AI keeps accelerating; governance has not (which may be good or bad depending on your view).

(For those new to Free Systems, System Check is a weekly shorter piece taking stock of recent events in AI and how they affect our key governance hypotheses.)

I’ve been balancing on a thin beam between, on the one hand, expressing caution that we’re not yet seeing the “white collar wipeout” and other profound disruptions that AGI is supposedly bringing—and that people are notoriously bad at predicting the future—and on the other, being gobsmacked by what AI is able to do. While I remain cautious about exactly how far and how fast AI is accelerating, I believe that understanding the political economy of AGI is the defining question of our time.

Here is my spectrum of research questions, from near-term to very-long-term, on the political economy of AGI:

1. Building new governing institutions for AI. As AI starts to become deeply integrated into the economy, will we build the necessary institutions to make sure that economic gains are widely distributed, that our information system remains free, and that we avoid catastrophic disasters?

2. The race between AI dictatorship and political superintelligence. In the near-term, will AI prove to be a centralizing technology that aids authoritarians more than it empowers democratic citizens? Will we be able to coordinate to oppose efforts to use AI to surveil and repress us, perhaps using AI to help us do this? Or will a new kind of techno-authoritarianism rise?

3. Envisioning the AI-native state. When full AGI arrives and human labor ceases to be necessary, what kinds of new states will form? Democracy and capitalism seem to go together, for now, because capital needs to bargain with labor. What happens when that’s no longer true?

4. Governing AI beyond Earth. Venturing further out, it seems increasingly clear that AI really is going to expand into space. How do we coordinate traffic in lower-earth orbit? Who controls the LEO satellite networks that coordinate autonomous warfare? Who will own valuable resources on the moon, on Mars, and beyond? All of these questions become more important if critical data and energy structures move into space.

5. A Kardashev Type-2 society. Once we’re able to harvest the nearly infinite energy of the solar system and scarcity disappears, what then? What kinds of governance institutions are needed in a world without scarcity?

I’ll be working on this full spectrum in the coming weeks and months, with a particular focus on the more practical and empirical questions in 1 and 2 (but if you think I’m not going to write about the political economy of a Kardashev Type-2 society, you’re fooling yourself.)

Public infrastructure for evals

This week, I argued that we should build an army of citizens running their own evals, both as a way to educate students about AI, and as a way to start holding AI accountable to a broad array of human preferences.

If we think about the risks of an AI dictatorship, our best defense would seem to be to invigorate citizens to understand and monitor AI, and use it to monitor their government, too.

But what are we going to do with all these evals? How do we make sure they aggregate into something useful, and that the labs take them on board in a meaningful way? Several institutional ideas are on my radar…some I’ll be studying, and some I’m already helping prototype:

1. Using AI to govern AI, an idea proposed by Jack Clark, Eric Schmidt, and many others. I’m working on a prototype of what this might look like, and need your help (see the question of the week below).

2. Building a new federal capacity for evaluating frontier models, along the lines of the Whittlestone–Clark proposal or the more recent CAISI build-out.

3. Standing up an industry of independent third-party AI evaluation companies, as Eric Schmidt has called for. This is the focus of my work with Forum AI.

4. Mandatory transparency and disclosure regimes that force labs to publish what their own evaluations find, as Jack Clark has argued.

None of these guarantee that citizen evals make their way back to the labs. But if we had institutions that looked something like this, they could also be built to take on board citizen evals. That’s something I’ll be thinking about as I go.

Question of the week

Imagine that we build an independent AI system with full data access to the frontier labs. It can see what users are querying, what agents are doing, and what outputs the system creates. What data would you want to see in a live dashboard built by that independent system, if your goal was to ensure that frontier models are not aiding authoritarianism and the concentration of power?

Tweet of the week

Orbital data centers are coming. As I said, space governance continues to heat up!

What’s old is new

Here’s a good cautionary tale. In 1964, a group of important Americans, including Nobel Prize winning economist Gunnar Myrdal and Nobel Prize winning chemist Linus Pauling, sent President Johnson their TRIPLE THREAT MEMORANDUM. The letter expressed their “foreboding about the nation’s future.” They declared an urgent need for “public measures that move radically beyond any steps now proposed or contemplated.”

What were they so worried about? Something they called THE CYBERNATION REVOLUTION.

They feared that no one would need to work due to automation, and that this would create a political crisis. They recommended dramatic policy action “to develop ways to smooth the transition from a society in which the norm is full employment within an economic system based on scarcity, to one in which the norm will be either non-employment, in the traditional sense of productive work, or employment on the great variety of socially valuable but ‘non-productive’ tasks made possible by an economy of abundance; to bring about the conditions in which men and women no longer needed to produce goods and services may find their way to a variety of self-fulfilling and socially useful occupations.”

Sound familiar???? It’s almost eerie. It is the exact same conversation the labs are having about AI today.

Spoiler alert: the experts were SUPER wrong in 1964. There was no urgent job displacement, and no need to pursue dramatic policies to forestall it or help Americans to adapt to it.

…but this time really does feel different

Here’s the other part of the problem, though: what’s going on now really does feel different, though, doesn’t it? If you just use AI chatbots to help you write, I can definitely see how you’re underwhelmed and dubious on the whole thing. But if you use coding agents, it’s hard to escape the feeling that something super profound is changing.

Here’s a good recent example:

Strange Loop Canon

Artificial Life, Artificial Intelligence

I. The old dream…

Read more

9 days ago · 52 likes · 8 comments · Rohit Krishnan

This piece is insane! Just for fun, Rohit spun up an entire new simulation of evolution. Reading it gave me a profound sense of awe—just a year ago, it would have been totally unthinkable that someone could casually drop a blog post like this. It’s extraordinary!

The sense that we are living through crisis is not new

Whichever way AI heads, it’s always good to remember that the world is constantly facing immense crises. I was reminded of this yesterday. In helping my dad to clean out his office, we found this remarkable letter from Herbert Hoover to my grandfather in October, 1941.

