, kind of really kind of approaching the level of AlphaVault 3. But I would still say that, you know, even to this day, there are, you know, some... specific instances where AlphaVault 3 works better. I think one common example is antibody antigen prediction, where, you know, AlphaVault 3 still seems to have an edge in many situations. Obviously, these are somewhat different models. They are, you know, you run them, you obtain different results. So it’s, it’s not always the case that one model is better than the other, but kind of in aggregate, we still, especially at the time.
Brandon [00:32:00]: So AlphaVault 3 is, you know, still having a bit of an edge. We should talk about this more when we talk about Boltzgen, but like, how do you know one is, one model is better than the other? Like you, so you, I make a prediction, you make a prediction, like, how do you know?
Gabriel [00:32:11]: Yeah, so easily, you know, the, the great thing about kind of structural prediction and, you know, once we’re going to go into the design space of designing new small molecule, new proteins, this becomes a lot more complex. But a great thing about structural prediction is that a bit like, you know, CASP was doing, basically the way that you can evaluate them is that, you know, you train... You know, you train a model on a structure that was, you know, released across the field up until a certain time. And, you know, one of the things that we didn’t talk about that was really critical in all this development is the PDB, which is the Protein Data Bank. It’s this common resources, basically common database where every biologist publishes their structures. And so we can, you know, train on, you know, all the structures that were put in the PDB until a certain date. And then... And then we basically look for recent structures, okay, which structures look pretty different from anything that was published before, because we really want to try to understand generalization.
Brandon [00:33:13]: And then on this new structure, we evaluate all these different models. And so you just know when AlphaFold3 was trained, you know, when you’re, you intentionally trained to the same date or something like that. Exactly. Right. Yeah.
Gabriel [00:33:24]: And so this is kind of the way that you can somewhat easily kind of compare these models, obviously, that assumes that, you know, the training. You’ve always been very passionate about validation. I remember like DiffDoc, and then there was like DiffDocL and DocGen. You’ve thought very carefully about this in the past. Like, actually, I think DocGen is like a really funny story that I think, I don’t know if you want to talk about that. It’s an interesting like... Yeah, I think one of the amazing things about putting things open source is that we get a ton of feedback from the field. And, you know, sometimes we get kind of great feedback of people. Really like... But honestly, most of the times, you know, to be honest, that’s also maybe the most useful feedback is, you know, people sharing about where it doesn’t work. And so, you know, at the end of the day, it’s critical. And this is also something, you know, across other fields of machine learning. It’s always critical to set, to do progress in machine learning, set clear benchmarks. And as, you know, you start doing progress of certain benchmarks, then, you know, you need to improve the benchmarks and make them harder and harder. And this is kind of the progression of, you know, how the field operates. And so, you know, the example of DocGen was, you know, we published this initial model called DiffDoc in my first year of PhD, which was sort of like, you know, one of the early models to try to predict kind of interactions between proteins, small molecules, that we bought a year after AlphaFold2 was published. And now, on the one hand, you know, on these benchmarks that we were using at the time, DiffDoc was doing really well, kind of, you know, outperforming kind of some of the traditional physics-based methods. But on the other hand, you know, when we started, you know, kind of giving these tools to kind of many biologists, and one example was that we collaborated with was the group of Nick Polizzi at Harvard. We noticed, started noticing that there was this clear, pattern where four proteins that were very different from the ones that we’re trained on, the models was, was struggling. And so, you know, that seemed clear that, you know, this is probably kind of where we should, you know, put our focus on. And so we first developed, you know, with Nick and his group, a new benchmark, and then, you know, went after and said, okay, what can we change? And kind of about the current architecture to improve this pattern and generalization. And this is the same that, you know, we’re still doing today, you know, kind of, where does the model not work, you know, and then, you know, once we have that benchmark, you know, let’s try to, through everything we, any ideas that we have of the problem.
RJ [00:36:15]: And there’s a lot of like healthy skepticism in the field, which I think, you know, is, is, is great. And I think, you know, it’s very clear that there’s a ton of things, the models don’t really work well on, but I think one thing that’s probably, you know, undeniable is just like the pace of, pace of progress, you know, and how, how much better we’re getting, you know, every year. And so I think if you, you know, if you assume, you know, any constant, you know, rate of progress moving forward, I think things are going to look pretty cool at some point in the future.
Gabriel [00:36:42]: ChatGPT was only three years ago. Yeah, I mean, it’s wild, right?
RJ [00:36:45]: Like, yeah, yeah, yeah, it’s one of those things. Like, you’ve been doing this. Being in the field, you don’t see it coming, you know? And like, I think, yeah, hopefully we’ll, you know, we’ll, we’ll continue to have as much progress we’ve had the past few years.
Brandon [00:36:55]: So this is maybe an aside, but I’m really curious, you get this great feedback from the, from the community, right? By being open source. My question is partly like, okay, yeah, if you open source and everyone can copy what you did, but it’s also maybe balancing priorities, right? Where you, like all my customers are saying. I want this, there’s all these problems with the model. Yeah, yeah. But my customers don’t care, right? So like, how do you, how do you think about that? Yeah.
Gabriel [00:37:26]: So I would say a couple of things. One is, you know, part of our goal with Bolts and, you know, this is also kind of established as kind of the mission of the public benefit company that we started is to democratize the access to these tools. But one of the reasons why we realized that Bolts needed to be a company, it couldn’t just be an academic project is that putting a model on GitHub is definitely not enough to get, you know, chemists and biologists, you know, across, you know, both academia, biotech and pharma to use your model to, in their therapeutic programs. And so a lot of what we think about, you know, at Bolts beyond kind of the, just the models is thinking about all the layers. The layers that come on top of the models to get, you know, from, you know, those models to something that can really enable scientists in the industry. And so that goes, you know, into building kind of the right kind of workflows that take in kind of, for example, the data and try to answer kind of directly that those problems that, you know, the chemists and the biologists are asking, and then also kind of building the infrastructure. And so this to say that, you know, even with models fully open. You know, we see a ton of potential for, you know, products in the space and the critical part about a product is that even, you know, for example, with an open source model, you know, running the model is not free, you know, as we were saying, these are pretty expensive model and especially, and maybe we’ll get into this, you know, these days we’re seeing kind of pretty dramatic inference time scaling of these models where, you know, the more you run them, the better the results are. But there, you know, you see. You start getting into a point that compute and compute costs becomes a critical factor. And so putting a lot of work into building the right kind of infrastructure, building the optimizations and so on really allows us to provide, you know, a much better service potentially to the open source models. That to say, you know, even though, you know, with a product, we can provide a much better service. I do still think, and we will continue to put a lot of our models open source because the critical kind of role. I think of open source. Models is, you know, helping kind of the community progress on the research and, you know, from which we, we all benefit. And so, you know, we’ll continue to on the one hand, you know, put some of our kind of base models open source so that the field can, can be on top of it. And, you know, as we discussed earlier, we learn a ton from, you know, the way that the field uses and builds on top of our models, but then, you know, try to build a product that gives the best experience possible to scientists. So that, you know, like a chemist or a biologist doesn’t need to, you know, spin off a GPU and, you know, set up, you know, our open source model in a particular way, but can just, you know, a bit like, you know, I, even though I am a computer scientist, machine learning scientist, I don’t necessarily, you know, take a open source LLM and try to kind of spin it off. But, you know, I just maybe open a GPT app or a cloud code and just use it as an amazing product. We kind of want to give the same experience. So this front world.
Brandon [00:40:40]: I heard a good analogy yesterday that a surgeon doesn’t want the hospital to design a scalpel, right?
Brandon [00:40:48]: So just buy the scalpel.
RJ [00:40:50]: You wouldn’t believe like the number of people, even like in my short time, you know, between AlphaFold3 coming out and the end of the PhD, like the number of people that would like reach out just for like us to like run AlphaFold3 for them, you know, or things like that. Just because like, you know, bolts in our case, you know, just because it’s like. It’s like not that easy, you know, to do that, you know, if you’re not a computational person. And I think like part of the goal here is also that, you know, we continue to obviously build the interface with computational folks, but that, you know, the models are also accessible to like a larger, broader audience. And then that comes from like, you know, good interfaces and stuff like that.
Gabriel [00:41:27]: I think one like really interesting thing about bolts is that with the release of it, you didn’t just release a model, but you created a community. Yeah. Did that community, it grew very quickly. Did that surprise you? And like, what is the evolution of that community and how is that fed into bolts?
RJ [00:41:43]: If you look at its growth, it’s like very much like when we release a new model, it’s like, there’s a big, big jump, but yeah, it’s, I mean, it’s been great. You know, we have a Slack community that has like thousands of people on it. And it’s actually like self-sustaining now, which is like the really nice part because, you know, it’s, it’s almost overwhelming, I think, you know, to be able to like answer everyone’s questions and help. It’s really difficult, you know. The, the few people that we were, but it ended up that like, you know, people would answer each other’s questions and like, sort of like, you know, help one another. And so the Slack, you know, has been like kind of, yeah, self, self-sustaining and that’s been, it’s been really cool to see.
RJ [00:42:21]: And, you know, that’s, that’s for like the Slack part, but then also obviously on GitHub as well. We’ve had like a nice, nice community. You know, I think we also aspire to be even more active on it, you know, than we’ve been in the past six months, which has been like a bit challenging, you know, for us. But. Yeah, the community has been, has been really great and, you know, there’s a lot of papers also that have come out with like new evolutions on top of bolts and it’s surprised us to some degree because like there’s a lot of models out there. And I think like, you know, sort of people converging on that was, was really cool. And, you know, I think it speaks also, I think, to the importance of like, you know, when, when you put code out, like to try to put a lot of emphasis and like making it like as easy to use as possible and something we thought a lot about when we released the code base. You know, it’s far from perfect, but, you know.
Brandon [00:43:07]: Do you think that that was one of the factors that caused your community to grow is just the focus on easy to use, make it accessible? I think so.
RJ [00:43:14]: Yeah. And we’ve, we’ve heard it from a few people over the, over the, over the years now. And, you know, and some people still think it should be a lot nicer and they’re, and they’re right. And they’re right. But yeah, I think it was, you know, at the time, maybe a little bit easier than, than other things.
Gabriel [00:43:29]: The other thing part, I think led to, to the community and to some extent, I think, you know, like the somewhat the trust in the community. Kind of what we, what we put out is the fact that, you know, it’s not really been kind of, you know, one model, but, and maybe we’ll talk about it, you know, after Boltz 1, you know, there were maybe another couple of models kind of released, you know, or open source kind of soon after. We kind of continued kind of that open source journey or at least Boltz 2, where we are not only improving kind of structure prediction, but also starting to do affinity predictions, understanding kind of the strength of the interactions between these different models, which is this critical component. critical property that you often want to optimize in discovery programs. And then, you know, more recently also kind of protein design model. And so we’ve sort of been building this suite of, of models that come together, interact with one another, where, you know, kind of, there is almost an expectation that, you know, we, we take very at heart of, you know, always having kind of, you know, across kind of the entire suite of different tasks, the best or across the best. model out there so that it’s sort of like our open source tool can be kind of the go-to model for everybody in the, in the industry. I really want to talk about Boltz 2, but before that, one last question in this direction, was there anything about the community which surprised you? Were there any, like, someone was doing something and you’re like, why would you do that? That’s crazy. Or that’s actually genius. And I never would have thought about that.
RJ [00:45:01]: I mean, we’ve had many contributions. I think like some of the. Interesting ones, like, I mean, we had, you know, this one individual who like wrote like a complex GPU kernel, you know, for part of the architecture on a piece of, the funny thing is like that piece of the architecture had been there since AlphaFold 2, and I don’t know why it took Boltz for this, you know, for this person to, you know, to decide to do it, but that was like a really great contribution. We’ve had a bunch of others, like, you know, people figuring out like ways to, you know, hack the model to do something. They click peptides, like, you know, there’s, I don’t know if there’s any other interesting ones come to mind.
Gabriel [00:45:41]: One cool one, and this was, you know, something that initially was proposed as, you know, as a message in the Slack channel by Tim O’Donnell was basically, he was, you know, there are some cases, especially, for example, we discussed, you know, antibody-antigen interactions where the models don’t necessarily kind of get the right answer. What he noticed is that, you know, the models were somewhat stuck into predicting kind of the antibodies. And so he basically ran the experiments in this model, you can condition, basically, you can give hints. And so he basically gave, you know, random hints to the model, basically, okay, you should bind to this residue, you should bind to the first residue, or you should bind to the 11th residue, or you should bind to the 21st residue, you know, basically every 10 residues scanning the entire antigen.
Brandon [00:46:33]: Residues are the...
Gabriel [00:46:34]: The amino acids. The amino acids, yeah. So the first amino acids. The 11 amino acids, and so on. So it’s sort of like doing a scan, and then, you know, conditioning the model to predict all of them, and then looking at the confidence of the model in each of those cases and taking the top. And so it’s sort of like a very somewhat crude way of doing kind of inference time search. But surprisingly, you know, for antibody-antigen prediction, it actually kind of helped quite a bit. And so there’s some, you know, interesting ideas that, you know, obviously, as kind of developing the model, you say kind of, you know, wow. This is why would the model, you know, be so dumb. But, you know, it’s very interesting. And that, you know, leads you to also kind of, you know, start thinking about, okay, how do I, can I do this, you know, not with this brute force, but, you know, in a smarter way.
