Vanishing Gradients

“It would be sad if local models were not an option and there were only proprietary models. It’s good to have alternatives. Competition is good for business.”
— Sebastian Raschka, on open-weight AI
Kimi K3’s weights landed about an hour before Hugo Bowne-Anderson and Sebastian Raschka went live. Sebastian had already updated his architecture diagram. That speed captures his approach to the current model wave: wait until the weights exist, run the model in the harness where it will actually work, then inspect the architecture closely enough to understand what changed.
The conversation arrived during a larger fight over who supplies the models underneath global software. Three days earlier, twenty-five companies including NVIDIA, Meta, Microsoft, Hugging Face, and IBM published Open Weights and American AI Leadership. Their argument closely matches Sebastian’s practical case for local models: open weights create competition, reduce dependence on a single provider, and let organizations choose a model at the right capability and cost.
Update: Four days after we recorded, DeepSeek released V4 Flash 0731, a re-post-trained API model for agentic coding. Developers are already reporting that it can debug multi-project codebases and stay on task across very long contexts.
You can find the full episode on Spotify, Apple Podcasts, and YouTube.
👉 The next cohort of Master Agentic Data Science starts August 4. Hugo is teaching it with Thomas Wiecki and Luca Fiaschi from PyMC Labs. You will build agents that explore data, run predictive and causal workflows, challenge one another’s conclusions, and produce analysis humans can inspect and reproduce. Podcast listeners get 20% off with code MADSAGC20. Join the cohort. 👈
In This Episode
* Kimi K3’s weights landed about an hour before we went live. Sebastian had already updated his architecture diagram, then walked us through how Kimi Delta Attention and its mixture-of-experts design turn Kimi Linear’s research architecture into a production model.
* When a local model is already good enough. Sebastian separates everyday coding and private workloads from problems where the strongest available model, extra reasoning time, and higher cost are justified.
* Why the model and harness must be evaluated together. The same open-weight model behaves differently across Qwen Code, Claude Code, Codex, Pi, and other harnesses because context handling, compaction, tool conventions, and post-training all affect the result.
* The harness should choose the model and its reasoning effort. Sebastian describes a router that uses the task, conversation history, context, and previous loops to make that decision automatically.
* How long agent contexts are changing model architecture. Sebastian moves through grouped-query attention, multi-head latent attention, sparse attention, mixture-of-experts, DeltaNet, and Mamba-style hybrid layers.
* The implementation does not lie. A from-scratch model can be checked against open reference weights tensor by tensor, and the first divergence exposes the missing RoPE scaling detail, RMSNorm variant, or architectural assumption that prose left out.
* When fine-tuning earns its cost. A specialized task performed thousands of times a day creates different economics from a task performed eight times, so the decision depends on quality, volume, latency, and token cost.
* Sebastian keeps the creative work and gives agents the tedious checks. He uses Codex and Claude to check links, enforce editorial conventions, and compare notebook code with the corresponding book manuscript.
About Sebastian Raschka
Sebastian Raschka is an independent AI researcher, author, and educator focused on large language models, reasoning models, deep learning, and practical machine-learning systems. He created Ahead of AI and is the author of Build a Reasoning Model (From Scratch) and Build a Large Language Model (From Scratch).
Resources
* Build a Reasoning Model (From Scratch): podcast listeners get 45% off the ebook or print edition with code 45VGraschka2.
* Sebastian Raschka’s LLM architecture gallery
* Kimi K3 in the architecture gallery
* Build a Large Language Model (From Scratch)
* Ahead of AI
* Using Local Coding Agents
* Open Weights and American AI Leadership
* Anthropic’s position on open-weight models
Listen or Watch
You can find the full episode on Spotify, Apple Podcasts, and YouTube.
👉 The next cohort of Master Agentic Data Science starts August 4. Hugo is teaching it with Thomas Wiecki and Luca Fiaschi from PyMC Labs. You will build agents that explore data, run predictive and causal workflows, challenge one another’s conclusions, and produce analysis humans can inspect and reproduce. Podcast listeners get 20% off with code MADSAGC20. Join the cohort. 👈
How You Can Support Vanishing Gradients
Vanishing Gradients is a podcast, workshop series, blog, and newsletter focused on what you can build with AI right now. More than 70 episodes with expert practitioners from Google DeepMind, Netflix, Stanford, and elsewhere. Hundreds of hours of free, hands-on workshops. All independent, all free.
If you want to help keep it going:
* Become a paid subscriber, from $8 per month
* Share this episode with a builder who would find it useful
* Subscribe to the Vanishing Gradients YouTube channel
* Join another Vanishing Gradients workshop