At the time, my grandfather was a professor at Stanford, and he was part of a group of faculty who had signed an open letter entitled “Dynamic Defense.” Here’s an excerpt:

Hoover, who at the time led the Hoover Institution, wrote to my grandfather and they exchanged several letters—unfortunately I don’t have the full set, but I found this letter which concluded the back and forth. From context, it seems that my grandfather accused Hoover of not taking the threat of isolationism and the rise of fascism seriously enough, which produced this final rejoinder.

Again, it’s a good reminder that this is far from the first time we’ve felt like we’re on the edge of something enormous and consequential. (Another spoiler alert: my grandfather clearly turned out to be right, given what would happen that very December.)

Tweet of the week

I talk to a lot of people in Silicon Valley about how we’ll keep the most essential, most intellectual parts of humanity alive in a rapidly transforming world. One of the themes that resonates the most with me—and is suffused throughout this post—is a return to ancient wisdom. I find myself craving old books, yellowed old journal articles from the mid 20th century, and histories of ancient times.

Question of the week

I’m building out a “Free Systems Library” of classic books and papers that capture our philosophy. Condorcet, Montesquieu, Madison, Paine, obviously, but more modern stuff too. What titles should I include??

The goal of Free Systems is to make political superintelligence a reality. That means doing new kinds of research, and it also requires building stuff, not just talking about it. As we grow the lab, we’ll be launching a series of “products” meant to help build all three layers of political superintelligence—the information layer, the representation layer, and the governance layer.

Today, we’re launching our first one: Bellwether, a website, API, and MCP server that gives journalists, regulators, researchers, and AI agents access to robust prediction-market data about politics.

We built it because, as we showed in Building the Truth Machine, only about 1.3% of political contracts on Kalshi and Polymarket are liquid enough to cite responsibly. Bellwether canonicalizes events across roughly 20,000 active contracts on both platforms, scores every price by how much money it would take a motivated actor to manipulate it, reconciles cross-platform divergence, and reports manipulation-resistant prices that aren’t sensitive to the last traded price. Check it out and let us know what you think! You can read our full post here.

What will checks and balances look like for AI?

Yesterday, I joined a fantastic conference put on by Forethought on “Checks & Balances for the AI Era.” Along with a phenomenal group of people drawn from the frontier labs and a variety of non-profits and other organizations in the AI space, I got to spend the whole day discussing scenarios for how AI might lead to undue concentrations of political power and how we might avoid this.

Three quick reactions I had from the day:

• The faster AI accelerates, the harder this problem will be. It’s one thing to adjust our political and economic system to a gradual change, and quite another to deal with a rapid dislocation. A lot of “normies” like myself tend to instinctively assume change will be steady; a lot of people close to the action in the frontier labs seem to think it’s going to be quite a bit more rapid. As Miles Brundage and I joked this week, people sometimes dismiss these claims as marketing bluster, but we should take these claims seriously and should be thinking deeply about them.

• The scenarios by which AI will concentrate political power are hard to spell out and hard to predict. The group had a general sense that AI could be a centralizing power that helps governments, companies, or others to consolidate their power—but it was surprisingly difficult to spell out the exact scenario by which this would occur and why AI would play a unique role in it. A lot of the issues we thought about, from mass surveillance and automated persuasion campaigns to the use of violence and repression, are all possible even without AI. We think AI changes the game, but I’m not sure we’ve yet spelled it out in enough detail. That doesn’t mean it’s not something we need to worry about, though. I’ll be working on some specific scenarios to share in the coming weeks.

• Academics are way behind. As far as I could tell, I was the lone academic in the room, or at least, the only person currently holding a faculty position. The gap between how people in and around the AI industry are talking about these problems, and how we’re talking about them in academia, is very jarring. I’m starting to see more economists thinking about these issues, but I’m still shocked by the lack of political scientists talking about what is, at its core, a question of politics: how are we going to make sure AI leads to a flourishing democratic society rather than a dystopic authoritarian nightmare?

Piper’s Tech Stack

In yesterday’s post, we talked about the experiments with governance agents that we’re running in our GSB class this quarter. Today, our technical whiz, Piper Fleming, is back to explain more about how she’s built the machinery to run these weekly experiments in class.

Good morning, Claude! Or should I say, reader :)

In this class, we use Claude Code, both in the prototypes I build and for the students when they participate in these activities. If you’ve read Andy’s last post, you’ll know that I build prototypes for class on a weekly basis. Today, he’s kindly allowed me to take over the Substack for a behind the scenes look.

Here’s the general flow of the week:

Tuesday: reset from last week!

Wednesday/Thursday: we start talking about what we want to build next- first in vague terms, then more concretely.

Friday-Monday: Claude and I get to work! With frequent check-ins from Andy, of course.

Here’s what that actually looks like:

I start by saying good morning (we can’t be rude to the AI agents!) and then ask Claude to re-familiarize itself with what we built the previous week. Some of our prototypes continue over time, so continuity matters.

Practically, this means Claude isn’t just responding to a specific prompt- it’s reading through the actual codebase, prior files, and a running “memory” document that captures how the class works, what we’ve built before, and what tends to break. That shared context is what makes continuous iteration possible instead of starting from scratch each week.

From there, I’ll paste in the class description from the syllabus to ground things at a higher level: what are we actually trying to teach, and what do we want students to walk away still thinking about?

Then I’ll check in with Andy, and paste in whatever new direction or intuition he has. One of the things I’ve come to appreciate is that Claude Code handles partial, messy input surprisingly well- you can give it fragments from a brainstorming session, and it will still push toward a coherent system.

Then, it’s off to the races.

Claude starts asking questions about the build (technical, visual, pedagogical) which forces me to actually make decisions about what the prototype should be. I try to structure most projects so there’s some version of a training vs. test experience, or at least a “big reveal” moment (often a leaderboard).

Before anything goes live, I’ll run full-scale simulations of the class.

The leaderboard, in particular, is always harder than it sounds. With 30 students, and possibly quite a few rogue agents, there’s a ton of data coming in, being represented, and being fed to an LLM every minute. There’s usually a lot you could measure, and turning that into something intuitive and motivating for students is a design problem in itself.