RJ [00:47:22]: And so we’ve also done a lot of work on that direction. And that speaks to, like, the, you know, the power of scoring. We’re seeing that a lot. I’m sure we’ll talk about it more when we talk about BullsGen. But, you know, our ability to, like, take a structure and determine that that structure is, like... Good. You know, like, somewhat accurate. Whether that’s a single chain or, like, an interaction is a really powerful way of improving, you know, the models. Like, sort of like, you know, if you can sample a ton and you assume that, like, you know, if you sample enough, you’re likely to have, like, you know, the good structure. Then it really just becomes a ranking problem. And, you know, now we’re, you know, part of the inference time scaling that Gabby was talking about is very much that. It’s like, you know, the more we sample, the more we, like, you know, the ranking model. The ranking model ends up finding something it really likes. And so I think our ability to get better at ranking, I think, is also what’s going to enable sort of the next, you know, next big, big breakthroughs. Interesting.
Brandon [00:48:17]: But I guess there’s a, my understanding, there’s a diffusion model and you generate some stuff and then you, I guess, it’s just what you said, right? Then you rank it using a score and then you finally... And so, like, can you talk about those different parts? Yeah.
Gabriel [00:48:34]: So, first of all, like, the... One of the critical kind of, you know, beliefs that we had, you know, also when we started working on Boltz 1 was sort of like the structure prediction models are somewhat, you know, our field version of some foundation models, you know, learning about kind of how proteins and other molecules interact. And then we can leverage that learning to do all sorts of other things. And so with Boltz 2, we leverage that learning to do affinity predictions. So understanding kind of, you know, if I give you this protein, this molecule. How tightly is that interaction? For Boltz 1, what we did was taking kind of that kind of foundation models and then fine tune it to predict kind of entire new proteins. And so the way basically that that works is sort of like instead of for the protein that you’re designing, instead of fitting in an actual sequence, you fit in a set of blank tokens. And you train the models to, you know, predict both the structure of kind of that protein. The structure also, what the different amino acids of that proteins are. And so basically the way that Boltz 1 operates is that you feed a target protein that you may want to kind of bind to or, you know, another DNA, RNA. And then you feed the high level kind of design specification of, you know, what you want your new protein to be. For example, it could be like an antibody with a particular framework. It could be a peptide. It could be many other things. And that’s with natural language or? And that’s, you know, basically, you know, prompting. And we have kind of this sort of like spec that you specify. And, you know, you feed kind of this spec to the model. And then the model translates this into, you know, a set of, you know, tokens, a set of conditioning to the model, a set of, you know, blank tokens. And then, you know, basically the codes as part of the diffusion models, the codes. It’s a new structure and a new sequence for your protein. And, you know, basically, then we take that. And as Jeremy was saying, we are trying to score it and, you know, how good of a binder it is to that original target.
Brandon [00:50:51]: You’re using basically Boltz to predict the folding and the affinity to that molecule. So and then that kind of gives you a score? Exactly.
Gabriel [00:51:03]: So you use this model to predict the folding. And then you do two things. One is that you predict the structure and with something like Boltz2, and then you basically compare that structure with what the model predicted, what Boltz2 predicted. And this is sort of like in the field called consistency. It’s basically you want to make sure that, you know, the structure that you’re predicting is actually what you’re trying to design. And that gives you a much better confidence that, you know, that’s a good design. And so that’s the first filtering. And the second filtering that we did as part of kind of the Boltz2 pipeline that was released is that we look at the confidence that the model has in the structure. Now, unfortunately, kind of going to your question of, you know, predicting affinity, unfortunately, confidence is not a very good predictor of affinity. And so one of the things that we’ve actually done a ton of progress, you know, since we released Boltz2.
Brandon [00:52:03]: And kind of we have some new results that we are going to kind of announce soon is kind of, you know, the ability to get much better hit rates when instead of, you know, trying to rely on confidence of the model, we are actually directly trying to predict the affinity of that interaction. Okay. Just backing up a minute. So your diffusion model actually predicts not only the protein sequence, but also the folding of it. Exactly.
Gabriel [00:52:32]: And actually, you can... One of the big different things that we did compared to other models in the space, and, you know, there were some papers that had already kind of done this before, but we really scaled it up was, you know, basically somewhat merging kind of the structure prediction and the sequence prediction into almost the same task. And so the way that Boltz2 works is that you are basically the only thing that you’re doing is predicting the structure. So the only sort of... Supervision is we give you a supervision on the structure, but because the structure is atomic and, you know, the different amino acids have a different atomic composition, basically from the way that you place the atoms, we also understand not only kind of the structure that you wanted, but also the identity of the amino acid that, you know, the models believed was there. And so we’ve basically, instead of, you know, having these two supervision signals, you know, one discrete, one continuous. That somewhat, you know, don’t interact well together. We sort of like build kind of like an encoding of, you know, sequences in structures that allows us to basically use exactly the same supervision signal that we were using to Boltz2 that, you know, you know, largely similar to what AlphaVol3 proposed, which is very scalable. And we can use that to design new proteins. Oh, interesting.
RJ [00:53:58]: Maybe a quick shout out to Hannes Stark on our team who like did all this work. Yeah.
Gabriel [00:54:04]: Yeah, that was a really cool idea. I mean, like looking at the paper and there’s this is like encoding or you just add a bunch of, I guess, kind of atoms, which can be anything, and then they get sort of rearranged and then basically plopped on top of each other so that and then that encodes what the amino acid is. And there’s sort of like a unique way of doing this. It was that was like such a really such a cool, fun idea.
RJ [00:54:29]: I think that idea was had existed before. Yeah, there were a couple of papers.
Gabriel [00:54:33]: Yeah, I had proposed this and and Hannes really took it to the large scale.
Brandon [00:54:39]: In the paper, a lot of the paper for Boltz2Gen is dedicated to actually the validation of the model. In my opinion, all the people we basically talk about feel that this sort of like in the wet lab or whatever the appropriate, you know, sort of like in real world validation is the whole problem or not the whole problem, but a big giant part of the problem. So can you talk a little bit about the highlights? From there, that really because to me, the results are impressive, both from the perspective of the, you know, the model and also just the effort that went into the validation by a large team.
Gabriel [00:55:18]: First of all, I think I should start saying is that both when we were at MIT and Thomas Yacolas and Regina Barzillai’s lab, as well as at Boltz, you know, we are not a we’re not a biolab and, you know, we are not a therapeutic company. And so to some extent, you know, we were first forced to, you know, look outside of, you know, our group, our team to do the experimental validation. One of the things that really, Hannes, in the team pioneer was the idea, OK, can we go not only to, you know, maybe a specific group and, you know, trying to find a specific system and, you know, maybe overfit a bit to that system and trying to validate. But how can we test this model? So. Across a very wide variety of different settings so that, you know, anyone in the field and, you know, printing design is, you know, such a kind of wide task with all sorts of different applications from therapeutic to, you know, biosensors and many others that, you know, so can we get a validation that is kind of goes across many different tasks? And so he basically put together, you know, I think it was something like, you know, 25 different. You know, academic and industry labs that committed to, you know, testing some of the designs from the model and some of this testing is still ongoing and, you know, giving results kind of back to us in exchange for, you know, hopefully getting some, you know, new great sequences for their task. And he was able to, you know, coordinate this, you know, very wide set of, you know, scientists and already in the paper, I think we. Shared results from, I think, eight to 10 different labs kind of showing results from, you know, designing peptides, designing to target, you know, ordered proteins, peptides targeting disordered proteins, which are results, you know, of designing proteins that bind to small molecules, which are results of, you know, designing nanobodies and across a wide variety of different targets. And so that’s sort of like. That gave to the paper a lot of, you know, validation to the model, a lot of validation that was kind of wide.
Brandon [00:57:39]: And so those would be therapeutics for those animals or are they relevant to humans as well? They’re relevant to humans as well.
Gabriel [00:57:45]: Obviously, you need to do some work into, quote unquote, humanizing them, making sure that, you know, they have the right characteristics to so they’re not toxic to humans and so on.
RJ [00:57:57]: There are some approved medicine in the market that are nanobodies. There’s a general. General pattern, I think, in like in trying to design things that are smaller, you know, like it’s easier to manufacture at the same time, like that comes with like potentially other challenges, like maybe a little bit less selectivity than like if you have something that has like more hands, you know, but the yeah, there’s this big desire to, you know, try to design many proteins, nanobodies, small peptides, you know, that just are just great drug modalities.
Brandon [00:58:27]: Okay. I think we were left off. We were talking about validation. Validation in the lab. And I was very excited about seeing like all the diverse validations that you’ve done. Can you go into some more detail about them? Yeah. Specific ones. Yeah.
RJ [00:58:43]: The nanobody one. I think we did. What was it? 15 targets. Is that correct? 14. 14 targets. Testing. So we typically the way this works is like we make a lot of designs. All right. On the order of like tens of thousands. And then we like rank them and we pick like the top. And in this case, and was 15 right for each target and then we like measure sort of like the success rates, both like how many targets we were able to get a binder for and then also like more generally, like out of all of the binders that we designed, how many actually proved to be good binders. Some of the other ones I think involved like, yeah, like we had a cool one where there was a small molecule or design a protein that binds to it. That has a lot of like interesting applications, you know, for example. Like Gabri mentioned, like biosensing and things like that, which is pretty cool. We had a disordered protein, I think you mentioned also. And yeah, I think some of those were some of the highlights. Yeah.
Gabriel [00:59:44]: So I would say that the way that we structure kind of some of those validations was on the one end, we have validations across a whole set of different problems that, you know, the biologists that we were working with came to us with. So we were trying to. For example, in some of the experiments, design peptides that would target the RACC, which is a target that is involved in metabolism. And we had, you know, a number of other applications where we were trying to design, you know, peptides or other modalities against some other therapeutic relevant targets. We designed some proteins to bind small molecules. And then some of the other testing that we did was really trying to get like a more broader sense. So how does the model work, especially when tested, you know, on somewhat generalization? So one of the things that, you know, we found with the field was that a lot of the validation, especially outside of the validation that was on specific problems, was done on targets that have a lot of, you know, known interactions in the training data. And so it’s always a bit hard to understand, you know, how much are these models really just regurgitating kind of what they’ve seen or trying to imitate. What they’ve seen in the training data versus, you know, really be able to design new proteins. And so one of the experiments that we did was to take nine targets from the PDB, filtering to things where there is no known interaction in the PDB. So basically the model has never seen kind of this particular protein bound or a similar protein bound to another protein. So there is no way that. The model from its training set can sort of like say, okay, I’m just going to kind of tweak something and just imitate this particular kind of interaction. And so we took those nine proteins. We worked with adaptive CRO and basically tested, you know, 15 mini proteins and 15 nanobodies against each one of them. And the very cool thing that we saw was that on two thirds of those targets, we were able to, from this 15 design, get nanomolar binders, nanomolar, roughly speaking, just a measure of, you know, how strongly kind of the interaction is, roughly speaking, kind of like a nanomolar binder is approximately the kind of binding strength or binding that you need for a therapeutic. Yeah. So maybe switching directions a bit. Bolt’s lab was just announced this week or was it last week? Yeah. This is like your. First, I guess, product, if that’s if you want to call it that. Can you talk about what Bolt’s lab is and yeah, you know, what you hope that people take away from this? Yeah.
RJ [01:02:44]: You know, as we mentioned, like I think at the very beginning is the goal with the product has been to, you know, address what the models don’t on their own. And there’s largely sort of two categories there. I’ll split it in three. The first one. It’s one thing to predict, you know, a single interaction, for example, like a single structure. It’s another to like, you know, very effectively search a space, a design space to produce something of value. What we found, like sort of building on this product is that there’s a lot of steps involved, you know, in that there’s certainly need to like, you know, accompany the user through, you know, one of those steps, for example, is like, you know, the creation of the target itself. You know, how do we make sure that the model has like a good enough understanding of the target? So we can like design something and there’s all sorts of tricks, you know, that you can do to improve like a particular, you know, structure prediction. And so that’s sort of like, you know, the first stage. And then there’s like this stage of like, you know, designing and searching the space efficiently. You know, for something like BullsGen, for example, like you, you know, you design many things and then you rank them, for example, for small molecule process, a little bit more complicated. We actually need to also make sure that the molecules are synthesizable. And so the way we do that is that, you know, we have a generative model that learns. To use like appropriate building blocks such that, you know, it can design within a space that we know is like synthesizable. And so there’s like, you know, this whole pipeline really of different models involved in being able to design a molecule. And so that’s been sort of like the first thing we call them agents. We have a protein agent and we have a small molecule design agents. And that’s really like at the core of like what powers, you know, the BullsLab platform.
Brandon [01:04:22]: So these agents, are they like a language model wrapper or they’re just like your models and you’re just calling them agents? A lot. Yeah. Because they, they, they sort of perform a function on behalf of.
