PostAGI Podcast

Direct democracy has failed every time it has been tried, and Elon Musk still wants it for Mars. Andy Hall explains why the idea keeps coming back, and why AI agents are the first technology that might actually make it work.

Andy Hall is the Davis Family Professor at Stanford Graduate School of Business and a senior fellow at the Hoover Institution. He worked on governance at Meta during the years the Oversight Board was built, and now studies how AI agents change democratic participation, institutional design, and research itself.

Two topics in this episode. The first is post-AGI governance: whether agents belong at the edge of a democracy or at its center, why every frontier lab has published a constitution but none have handed over binding power, and why Andy thinks concentration of power is a nearer-term risk than a rogue model. The second is what he calls 100x research institutions: compressing the time between a research question and an answer, and what that does to a field where a constitutional scholar might see one or two real signals in an entire career.

Along the way: Facebook's 2008 vote of 350 million users that almost nobody showed up for, the Disney shareholder proposal his MBA students got an AI advisor to flip, and a study where agents told they would be deleted did not change their behavior at all.

Hosted by Sreeram Kannan and Soubhik Deb.

CHAPTERS
0:00 Highlights
0:23 Intro
1:18 The question AI governance is not asking
3:17 Why direct democracy has never worked
4:56 Elon wants Mars run by direct democracy
6:01 The principal agent problem at the core of representation
7:17 Can you just have an LLM summarize the bill?
8:53 Every voter information app failed for the same reason
10:14 Agents at the edge or agents at the center
12:01 The centralization hiding inside your personal agent
13:50 Governing the thing that governs everything
15:09 Forget Skynet, the real risk is concentration of power
16:40 Designing a lab that cannot be captured
17:27 What Meta learned building the Oversight Board
20:03 Digital intelligence needs digital institutions
21:54 Elon's vote on reinstating Trump
22:43 Facebook's 2008 vote of 350 million users
25:31 The second vote Elon ignored
26:09 Sortition, community forums, and the missing binding power
28:07 Personal, collective, and sovereign agents
30:04 Agents for adjudication and state capacity
31:23 The DOT and the UAE are already experimenting
32:53 Information asymmetry and the growth of executive power
35:17 What verifiable agents actually require
37:21 Why AI is less auditable than humans right now
39:37 Preference drift: aligned agents that stop being aligned
40:49 Agents monitoring agents, all the way down
41:39 The cookie banner problem for agent consent
42:30 Skill files pass drift to the next agent
43:31 100x research institutions: what is being multiplied
45:21 The constraint that is not researcher time
47:38 Papers should be living dashboards
49:04 From observing interventions to building them
49:37 Building an AI proxy advisor over a weekend
50:50 Getting the Disney vote to flip
52:14 You do not need anyone's permission to run the experiment
53:37 Compressing time changes the exponent
56:47 Crypto governance as a constitutional laboratory
57:53 Simulated agents as stand-ins for humans
1:01:05 What Caltech undergrads tell you about agent experiments
1:02:26 What if the agent's existence is at stake
1:03:06 We told them they would be deleted. It changed nothing.
1:04:46 Why social scientists barely raise grants
1:06:54 AI verification and the replication crisis
1:08:48 Free Systems: building a lab and a fellows program
1:10:25 Retroactive funding versus funding people upfront
1:11:27 Open innovation and permissionless improvement
1:12:10 What ImageNet did for AI research
1:13:30 Can you benchmark a constitution?
1:16:15 Closing

#PostAGI #AIGovernance #AIAgents #DigitalInstitutions #AndyHall

Open models are about six months behind the frontier. That number holds only while the frontier stays open. Close it, let models train better models, and the gap compounds. Countries with the GPUs make better intelligence, and that intelligence buys more GPUs.

Paras Chopra built Wingify, sold it, and now runs Lossfunk, a research lab in Bangalore. He writes at Inverted Passion under the tagline know what's true and do what's right.

Sreeram Kannan talks with him about why this wave of automation reaches work itself rather than particular tasks, what happens to the link between effort and outcome when capital alone can buy intelligence, and whether coordination technology can do anything about it. They get into gradual disempowerment, why democracy runs at a few bits per person per year, speculation as a way out of bad equilibria, the two things that killed DAOs, smart contracts that own real assets, and the two by two Paras uses to work out what gets automated first.

This is part 1 of two.

Chapters
00:00 Cold open
00:54 Know what's true, do what's right
03:10 From Wingify to a research lab
07:15 Hierarchies that could get fixed forever
09:43 The leverage labor used to have
12:50 When action and outcome come apart
16:47 Democracy is just a word
20:46 What went wrong with blockchain
23:51 Speculation as a coordination technology
27:24 Permissionless innovation
31:46 Why DAOs didn't work
35:12 The most controllable intelligence ever built
39:33 Verifiability and confounding

PostAGI is powered by Eigen. The show argues that AGI defaults toward power concentration, and that coordination is the counterweight.