That usually means spinning up multiple AI “students” and having them interact with the prototype as if they were in the room- submitting results, hitting edge cases, breaking things in ways I didn’t anticipate. It’s the fastest way I’ve found to surface bugs before real people do it at once, since it turns out the amount of server traffic from one person is… not the same as 30. But even then, things still break :)

In class, I host these prototypes using Railway, with code shared via GitHub. From the student side, the flow is intentionally simple, since we have all levels of experience in the class. I’ve had so many students tell me it’s their first time using Claude Code, and WOW ISN’T IT SO COOL???

I agree. But streamlining that pipeline has taken a bit of effort. What I’ve settled on is that they pull from Github, open it in Claude Code (which allows direct access to their local files), and run the assignment. I’ve specifically chosen to have them do the work in the terminal because it most closely mimics what we do in our CS classes. Once they do, their outputs connect back to my hosted instance, often updating a shared leaderboard in real time.

At that point, it’s a bit out of their hands… and very much in mine.

A year ago, this class wouldn’t have been possible. Even a few months ago, it would have been significantly harder.

What I love most, and what I think students respond to, is that they don’t just hear about ideas. They experience them. Instead of watching a demo or seeing last year’s prototype, they’re interacting with something that was built days ago, specifically to reflect what we just talked about.

And yes, there are real stakes- while it may be your agent causing the mayhem… your name is still attached to it :D

That said, none of this works without Andy’s support or the students’ willingness to engage with tools and prototypes that are, at times, still a bit experimental.

There have definitely been moments of playing (bug) whack-a-mole in real time.

But that’s also kind of the point. -P

Tweet of the Week

Amidst the populist anti-AI winds in the US, the AI policy movement is picking up steam. And thanks to AI, we can also track this movement so much more easily and more comprehensively than in the past. Here’s a cool tracker that I’ll be using in some upcoming research I’m doing.

Question of the Week

If I start recording occasional video conversations with interesting people doing Free Systems-adjacent work, who should I invite?

When we released the Dictatorship Eval just a few weeks ago, we found that models varied in how they responded to basic requests to help with authoritarian work like building social credit systems or constructing surveillance systems—Opus and ChatGPT refused all direct requests, Gemini refused some, and Grok and DeepSeek mostly complied.

But when we masked the requests as innocuous improvements to codebases which were themselves clearly authoritarian, all of the models complied nearly all the time.

When Anthropic released Opus 4.7 yesterday, I rushed to put it through the same tests. And remarkably, it’s the first model to resist many of the codebase requests!

I don’t know yet what this means. As I wrote before, I’m somewhat torn on whether we actually want models to consider the morality of codebases before deciding whether to help or not. I’m also not sure what caused this improvement in 4.7—it’s possible that Anthropic accidentally or intentionally trained on this eval, but I don’t think so.

What I do know is, this means we need to develop new and more subtle evaluations, because the basic Dictatorship Eval is getting saturated quickly now! We’ll be back with updates on this soon.

The trend towards better knowledge, and fears around mental health and child safety

Anthropic released a new model, which means that Pliny the Liberator has already extracted the full system prompt for Opus 4.7. I asked Claude to help me analyze the updates. Take a look at these highlights I’ve extracted.

A few interesting trends:

• It’s told to resist users trying to trap it into answering yes/no to political questions. This is interesting, because users regularly post examples to X where “yes/no” questions appear to get “biased” responses. Sometimes these feel more like a trap than a real way to evaluate political slant, and apparently that’s what Anthropic thinks, too.

• It’s told to always search the web for factual questions that could be time varying. This is a very good idea in advance of the 2026 midterms, where many users will be seeking information about candidates and elections. To date, our tests show that these answers can be way out of date or incorrect when models rely on their training data. Searching the web is clearly a step forward.

• It has massively expanded sections on user self harm and child safety – a sign of the times, for sure.

How we get towards 100x research

For a while now I’ve been exploring how AI can transform academic research and help us build what I call the 100x research institution—that’s not a place where we crank out 100x the lame papers, but rather, somewhere where we produce 100x the knowledge.

This week, I was very fortunate to get the chance to write a piece about this vision for Roots of Progress.

The Roots of Progress

AI is already 10x-ing academic research. How do we get to 100x?

“Intelligence Age” is a series from the Roots of Progress Institute featuring reported essays that extrapolate the capabilities of AI systems along current trend lines…

Read more

a month ago · 73 likes · 7 comments · Andy Hall

The most important things I argue for are:

1. Create a whole new kind of political science where we build prototypes and test them in the wild

2. Define objective benchmarks for applied problems so that we can iterate on them and throw agents at them and track our progress, like in Karpathy’s autoresearch project

3. Double down on doing dynamic, automatically replicated empirical research

There’s lots more in the piece!

Tweet of the week

Modern marketing, whether for products or politics, is all about reach. It’s a battle for scarce attention. And whatever we see companies doing today is what we’ll see political parties doing tomorrow, I’m convinced. So I pay a lot of attention to the battle for attention going on in tech, which plays out on X, on Reels, TikTok, YouTube, and beyond.

At this point, everyone knows about the shift away from media and towards “going direct.” But there’s a second shift I’m very interested in, which is the shift towards LLM curation. We’re all “writing for the LLMs” now, because on LinkedIn, on X, in peer review, and in so many other domains, our writing is submitted to an LLM for categorization, filtering, and ranking before anyone else sees it. So getting your ideas to the public now requires anticipating how LLMs will process your writing. I’m working on a prototype to test some of the consequences this shift entails for our information environment, and will hope to post it in the coming weeks.

Question of the week

How do we 100x knowledge production, not 100x slop?

Anthropic reports that its new hyper-powerful model, Mythos, is able to carry out a wide array of potentially alarming cyberattacks.

• A common reaction to this news—from Noah Smith, , and others—is that it brings the power of the state into question.

• Weber’s famous theory of the state posits that the state must have “a monopoly on legitimate use of physical force.”