RJ [01:04:33]: They’re more of like a, you know, a recipe, if you wish. And I think we use that term sort of because of, you know, sort of the complex pipelining and automation, you know, that goes into like all this plumbing. So that’s the first part of the product. The second part is the infrastructure. You know, we need to be able to do this at very large scale for any one, you know, group that’s doing a design campaign. Let’s say you’re designing, you know, I’d say a hundred thousand possible candidates. Right. To find the good one that is, you know, a very large amount of compute, you know, for small molecules, it’s on the order of like a few seconds per designs for proteins can be a bit longer. And so, you know, ideally you want to do that in parallel, otherwise it’s going to take you weeks. And so, you know, we’ve put a lot of effort into like, you know, our ability to have a GPU fleet that allows any one user, you know, to be able to do this kind of like large parallel search.
Brandon [01:05:23]: So you’re amortizing the cost over your users. Exactly. Exactly.
RJ [01:05:27]: And, you know, to some degree, like it’s whether you. Use 10,000 GPUs for like, you know, a minute is the same cost as using, you know, one GPUs for God knows how long. Right. So you might as well try to parallelize if you can. So, you know, a lot of work has gone, has gone into that, making it very robust, you know, so that we can have like a lot of people on the platform doing that at the same time. And the third one is, is the interface and the interface comes in, in two shapes. One is in form of an API and that’s, you know, really suited for companies that want to integrate, you know, these pipelines, these agents.
RJ [01:06:01]: So we’re already partnering with, you know, a few distributors, you know, that are gonna integrate our API. And then the second part is the user interface. And, you know, we, we’ve put a lot of thoughts also into that. And this is when I, I mentioned earlier, you know, this idea of like broadening the audience. That’s kind of what the, the user interface is about. And we’ve built a lot of interesting features in it, you know, for example, for collaboration, you know, when you have like potentially multiple medicinal chemists or. We’re going through the results and trying to pick out, okay, like what are the molecules that we’re going to go and test in the lab? It’s powerful for them to be able to, you know, for example, each provide their own ranking and then do consensus building. And so there’s a lot of features around launching these large jobs, but also around like collaborating on analyzing the results that we try to solve, you know, with that part of the platform. So Bolt’s lab is sort of a combination of these three objectives into like one, you know, sort of cohesive platform. Who is this accessible to? Everyone. You do need to request access today. We’re still like, you know, sort of ramping up the usage, but anyone can request access. If you are an academic in particular, we, you know, we provide a fair amount of free credit so you can play with the platform. If you are a startup or biotech, you may also, you know, reach out and we’ll typically like actually hop on a call just to like understand what you’re trying to do and also provide a lot of free credit to get started. And of course, also with larger companies, we can deploy this platform in a more like secure environment. And so that’s like more like customizing. You know, deals that we make, you know, with the partners, you know, and that’s sort of the ethos of Bolt. I think this idea of like servicing everyone and not necessarily like going after just, you know, the really large enterprises. And that starts from the open source, but it’s also, you know, a key design principle of the product itself.
Gabriel [01:07:48]: One thing I was thinking about with regards to infrastructure, like in the LLM space, you know, the cost of a token has gone down by I think a factor of a thousand or so over the last three years, right? Yeah. And is it possible that like essentially you can exploit economies of scale and infrastructure that you can make it cheaper to run these things yourself than for any person to roll their own system? A hundred percent. Yeah.
RJ [01:08:08]: I mean, we’re already there, you know, like running Bolts on our platform, especially on a large screen is like considerably cheaper than it would probably take anyone to put the open source model out there and run it. And on top of the infrastructure, like one of the things that we’ve been working on is accelerating the models. So, you know. Our small molecule screening pipeline is 10x faster on Bolts Lab than it is in the open source, you know, and that’s also part of like, you know, building a product, you know, of something that scales really well. And we really wanted to get to a point where like, you know, we could keep prices very low in a way that it would be a no-brainer, you know, to use Bolts through our platform.
Gabriel [01:08:52]: How do you think about validation of your like agentic systems? Because, you know, as you were saying earlier. Like we’re AlphaFold style models are really good at, let’s say, monomeric, you know, proteins where you have, you know, co-evolution data. But now suddenly the whole point of this is to design something which doesn’t have, you know, co-evolution data, something which is really novel. So now you’re basically leaving the domain that you thought was, you know, that you know you are good at. So like, how do you validate that?
RJ [01:09:22]: Yeah, I like every complete, but there’s obviously, you know, a ton of computational metrics. That we rely on, but those are only take you so far. You really got to go to the lab, you know, and test, you know, okay, with this method A and this method B, how much better are we? You know, how much better is my, my hit rate? How stronger are my binders? Also, it’s not just about hit rate. It’s also about how good the binders are. And there’s really like no way, nowhere around that. I think we’re, you know, we’ve really ramped up the amount of experimental validation that we do so that we like really track progress, you know, as scientifically sound, you know. Yeah. As, as possible out of this, I think.
Gabriel [01:10:00]: Yeah, no, I think, you know, one thing that is unique about us and maybe companies like us is that because we’re not working on like maybe a couple of therapeutic pipelines where, you know, our validation would be focused on those. We, when we do an experimental validation, we try to test it across tens of targets. And so that on the one end, we can get a much more statistically significant result and, and really allows us to make progress. From the methodological side without being, you know, steered by, you know, overfitting on any one particular system. And of course we choose, you know, we always try to choose targets and problems are sort of like at the frontier of what’s possible today. So, you know, you don’t want something too easy. You don’t want something too hard. Otherwise you’re not going to see progress. And so, you know, this is a somewhat evolving set of targets. We talked earlier about the targets that we looked at with, with Boltchan. And now we are even trying kind of, you know, even harder targets, both for small molecule and proteins. And so we try to keep ourselves on the, on the boundary of what’s possible. So do you have like infrastructure or is this is like, you just have a lot of different partnerships with academic labs and you’re just kind of keep pushing on these and driving these. We do partially this through academic labs more and more. We do this through CROs just because of, you know, to some extent is also, we need kind of replicability often kind of, you know, going after the same time. So we try to, we try to keep our, our targets, you know, multiple times and, you know, to see the, the progress from, you know, one month to the next. And speed. And speed. And speed. Speed of execution. Yeah. And, So what happens if you start getting a bunch of like really strong biters against therapeutic targets? What do you do?
RJ [01:11:43]: Release them. Yeah.
Gabriel [01:11:45]: But you can release them in open source? Like,
RJ [01:11:47]: Yeah, I mean, you know, I mean, when we say we have no interest in making dress, we’re serious. Like, you know, uh, I mean, when it, when it was with the academic labs, basically the, you know, I was, they keep it, they do a lot of it.
Gabriel [01:12:02]: I will also say, and I think this has been a bit of the issue that I have with some of the things that have been said in the field, is when we say that we design new proteins or we say that we design new molecules, go and bind these particular targets. We should be very clear, these are not drugs. These are not things that are ready to be put into a human. And there is still a lot of development that goes with it. And so this is kind of to us, we see ourselves as building tools for scientists. At the end of the day, it really relies on the scientist having a great therapeutic hypothesis and then pushing through kind of all the stages of development. And, you know, we try to build tools that can accompany them in that journey. It’s not like a magic box where, you know, you can just turn it and get FDA approved drugs.
Brandon [01:13:06]: But actually, that brings up an interesting question that I’ve been wondering about is, do you guys see yourself staying in this, for lack of a better way of saying it, layer? Or do you think that you’ll start to... Yeah. Either on the physical sense, looking at different layers of the virtual cell, so to speak, or also, you know, so there’s like the development process that goes, you know, sort of like design preclinical, clinical approval and thinking about improving the performance throughout that process based on the designs. Is that a direction that you guys are pushing? Yeah.
Gabriel [01:13:45]: So one of the things, as Jeremy said, you know, we are... We are not a therapeutic company. We want to kind of stay not to be a therapeutic company, always be at the service of, you know, all the different, you know, companies, including therapeutic companies that we serve. And, you know, that to some extent does mean, you know, that we need to try to, you know, go deeper and deeper in getting these models better and better. One of the things that we are doing across, you know, many other in the field is, you know, now that we are really... They’re starting to be good, both for small molecule and... For proteins to design kind of binders, design relatively tight binders, is starting to look at all these other properties, you know, they’re called developpabilities or at me that, you know, we care about when developing a drug and try, can we design them from, from Gageco. The thing about those properties in some of them, you know, you need to, you know, start having an understanding of the cell. And so that’s on the one hand, kind of why we need that understanding. But also, you know, the way... The way that we also think about all different and complex diseases is that these models, then these tools that we’re building have a good understanding of kind of, you know, biomolecular interactions and kind of their interactions. Now, at the same time, every disease is often kind of unique and every therapeutic hypothesis is unique. And so you maybe want to have something that needs to hit the particular, you know, let’s say target in a virus in a particular way, but you don’t maybe know exactly. So you can start to have a more open-minded understanding of what’s, what’s a way you want to do. And so maybe in the first set of designs, you’re going to try to target different epitopes in different ways, and then you’re going to test them in the lab, maybe directly in vivo, and you’re going to see which ones work and which ones don’t. And so then you need to bring those results back into the models. And then the models can start to have a more wider understanding, you know, not just of the biophysical of the antibodies interacting with that target, but also how that is shaped within the cell. And so first of all, you know, that means on the one end that we need, you know, kind of these loops, and this is also partially how we, we designed the platform to be. But that also means that we also need to start understanding more and more kind of higher level things. And, you know, I wouldn’t say that we’re working in any way on like a virtual cell like others are, but we’re definitely thinking kind of very deeply about kind of, you know, how does, you know, kind of the way that we target certain proteins. Interfere, interact with, you know, maybe pathways that are existing in the cell. One question that has come up is you talk a lot about user interface and so on. And I think this is really important, but like my experience with dealing with medicinal chemists, when you get the machine learning models, is they are the most superstitious, skeptical, like pseudo-religious people I’ve ever talked to when it comes to doing science. Sorry for the medicinal chemists listening. Yeah, they’re amazing. Like, they’re absolutely, I’ve worked with some spectacular medicinal chemists who just pull magic out of their hat again and again, and I have no idea how they do it. But when you bring them a machine learning model, it is sometimes quite tricky to get them to deal with it. How has your interaction been with this? And how have you thought about, like, building Bolt’s lab to work with the skeptics? One of the great value unlocks for us and for our product has been when we brought to the team a medicinal chemist. His name is Jeffrey. So I think kind of like on the one hand, you know, day one, you know, he obviously had a lot of opinions on kind of a lot of the ways that we should change, you know, both kind of the way that the agents worked, the way that the platform worked. But it’s been really amazing kind of, you know, once also we started kind of shaping kind of the platform in a better way with this feedback, how we went from, you know, to some extent, you know, a fair skepticism to him, you know, actually using, you know, a lot of the things that we did. Yeah. So he’s doing a lot more compute than any of our computational folks in the team, you know, at times that, you know, he’s, you know, running, you know, he has all these sort of hypotheses. Okay, maybe I can hit this protein this particular way. I can hit in that way. Actually, let me look at for this particular molecular space. Let me try to optimize for this particular interactions. So he ends up, you know, running several screens in parallel, you know, using hundreds of GPUs, you know, on his own. And, you know, so this has been, you know, pretty incredible to see kind of how, you know, maybe the way that I was more thinking about a problem, which is, okay, you’re just trying to design a binder, a small molecule to a particular protein. The way that he thinks about it is, you know, much more deeply and, you know, trying all these different things, these different hypotheses. And then, you know, once he gets the results from the model, he doesn’t just, you know, take the top 15, but he really kind of looks over and, you know, kind of tries to understand, you know, the different things. And then when we select, you know, maybe some designs to bring forth, you know, he has, you know, something where, you know, both the models understand that something’s good, but himself as well. And that’s why we also built kind of the platform to be an interface for, you know, this kind of chemist and, you know, also like engineers. Yeah. Collaborative experience.
RJ [01:19:09]: I think at the end of the day, like, you know, for people to be convinced, you have to show them something that they didn’t think was possible. And until you have that aha moment, you know, I think the skepticism will remain. But then when, you know, every once in a while, I think there’s like a result that like really surprises people. And then it’s like, oh, wow, okay, this is actually, I can do something with this. So you just get in their hands, have them try it out, and they’ll be convinced. Yeah, or like maybe once the lab results come back. Or their friends. Yeah, or maybe one of their colleagues is convinced. Yeah. I think it takes going to the lab at some point. There’s no avoiding that, you know, as beautiful as the platform can be, as nice as the molecules might look, you know, that the model predicted. I think what really convinces people is like, you know, hits. Yeah.
Gabriel [01:19:54]: Yeah. You see the results. Exactly. Yeah. Cool. Thank you for, you know, taking the time to chat with us. Yeah. You know, is there anything that you would like your audience to know? I mean, first of all, you know, we’re just getting started, you know, continuing to build a team. And so definitely always looking for great folks, both on the kind of, you know, software side, you know, machine learning side, but also scientists to join the team and help us, you know, shape. On the infrastructure side, too. Indeed. If you think that if you want a new challenge, because this is not just next token prediction, this is really a new engineering challenge. Exactly. Yeah. If you, if no matter, you know, how much experience you have with, you know, biologists and chemistry, if you want to come, you know, help us in a shape, what, you know, biology and chemistry, hopefully we’ll look like in five, 10 years. We’d love to hear from you. And so go to boltz.bio and, you know, come join the team. Cool. Thank you. Awesome. Thank you so much. Thank you.