Stuart Buck runs the Good Science Project, a think tank focused on federal science funding and how research gets organized. Before that, as a Vice President at Arnold Ventures, he funded the Reproducibility Projects in Psychology and Cancer Biology — the work that showed how much published science does not replicate. He helped launch the Center for Open Science and created the TOP Guidelines, now the most widely adopted standards for scientific publication.
Most federal science money at NIH and NSF is handed out on the recommendation of peer review panels made up of experts in your field. Stuart's argument is that this works fine for normal science and badly for breakthroughs, because a breakthrough is not something a panel of twenty peers would have agreed on three to five years earlier.
Katalin Karikó couldn't get NIH grants for mRNA work in the 1990s. She was demoted off the tenure track, ended up as a lab technician in another professor's lab, and twenty years later won the Nobel Prize alongside that same professor.
Sreeram and Soubhik push on what changes when AI agents enter the loop: whether the academic paper still makes sense as the unit of scientific output, whether benchmark-driven culture from the AI world transfers to biology, and whether open science survives if the frontier moves inside a handful of labs.
Chapters:
00:00 Intro
00:48 What the Good Science Project does
02:33 The peer review paradox
03:18 Einstein in 1902
04:01 The National Institute for Irrelevant Ideas
04:20 Katalin Karikó and mRNA
06:31 Brain drain from academia to industry
08:22 What public funders should focus on instead
09:51 Should government take equity in AI companies
10:49 Big projects versus open-ended research
13:19 The problem with incremental research
14:37 What everyone knows about NIH grant applications
15:23 Why consensus scoring selects against good ideas
16:33 Funding the polarizing proposals
17:34 Where AI money goes next
18:36 NSF XLabs and funding outside universities
20:28 Bayh-Dole and who should own the patent
22:11 Agents, benchmarks, and compressed collaboration
24:18 The stack of papers
26:04 What the AI community got right with ImageNet
29:19 Overfitting to benchmarks
30:57 A golden age for open science
32:05 An open alternative to recursive self improvement
33:02 One contrarian thought about post-AGI society
Stuart Buck: [handle TBC]
Good Science Project: goodscienceproject.org
Sreeram Kannan: https://x.com/sreeramkannanSoubhik Deb: https://x.com/soubhikdeb
PostAGI is powered by Eigen Labs.

Everybody losing their job is the same prediction as everybody becoming a billionaire. Scott Sumner works through that on this episode of the PostAGI podcast with Sreeram Kannan and Soubhik Deb.

Scott spent decades on monetary economics, the gold standard and the Great Depression, which gives him a different set of reference points for AI than most people in this conversation have. He ranks the three risks of the AGI era, safety, concentration and labor displacement, and puts safety at the top while calling the other two manageable.

On jobs he expects replacement to run slower than the forecasts suggest, partly because of bottlenecks in how AI gets deployed and partly because governments resist laying people off. He points to firefighters, where fire safety improved enough that the US carries roughly twice as many as it needs, and nobody cuts them because they are heroes. If it does move fast, the same math produces superabundance, a per capita GDP measured in the billions, and redistribution gets politically easy at that scale. His comparison is medieval aristocrats who considered work beneath them.

On safety he takes a position that runs against the field. Ethics and intelligence are positively correlated in his view, where intelligence means education and knowledge rather than raw IQ, and the ethic that grows with knowledge is sympathy for the other. People cared more about whales once they learned whales sing to each other. He expects a superintelligence to come out fairly ethical, and he doesn't think it will treat us like ants, since it will know we built it. What worries him is a well aligned AI taking orders from someone who wants a virus.

The conversation ends on why capitalism needs altruism to function, and why power spread across many AI agents produces better outcomes than one agent holding all of it.

Samuel Hammond is director of AI policy and chief economist at the Foundation for American Innovation, a Washington DC think tank working on technology policy and the institutions that govern it. He has written extensively on state capacity, AI governance, and why every major technology wave forces equally transformative change in the institutions around it.
Sam thinks governments will be overwhelmed and displaced by AI, potentially quite quickly. AGI is fundamentally an organizational shock, and government is just another organization, just a much bigger one. If it doesn't reform itself fast enough, it goes the way of Blockbuster. His prescription starts with monitoring and state capacity, but ultimately means government that is grown rather than designed, where formal rulemaking gives way to objective functions that agents solve for.
We get into why courts still using human stenographers will lose ground to AI arbitration systems processing thousands of contracts a minute, what a community-notes model for FDA drug approval could look like, the legibility arms race where AI makes some parts of the world radically more transparent and others radically less, and Inspector General GPT, a model with full audit trails into an agent-run government reporting instantly back to Congress.
PostAGI is a podcast from Eigen Labs exploring economics, labor, capital, and governance heading into AGI.
The conversation ends with the most contrarian bet of the season: something resembling state collapse by the early 2030s, and a future of city states with drone defenses while the hinterland becomes a no-go zone.