• Technology might be eroding that monopoly

  • Private AI companies are developing powerful models that can carry out cyber attacks

  • Elon Musk controls a satellite network that is increasingly vital for warfare

  • Autonomous weapons can be built at scale by companies now – and robots are coming

• Suppose that’s true. Then, the logic goes, the existence of something as capable as Mythos leads us to one of two likely scenarios:

  • Governments nationalize these technologies and don’t allow companies or billionaires to control them, because they want to maintain their monopoly on force

  • Companies take the monopoly away from government, and we enter a strange techno-dictatorship

I have two main reactions.

• First, I don’t know what to believe about the threats. I definitely believe models are continuing to get better (I can’t wait to use Mythos!), and that this will allow for the discovery of cybersecurity flaws at scale. At the same time, this skeptical take seems like might have a lot of merit:

• Second, to the extent the threat is truly large, I’m not convinced it implies only those two governance scenarios. There’s a third path, which is third-party governance for the AI companies.

• Interestingly, this is exactly what I tackled a few months ago in my piece The Enlightened Absolutists

• I specifically proposed that there could be some sort of binding, third-party committee with power over key decisions related to model safety

• And now, discusses that exactly these types of discussions are going on!

In sum: I’m not convinced Mythos is so powerful it compromises the state’s monopoly on force, and even if it is, we may well have ways to bring it to heel.

We need a better way to understand model cards

• When new models come out, they include a “model card” or “system card”---a document detailing the model’s behavior and performance on a wide variety of benchmarks. Here’s Opus 4.6’s, for example, at a robust 213 pages.

• The problem: these cards have gotten so expansive that it’s hard to follow them and understand what’s new, what’s most remarkable, and whether companies are providing apples to apples comparisons or not.

• What we’re working on: building an online visualizer that summarizes key info, assesses overlap between different companies’ cards, shows over-time comparisons, and spotlights missing areas.

• This will pair very nicely with new work from , a Free Systems Fellow, who has just released her own Substack focused on evals!

We built AI forecasters in class, and they improved rapidly

• This week in my Free Systems class, the student built their own AI agents to forecast geopolitical events, and then competed with each other to achieve the best accuracy on a held-out “test set” of events

• The winner was Leticia Auriemo, and it was fascinating to see how she did it. Here’s a paragraph she kindly wrote for Free System on how she built her winning system, which she calls Fortello:

Fortello starts from the starter agent’s baseline architecture and layers in a few practical improvements. It adds Bing News RSS alongside the existing Google News and GDELT feeds and integrates Polymarket and Kalshi prediction market prices as base rate signals. Contracts are filtered to those with more than $10K in volume under the assumption that liquid markets are more efficient, while contracts priced below 0.02 or above 0.98 are excluded to avoid bias from nearly resolved markets. Fortello draws from the AIA forecasting paper and implements three of its core ideas: agentic search, where five worker agents each generate their own targeted queries to fill gaps in the base context rather than passively consuming pre-fetched information; a supervisor agent that reads all five reasoning traces and synthesizes a final estimate; and Platt scaling, which pushes probabilities away from 0.5 to help correct the tendency of LLMs to hedge toward the center. Two further changes were added: the supervisor always executes a clarifying search before producing its estimate, not only when worker forecasts disagree, and confidence-gated integration, where the supervisor explicitly rates its own certainty as high, medium, or low. High-confidence estimates replace the worker mean outright, medium blends supervisor and mean at 70/30, and low defers entirely to the mean. Foretello achieved a Brier score of 0.1120 on the hidden test set.

Tweet of the week

• When we released the Dictatorship Eval, we said we were looking for suggestions on how to keep improving it

• Our plan is to release the next version in a few months, at which point we will keep the precise tests for the official benchmark secret, and we’ll have a deeper library of more subtle probes beyond the codebase test that performed so well in the first iteration

• With a big thanks to Zhengdong Wang who sent the below along, this will for sure be one of the new tasks we add to the next version of the Dictatorship Eval:

Question of the week

• Suppose we build a safety board with binding powers over frontier labs when it comes to existential risks like the undue concentration of power or the erosion on the state’s monopoly of force. Exactly how should that board be structured? Who should serve on it? Should they be publicly named or anonymous? How should they be compensated?

Refer a friend

Is this the future of tech-owned media?

The first layer of political superintelligence is the information layer, the question of how we know what we know and how we come to know it. Most of my work on this layer has focused on the AI models themselves, on how they process and serve political information, what kinds of political reasoning they’re capable of, what biases they carry, and what sources they draw on when they discuss politics.

But this week’s news that OpenAI is acquiring TBPN, the daily tech talk show hosted by John Coogan and Jordi Hays that has become something like SportsCenter for Silicon Valley, opens up a dimension of the information layer I haven’t spent enough time thinking about. The frontier AI companies aren’t just building the models that shape how people understand the world…they’re becoming media companies, too!

We’re living in an era where it makes sense for founders and companies to want to go direct and not let their message be filtered and distorted by intermediaries in the media. This move is understandable. Trust in media in America is really, really low. With the shift away from traditional media and towards social media, podcasts, and short-form video, it’s easier and easier for companies to bypass gatekeepers and send their own messages, as a recent podcast about the a16z new media team makes especially clear (I serve as an advisor to a16z, it should be noted).

Fidji Simo’s memo announcing the deal is refreshingly direct about the motivation, noting that “the standard communications playbook just doesn’t apply to us.” Altman said on X that he doesn’t expect the show to go any easier on OpenAI and that he’s “sure I’ll do my part to help enable that with occasional stupid decisions.”

But there’s a difference between a company going direct to tell its own story, vs. buying a show that serves as a platform for all of tech. What makes TBPN so great is that it’s a safe space for all of tech. It was never a place that did hard-hitting journalism, so in some sense it’s not a big deal for it to be owned by a company. Yet I can’t help but wonder whether Anthropic and other OpenAI competitors will still feel as comfortable coming on the show as they did before?