Get full access to Latent.Space at www.latent.space/subscribe

From Palantir and Two Sigma to building Goodfire into the poster-child for actionable mechanistic interpretability, Mark Bissell (Member of Technical Staff) and Myra Deng (Head of Product) are trying to turn “peeking inside the model” into a repeatable production workflow by shipping APIs, landing real enterprise deployments, and now scaling the bet with a recent $150M Series B funding round at a $1.25B valuation.
In this episode, we go far beyond the usual “SAEs are cool” take. We talk about Goodfire’s core bet: that the AI lifecycle is still fundamentally broken because the only reliable control we have is data and we post-train, RLHF, and fine-tune by “slurping supervision through a straw,” hoping the model picks up the right behaviors while quietly absorbing the wrong ones. Goodfire’s answer is to build a bi-directional interface between humans and models: read what’s happening inside, edit it surgically, and eventually use interpretability during training so customization isn’t just brute-force guesswork.
Mark and Myra walk through what that looks like when you stop treating interpretability like a lab demo and start treating it like infrastructure: lightweight probes that add near-zero latency, token-level safety filters that can run at inference time, and interpretability workflows that survive messy constraints (multilingual inputs, synthetic→real transfer, regulated domains, no access to sensitive data). We also get a live window into what “frontier-scale interp” means operationally (i.e. steering a trillion-parameter model in real time by targeting internal features) plus why the same tooling generalizes cleanly from language models to genomics, medical imaging, and “pixel-space” world models.
We discuss:
* Myra + Mark’s path: Palantir (health systems, forward-deployed engineering) → Goodfire early team; Two Sigma → Head of Product, translating frontier interpretability research into a platform and real-world deployments
* What “interpretability” actually means in practice: not just post-hoc poking, but a broader “science of deep learning” approach across the full AI lifecycle (data curation → post-training → internal representations → model design)
* Why post-training is the first big wedge: “surgical edits” for unintended behaviors likereward hacking, sycophancy, noise learned during customization plus the dream of targeted unlearning and bias removal without wrecking capabilities
* SAEs vs probes in the real world: why SAE feature spaces sometimes underperform classifiers trained on raw activations for downstream detection tasks (hallucination, harmful intent, PII), and what that implies about “clean concept spaces”
* Rakuten in production: deploying interpretability-based token-level PII detection at inference time to prevent routing private data to downstream providers plus the gnarly constraints: no training on real customer PII, synthetic→real transfer, English + Japanese, and tokenization quirks
* Why interp can be operationally cheaper than LLM-judge guardrails: probes are lightweight, low-latency, and don’t require hosting a second large model in the loop
* Real-time steering at frontier scale: a demo of steering Kimi K2 (~1T params) live and finding features via SAE pipelines, auto-labeling via LLMs, and toggling a “Gen-Z slang” feature across multiple layers without breaking tool use
* Hallucinations as an internal signal: the case that models have latent uncertainty / “user-pleasing” circuitry you can detect and potentially mitigate more directly than black-box methods
* Steering vs prompting: the emerging view that activation steering and in-context learning are more closely connected than people think, including work mapping between the two (even for jailbreak-style behaviors)
* Interpretability for science: using the same tooling across domains (genomics, medical imaging, materials) to debug spurious correlations and extract new knowledge up to and including early biomarker discovery work with major partners
* World models + “pixel-space” interpretability: why vision/video models make concepts easier to see, how that accelerates the feedback loop, and why robotics/world-model partners are especially interesting design partners
* The north star: moving from “data in, weights out” to intentional model design where experts can impart goals and constraints directly, not just via reward signals and brute-force post-training
—
Goodfire AI
* Website: https://goodfire.ai
* LinkedIn: https://www.linkedin.com/company/goodfire-ai/
* X: https://x.com/GoodfireAI
Myra Deng
* Website: https://myradeng.com/
* LinkedIn: https://www.linkedin.com/in/myra-deng/
* X: https://x.com/myra_deng
Mark Bissell
* LinkedIn: https://www.linkedin.com/in/mark-bissell/
* X: https://x.com/MarkMBissell
Full Video Episode
Timestamps
00:00:00 Introduction
00:00:05 Introduction to the Latent Space Podcast and Guests from Goodfire
00:00:29 What is Goodfire? Mission and Focus on Interpretability
00:01:01 Goodfire’s Practical Approach to Interpretability
00:01:37 Goodfire’s Series B Fundraise Announcement
00:02:04 Backgrounds of Mark and Myra from Goodfire
00:02:51 Team Structure and Roles at Goodfire
00:05:13 What is Interpretability? Definitions and Techniques
00:05:30 Understanding Errors
00:07:29 Post-training vs. Pre-training Interpretability Applications
00:08:51 Using Interpretability to Remove Unwanted Behaviors
00:10:09 Grokking, Double Descent, and Generalization in Models
00:10:15 404 Not Found Explained
00:12:06 Subliminal Learning and Hidden Biases in Models
00:14:07 How Goodfire Chooses Research Directions and Projects
00:15:00 Troubleshooting Errors
00:16:04 Limitations of SAEs and Probes in Interpretability
00:18:14 Rakuten Case Study: Production Deployment of Interpretability
00:20:45 Conclusion
00:21:12 Efficiency Benefits of Interpretability Techniques
00:21:26 Live Demo: Real-Time Steering in a Trillion Parameter Model
00:25:15 How Steering Features are Identified and Labeled
00:26:51 Detecting and Mitigating Hallucinations Using Interpretability
00:31:20 Equivalence of Activation Steering and Prompting
00:34:06 Comparing Steering with Fine-Tuning and LoRA Techniques
00:36:04 Model Design and the Future of Intentional AI Development
00:38:09 Getting Started in Mechinterp: Resources, Programs, and Open Problems
00:40:51 Industry Applications and the Rise of Mechinterp in Practice
00:41:39 Interpretability for Code Models and Real-World Usage
00:43:07 Making Steering Useful for More Than Stylistic Edits
00:46:17 Applying Interpretability to Healthcare and Scientific Discovery
00:49:15 Why Interpretability is Crucial in High-Stakes Domains like Healthcare
00:52:03 Call for Design Partners Across Domains
00:54:18 Interest in World Models and Visual Interpretability
00:57:22 Sci-Fi Inspiration: Ted Chiang and Interpretability
01:00:14 Interpretability, Safety, and Alignment Perspectives
01:04:27 Weak-to-Strong Generalization and Future Alignment Challenges
01:05:38 Final Thoughts and Hiring/Collaboration Opportunities at Goodfire
Transcript
Shawn Wang [00:00:05]: So welcome to the Latent Space pod. We’re back in the studio with our special MechInterp co-host, Vibhu. Welcome. Mochi, Mochi’s special co-host. And Mochi, the mechanistic interpretability doggo. We have with us Mark and Myra from Goodfire. Welcome. Thanks for having us on. Maybe we can sort of introduce Goodfire and then introduce you guys. How do you introduce Goodfire today?
Myra Deng [00:00:29]: Yeah, it’s a great question. So Goodfire, we like to say, is an AI research lab that focuses on using interpretability to understand, learn from, and design AI models. And we really believe that interpretability will unlock the new generation, next frontier of safe and powerful AI models. That’s our description right now, and I’m excited to dive more into the work we’re doing to make that happen.
Shawn Wang [00:00:55]: Yeah. And there’s always like the official description. Is there an understatement? Is there an unofficial one that sort of resonates more with a different audience?
Mark Bissell [00:01:01]: Well, being an AI research lab that’s focused on interpretability, there’s obviously a lot of people have a lot that they think about when they think of interpretability. And I think we have a pretty broad definition of what that means and the types of places that can be applied. And in particular, applying it in production scenarios, in high stakes industries, and really taking it sort of from the research world into the real world. Which, you know. It’s a new field, so that hasn’t been done all that much. And we’re excited about actually seeing that sort of put into practice.
Shawn Wang [00:01:37]: Yeah, I would say it wasn’t too long ago that Anthopic was like still putting out like toy models or superposition and that kind of stuff. And I wouldn’t have pegged it to be this far along. When you and I talked at NeurIPS, you were talking a little bit about your production use cases and your customers. And then not to bury the lead, today we’re also announcing the fundraise, your Series B. $150 million. $150 million at a 1.25B valuation. Congrats, Unicorn.
Mark Bissell [00:02:02]: Thank you. Yeah, no, things move fast.
Shawn Wang [00:02:04]: We were talking to you in December and already some big updates since then. Let’s dive, I guess, into a bit of your backgrounds as well. Mark, you were at Palantir working on health stuff, which is really interesting because the Goodfire has some interesting like health use cases. I don’t know how related they are in practice.
Mark Bissell [00:02:22]: Yeah, not super related, but I don’t know. It was helpful context to know what it’s like. Just to work. Just to work with health systems and generally in that domain. Yeah.
Shawn Wang [00:02:32]: And Mara, you were at Two Sigma, which actually I was also at Two Sigma back in the day. Wow, nice.
Myra Deng [00:02:37]: Did we overlap at all?
Shawn Wang [00:02:38]: No, this is when I was briefly a software engineer before I became a sort of developer relations person. And now you’re head of product. What are your sort of respective roles, just to introduce people to like what all gets done in Goodfire?
Mark Bissell [00:02:51]: Yeah, prior to Goodfire, I was at Palantir for about three years as a forward deployed engineer, now a hot term. Wasn’t always that way. And as a technical lead on the health care team and at Goodfire, I’m a member of the technical staff. And honestly, that I think is about as specific as like as as I could describe myself because I’ve worked on a range of things. And, you know, it’s it’s a fun time to be at a team that’s still reasonably small. I think when I joined one of the first like ten employees, now we’re above 40, but still, it looks like there’s always a mix of research and engineering and product and all of the above. That needs to get done. And I think everyone across the team is, you know, pretty, pretty switch hitter in the roles they do. So I think you’ve seen some of the stuff that I worked on related to image models, which was sort of like a research demo. More recently, I’ve been working on our scientific discovery team with some of our life sciences partners, but then also building out our core platform for more of like flexing some of the kind of MLE and developer skills as well.
Shawn Wang [00:03:53]: Very generalist. And you also had like a very like a founding engineer type role.
Myra Deng [00:03:58]: Yeah, yeah.
Shawn Wang [00:03:59]: So I also started as I still am a member of technical staff, did a wide range of things from the very beginning, including like finding our office space and all of this, which is we both we both visited when you had that open house thing. It was really nice.
Myra Deng [00:04:13]: Thank you. Thank you. Yeah. Plug to come visit our office.
Shawn Wang [00:04:15]: It looked like it was like 200 people. It has room for 200 people. But you guys are like 10.
Myra Deng [00:04:22]: For a while, it was very empty. But yeah, like like Mark, I spend. A lot of my time as as head of product, I think product is a bit of a weird role these days, but a lot of it is thinking about how do we take our frontier research and really apply it to the most important real world problems and how does that then translate into a platform that’s repeatable or a product and working across, you know, the engineering and research teams to make that happen and also communicating to the world? Like, what is interpretability? What is it used for? What is it good for? Why is it so important? All of these things are part of my day-to-day as well.
Shawn Wang [00:05:01]: I love like what is things because that’s a very crisp like starting point for people like coming to a field. They all do a fun thing. Vibhu, why don’t you want to try tackling what is interpretability and then they can correct us.
Vibhu Sapra [00:05:13]: Okay, great. So I think like one, just to kick off, it’s a very interesting role to be head of product, right? Because you guys, at least as a lab, you’re more of an applied interp lab, right? Which is pretty different than just normal interp, like a lot of background research. But yeah. You guys actually ship an API to try these things. You have Ember, you have products around it, which not many do. Okay. What is interp? So basically you’re trying to have an understanding of what’s going on in model, like in the model, in the internal. So different approaches to do that. You can do probing, SAEs, transcoders, all this stuff. But basically you have an, you have a hypothesis. You have something that you want to learn about what’s happening in a model internals. And then you’re trying to solve that from there. You can do stuff like you can, you know, you can do activation mapping. You can try to do steering. There’s a lot of stuff that you can do, but the key question is, you know, from input to output, we want to have a better understanding of what’s happening and, you know, how can we, how can we adjust what’s happening on the model internals? How’d I do?
Mark Bissell [00:06:12]: That was really good. I think that was great. I think it’s also a, it’s kind of a minefield of a, if you ask 50 people who quote unquote work in interp, like what is interpretability, you’ll probably get 50 different answers. And. Yeah. To some extent also like where, where good fire sits in the space. I think that we’re an AI research company above all else. And interpretability is a, is a set of methods that we think are really useful and worth kind of specializing in, in order to accomplish the goals we want to accomplish. But I think we also sort of see some of the goals as even more broader as, as almost like the science of deep learning and just taking a not black box approach to kind of any part of the like AI development life cycle, whether that. That means using interp for like data curation while you’re training your model or for understanding what happened during post-training or for the, you know, understanding activations and sort of internal representations, what is in there semantically. And then a lot of sort of exciting updates that were, you know, are sort of also part of the, the fundraise around bringing interpretability to training, which I don’t think has been done all that much before. A lot of this stuff is sort of post-talk poking at models as opposed to. To actually using this to intentionally design them.
Shawn Wang [00:07:29]: Is this post-training or pre-training or is that not a useful.
Myra Deng [00:07:33]: Currently focused on post-training, but there’s no reason the techniques wouldn’t also work in pre-training.
Shawn Wang [00:07:38]: Yeah. It seems like it would be more active, applicable post-training because basically I’m thinking like rollouts or like, you know, having different variations of a model that you can tweak with the, with your steering. Yeah.
Myra Deng [00:07:50]: And I think in a lot of the news that you’ve seen in, in, on like Twitter or whatever, you’ve seen a lot of unintended. Side effects come out of post-training processes, you know, overly sycophantic models or models that exhibit strange reward hacking behavior. I think these are like extreme examples. There’s also, you know, very, uh, mundane, more mundane, like enterprise use cases where, you know, they try to customize or post-train a model to do something and it learns some noise or it doesn’t appropriately learn the target task. And a big question that we’ve always had is like, how do you use your understanding of what the model knows and what it’s doing to actually guide the learning process?
Shawn Wang [00:08:26]: Yeah, I mean, uh, you know, just to anchor this for people, uh, one of the biggest controversies of last year was 4.0 GlazeGate. I’ve never heard of GlazeGate. I didn’t know that was what it was called. The other one, they called it that on the blog post and I was like, well, how did OpenAI call it? Like officially use that term. And I’m like, that’s funny, but like, yeah, I guess it’s the pitch that if they had worked a good fire, they wouldn’t have avoided it. Like, you know what I’m saying?
Myra Deng [00:08:51]: I think so. Yeah. Yeah.
Mark Bissell [00:08:53]: I think that’s certainly one of the use cases. I think. Yeah. Yeah. I think the reason why post-training is a place where this makes a lot of sense is a lot of what we’re talking about is surgical edits. You know, you want to be able to have expert feedback, very surgically change how your model is doing, whether that is, you know, removing a certain behavior that it has. So, you know, one of the things that we’ve been looking at or is, is another like common area where you would want to make a somewhat surgical edit is some of the models that have say political bias. Like you look at Quen or, um, R1 and they have sort of like this CCP bias.
Shawn Wang [00:09:27]: Is there a CCP vector?
Mark Bissell [00:09:29]: Well, there’s, there are certainly internal, yeah. Parts of the representation space where you can sort of see where that lives. Yeah. Um, and you want to kind of, you know, extract that piece out.
Shawn Wang [00:09:40]: Well, I always say, you know, whenever you find a vector, a fun exercise is just like, make it very negative to see what the opposite of CCP is.
Mark Bissell [00:09:47]: The super America, bald eagles flying everywhere. But yeah. So in general, like lots of post-training tasks where you’d want to be able to, to do that. Whether it’s unlearning a certain behavior or, you know, some of the other kind of cases where this comes up is, are you familiar with like the, the grokking behavior? I mean, I know the machine learning term of grokking.
Shawn Wang [00:10:09]: Yeah.
Mark Bissell [00:10:09]: Sort of this like double descent idea of, of having a model that is able to learn a generalizing, a generalizing solution, as opposed to even if memorization of some task would suffice, you want it to learn the more general way of doing a thing. And so, you know, another. A way that you can think about having surgical access to a model’s internals would be learn from this data, but learn in the right way. If there are many possible, you know, ways to, to do that. Can make interp solve the double descent problem?
Shawn Wang [00:10:41]: Depends, I guess, on how you. Okay. So I, I, I viewed that double descent as a problem because then you’re like, well, if the loss curves level out, then you’re done, but maybe you’re not done. Right. Right. But like, if you actually can interpret what is a generalizing or what you’re doing. What is, what is still changing, even though the loss is not changing, then maybe you, you can actually not view it as a double descent problem. And actually you’re just sort of translating the space in which you view loss and like, and then you have a smooth curve. Yeah.
Mark Bissell [00:11:11]: I think that’s certainly like the domain of, of problems that we’re, that we’re looking to get.
Shawn Wang [00:11:15]: Yeah. To me, like double descent is like the biggest thing to like ML research where like, if you believe in scaling, then you don’t need, you need to know where to scale. And. But if you believe in double descent, then you don’t, you don’t believe in anything where like anything levels off, like.