Get full access to Vanishing Gradients at hugobowne.substack.com/subscribe

When ML/AI Engineer William Horton last joined me, Maven Assistant had reached its first external users the day before. The healthcare AI agent was available to 20 percent of Maven Clinic’s users, and the team had deliberately withheld answers about benefits. A wrong response could shape a decision involving $15,000 of fertility coverage, and the evals had not earned the right to ship it.
Four months later, Maven Assistant is available to 100 percent of users, benefits answering is live, and weekly conversation volume has grown by roughly ten times. Real usage also overturned part of the roadmap. The team had invested heavily in provider search and appointment tools, but 50 to 60 percent of early conversations were basic health questions such as whether someone could eat tuna while pregnant.
Production changed the engineering system too. An emergency guardrail told someone already in the ER to go to the ER. Zendesk content told people already using the Maven app to open the app. A newer model failed an upcoming-appointments eval because it correctly noticed that the mocked appointments were in the past.
William explains how Maven turns those failures into deterministic tests, LLM judges, synthetic negatives, and manual review. He also walks through the move from Gemini Flash models toward newer OpenAI models, what GPT-5.6 and Fable mean for a production agent, why model upgrades can make old prompt instructions obsolete, and why open-weight models still have to justify their GPU, infrastructure, and engineering costs.
“If anybody tells you that they’ve got their evaluations so good that they can just swap a model and know, with no manual review, that it’s going to be better, that person is probably lying, or they work at one of three places in the world.”
— William Horton, Staff Machine Learning Engineer, Maven Clinic
You can also find the full episode on Spotify, Apple Podcasts, and YouTube.
👉 Want to build agents from the ground up? Registration is open for Build AI Agents from First Principles, a live workshop on the loops, tools, context, harnesses, and engineering decisions behind useful AI agents. You’ll learn how to design agent systems from first principles, with enough structure to decide which harness patterns your product actually needs. Sign up today with code production10 for 10% off. 👈

In This Episode
* What changed between 20 percent and 100 percent rollout. Benefits answering cleared its release bar, Maven Assistant reached the remaining users, and weekly conversation volume grew by roughly ten times.
* Why real usage beat the original roadmap. The complex provider and appointment agents received less traffic than expected, while 50 to 60 percent of early conversations were basic health questions.
* How production failures enter the evaluation system. William estimates that the deterministic tool-use layer now involves more than 1,000 test scenarios, while clinical quality, empathy, and completeness still require judges, human calibration, and manual work.
* When the model is right and the eval is broken. GPT-5.6 Terra rejected an “upcoming” appointment that had already happened, exposing a bad test fixture that the previous model had accepted.
* What happens when GPT-5.6 or Fable arrives. A model swap can remove old prompt instructions, add new behavioral failures, change latency and cost, or reveal that yesterday’s harness is constraining a more capable model.
* Why Maven moved beyond Gemini Flash 2.5. William discusses adopting newer OpenAI models, keeping real-time chat on smaller models and low reasoning settings, and changing the model without simultaneously rewriting the prompt.
* The economics of open-weight models. A smaller self-hosted model still needs an always-on GPU, infrastructure, and engineering attention that could otherwise go into the product.
* What William would rebuild today. Provider search and appointment booking probably belong in one agent, and model experiments should begin before the original choice hardens into the architecture.
Start With the First Episode
William first joined Vanishing Gradients the day after Maven Assistant reached external users. In Building an Enterprise AI Agent for Healthcare, he explains the original architecture, how failures become regression cases, why deterministic checks should come before LLM judges, and how the consequence of a wrong answer sets the release bar.
Resources
* Maven Clinic
* Building an Enterprise AI Agent for Healthcare
* Stop Overengineering Your Agent Harness
* Build AI Agents from First Principles
* All Vanishing Gradients workshops
👉 Want to build agents from the ground up? Registration is open for Build AI Agents from First Principles, a live workshop on the loops, tools, context, harnesses, and engineering decisions behind useful AI agents. You’ll learn how to design agent systems from first principles, with enough structure to decide which harness patterns your product actually needs. Sign up today with code production10 for 10% off. 👈
How You Can Support Vanishing Gradients
Vanishing Gradients is a podcast, workshop series, blog, and newsletter focused on what you can build with AI right now. More than 70 episodes with expert practitioners from Google DeepMind, Netflix, Stanford, and elsewhere. Hundreds of hours of free, hands-on workshops. All independent, all free.
If you want to help keep it going:
* Become a paid subscriber, from $8 per month
* Share this episode with a builder who would find it useful
* Subscribe to the Vanishing Gradients YouTube channel
* Join another Vanishing Gradients workshop