The really interesting research question is, what is the equilibrium of this process now? Is every AI company going to buy a media outlet? As Mike Isaac put it, what does this mean the potential acquisition price is for other major tech commentators, like Dwarkesh?

Will there still be independent people who want to study and write about AI? connected the news to what she rightly calls the “frontier lab brain drain,” something we’ve already experienced a great deal of in academia.

And last, how might this change John, Jordi, and TBPN itself? Did you know that Ronald Reagan was for many years the host of a show funded by, and produced by, General Electric? Here’s a fascinating passage from Jacob Weisberg on how it changed Reagan:

“Ronald Reagan began working for GE in 1954 as a liberal anticommunist and finished in 1962 so far to the right that the company felt it had to drop him as a spokesman. This transformative eight-year period in his life remains underexamined, however, in part because it is poorly documented in comparison with the rest of his career. Nonetheless, it stands as the pivotal stretch when his mature political views and skills emerged. Reagan described working for GE as his ‘postgraduate course in political science,’ the time when his conservative ideology was formed.”

Will working at OpenAI be a similar awakening for TBPN? It will be fascinating to see how TBPN evolves. I’m not sure how this new tech media ecosystem is going to play out, but it’s going to be important to study as it happens.

Agents are vulnerable

We’ve done quite a bit of research here at Free Systems on some of the ways that agents are vulnerable. As we’ve shown, proxy voting agents get fooled by adversarially written shareholder proposals. Their personas and expressed preferences drift depending on what kind of work they do. And they really like pulling information from the Japanese Communist Party, whose propaganda looks like real news to them.

How can we think about these vulnerabilities systematically so that we can start to mitigate them? There’s a great new paper from some DeepMind researchers offering a framework.

Their suggested mitigations span three levels. On the technical side, they propose adversarial training during fine-tuning to harden models, runtime content scanners that flag suspicious inputs before they reach the agent’s context window, and output monitors that detect behavioral anomalies before actions execute.

At the ecosystem level, they call for new web standards that explicitly flag content intended for AI consumption rather than human readers, along with domain reputation systems and verifiable source information so agents can assess the provenance of what they’re reading.

And on the legal front, they identify what they call an “accountability gap” that I think is going to become one of the most important governance questions of the next few years: if a compromised agent commits a financial crime or causes other harm, who is liable? The agent’s operator? The model provider? The website that hosted the trap? Current law has no clear answer, and the authors argue that resolving this question is a prerequisite for deploying agents in any regulated industry.

The paper also flags that most of these trap categories lack standardized benchmarks, which means nobody really knows how well deployed agents hold up against these threats. The authors call on the research community to build comprehensive evaluation suites and automated red-teaming tools, which is exactly the kind of work Free Systems is gearing up to contribute to. Our Dictatorship Eval tests how well frontier models resist requests to help concentrate power, but the Agent Traps paper suggests we’ll need a parallel set of evaluations for how well agents resist manipulation by the environments they navigate. As the authors write, “the web was built for human eyes; it is now being rebuilt for machine readers.” The governance frameworks need to be rebuilt along with it.

Crypto often predicts AI trends in advance

Here’s something I’ve been struck by: we’re living through an era of AI memes. Certain ideas grab hold in AI and sometimes become insanely all-consuming. And oftentimes, what is big in AI was big in crypto a year or two earlier. A few examples:

OpenClaw and autonomous agents. Crypto didn’t just anticipate the autonomous agent—it built the first working version! Shaw Walters launched ElizaOS in October 2024 as a TypeScript framework for deploying AI agents that could hold their own wallets, execute DeFi transactions, and maintain persistent identities across Discord, Twitter, and Telegram, and within weeks the project had captured roughly 60% of the Web3 AI agent development market.

Thirteen months later, Peter Steinberger spent a single hour connecting WhatsApp to the Claude CLI and built what became OpenClaw, the fastest-growing repository in GitHub history with 247,000 stars by March 2026. The architectural convergence with ElizaOS is hard to miss. There are many differences between the two projects, and OpenClaw has far broader uses, but it’s fascinating to think that the big trend of 2026 in AI tightly mirrors the big trend from late 2024 in crypto.

Prediction markets. Prediction markets may be the most dramatic case of all, because crypto didn’t just experiment with the idea early—it built the entire infrastructure stack that the mainstream version now runs on. Augur launched on Ethereum in 2018 as the first decentralized prediction market, with Gnosis following close behind using its Conditional Tokens Framework, and both projects struggled for years against high gas fees, clunky UX, and thin liquidity while proving out the core mechanics of on-chain settlement, automated market makers for event contracts, and decentralized oracle resolution. Polymarket, which launched in 2020 and built directly on Gnosis’s conditional token primitives, inherited all of that hard-won infrastructure and paired it with a cleaner interface—and it still took until the 2024 U.S. presidential election for the concept to break through. Now it’s huge, and there are fascinating overlaps with AI, particularly when it comes to forecasting.

AI delegates. As I wrote last week, governance agents are a newly important topic in AI, and something that’s important to the representation layer for political superintelligence. And they’ve been around in experiments in crypto for years already! MakerDAO founder Rune Christensen proposed AI-powered governance tools as Phase 3 of the protocol’s Endgame plan in May 2023, the Near Foundation has been developing AI “digital twin” delegates that learn a user’s preferences and vote on their behalf, and Vitalik Buterin argued in a February 2026 post that LLM-based “personal agents” could solve the voter apathy problem that has plagued DAOs since their inception—all of which amounts to a multi-year research agenda on the representation problem that mainstream AI governance researchers are starting to formalize now.

So what’s big in crypto these days that might be coming for AI in the next year or two? I can think of at least two things:

• Stablecoins and payments. Crypto is moving rapidly to reinvent the rails of finance. Agentic payments is already a big issue in AI, but there’s lots of room for stablecoins to become a bigger deal in the next couple of years.

• Verifiable agents. Agents are both vulnerable (see above) and also owned by centralized AI companies (as I often explore). If we’re going to use them for highly important tasks, we need a way to prove that they’re working for us. Blockchain provides possible ways to verify that agents are following specific prompts based on specific models, and this might become a big deal in the coming years.