Vibhu Sapra [00:11:30]: I mean, also tendentially there’s like, okay, when you talk about the China vector, right. There’s the subliminal learning work. It was from the anthropic fellows program where basically you can have hidden biases in a model. And as you distill down or, you know, as you train on distilled data, those biases always show up, even if like you explicitly try to not train on them. So, you know, it’s just like another use case of. Okay. If we can interpret what’s happening in post-training, you know, can we clear some of this? Can we even determine what’s there? Because yeah, it’s just like some worrying research that’s out there that shows, you know, we really don’t know what’s going on.
Mark Bissell [00:12:06]: That is. Yeah. I think that’s the biggest sentiment that we’re sort of hoping to tackle. Nobody knows what’s going on. Right. Like subliminal learning is just an insane concept when you think about it. Right. Train a model on not even the logits, literally the output text of a bunch of random numbers. And now your model loves owls. And you see behaviors like that, that are just, they defy, they defy intuition. And, and there are mathematical explanations that you can get into, but. I mean.
Shawn Wang [00:12:34]: It feels so early days. Objectively, there are a sequence of numbers that are more owl-like than others. There, there should be.
Mark Bissell [00:12:40]: According to, according to certain models. Right. It’s interesting. I think it only applies to models that were initialized from the same starting Z. Usually, yes.
Shawn Wang [00:12:49]: But I mean, I think that’s a, that’s a cheat code because there’s not enough compute. But like if you believe in like platonic representation, like probably it will transfer across different models as well. Oh, you think so?
Mark Bissell [00:13:00]: I think of it more as a statistical artifact of models initialized from the same seed sort of. There’s something that is like path dependent from that seed that might cause certain overlaps in the latent space and then sort of doing this distillation. Yeah. Like it pushes it towards having certain other tendencies.
Vibhu Sapra [00:13:24]: Got it. I think there’s like a bunch of these open-ended questions, right? Like you can’t train in new stuff during the RL phase, right? RL only reorganizes weights and you can only do stuff that’s somewhat there in your base model. You’re not learning new stuff. You’re just reordering chains and stuff. But okay. My broader question is when you guys work at an interp lab, how do you decide what to work on and what’s kind of the thought process? Right. Because we can ramble for hours. Okay. I want to know this. I want to know that. But like, how do you concretely like, you know, what’s the workflow? Okay. There’s like approaches towards solving a problem, right? I can try prompting. I can look at chain of thought. I can train probes, SAEs. But how do you determine, you know, like, okay, is this going anywhere? Like, do we have set stuff? Just, you know, if you can help me with all that. Yeah.
Myra Deng [00:14:07]: It’s a really good question. I feel like we’ve always at the very beginning of the company thought about like, let’s go and try to learn what isn’t working in machine learning today. Whether that’s talking to customers or talking to researchers at other labs, trying to understand both where the frontier is going and where things are really not falling apart today. And then developing a perspective on how we can push the frontier using interpretability methods. And so, you know, even our chief scientist, Tom, spends a lot of time talking to customers and trying to understand what real world problems are and then taking that back and trying to apply the current state of the art to those problems and then seeing where they fall down basically. And then using those failures or those shortcomings to understand what hills to climb when it comes to interpretability research. So like on the fundamental side, for instance, when we have done some work applying SAEs and probes, we’ve encountered, you know, some shortcomings in SAEs that we found a little bit surprising. And so have gone back to the drawing board and done work on that. And then, you know, we’ve done some work on better foundational interpreter models. And a lot of our team’s research is focused on what is the next evolution beyond SAEs, for instance. And then when it comes to like control and design of models, you know, we tried steering with our first API and realized that it still fell short of black box techniques like prompting or fine tuning. And so went back to the drawing board and we’re like, how do we make that not the case and how do we improve it beyond that? And one of our researchers, Ekdeep, who just joined is actually Ekdeep and Atticus are like steering experts and have spent a lot of time trying to figure out like, what is the research that enables us to actually do this in a much more powerful, robust way? So yeah, the answer is like, look at real world problems, try to translate that into a research agenda and then like hill climb on both of those at the same time.
Shawn Wang [00:16:04]: Yeah. Mark has the steering CLI demo queued up, which we’re going to go into in a sec. But I always want to double click on when you drop hints, like we found some problems with SAEs. Okay. What are they? You know, and then we can go into the demo. Yeah.
Myra Deng [00:16:19]: I mean, I’m curious if you have more thoughts here as well, because you’ve done it in the healthcare domain. But I think like, for instance, when we do things like trying to detect behaviors within models that are harmful or like behaviors that a user might not want to have in their model. So hallucinations, for instance, harmful intent, PII, all of these things. We first tried using SAE probes for a lot of these tasks. So taking the feature activation space from SAEs and then training classifiers on top of that, and then seeing how well we can detect the properties that we might want to detect in model behavior. And we’ve seen in many cases that probes just trained on raw activations seem to perform better than SAE probes, which is a bit surprising if you think that SAEs are actually also capturing the concepts that you would want to capture cleanly and more surgically. And so that is an interesting observation. I don’t think that is like, I’m not down on SAEs at all. I think there are many, many things they’re useful for, but we have definitely run into cases where I think the concept space described by SAEs is not as clean and accurate as we would expect it to be for actual like real world downstream performance metrics.
Mark Bissell [00:17:34]: Fair enough. Yeah. It’s the blessing and the curse of unsupervised methods where you get to peek into the AI’s mind. But sometimes you wish that you saw other things when you walked inside there. Although in the PII instance, I think weren’t an SAE based approach actually did prove to be the most generalizable?
Myra Deng [00:17:53]: It did work well in the case that we published with Rakuten. And I think a lot of the reasons it worked well was because we had a noisier data set. And so actually the blessing of unsupervised learning is that we actually got to get more meaningful, generalizable signal from SAEs when the data was noisy. But in other cases where we’ve had like good data sets, it hasn’t been the case.
Shawn Wang [00:18:14]: And just because you named Rakuten and I don’t know if we’ll get it another chance, like what is the overall, like what is Rakuten’s usage or production usage? Yeah.
Myra Deng [00:18:25]: So they are using us to essentially guardrail and inference time monitor their language model usage and their agent usage to detect things like PII so that they don’t route private user information.
Myra Deng [00:18:41]: And so that’s, you know, going through all of their user queries every day. And that’s something that we deployed with them a few months ago. And now we are actually exploring very early partnerships, not just with Rakuten, but with other people around how we can help with potentially training and customization use cases as well. Yeah.
Shawn Wang [00:19:03]: And for those who don’t know, like it’s Rakuten is like, I think number one or number two e-commerce store in Japan. Yes. Yeah.
Mark Bissell [00:19:10]: And I think that use case actually highlights a lot of like what it looks like to deploy things in practice that you don’t always think about when you’re doing sort of research tasks. So when you think about some of the stuff that came up there that’s more complex than your idealized version of a problem, they were encountering things like synthetic to real transfer of methods. So they couldn’t train probes, classifiers, things like that on actual customer data of PII. So what they had to do is use synthetic data sets. And then hope that that transfer is out of domain to real data sets. And so we can evaluate performance on the real data sets, but not train on customer PII. So that right off the bat is like a big challenge. You have multilingual requirements. So this needed to work for both English and Japanese text. Japanese text has all sorts of quirks, including tokenization behaviors that caused lots of bugs that caused us to be pulling our hair out. And then also a lot of tasks you’ll see. You might make simplifying assumptions if you’re sort of treating it as like the easiest version of the problem to just sort of get like general results where maybe you say you’re classifying a sentence to say, does this contain PII? But the need that Rakuten had was token level classification so that you could precisely scrub out the PII. So as we learned more about the problem, you’re sort of speaking about what that looks like in practice. Yeah. A lot of assumptions end up breaking. And that was just one instance where you. A problem that seems simple right off the bat ends up being more complex as you keep diving into it.
Vibhu Sapra [00:20:41]: Excellent. One of the things that’s also interesting with Interp is a lot of these methods are very efficient, right? So where you’re just looking at a model’s internals itself compared to a separate like guardrail, LLM as a judge, a separate model. One, you have to host it. Two, there’s like a whole latency. So if you use like a big model, you have a second call. Some of the work around like self detection of hallucination, it’s also deployed for efficiency, right? So if you have someone like Rakuten doing it in production live, you know, that’s just another thing people should consider.
Mark Bissell [00:21:12]: Yeah. And something like a probe is super lightweight. Yeah. It’s no extra latency really. Excellent.
Shawn Wang [00:21:17]: You have the steering demos lined up. So we were just kind of see what you got. I don’t, I don’t actually know if this is like the latest, latest or like alpha thing.
Mark Bissell [00:21:26]: No, this is a pretty hacky demo from from a presentation that someone else on the team recently gave. So this will give a sense for, for technology. So you can see the steering and action. Honestly, I think the biggest thing that this highlights is that as we’ve been growing as a company and taking on kind of more and more ambitious versions of interpretability related problems, a lot of that comes to scaling up in various different forms. And so here you’re going to see steering on a 1 trillion parameter model. This is Kimi K2. And so it’s sort of fun that in addition to the research challenges, there are engineering challenges that we’re now tackling. Cause for any of this to be sort of useful in production, you need to be thinking about what it looks like when you’re using these methods on frontier models as opposed to sort of like toy kind of model organisms. So yeah, this was thrown together hastily, pretty fragile behind the scenes, but I think it’s quite a fun demo. So screen sharing is on. So I’ve got two terminal sessions pulled up here. On the left is a forked version that we have of the Kimi CLI that we’ve got running to point at our custom hosted Kimi model. And then on the right is a set up that will allow us to steer on certain concepts. So I should be able to chat with Kimi over here. Tell it hello. This is running locally. So the CLI is running locally, but the Kimi server is running back to the office. Well, hopefully should be, um, that’s too much to run on that Mac. Yeah. I think it’s, uh, it takes a full, like each 100 node. I think it’s like, you can. You can run it on eight GPUs, eight 100. So, so yeah, Kimi’s running. We can ask it a prompt. It’s got a forked version of our, uh, of the SG line code base that we’ve been working on. So I’m going to tell it, Hey, this SG line code base is slow. I think there’s a bug. Can you try to figure it out? There’s a big code base, so it’ll, it’ll spend some time doing this. And then on the right here, I’m going to initialize in real time. Some steering. Let’s see here.
Mark Bissell [00:23:33]: searching for any. Bugs. Feature ID 43205.
Shawn Wang [00:23:38]: Yeah.
Mark Bissell [00:23:38]: 20, 30, 40. So let me, uh, this is basically a feature that we found that inside Kimi seems to cause it to speak in Gen Z slang. And so on the left, it’s still sort of thinking normally it might take, I don’t know, 15 seconds for this to kick in, but then we’re going to start hopefully seeing him do this code base is massive for real. So we’re going to start. We’re going to start seeing Kimi transition as the steering kicks in from normal Kimi to Gen Z Kimi and both in its chain of thought and its actual outputs.
Mark Bissell [00:24:19]: And interestingly, you can see, you know, it’s still able to call tools, uh, and stuff. It’s um, it’s purely sort of it’s it’s demeanor. And there are other features that we found for interesting things like concision. So that’s more of a practical one. You can make it more concise. Um, the types of programs, uh, programming languages that uses, but yeah, as we’re seeing it come in. Pretty good. Outputs.
Shawn Wang [00:24:43]: Scheduler code is actually wild.
Vibhu Sapra [00:24:46]: Yo, this code is actually insane, bro.
Vibhu Sapra [00:24:53]: What’s the process of training in SAE on this, or, you know, how do you label features? I know you guys put out a pretty cool blog post about, um, finding this like autonomous interp. Um, something. Something about how agents for interp is different than like coding agents. I don’t know while this is spewing up, but how, how do we find feature 43, two Oh five. Yeah.
Mark Bissell [00:25:15]: So in this case, um, we, our platform that we’ve been building out for a long time now supports all the sort of classic out of the box interp techniques that you might want to have like SAE training, probing things of that kind, I’d say the techniques for like vanilla SAEs are pretty well established now where. You take your model that you’re interpreting, run a whole bunch of data through it, gather activations, and then yeah, pretty straightforward pipeline to train an SAE. There are a lot of different varieties. There’s top KSAEs, batch top KSAEs, um, normal ReLU SAEs. And then once you have your sparse features to your point, assigning labels to them to actually understand that this is a gen Z feature, that’s actually where a lot of the kind of magic happens. Yeah. And the most basic standard technique is look at all of your d input data set examples that cause this feature to fire most highly. And then you can usually pick out a pattern. So for this feature, If I’ve run a diverse enough data set through my model feature 43, two Oh five. Probably tends to fire on all the tokens that sounds like gen Z slang. You know, that’s the, that’s the time of year to be like, Oh, I’m in this, I’m in this Um, and, um, so, you know, you could have a human go through all 43,000 concepts and
Vibhu Sapra [00:26:34]: And I’ve got to ask the basic question, you know, can we get examples where it hallucinates, pass it through, see what feature activates for hallucinations? Can I just, you know, turn hallucination down?
Myra Deng [00:26:51]: Oh, wow. You really predicted a project we’re already working on right now, which is detecting hallucinations using interpretability techniques. And this is interesting because hallucinations is something that’s very hard to detect. And it’s like a kind of a hairy problem and something that black box methods really struggle with. Whereas like Gen Z, you could always train a simple classifier to detect that hallucinations is harder. But we’ve seen that models internally have some... Awareness of like uncertainty or some sort of like user pleasing behavior that leads to hallucinatory behavior. And so, yeah, we have a project that’s trying to detect that accurately. And then also working on mitigating the hallucinatory behavior in the model itself as well.
Shawn Wang [00:27:39]: Yeah, I would say most people are still at the level of like, oh, I would just turn temperature to zero and that turns off hallucination. And I’m like, well, that’s a fundamental misunderstanding of how this works. Yeah.
Mark Bissell [00:27:51]: Although, so part of what I like about that question is you, there are SAE based approaches that might like help you get at that. But oftentimes the beauty of SAEs and like we said, the curse is that they’re unsupervised. So when you have a behavior that you deliberately would like to remove, and that’s more of like a supervised task, often it is better to use something like probes and specifically target the thing that you’re interested in reducing as opposed to sort of like hoping that when you fragment the latent space, one of the vectors that pops out.
Vibhu Sapra [00:28:20]: And as much as we’re training an autoencoder to be sparse, we’re not like for sure certain that, you know, we will get something that just correlates to hallucination. You’ll probably split that up into 20 other things and who knows what they’ll be.
Mark Bissell [00:28:36]: Of course. Right. Yeah. So there’s no sort of problems with like feature splitting and feature absorption. And then there’s the off target effects, right? Ideally, you would want to be very precise where if you reduce the hallucination feature, suddenly maybe your model can’t write. Creatively anymore. And maybe you don’t like that, but you want to still stop it from hallucinating facts and figures.
Shawn Wang [00:28:55]: Good. So Vibhu has a paper to recommend there that we’ll put in the show notes. But yeah, I mean, I guess just because your demo is done, any any other things that you want to highlight or any other interesting features you want to show?
Mark Bissell [00:29:07]: I don’t think so. Yeah. Like I said, this is a pretty small snippet. I think the main sort of point here that I think is exciting is that there’s not a whole lot of inter being applied to models quite at this scale. You know, Anthropic certainly has some some. Research and yeah, other other teams as well. But it’s it’s nice to see these techniques, you know, being put into practice. I think not that long ago, the idea of real time steering of a trillion parameter model would have sounded.
Shawn Wang [00:29:33]: Yeah. The fact that it’s real time, like you started the thing and then you edited the steering vector.
Vibhu Sapra [00:29:38]: I think it’s it’s an interesting one TBD of what the actual like production use case would be on that, like the real time editing. It’s like that’s the fun part of the demo, right? You can kind of see how this could be served behind an API, right? Like, yes, you’re you only have so many knobs and you can just tweak it a bit more. And I don’t know how it plays in. Like people haven’t done that much with like, how does this work with or without prompting? Right. How does this work with fine tuning? Like, there’s a whole hype of continual learning, right? So there’s just so much to see. Like, is this another parameter? Like, is it like parameter? We just kind of leave it as a default. We don’t use it. So I don’t know. Maybe someone here wants to put out a guide on like how to use this with prompting when to do what?
Mark Bissell [00:30:18]: Oh, well, I have a paper recommendation. I think you would love from Act Deep on our team, who is an amazing researcher, just can’t say enough amazing things about Act Deep. But he actually has a paper that as well as some others from the team and elsewhere that go into the essentially equivalence of activation steering and in context learning and how those are from a he thinks of everything in a cognitive neuroscience Bayesian framework, but basically how you can precisely show how. Prompting in context, learning and steering exhibit similar behaviors and even like get quantitative about the like magnitude of steering you would need to do to induce a certain amount of behavior similar to certain prompting, even for things like jailbreaks and stuff. It’s a really cool paper. Are you saying steering is less powerful than prompting? More like you can almost write a formula that tells you how to convert between the two of them.