Get full access to Vanishing Gradients at hugobowne.substack.com/subscribe

Every capability in an agent needs its own evidence and release bar. A model-provider slip, an incorrect tool call, and a wrong fertility-benefits answer should not be held to the same pass rate.
William Horton, Staff AI Engineer at Maven Clinic, joined us the day after Maven Assistant reached its first external users. The agent helps members inside Maven Clinic’s women’s and family healthcare platform find providers, manage appointments, navigate Maven, and get basic health information. William had spent much of launch day reading chat traces and turning the surprises into product decisions and tests.
William shows how a production failure moves through Maven’s system: the trace becomes a regression case, code handles deterministic checks, and LLM judges cover behavior that cannot be reduced to exact outputs. Human labels calibrate those judges, while the consequence of a wrong answer determines whether the capability ships. You can apply the same release workflow to the agent you are building now.
“For a lot of our tool-call evaluation, I’ll accept that it runs ten times and passes nine times. Going for that ten out of ten is just not worth the effort.”
— William Horton, Staff AI Engineer, Maven Clinic
You can also find the full episode on Spotify, Apple Podcasts, and YouTube.
👉 Want to build agents from the ground up? Registration is open for Build AI Agents from First Principles, a live workshop on the loops, tools, context, harnesses, and engineering decisions behind useful AI agents. You'll learn how to design agent systems from first principles, with enough structure to decide which harness patterns your product actually needs. Sign up today with vg-code for 10% off 👈

In This Episode
* The architecture behind Maven Assistant. A stronger lead agent routes requests to four narrower specialists for appointments, provider search, health questions, and Maven support. Hard guardrails run before dynamic routing.
* Why an enterprise healthcare assistant only needs 15 to 20 tools. Maven divides a manageable toolset across its specialists instead of exposing one model to hundreds of choices. Existing APIs become safer agent tools, with user identity and other application state injected by code.
* Turn failures into the cheapest reliable eval. A response claiming the agent was “made by Google” became a string check, tool calls are verified deterministically, and LLM judges handle clinical accuracy and other qualitative behavior.
* Set release thresholds from the consequences. Nine passes in ten can be acceptable for a cheap failure. Maven withheld benefits answers that could influence tens of thousands of dollars and routes self-harm language directly to human support.
* Let production change the product and the test set. Early chats changed the roadmap, became regression cases, exposed weaknesses in the judges, and supplied realistic opening messages for simulated users.
Join the Four-Month Follow-Up
This episode was recorded live inside our Building AI Applications course the day after Maven Assistant reached its first external users. By the follow-up four months later, Maven will have a much larger body of real conversations. William will return to compare the launch assumptions with what members actually used, which evals changed, and how newer models altered the system.
Register to join the livestream or receive the recording afterwards.
Resources
* Maven Clinic
* Maven introduces Maven Intelligence
* Google Agent Development Kit