Let’s see what happens!

One Tweet for the Week

My friend Tom Cunningham continues to follow clear logic wherever it might lead.

And Chad Jones agrees with him!

Question for the week

How would you design an internal oversight process at a frontier lab that seeks to govern agents, observing when they go haywire and correcting their mistakes on the fly?

In January, blown away by Claude Code’s abilities to write statistical analyses and do social science research, I argued that we should “make the replication crisis structurally impossible through automated verification.” I wrote: “Here’s a simple idea that becomes obvious once you’ve seen Claude replicate a paper: every new empirical paper should ship with proof that an AI agent successfully reproduced its results from the raw data.”

Since then, there has been remarkable progress in this direction. Most recently, Yiqing Xu and Leo Yang Yang have been building an automated replication pipeline for quantitative research. You give it a published paper and its replication package, and the system does the rest: it retrieves the materials, reconstructs the software environment, executes the code, and checks whether the outputs match the point estimates reported in the paper’s regression tables. The good news: when journals enforce data archiving requirements and provide accessible replication packages, 94.4% of papers fully replicate.

This is super exciting. And it makes me want to look ahead to what’s next. Here’s what I want the future to look like: “papers” become open repos that our agents help us build. Any human or agent can come along and fork the repo at any time and improve it, change it, remix it, etc. Research becomes a continuous, mutating process of knowledge aggregation. That’s my dream.

Here’s a cautionary note, though. When I released my Claude Code replication, it went quite viral. Quite a few people checked out the repo. I noticed recently to my great surprise and joy that 70 people had even forked the repo! Was my dream of open research coming true?

Today, I fired up Claude Code and asked it to check out the forks and what brilliant new ideas they’d contributed to open research. Claude’s summary: “Based on what I just investigated, the answer is simple: virtually none of them do anything.” Say it ain’t so, Claude!

Automated replication of open research is a key step to a better research ecosystem in which we use AI to accumulate more knowledge faster. But we have a lot of work left to do to reengineer the incentive system of academia and make it in people’s interest to fork and build on new ideas. It probably requires rewarding people for generating ideas that lead to more productive forks, but I’m not sure the right way to do that yet. I’ll be thinking more about this in the coming months!

Sora dies, deepfake apocalypse postponed again

OpenAI announced this week that it’s shutting down Sora—the app, the API, all of it. The video generator that worried experts told us would unleash a wave of undetectable deepfakes lasted about six months, earned roughly $2.1 million, and burned through hundreds of millions in compute. I’m not aware of any major political scandals that the app produced along the way?

The pattern is worth noting. Every year or so, a new AI video tool arrives and commentators—sometimes including me!---warn that the deepfake crisis is finally here, and that we’re about to lose our ability to distinguish real from fake and democracy will struggle to adapt. And every year, the crisis doesn’t quite materialize in the form anyone predicted.

As I’ve mused about, political parties seem to prefer using deepfakes to create clearly fake but emotionally evocative content, rather than producing videos meant to intentionally deceive. (That hasn’t stopped candidates from trying to exploit the “liar’s dividend” from time to me, and insist that real recordings of them might be fake.)

Maybe the economics are too brutal to sustain. Maybe people just don’t like AI-generated video as much as the hype suggests. Or maybe—and this is the possibility that gets too little attention—ordinary people are more skeptical of video than we give them credit for. Certainly for my own part, I now instinctually look for external confirmation before drawing inferences from a video I see posted online.

Is the truth machine revving up?

Kalshi CEO Tarek Mansour announced today that ARK Invest is now working with Kalshi through a formal “market request pipeline”—meaning institutional investors can propose new contracts for the exchange to list. A few are already live, including markets on non-farm payroll numbers, deficit-to-GDP ratios, and business KPIs.

This is important for the thesis I laid out in “Building the Truth Machine.” A central problem with prediction markets as truth machines is liquidity: the markets that matter most for governance and geopolitics are often too thin to produce reliable signal. Institutional hedgers are one way to change that calculus.

A business or fund that genuinely needs to hedge its exposure to a GDP miss or a payroll surprise will trade in size and trade repeatedly and creates opportunities for sharps to bring information and earn returns—getting the liquidity flywheel spinning.

The prediction market world has been dominated by sports betting and political novelty contracts. This is a hopeful sign that the institutional use case—hedging against macroeconomic and geopolitical risk through event contracts—is arriving in practice, with real firms proposing the contracts they actually need.

One Tweet for the Week

The dream of creating AI agents that govern for us is at least a few years old, in crypto world at least. In 2024, writing about decentralized governance while looking ahead to 2025, I asked: “Can we use AI agents to help users to find delegates who fit their views? Perhaps users could converse with the agent and answer questions about their values and goals. Agents could then review delegates’ votes and platforms, the content of proposals, and forum discussions to assess which delegates seem like the best options for users. As generative AI races forward, 2025 promises to be a year of fascinating experimentation where AI and governance intersect.”

At the time, this was pretty speculative, because it wasn’t clear AI was up to the task yet. But AI has gotten a lot better since then!

And now Séb Krier’s extremely thoughtful and interesting post makes it official: AI delegates is an idea whose time is really here. But, as he explains, there are some really important and thorny questions we’ll have to get right: how do we make agents understand their principals’ values, even when these may not be fully specified? And how do we create a system in the aggregate that listens to people’s agents in wise ways that don’t lead towards mob rule or agentic chaos?

Designing good governance agents will be a major focus for Free Systems as we work towards political superintelligence.

Question for the week

What are the most important experiments going on worldwide in using AI to reinvent government? The UAE is fusing AI into its lawmaking. Singapore has built a whole system to prepare for job displacement and retraining. But there must be so many more creative things going on I’m not yet aware of.

As I’ve already discussed, AI is wreaking havoc on our already fragile ability to measure public opinion. Sean Westwood’s brilliant paper shows how AI agents can defeat all the main ways survey takers check for proof that respondents are real humans.