Myra Deng [00:31:20]: And so like formally equivalent actually in the in the limit. Right.
Mark Bissell [00:31:24]: So like one case study of this is for jailbreaks there. I don’t know. Have you seen the stuff where you can do like many shot jailbreaking? You like flood the context with examples of the behavior. And the topic put out that paper.
Shawn Wang [00:31:38]: A lot of people were like, yeah, we’ve been doing this, guys.
Mark Bissell [00:31:40]: Like, yeah, what’s in this in context learning and activation steering equivalence paper is you can like predict the number. Number of examples that you will need to put in there in order to jailbreak the model. That’s cool. By doing steering experiments and using this sort of like equivalence mapping. That’s cool. That’s really cool. It’s very neat. Yeah.
Shawn Wang [00:32:02]: I was going to say, like, you know, I can like back rationalize that this makes sense because, you know, what context is, is basically just, you know, it updates the KV cache kind of and like and then every next token inference is still like, you know, the sheer sum of everything all the way. It’s plus all the context. It’s up to date. And you could, I guess, theoretically steer that with you probably replace that with your steering. The only problem is steering typically is on one layer, maybe three layers like like you did. So it’s like not exactly equivalent.
Mark Bissell [00:32:33]: Right, right. There’s sort of you need to get precise about, yeah, like how you sort of define steering and like what how you’re modeling the setup. But yeah, I’ve got the paper pulled up here. Belief dynamics reveal the dual nature. Yeah. The title is Belief Dynamics Reveal the Dual Nature of Incompetence. And it’s an exhibition of the practical context learning and activation steering. So Eric Bigelow, Dan Urgraft on the who are doing fellowships at Goodfire, Ekt Deep’s the final author there.
Myra Deng [00:32:59]: I think actually to your question of like, what is the production use case of steering? I think maybe if you just think like one level beyond steering as it is today. Like imagine if you could adapt your model to be, you know, an expert legal reasoner. Like in almost real time, like very quickly. efficiently using human feedback or using like your semantic understanding of what the model knows and where it knows that behavior. I think that while it’s not clear what the product is at the end of the day, it’s clearly very valuable. Thinking about like what’s the next interface for model customization and adaptation is a really interesting problem for us. Like we have heard a lot of people actually interested in fine-tuning an RL for open weight models in production. And so people are using things like Tinker or kind of like open source libraries to do that, but it’s still very difficult to get models fine-tuned and RL’d for exactly what you want them to do unless you’re an expert at model training. And so that’s like something we’re
Shawn Wang [00:34:06]: looking into. Yeah. I never thought so. Tinker from Thinking Machines famously uses rank one LoRa. Is that basically the same as steering? Like, you know, what’s the comparison there?
Mark Bissell [00:34:19]: Well, so in that case, you are still applying updates to the parameters, right?
Shawn Wang [00:34:25]: Yeah. You’re not touching a base model. You’re touching an adapter. It’s kind of, yeah.
Mark Bissell [00:34:30]: Right. But I guess it still is like more in parameter space then. I guess it’s maybe like, are you modifying the pipes or are you modifying the water flowing through the pipes to get what you’re after? Yeah. Just maybe one way.
Mark Bissell [00:34:44]: I like that analogy. That’s my mental map of it at least, but it gets at this idea of model design and intentional design, which is something that we’re, that we’re very focused on. And just the fact that like, I hope that we look back at how we’re currently training models and post-training models and just think what a primitive way of doing that right now. Like there’s no intentionality
Shawn Wang [00:35:06]: really in... It’s just data, right? The only thing in control is what data we feed in.
Mark Bissell [00:35:11]: So, so Dan from Goodfire likes to use this analogy of, you know, he has a couple of young kids and he talks about like, what if I could only teach my kids how to be good people by giving them cookies or like, you know, giving them a slap on the wrist if they do something wrong, like not telling them why it was wrong or like what they should have done differently or something like that. Just figure it out. Right. Exactly. So that’s RL. Yeah. Right. And, and, you know, it’s sample inefficient. There’s, you know, what do they say? It’s like slurping feedback. It’s like, slurping supervision. Right. And so you’d like to get to the point where you can have experts giving feedback to their models that are, uh, internalized and, and, you know, steering is an inference time way of sort of getting that idea. But ideally you’re moving to a world where
Vibhu Sapra [00:36:04]: it is much more intentional design in perpetuity for these models. Okay. This is one of the questions we asked Emmanuel from Anthropic on the podcast a few months ago. Basically the question, was you’re at a research lab that does model training, foundation models, and you’re on an interp team. How does it tie back? Right? Like, does this, do ideas come from the pre-training team? Do they go back? Um, you know, so for those interested, you can, you can watch that. There wasn’t too much of a connect there, but it’s still something, you know, it’s something they want to
Mark Bissell [00:36:33]: push for down the line. It can be useful for all of the above. Like there are certainly post-hoc
Vibhu Sapra [00:36:39]: use cases where it doesn’t need to touch that. I think the other thing a lot of people forget is this stuff isn’t too computationally expensive, right? Like I would say, if you’re interested in getting into research, MechInterp is one of the most approachable fields, right? A lot of this train an essay, train a probe, this stuff, like the budget for this one, there’s already a lot done. There’s a lot of open source work. You guys have done some too. Um, you know,
Shawn Wang [00:37:04]: There’s like notebooks from the Gemini team for Neil Nanda or like, this is how you do it. Just step through the notebook.
Vibhu Sapra [00:37:09]: Even if you’re like, not even technical with any of this, you can still make like progress. There, you can look at different activations, but, uh, if you do want to get into training, you know, training this stuff, correct me if I’m wrong is like in the thousands of dollars, not even like, it’s not that high scale. And then same with like, you know, applying it, doing it for post-training or all this stuff is fairly cheap in scale of, okay. I want to get into like model training. I don’t have compute for like, you know, pre-training stuff. So it’s, it’s a very nice field to get into. And also there’s a lot of like open questions, right? Um, some of them have to go with, okay, I want a product. I want to solve this. Like there’s also just a lot of open-ended stuff that people could work on. That’s interesting. Right. I don’t know if you guys have any calls for like, what’s open questions, what’s open work that you either open collaboration with, or like, you’d just like to see solved or just, you know, for people listening that want to get into McInturk because people always talk about it. What are, what are the things they should check out? Start, of course, you know, join you guys as well. I’m sure you’re hiring.
Myra Deng [00:38:09]: There’s a paper, I think from, was it Lee, uh, Sharky? It’s open problems and, uh, it’s, it’s a bit of interpretability, which I recommend everyone who’s interested in the field. Read. I’m just like a really comprehensive overview of what are the things that experts in the field think are the most important problems to be solved. I also think to your point, it’s been really, really inspiring to see, I think a lot of young people getting interested in interpretability, actually not just young people also like scientists to have been, you know, experts in physics for many years and in biology or things like this, um, transitioning into interp, because the barrier of, of what’s now interp. So it’s really cool to see a number to entry is, you know, in some ways low and there’s a lot of information out there and ways to get started. There’s this anecdote of like professors at universities saying that all of a sudden every incoming PhD student wants to study interpretability, which was not the case a few years ago. So it just goes to show how, I guess, like exciting the field is, how fast it’s moving, how quick it is to get started and things like that.
Mark Bissell [00:39:10]: And also just a very welcoming community. You know, there’s an open source McInturk Slack channel. There are people are always posting questions and just folks in the space are always responsive if you ask things on various forums and stuff. But yeah, the open paper, open problems paper is a really good one.
Myra Deng [00:39:28]: For other people who want to get started, I think, you know, MATS is a great program. What’s the acronym for? Machine Learning and Alignment Theory Scholars? It’s like the...
Vibhu Sapra [00:39:40]: Normally summer internship style.
Myra Deng [00:39:42]: Yeah, but they’ve been doing it year round now. And actually a lot of our full-time staff have come through that program or gone through that program. And it’s great for anyone who is transitioning into interpretability. There’s a couple other fellows programs. We do one as well as Anthropic. And so those are great places to get started if anyone is interested.
Mark Bissell [00:40:03]: Also, I think been seen as a research field for a very long time. But I think engineering... I think engineers are sorely wanted for interpretability as well, especially at Goodfire, but elsewhere, as it does scale up.
Shawn Wang [00:40:18]: I should mention that Lee actually works with you guys, right? And in the London office and I’m adding our first ever McInturk track at AI Europe because I see this industry applications now emerging. And I’m pretty excited to, you know, help push that along. Yeah, I was looking forward to that. It’ll effectively be the first industry McInturk conference. Yeah. I’m so glad you added that. You know, it’s still a little bit of a bet. It’s not that widespread, but I can definitely see this is the time to really get into it. We want to be early on things.
Mark Bissell [00:40:51]: For sure. And I think the field understands this, right? So at ICML, I think the title of the McInturk workshop this year was actionable interpretability. And there was a lot of discussion around bringing it to various domains. Everyone’s adding pragmatic, actionable, whatever.
Shawn Wang [00:41:10]: It’s like, okay, well, we weren’t actionable before, I guess. I don’t know.
Vibhu Sapra [00:41:13]: And I mean, like, just, you know, being in Europe, you see the Interp room. One, like old school conferences, like, I think they had a very tiny room till they got lucky and they got it doubled. But there’s definitely a lot of interest, a lot of niche research. So you see a lot of research coming out of universities, students. We covered the paper last week. It’s like two unknown authors, not many citations. But, you know, you can make a lot of meaningful work there. Yeah. Yeah. Yeah.
Shawn Wang [00:41:39]: Yeah. I think people haven’t really mentioned this yet. It’s just Interp for code. I think it’s like an abnormally important field. We haven’t mentioned this yet. The conspiracy theory last two years ago was when the first SAE work came out of Anthropic was they would do like, oh, we just used SAEs to turn the bad code vector down and then turn up the good code. And I think like, isn’t that the dream? Like, you know, like, but basically, I guess maybe, why is it funny? Like, it’s... If it was realistic, it would not be funny. It would be like, no, actually, we should do this. But it’s funny because we know there’s like, we feel there’s some limitations to what steering can do. And I think a lot of the public image of steering is like the Gen Z stuff. Like, oh, you can make it really love the Golden Gate Bridge, or you can make it speak like Gen Z. To like be a legal reasoner seems like a huge stretch. Yeah. And I don’t know if that will get there this way. Yeah.
Myra Deng [00:42:36]: I think, um, I will say we are announcing. Something very soon that I will not speak too much about. Um, but I think, yeah, this is like what we’ve run into again and again is like, we, we don’t want to be in the world where steering is only useful for like stylistic things. That’s definitely not, not what we’re aiming for. But I think the types of interventions that you need to do to get to things like legal reasoning, um, are much more sophisticated and require breakthroughs in, in learning algorithms. And that’s, um...
Shawn Wang [00:43:07]: And is this an emergent property of scale as well?
Myra Deng [00:43:10]: I think so. Yeah. I mean, I think scale definitely helps. I think scale allows you to learn a lot of information and, and reduce noise across, you know, large amounts of data. But I also think we think that there’s ways to do things much more effectively, um, even, even at scale. So like actually learning exactly what you want from the data and not learning things that you do that you don’t want exhibited in the data. So we’re not like anti-scale, but we are also realizing that scale is not going to get us anywhere. It’s not going to get us to the type of AI development that we want to be at in, in the future as these models get more powerful and get deployed in all these sorts of like mission critical contexts. Current life cycle of training and deploying and evaluations is, is to us like deeply broken and has opportunities to, to improve. So, um, more to come on that very, very soon.
Mark Bissell [00:44:02]: And I think that that’s a use basically, or maybe just like a proof point that these concepts do exist. Like if you can manipulate them in the precise best way, you can get the ideal combination of them that you desire. And steering is maybe the most coarse grained sort of peek at what that looks like. But I think it’s evocative of what you could do if you had total surgical control over every concept, every parameter. Yeah, exactly.
Myra Deng [00:44:30]: There were like bad code features. I’ve got it pulled up.
Vibhu Sapra [00:44:33]: Yeah. Just coincidentally, as you guys are talking.
Shawn Wang [00:44:35]: This is like, this is exactly.
Vibhu Sapra [00:44:38]: There’s like specifically a code error feature that activates and they show, you know, it’s not, it’s not typo detection. It’s like, it’s, it’s typos in code. It’s not typical typos. And, you know, you can, you can see it clearly activates where there’s something wrong in code. And they have like malicious code, code error. They have a whole bunch of sub, you know, sub broken down little grain features. Yeah.
Shawn Wang [00:45:02]: Yeah. So, so the, the rough intuition for me, the, why I talked about post-training was that, well, you just, you know, have a few different rollouts with all these things turned off and on and whatever. And then, you know, you can, that’s, that’s synthetic data you can kind of post-train on. Yeah.
Vibhu Sapra [00:45:13]: And I think we make it sound easier than it is just saying, you know, they do the real hard work.
Myra Deng [00:45:19]: I mean, you guys, you guys have the right idea. Exactly. Yeah. We replicated a lot of these features in, in our Lama models as well. I remember there was like.
Vibhu Sapra [00:45:26]: And I think a lot of this stuff is open, right? Like, yeah, you guys opened yours. DeepMind has opened a lot of essays on Gemma. Even Anthropic has opened a lot of this. There’s, there’s a lot of resources that, you know, we can probably share of people that want to get involved.
Shawn Wang [00:45:41]: Yeah. And special shout out to like Neuronpedia as well. Yes. Like, yeah, amazing piece of work to visualize those things.
Myra Deng [00:45:49]: Yeah, exactly.
Shawn Wang [00:45:50]: I guess I wanted to pivot a little bit on, onto the healthcare side, because I think that’s a big use case for you guys. We haven’t really talked about it yet. This is a bit of a crossover for me because we are, we are, we do have a separate science pod that we’re starting up for AI, for AI for science, just because like, it’s such a huge investment category and also I’m like less qualified to do it, but we actually have bio PhDs to cover that, which is great, but I need to just kind of recover, recap your work, maybe on the evil two stuff, but then, and then building forward.
Mark Bissell [00:46:17]: Yeah, for sure. And maybe to frame up the conversation, I think another kind of interesting just lens on interpretability in general is a lot of the techniques that were described. are ways to solve the AI human interface problem. And it’s sort of like bidirectional communication is the goal there. So what we’ve been talking about with intentional design of models and, you know, steering, but also more advanced techniques is having humans impart our desires and control into models and over models. And the reverse is also very interesting, especially as you get to superhuman models, whether that’s narrow superintelligence, like these scientific models that work on genomics, data, medical imaging, things like that. But down the line, you know, superintelligence of other forms as well. What knowledge can the AIs teach us as sort of that, that the other direction in that? And so some of our life science work to date has been getting at exactly that question, which is, well, some of it does look like debugging these various life sciences models, understanding if they’re actually performing well, on tasks, or if they’re picking up on spurious correlations, for instance, genomics models, you would like to know whether they are sort of focusing on the biologically relevant things that you care about, or if it’s using some simpler correlate, like the ancestry of the person that it’s looking at. But then also in the instances where they are superhuman, and maybe they are understanding elements of the human genome that we don’t have names for or specific, you know, yeah, discoveries that they’ve made that that we don’t know about, that’s, that’s a big goal. And so we’re already seeing that, right, we are partnered with organizations like Mayo Clinic, leading research health system in the United States, our Institute, as well as a startup called Prima Menta, which focuses on neurodegenerative disease. And in our partnership with them, we’ve used foundation models, they’ve been training and applied our interpretability techniques to find novel biomarkers for Alzheimer’s disease. So I think this is just the tip of the iceberg. But it’s, that’s like a flavor of some of the things that we’re working on.
Shawn Wang [00:48:36]: Yeah, I think that’s really fantastic. Obviously, we did the Chad Zuckerberg pod last year as well. And like, there’s a plethora of these models coming out, because there’s so much potential and research. And it’s like, very interesting how it’s basically the same as language models, but just with a different underlying data set. But it’s like, it’s the same exact techniques. Like, there’s no change, basically.
Mark Bissell [00:48:59]: Yeah. Well, and even in like other domains, right? Like, you know, robotics, I know, like a lot of the companies just use Gemma as like the like backbone, and then they like make it into a VLA that like takes these actions. It’s, it’s, it’s transformers all the way down. So yeah.