👉 Want to build agents from the ground up? Registration is open for Build AI Agents from First Principles, a live workshop on the loops, tools, context, harnesses, and engineering decisions behind useful AI agents. You'll learn how to design agent systems from first principles, with enough structure to decide which harness patterns your product actually needs. Sign up today with vg-code for 10% off 👈
How You Can Support Vanishing Gradients
Vanishing Gradients is a podcast, workshop series, blog, and newsletter focused on what you can build with AI right now. Over 70 episodes with expert practitioners from Google DeepMind, Netflix, Stanford, and elsewhere. Hundreds of hours of free, hands-on workshops. All independent, all free.
If you want to help keep it going:
* Become a paid subscriber, from $8/month
* Share this with a builder who’d find it useful
* Subscribe to our YouTube channel
* Join one of our other workshops here


Get full access to Vanishing Gradients at hugobowne.substack.com/subscribe

Nicolay Gerold works all day and night on AMP, one of the most interesting coding-agent harnesses out there.
If you’re building with coding agents, this conversation will help you understand:
* when to trust the model,
* when to build harnesses around it,
* which model is worth paying for,
* which programming languages gives the agent better feedback, and
* when to take the keyboard back.
Coding-agent products are living inside a blender. Opus 4.8 to Fable changes what the model can be trusted with, eats a workflow, and suddenly the best product decision is to delete code.
AMP had handoff because long agent threads used to get messy. Compaction would lose the plot, the model would make worse decisions, and the product needed a way to move the work somewhere cleaner. Then compaction got better. The model ate the feature. AMP killed it.
Builders inherit the annoying product test: does this harness code help inspect, verify, recover, or merge model work, or is it just babysitting yesterday’s model?
Nico and Hugo riff on why loop engineering is overrated (and when to use it), why Fable is the first model with real engineering taste, and why you should stop writing Python code today and start writing TypeScript and Rust for all your AI Engineering workflows.
You can also find the full episode on Spotify, Apple Podcasts, and YouTube.
👉 Want to build agents from the ground up? Registration is open for Build AI Agents from First Principles, a live workshop on the loops, tools, context, harnesses, and engineering decisions behind useful AI agents. You'll learn how to design agent systems from first principles, with enough structure to decide which harness patterns your product actually needs. Sign up today with vg-code for 10% off 👈

In This Episode
* Coding-agent harnesses today: compaction, sandboxes, review flows, and the features frontier models are starting to absorb.
* Why AMP keeps deleting its own features when models get better.
* The test for every harness feature: does it make the agent’s work easier to inspect, verify, or recover from?
* Local agents, cloud sandboxes, and where each fits when bugs, issues, logs, or customer feedback turn into code changes.
* Background agents without auto-merge fantasy: how useful work comes back as branches, checkouts, or review candidates.
* Loop engineering in practice: tight loops with clear objectives, broad loops that create review overload, and where builders should draw the line.
* When deterministic code beats an AI step, and when a single agent with the right tools can replace brittle orchestration.
* The TikTok problem for coding: hundreds of agent threads, fragmented attention, and why loop engineering can become a trap.- The TikTok problem for coding: hundreds of agent threads, fragmented attention, and why loop engineering can become a trap.
Resources
* AMP
* AMP Owner’s Manual
* Nicolay Gerold’s Show Us Your Agent Skills dossier
* Clio: Privacy-Preserving Insights into Real-World AI Use
* TigerBeetle TigerStyle
* How to Build A Coding Agent with Nico and Hugo
Build AI Agents From First Principles
👉 Want to build agents from the ground up? Registration is open for Build AI Agents from First Principles, a live workshop on the loops, tools, context, harnesses, and engineering decisions behind useful AI agents. You’ll learn how to design agent systems from first principles, with enough structure to decide which harness patterns your product actually needs. Sign up today with vg-code for 10% off. 👈
How You Can Support Vanishing Gradients
Vanishing Gradients is a podcast, workshop series, blog, and newsletter focused on what you can build with AI right now. Over 70 episodes with expert practitioners from Google DeepMind, Netflix, Stanford, and elsewhere. Hundreds of hours of free, hands-on workshops. All independent, all free.
If you want to help keep it going:
* Become a paid subscriber, from $8/month
* Share this with a builder who’d find it useful
* Subscribe to our YouTube channel
* Join one of our other workshops here