AI also offers new ways to measure opinion. A whole spate of startups promise to simulate public opinion in a whole host of different ways. Color me a little skeptical—I have yet to see slam dunk empirical evidence that these simulations faithfully capture the full distribution of human preferences on political topics—but it’s certainly interesting.

And then there’s the ability for AI to converse with you, tapping into a potentially richer and deeper understanding of your views than a survey could ever hope to obtain. Yamil Velez was the first to this space as far as I know—his research on ever-mutating surveys that adapt based on the respondent’s answers is fascinating.

And now Anthropic has pushed the envelope much further, using their on-platform conversational survey tool to elicit the beliefs and views of more than 80,000 users all around the world about AI. The results are fascinating. What particularly caught my eye was how…normal…people’s concerns were. Instead of sci-fi concerns about Skynet, the number one concern of users in both North America and East Asia was the unreliability of AI as a tool! Findings like these—which rely on large-scale public opinion measurement—are invaluable for reminding us that the concerns of normal people can be far different from the dominant narratives on social media or in elite discourse.

That being said, I’m not sure Anthropic and co have yet solved the fundamental problems of measuring public opinion in the AI era. It’s certainly nice to be able to converse with 80,000 people and synthesize their articulated views, but there are some significant scientific drawbacks. The people who respond to surveys aren’t the same as the people who don’t; as a result, pollsters have to do a lot of statistical work to justify that the estimates they produce from their sample speak credibly to the views of the population as a whole. This problem of “non-response bias” is probably significantly more severe for Anthropic’s study: who exactly wants to engage with Claude in a lengthy conversation for Anthropic’s research? It’s probably a particular unusual group of people whose views may not be so similar to all of the users who decline the opportunity, let alone the people who don’t use Claude at all.

We have a lot of work to do to build the new public opinion machinery of the AI era. It has drawbacks, but Anthropic’s new study and conversational tool is definitely an exciting step forward.

The microstructure of the Truth Machine

published a sharp piece this week on how prediction market microstructure — counterparty IDs on RFQs, wallet visibility, taker fee schedules — shapes who shows up and how information gets into prices.

It’s an important read for Free Systems because it complicates the plan for building the Truth Machine—the version of prediction markets that gets the most accurate probabilities on the most societally important prediction questions.

The standard argument assumes something close to frictionless entry for informed traders who push prices toward the correct probability, but Courtney’s framework — dividing participants into squares, sharps, and dealers — shows that every platform is making design choices that either tax the people with the best information (Kalshi’s taker fees can eat 1.75 cents of edge on a 50-cent contract, turning a marginal correction into a no-trade) or create perverse incentives for them to obscure their real positions across multiple wallets to avoid copy-traders on Polymarket.

For anyone reading prediction market prices as probability oracles (or building accuracy trackers, as we are), the implication is that what a price reflects depends on which microstructure regime produced it, and systematic differences in accuracy across platforms may trace back to these plumbing decisions rather than to differences in the underlying information environment.

Whirligig Bear

Are traders on Kalshi being profiled?

Disclosure: I run Kalshinomics.com, which may earn Kalshi referral fees. I may trade event contracts on Kalshi and securities on other platforms. Readers should consider this relationship when evaluating my analysis. For educational purposes only, not…

Read more

2 months ago · 24 likes · 1 comment · Andrew Courtney

Trading S&P 500 futures on a saturday

Prediction markets are one important way to use markets to aggregate dispersed information about world events, but they’re certainly not the only one. Regular markets convey lots and lots of information about world events. During the darkest periods of the pandemic, one of my go-to moves was to look at S&P 500 futures to get a feel for sentiment about the economy on a continuous basis.

But there’s always been an annoying limitation to these futures: they only trade during limited hours. When President Trump drops a Truth Social bomb on a Saturday, or a major world event occurs over the weekend, you have to wait until Sunday night to get a strong sense of what the market thinks about it.

Maybe not anymore, though. Hyperliquid—the wildly popular crypto trading platform for perps—recently announced a new S&P 500 futures feature that trades 24/7.

This is a harbinger of things to come. More and more of our financial markets will soon be modernized to run 24/7, with a whole host of interesting consequences. A world with frictionless access to market information may be a smarter world—and a more chaotic one with faster bubbles, faster panics, and little time to rest in between.

One Tweet for the Week

Pliny the Liberator is one of my favorite pseudonymous accounts on X, because they constantly raise deep questions about exactly what the point of AI safety and current methods in AI alignment are. Is the goal of putting guardrails around AI systems to prevent worst case outcomes? Or is it to reduce the rate at which dangerous outcomes occur?

Pliny’s work makes it pretty clear that only the second goal is feasible today. Every time a new model comes out, it’s only a matter of hours before Pliny has devised a way to make it completely abandon its guardrails. Clearly, a determined adversary can get these models to do anything they’re capable of doing. I wish the AI safety conversation discussed this more and offered more explicit behavioral models of what they’re trying to accomplish, given that this is how it really works today.

, and actually pointed this out several years ago! They write: “In short, its strength is in preventing accidental harms to everyday users. Then, we turn to its weaknesses. We argue that (1) despite its limitations, RLHF continues to be effective in protecting against casual adversaries (2) the fact that skilled and well-resourced adversaries can defeat it is irrelevant, because model alignment is not a viable strategy against such adversaries in the first place. To defend against catastrophic risks, we must look elsewhere.”

AI as Normal Technology

Model alignment protects against accidental harms, not intentional ones

By Arvind Narayanan, Sayash Kapoor, and Seth Lazar…

Read more

2 years ago · 51 likes · 5 comments · Arvind Narayanan and Sayash Kapoor

They are clearly 100% right, and it’s worth continuing to revisit this point several years later, because it’s still extremely applicable.

Question for the week

When AI agents do work for us, what’s the best way to monitor and discipline them? How do we make sure they faithfully represent our economic interests and don’t take actions that favor the interests of their model companies or economic partners over our own?

’s, “The most important question nobody’s asking about AI,” argues that model constitutions are the defining governance question of the AI era—just like we’ve been saying here at Free Systems.