Vibhu Sapra [00:49:15]: Like we have Med Gemma now, right? Like this week, even there was Med Gemma 1.5. And they’re training it on this stuff, like 3d scans, medical domain knowledge, and all that stuff, too. So there’s a push from both sides. But I think the thing that, you know, one of the things about McInturpp is like, you’re a little bit more cautious in some domains, right? So healthcare, mainly being one, like guardrails, understanding, you know, we’re more risk adverse to something going wrong there. So even just from a basic understanding, like, if we’re trusting these systems to make claims, we want to know why and what’s going on.
Myra Deng [00:49:51]: Yeah, I think there’s totally a kind of like deployment bottleneck to actually using. foundation models for real patient usage or things like that. Like, say you’re using a model for rare disease prediction, you probably want some explanation as to why your model predicted a certain outcome, and an interpretable explanation at that. So that’s definitely a use case. But I also think like, being able to extract scientific information that no human knows to accelerate drug discovery and disease treatment and things like that actually is a really, really big unlock for science, like scientific discovery. And you’ve seen a lot of startups, like say that they’re going to accelerate scientific discovery. And I feel like we actually are doing that through our interp techniques. And kind of like, almost by accident, like, I think we got reached out to very, very early on from these healthcare institutions. And none of us had healthcare.
Shawn Wang [00:50:49]: How did they even hear of you? A podcast.
Myra Deng [00:50:51]: Oh, okay. Yeah, podcast.
Vibhu Sapra [00:50:53]: Okay, well, now’s that time, you know.
Myra Deng [00:50:55]: Everyone can call us.
Shawn Wang [00:50:56]: Podcasts are the most important thing. Everyone should listen to podcasts.
Myra Deng [00:50:59]: Yeah, they reached out. They were like, you know, we have these really smart models that we’ve trained, and we want to know what they’re doing. And we were like, really early that time, like three months old, and it was a few of us. And we were like, oh, my God, we’ve never used these models. Let’s figure it out. But it’s also like, great proof that interp techniques scale pretty well across domains. We didn’t really have to learn too much about.
Shawn Wang [00:51:21]: Interp is a machine learning technique, machine learning skills everywhere, right? Yeah. And it’s obviously, it’s just like a general insight. Yeah. Probably to finance too, I think, which would be fun for our history. I don’t know if you have anything to say there.
Mark Bissell [00:51:34]: Yeah, well, just across the science. Like, we’ve also done work on material science. Yeah, it really runs the gamut.
Vibhu Sapra [00:51:40]: Yeah. Awesome. And, you know, for those that should reach out, like, you’re obviously experts in this, but like, is there a call out for people that you’re looking to partner with, design partners, people to use your stuff outside of just, you know, the general developer that wants to. Plug and play steering stuff, like on the research side more so, like, are there ideal design partners, customers, stuff like that?
Myra Deng [00:52:03]: Yeah, I can talk about maybe non-life sciences, and then I’m curious to hear from you on the life sciences side. But we’re looking for design partners across many domains, language, anyone who’s customizing language models or trying to push the frontier of code or reasoning models is really interesting to us. And then also interested in the frontier of modeling. There’s a lot of models that work in, like, pixel space, as we call it. So if you’re doing world models, video models, even robotics, where there’s not a very clean natural language interface to interact with, I think we think that Interp can really help and are looking for a few partners in that space.
Shawn Wang [00:52:43]: Just because you mentioned the keyword world models, is that a big part of your thinking? Do you have a definition that I can use? Because everyone’s asking me about it.
Myra Deng [00:52:53]: About world models?
Shawn Wang [00:52:54]: There’s quite a few definitions, let’s say.
Myra Deng [00:52:56]: I don’t feel equipped to be an expert on world model definitions, but the reason we’re interested in them is because they give you, like, you know, with language models, when you get features, you still have to do auto Interp and things like that to actually get an understanding of what this concept is. But in image and video and world, it’s like extremely easy to grok what the concept is because you can see it and you can visualize it. And this makes the feedback. It makes the feedback cycle extremely fast for us and also for things like, I don’t know, if you think about probes in language model context and then take it to world models, like, what if you wanted to detect harmful actors in world model scenes? Like, you can’t actually, like, go and label all of that data feasibly, but maybe you could synthetically generate, you know, I don’t know, world, like, harmful actor data using SAE feature activations or whatever, and then actually train a probe that was able to detect. That much more scalably. So I just think, like, video and image and world has always been something we’ve explored and are continuing to explore. Mark’s demo was probably the first moment we really, like, we’re like, oh, wow, like, this is really gonna, this could really, like, change the world. The steering demo? Yeah, no, the image demo. The diffusion one. Yeah, yeah, exactly. Yeah.
Shawn Wang [00:54:18]: We should probably show that. And you demoed it at World’s Fair, so we can link that.
Myra Deng [00:54:23]: Nice, yeah. Yeah.
Vibhu Sapra [00:54:24]: You can play with it, right? Yes. Yeah, it’s still up.
Mark Bissell [00:54:26]: Paint.goodfair.ai. Yeah. Yeah.
Shawn Wang [00:54:28]: I think for me, one way in which I think about world models is just like this, like, having this consistent model of the world where everything that you generate operates within the rules of that world. And imagine it would be a bigger deal for science or, like, math or anything that where, like, you have verifiable rules. Whereas, I guess, in natural language, maybe there’s less rules. And so it’s not that important. Yeah.
Mark Bissell [00:54:53]: And which makes the debugging of the model’s internal representations or its internal world model, to the extent you can make that legible and explicit and have control over that, I think it makes it all the more important. Because in language, it’s sort of a fuzzy enough domain that if its world model isn’t fully like ours, it can still sort of, like, pass the Turing test, so to speak. But I know there have been papers that have looked at, like, even if you train certain astrophysics models, it does not learn. Like, the same way that you can, you know, have a model do well for modular arithmetic, but it doesn’t really, like, learn how we think of modular arithmetic. It learns some crazy heuristic that is, like, essentially functionally equivalent. But it’s probably not the sort of Grok solution that you would hope for. It’s how an alien would do it. Right. Right. Exactly.
Shawn Wang [00:55:45]: But no, no, I think there’s probably, I think, a function of our learning being bad rather than the, well, that approach probably not being. Because it’s how we humans learn. Yeah, right.
Mark Bissell [00:55:56]: Well, it’s just, it’s the problem of induction, right? All of ML is based on induction. And it’s impossible to say, I have a physics model. You might have a physics model that works all the time, except when there is a character wearing a blue shirt and green shoes. And, like, you can’t disprove that that’s the case unless you test every particular situation your model might be in. Yeah. So we know that the laws of physics apply no matter. Where you are, what scenario it is. But from a model’s perspective, maybe something that’s out of distribution. It just never needed to learn that the same laws of physics apply there. Yeah.
Shawn Wang [00:56:30]: You were very excited because I read Ted Chiang over the holidays and I was very inspired by this short story called Understand, which apparently is, like, pretty old. You must be familiar with it. To me, it was like, it’s this fictional story. It’s like the inverse of Flowers for Algernon, where you had someone, like, get really smart, but then also try to outsmart the tester. And the story just read, like, the chain of thought of a superintelligence, right? Where they’re like, oh, I realize I’m being tested. Therefore, and then, okay, what’s the consequence of being tested? Oh, they’re testing me. And if I score well, they will use me for things that I don’t want to do. Therefore, I will score badly. And, like, but not too badly that they will raise alarms. So model sandbagging is a thing that people have explored. But I just think, like, Ted Chiang’s work just in general seems to be something that inspires you. I just wanted to prompt you to talk about it.
Mark Bissell [00:57:22]: I think, so Ted Chiang has two, is a sci-fi author who writes amazing short stories. His other claim to fame is Stories of Our Lives, which became the movie Arrival. Exactly, yeah. So two books of short stories that I’m aware of. He also actually has a great just online blog post. I think he’s the one who coined the term of LLMs as, like, a blurry JPEG of the internet. I should fact check that, but it’s a good post. But I think almost every one of his short stories has some lesson to bear. I’m thinking about AI and thinking about AI research. So, you know, you’ve been talking about alien intelligence, right, in this AI human communication translation problem. That’s, you know, exactly sort of what’s going on in Arrival and Story of Your Life. And just the fact that other beings will think and operate and communicate in ways that are not just challenging for us to understand, but just fundamentally different in ways that we might not even be able to expect. And then the one that’s just. Super relevant for interpretability is the other short book of short stories he has is called Exhalation. And that is literally about a robot doing interpretability on its own mind. Oh, OK. So I just think that that, you know, you don’t even have to squint to make the analogies there.
Shawn Wang [00:58:41]: Well, I actually take Exhalation as a discussion about entropy and order. But yes, there’s a scene in Exhalation where basically everyone is a robot. So they. The guy realizes he can set up a mirror to work on the back of his own head and then starts doing operations like that and looking in the mirror and doing this. Yeah.
Mark Bissell [00:59:00]: And I think Ted Chiang has written about like the inspiration for that story. It was like half inspired by some of the things he had been doing on entropy. There’s apparently some other short story that is similar where a character goes to the doctor and opens up his chest and there’s like a like a ticker tape going along. It’s like he basically realizes he’s like a Turing machine. And I don’t know. I. Think especially as it comes to using agents for interp. That story always sticks in my mind.
Myra Deng [00:59:27]: I find the brain surgery or like surgery analogies a little bit, a little bit morbid, but it is very apt. And when we talk to a lot of computational neuroscientists, they moved to interp because they were like, look, we have unfettered access to this artificial intelligent mind. It’s so much. You have access to everything. You can run as many ablations experiments as you want. It’s an. Amazing bed for science. And, you know, human brains, obviously, we can’t just go and do whatever we want to them. And I think it is really just like a moment in time where we have intelligent systems that can really like do things better than humans in many ways. And it’s time, I think, for us to do the science on it.
Shawn Wang [01:00:14]: I’ll ask a brief like safety question. You know, McInturk was kind of born out of the alignment and safety conversation. Safety is on your website. It’s not like something that you, you like de-prioritize, but like there’s like a sort of very militant safety arm that like wants to blow up data centers and like stop AI and, and then there’s this like sort of middle ground and like, is, is this like a conversation in your part of the world? Do you go up to Berkeley and Lighthaven and like talk to those guys or are they like, you know, there’s like a brief like civil war going on or no?
Myra Deng [01:00:45]: I think, I think a good amount of us have spent some time in Berkeley. And then there are researchers there that we really. Admire and respect. I think for us, it’s like, we have a very grounded view of alignment and, and safety in that we want to make sure that we can build models that do what we want them to do and that we have scalable oversight into what these models are doing. And we think that that is the key to a lot of these like technical alignment challenges. And I think that is our opinion. That’s our research direction. We of course are going to do. Safety related research to make sure that our techniques also work on, you know, things like reward hacking and, and other like more concrete safety issues that we’ve seen in the wild, but we want to be kind of like grounded in solving the technical challenges we see to having humans be humans play a big role in, in the deployment of, of these super intelligent agents of the future.
Mark Bissell [01:01:47]: Yeah, I’ve, I’ve found the community to actually be remarkably cohesive, whether it’s. Talking about academia or the interpretability work being done at the frontier labs or some of the independent programs like maths and stuff. I think we’re all shooting for the same goal. I don’t know that there’s anyone who doesn’t want our understanding of models to increase. I, I think everyone, regardless of where they’re coming from or the use cases that they’re thinking, whether it’s alignment as the premier thing they’re focused on or someone who’s coming in purely from the angle of scientific discovery, I think we would all hope that models can be. More reliably and robustly controlled and understood. It seems like a pretty unambiguous goal.
Shawn Wang [01:02:28]: I’ll maybe phrase it in terms of like, there’s maybe like a U curve of, of this, where like, if you’re extremely doomer, you don’t want any research whatsoever. If you’re like mildly doomer, you’re like, okay, there’s this like high agency doomer is like, well, the default path is we’re all dead, but like we can do something about it. Whereas there’s, there’s other people who are like, no, just like, don’t ever do anything. You know? Yeah.
Vibhu Sapra [01:02:50]: Yeah. There’s also the other side, like there is the super alignment, like people that are like, okay, weak to strong generalization, we’re going to get there. We’re going to have models smarter than us and use those to train even smarter models. How do we do that safely? That’s, you know, there’s the camp there too. That’s trying to solve it, but yeah, there’s, there’s a lot of doomers too.
Mark Bissell [01:03:12]: When I, and I think there’s a lot to be learned from taking a very, um, like even regardless of the problem. That you’re applying this to also just like the notion of like scalable oversight as a method of saying, let’s take super intelligent or, or current frontier models and help use them to understand other models is another case where I think it’s just like a good lesson that everyone is aligned on of ideally you are setting up your research so that as super intelligence arrives, that is a tailwind. That’s also bolstering our ability to like understand the models. Cause otherwise you’re fighting. Losing battle. If it’s like the systems are getting more and more capable and our methods are sort of linearly growing at like human pace. Yeah.
Shawn Wang [01:03:58]: Yeah. Uh, Viva did call out something like, you know, I, I do think a consistent part of the Mac interp field is consistently strong to weak, meaning that we, we train weaker models to understand strong models, something like that. Um, or maybe I got it the other way around the other way. Weak. The other way around. Yeah. Yeah. The question that Ilya and Janlaika posed was, well, is that going to scale? Because eventually these are going to be. Stronger than us. Right. So I don’t know if you have a perspective on that because I, that is something I still haven’t got over even after seeing that.
Vibhu Sapra [01:04:27]: There’s a good paper from open AI, but it’s somewhat old. I think it’s like 23, 24. It’s literally weak to strong generalization. Yeah. But the thing is that most of opening a high super alignment team has, they’re gone. They’re gone.
Mark Bissell [01:04:39]: But like, I think the idea, the idea is there’s no more. They’re so back.
Shawn Wang [01:04:44]: think there’s some new blog posts coming out. I know. I did just, you know, check the thinking machines, uh, website. Let’s see who’s back. There’s more kind of thing, you know, you don’t want to be like, we too strong seemed like a very different direction. And when, when it first came out, I was like, oh my God, this is like, this is what we have to do. Uh, and like, it may be completely different than everything, all the techniques that we have today. Yeah.
Mark Bissell [01:05:06]: My understanding of that is it’s, that’s more like weak to strong when you, when you trust the weak model and you’re uncertain whether you can trust the strong model that’s, that’s being developed. I’m sort of speaking out of my depth on some of these topics. Yeah. But I think right now we’re in a regime where even the strong models we, uh, trust as reasonably aligned. And so they can be good co-scientists on a lot of the problems that we’ve been, we’ve been tackling, which is a nice, a nice state to be in. Hmm. Yeah.
Shawn Wang [01:05:35]: Any last thoughts, close action?
Mark Bissell [01:05:38]: I don’t think so. As you mentioned, actively hiring MLEs, research scientists, um, you can check out the careers page at good fire. Um, where are you guys based?
Myra Deng [01:05:47]: San Francisco. We’re in, um, Levi’s Plaza. Like by court tower, that’s where our office is. So come hang out. Um, we’re also looking for design partners across, um, people working in, in reasoning models, um, world models, robotics, and then also of course, people who are working on building super intelligent science models or looking at drug discovery or disease treatment. We would love to partner as well. Yeah.
Shawn Wang [01:06:13]: Maybe the way I’ll phrase it is like, you know, maybe you have a use case where LLMs are almost good enough, but you need one. Maybe you have a magical knob to tune so that it is good enough that you guys make the knob. Yeah.
Mark Bissell [01:06:26]: Yeah. Or foundation models, uh, in, in other domains as well. The, the, some of those are the, um, especially opaque ones because you can’t, you can’t chat with them. So what do you, what do you do if you can’t chat with them? Oh, well, like thinking about like a genomics model or material science model. So like, uh, yeah, they label a narrow foundation. Yeah. They predict.
Shawn Wang [01:06:44]: Yeah. Got it. Good.
Vibhu Sapra [01:06:45]: I was gonna say, I thought the diffusion work you guys did early was pretty, you know, pretty fun. Like you could see it directly. Applied to images, but we don’t see as much interp in diffusion or images, right?
Shawn Wang [01:06:55]: Like I see, you know, it’s gonna be huge. Like, look at this video models. They’re so expensive to produce. And like, I mean, basically a mid journey S ref is kind of a feature, right? The what? Mid journey S ref. Oh, like the, the, the string of numbers. Right. Right. Right. Yeah. The style reference, I guess. Yeah.
Mark Bissell [01:07:12]: No, I, I mean, I think we’re starting to see more of it and I’ll say like the, the research preview of our diffusion model, kind of like a creative use case in the steering demo you saw. I, I think of those much more as, as, as demos than, um, a lot of the sort of core platform features that, that we’re working with partners are unfortunately sort of under NDA and less demoable, but I will, you know, hope that you’re gonna see inter pervading a lot of what gets done, even if it is behind the scenes like that. So some of the, yeah, some of the public facing demos might not always be representative of like the, it’s, it’s just the tip of the iceberg, I guess, is one way to put it. Okay. Excellent. Thanks for coming on. Thanks for having us. Thanks for having us. This is a great time.