Get full access to Vanishing Gradients at hugobowne.substack.com/subscribe

So I think we’re really at a historical moment, and the opportunity is massive. Almost 15 years ago, we were promised that data science was going to be this incredible thing and create all this value for people. And I think nowadays it’s mostly viewed as a cost center in most companies. I think we can really now fulfill that original promise with agentic data science.
Thomas Wiecki, Co-creator of PyMC and Founder at PyMC Labs, joins Hugo to talk about how agentic data science is finally fulfilling the promise of Decision Intelligence.
We Discuss:
* Decision Engines: Transform data science from a cost center providing cryptic answers into a real-time decision intelligence hub delivering actionable outcomes;
* Tame the “Garden of Forking Paths”: Overcome human shortcuts by running parallel analyses to provide an honesty check, revealing the true robustness of business conclusions;
* Multiplayer Data Science: Foster organizational learning by moving agents into team chats, democratizing “what-if” questions and reducing context-switching friction;
* The Full Agentic Data Science Stack: Beyond harness and skills, the full stack includes orchestration for parallel analyses and a causal eval layer to measure actual outcome improvement;
* Agentic Dashboards: Move beyond static BI; use chat interfaces to inquire into models and generate real-time, custom visualizations for specific follow-up questions;
* Encode Professional Judgment as Skills: Elevate agent performance by encoding expert domain standards and high-fidelity workflows into specific Agent Skills, rather than relying on LLM pre-training;
* Ground Decisions in Generative Processes: Prevent hallucinations by forcing agents to model underlying physical or behavioral processes, providing a programmatic guardrail aligned with market realities;
* Scripted Causal-Bayesian Workflows: Their methodologically structured nature—from prior elicitation to posterior predictive checks—makes Causal-Bayesian workflows inherently automatable for agents;
* Iterative Autonomy via Skills: Achieve autonomy iteratively: verify workflows with human oversight, then encode verifiable parts as skills to hand off trusted tasks;
You can also find the full episode on Spotify, Apple Podcasts, and YouTube.
You can also interact directly with the transcript here in NotebookLM: If you do so, let us know anything you find in the comments!

👉Want to learn how to apply agentic engineering to the world of data science? Come build the future of Agentic Data Science with us in our upcoming course. It’s a live cohort with hands on exercises, capstones, and reusable agent skills, OSS code, and notebooks that will 10x your data science projects. Sign up here and use the code ADSVG10 for 10% off. Hit reply to enquire about group discounts.👈
LINKS
* Thomas Wiecki on LinkedIn
* PyMC Labs
* Open-Sourcing Decision Lab: Scaling AI Judgment in Data Science (PyMC Labs blog)
* Decision AI Discord
* Many Analysts, One Data Set: Making Transparent How Variations in Analytic Choices Affect Results (Sage Journals)
* The Agent Harness Reading List
* Show Us Your Agent Skills (GitHub)
* Agentic Data Science course with Hugo, Thomas, and Luca (10% off with code ADSVG10)
* Upcoming Events on Luma
* Vanishing Gradients on YouTube
* Watch the podcast video on YouTube

👉Want to learn how to apply agentic engineering to the world of data science? Come build the future of Agentic Data Science with us in our upcoming course. It’s a live cohort with hands on exercises, capstones, and reusable agent skills, OSS code, and notebooks that will 10x your data science projects. Sign up here and use the code ADSVG10 for 10% off. Hit reply to enquire about group discounts👈


Get full access to Vanishing Gradients at hugobowne.substack.com/subscribe