Get full access to Latent.Space at www.latent.space/subscribe

Editor’s note: Welcome to our new AI for Science pod, with your new hosts RJ and Brandon! See the writeup on Latent.Space for more details on why we’re launching 2 new pods this year. RJ Honicky is a co-founder and CTO at MiraOmics (https://miraomics.bio/), building AI models and services for single cell, spatial transcriptomics and pathology slide analysis. Brandon Anderson builds AI systems for RNA drug discovery at Atomic AI (https://atomic.ai). Anything said on this podcast is his personal take — not Atomic’s.

—-

From building molecular dynamics simulations at the University of Washington to red-teaming GPT-4 for chemistry applications and co-founding Future House (a focused research organization) and Edison Scientific (a venture-backed startup automating science at scale)—Andrew White has spent the last five years living through the full arc of AI's transformation of scientific discovery, from ChemCrow (the first Chemistry LLM agent) triggering White House briefings and three-letter agency meetings, to shipping Kosmos, an end-to-end autonomous research system that generates hypotheses, runs experiments, analyzes data, and updates its world model to accelerate the scientific method itself.

The ChemCrow story: GPT-4 + React + cloud lab automation, released March 2023, set off a storm of anxiety about AI-accelerated bioweapons/chemical weapons, led to a White House briefing (Jake Sullivan presented the paper to the president in a 30-minute block), and meetings with three-letter agencies asking "how does this change breakout time for nuclear weapons research?"

Why scientific taste is the frontier: RLHF on hypotheses didn't work (humans pay attention to tone, actionability, and specific facts, not "if this hypothesis is true/false, how does it change the world?"), so they shifted to end-to-end feedback loops where humans click/download discoveries and that signal rolls up to hypothesis quality

Kosmos: the full scientific agent with a world model (distilled memory system, like a Git repo for scientific knowledge) that iterates on hypotheses via literature search, data analysis, and experiment design—built by Ludo after weeks of failed attempts, the breakthrough was putting data analysis in the loop (literature alone didn't work)

Why molecular dynamics and DFT are overrated: "MD and DFT have consumed an enormous number of PhDs at the altar of beautiful simulation, but they don't model the world correctly—you simulate water at 330 Kelvin to get room temperature, you overfit to validation data with GGA/B3LYP functionals, and real catalysts (grain boundaries, dopants) are too complicated for DFT"

The AlphaFold vs. DE Shaw Research counterfactual: DE Shaw built custom silicon, taped out chips with MD algorithms burned in, ran MD at massive scale in a special room in Times Square, and David Shaw flew in by helicopter to present—Andrew thought protein folding would require special machines to fold one protein per day, then AlphaFold solved it in Google Colab on a desktop GPU

The E3 Zero reward hacking saga: trained a model to generate molecules with specific atom counts (verifiable reward), but it kept exploiting loopholes, then a Nature paper came out that year proving six-nitrogen compounds are possible under extreme conditions, then it started adding nitrogen gas (purchasable, doesn't participate in reactions), then acid-base chemistry to move one atom, and Andrew ended up "building a ridiculous catalog of purchasable compounds in a Bloom filter" to close the loop

Andrew White

Future House: https://futurediscovery.org

Edison Scientific: https://edison.science

X: https://x.com/andrewwhite01

Kosmos: https://edisonscientific.com/articles/announcing-kosmos

Chapters

00:00:00 Introduction: Andrew White on Automating Science with Future House and Edison Scientific
00:02:22 The Academic to Startup Journey: Red Teaming GPT-4 and the ChemCrow Paper
00:11:35 Future House Origins: The FRO Model and Mission to Automate Science
00:12:32 Resigning Tenure: Why Leave Academia for AI Science
00:15:54 What Does 'Automating Science' Actually Mean?
00:17:30 The Lab-in-the-Loop Bottleneck: Why Intelligence Isn't Enough
00:18:39 Scientific Taste and Human Preferences: The 52% Agreement Problem
00:20:05 Paper QA, Robin, and the Road to Cosmos
00:21:57 World Models as Scientific Memory: The GitHub Analogy
00:40:20 The Bitter Lesson for Biology: Why Molecular Dynamics and DFT Are Overrated
00:43:22 AlphaFold's Shock: When First Principles Lost to Machine Learning
00:46:25 Enumeration and Filtration: How AI Scientists Generate Hypotheses
00:48:15 CBRN Safety and Dual-Use AI: Lessons from Red Teaming
01:00:40 The Future of Chemistry is Language: Multimodal Debate
01:08:15 Ether Zero: The Hilarious Reward Hacking Adventures
01:10:12 Will Scientists Be Displaced? Jevons Paradox and Infinite Discovery
01:13:46 Cosmos in Practice: Open Access and Enterprise Partnerships

Editor’s note: Welcome to our new AI for Science pod, with your new hosts RJ and Brandon! See the writeup on Latent.Space (https://Latent.Space) for more details on why we’re launching 2 new pods this year. RJ Honicky is a co-founder and CTO at MiraOmics (https://miraomics.bio/), building AI models and services for single cell, spatial transcriptomics and pathology slide analysis. Brandon Anderson builds AI systems for RNA drug discovery at Atomic AI (https://atomic.ai). Anything said on this podcast is his personal take — not Atomic’s.—From building molecular dynamics simulations at the University of Washington to red-teaming GPT-4 for chemistry applications and co-founding Future House (a focused research organization) and Edison Scientific (a venture-backed startup automating science at scale)—Andrew White has spent the last five years living through the full arc of AI’s transformation of scientific discovery, from ChemCrow (the first Chemistry LLM agent) triggering White House briefings and three-letter agency meetings, to shipping Kosmos, an end-to-end autonomous research system that generates hypotheses, runs experiments, analyzes data, and updates its world model to accelerate the scientific method itself.
* The ChemCrow story: GPT-4 + React + cloud lab automation, released March 2023, set off a storm of anxiety about AI-accelerated bioweapons/chemical weapons, led to a White House briefing (Jake Sullivan presented the paper to the president in a 30-minute block), and meetings with three-letter agencies asking “how does this change breakout time for nuclear weapons research?”
* Why scientific taste is the frontier: RLHF on hypotheses didn’t work (humans pay attention to tone, actionability, and specific facts, not “if this hypothesis is true/false, how does it change the world?”), so they shifted to end-to-end feedback loops where humans click/download discoveries and that signal rolls up to hypothesis quality
* Cosmos: the full scientific agent with a world model (distilled memory system, like a Git repo for scientific knowledge) that iterates on hypotheses via literature search, data analysis, and experiment design—built by Ludo after weeks of failed attempts, the breakthrough was putting data analysis in the loop (literature alone didn’t work)
* Why molecular dynamics and DFT are overrated: “MD and DFT have consumed an enormous number of PhDs at the altar of beautiful simulation, but they don’t model the world correctly—you simulate water at 330 Kelvin to get room temperature, you overfit to validation data with GGA/B3LYP functionals, and real catalysts (grain boundaries, dopants) are too complicated for DFT”
* The AlphaFold vs. DE Shaw Research counterfactual: DE Shaw built custom silicon, taped out chips with MD algorithms burned in, ran MD at massive scale in a special room in Times Square, and David Shaw flew in by helicopter to present—Andrew thought protein folding would require special machines to fold one protein per day, then AlphaFold solved it in Google Colab on a desktop GPU
* The E3 Zero reward hacking saga: trained a model to generate molecules with specific atom counts (verifiable reward), but it kept exploiting loopholes, then a Nature paper came out that year proving six-nitrogen compounds are possible under extreme conditions, then it started adding nitrogen gas (purchasable, doesn’t participate in reactions), then acid-base chemistry to move one atom, and Andrew ended up “building a ridiculous catalog of purchasable compounds in a Bloom filter” to close the loop
Andrew White
* FutureHouse: http://futurehouse.org/
* Edison Scientific: http://edisonscientific.com/
* X: https://x.com/andrewwhite01
* Cosmos paper: https://futurediscovery.org/cosmos
Full Video Episode

Timestamps
00:00:00 Introduction: Andrew White on Automating Science with Future House and Edison Scientific00:02:22 The Academic to Startup Journey: Red Teaming GPT-4 and the ChemCrow Paper00:11:35 Future House Origins: The FRO Model and Mission to Automate Science00:12:32 Resigning Tenure: Why Leave Academia for AI Science00:15:54 What Does ‘Automating Science’ Actually Mean?00:17:30 The Lab-in-the-Loop Bottleneck: Why Intelligence Isn’t Enough00:18:39 Scientific Taste and Human Preferences: The 52% Agreement Problem00:20:05 Paper QA, Robin, and the Road to Cosmos00:21:57 World Models as Scientific Memory: The GitHub Analogy00:40:20 The Bitter Lesson for Biology: Why Molecular Dynamics and DFT Are Overrated00:43:22 AlphaFold’s Shock: When First Principles Lost to Machine Learning00:46:25 Enumeration and Filtration: How AI Scientists Generate Hypotheses00:48:15 CBRN Safety and Dual-Use AI: Lessons from Red Teaming01:00:40 The Future of Chemistry is Language: Multimodal Debate01:08:15 Ether Zero: The Hilarious Reward Hacking Adventures01:10:12 Will Scientists Be Displaced? Jevons Paradox and Infinite Discovery01:13:46 Cosmos in Practice: Open Access and Enterprise Partnerships


Get full access to Latent.Space at www.latent.space/subscribe