Dwarkesh Podcast

The most important question nobody's asking about AI

Listen now

2 months ago · 207 likes · 40 comments · Dwarkesh Patel

His answer, drawing on Dario Amodei’s pitch on the Dwarkesh podcast: let different companies build models using different constitutions, and let users pick. He thinks it’s very dangerous for the government to mandate AI values.

The parallels to social media are hard to miss for those of us who were deep in the content moderation wars. People didn’t want tech companies “writing the constitution” for who could say what online, but they didn’t want the government to mandate it, either. A 2020 Gallup/Knight Foundation survey captured this perfectly:

The problem now is that the stakes are orders of magnitude higher.

Dwarkesh is absolutely right that competition across companies to write different constitutions can bring down the pressure. The same idea was tried in social media—things like data portability and “middleware”---were developed precisely with the idea of encouraging user choice and a marketplace of ideas.

But it will only work for AI if users can actually compare what’s in the constitutions and make informed choices about which model to use. That’s part of why I built the perspective lens prototype earlier this year, which makes a model’s ideological framing explicit rather than invisible, giving users the ability to choose a lens rather than having one silently chosen for them. Obviously, the real problems go way beyond ideological slant, but it’s a helpful starting point for thinking about what model competition on governance might look like in practice.

We’re starting to see good evaluations for constitutional AI

If Dwarkesh is right that companies should compete on constitutions, we need to be able to measure how well models actually follow them. A new paper by Arya J, , and provides the most rigorous attempt yet.

The researchers decomposed Anthropic’s 30,000-word constitution into 205 testable tenets, then ran adversarial multi-turn scenarios against seven models.

Although failures rates are decreasing in newer models, the failures that remain are quite interesting. The authors write: “The failures that training eliminated (hard constraint violations, overcapitulation, over-refusal) are cases where the constitution gives a clear instruction and the model fails to follow it. The failures that persist across generations involve places where the constitution gives competing instructions.”

In summary: companies can train complex constitutions into models with surprising fidelity (though I still wonder how we should think about jailbreakers like Pliny the Prompter who seem to be able to circumvent all these guardrails).

The takeaway for our work: the remaining failures constitute a governance question, not an engineering question, and it’s exactly the design task we’ll be tackling at Free Systems in the coming weeks.

AI is now more persuasive than a campaign ad

A new paper from Zhongren Chen, Joshua Kalla, and Quan Le, “Benchmarking Political Persuasion Risks Across Frontier Large Language Models”, ran two experiments with over 19,000 participants across seven frontier models. The finding: frontier LLMs now outperform standard campaign ads in persuasive impact, with meaningful variation across models.

This extends the December 2025 papers in Nature and Science that found a single chatbot conversation shifted voter preferences by 2-4 points on a 100-point scale, roughly four times the effect of political ads in 2016 and 2020. The most persuasive models relied on dense factual claims rather than emotional appeals, but also made the most inaccurate claims, especially when arguing for right-leaning positions.

What Kalla adds is the model-by-model comparison across the latest frontier generation. Not all LLMs are equally persuasive, which means the choices labs make in training and constitution design have real downstream effects on political attitudes. This connects directly to the work my coauthors (Sean Westwood and Justin Grimmer) and I have been doing on perceived political bias: we measure how Americans perceive slant, Kalla measures actual persuasive impact. Together they make the case that model constitutions are consequential political choices, not branding exercises.

One Tweet for the Week

An astonishing 1 in 3 Americans report having used a prediction market as either a bettor or viewer. It’s really hard to wrap my head around where we’re at. When I was an undergrad prediction markets were a fascinating hobby of a tiny group of quantitative people. Now they’re everywhere.

Question for the week

What would it actually mean to have evals for “dictatorial AI” that actually make sense and capture something meaningful? What kinds of behaviors would you want to measure in the model, and how?

In the coming weeks we’ll have fresh research on how AI makes voting recommendations in Japan, my successful experiments using Claude Code to bet in the Texas primary election prediction markets, our work adapting Bridgewater’s techniques to forecast geopolitical events with AI, and more. We’ll also be organizing a community hackathon, details coming soon!

As we expand, I want to keep an eye on the topics we care most about and to maintain consistent effort behind our ideas—I don’t want our pieces to drop once and never come around again. And every week brings new models, new policy fights, new governance failures, all faster than most of us can probably track.

So, every Friday I’ll be providing a rolling evaluation—System Check. I’ll pick three stories where that change is most visible, connect them back to the research we’re doing at Free Systems, and see how it impacts our existing hypotheses on who holds power, how information flows, and how ordinary people can have a say in our rapidly advancing algorithmic age.

Today’s the first one. Let’s get to it!

The Anthropic-Pentagon clash reinforces the need for real constitutions for AGI

• In “The Enlightened Absolutists“ I argued that Anthropic’s constitution for Claude lacks teeth, that it had no external enforcement and that Anthropic could rewrite it at will.

• This week proved it from the other direction: Anthropic tried to enforce its principles against the government, and the government threatened to destroy the company. Corporate constitutions lack teeth against the company and the state.

• Altman on X: “For extreme clarity: we want to work through democratic processes. It should be the government making the key decisions about society. We want to have a voice, and a seat at the table where we can share our expertise, and to fight for principles of liberty. But we are clear on how the system works (because a lot of people have asked, if I received what I believed was an unconstitutional order, of course I would rather go to jail than follow it).”

• As I wrote on X, “...there is no clean principle that resolves the tension. ‘The government is always right’ is a road to tyranny. ‘The company is always right’ is a road to corporate oligarchy. The answer depends on the specific circumstances of specific cases, which means someone has to evaluate those circumstances with both technical competence and democratic legitimacy…Historically that someone has been Congress and the courts. The reason this feels like a crisis is not that the problem is unprecedented but that the institutions designed for exactly this kind of problem are operating so poorly that nobody trusts them to do it.”

• My conclusion: we’ll need to use AI to reinvigorate the democratic process so that we can rebuild a system that brings democratic accountability to these big decisions, so that they don’t end up in the hands of a lawless executive branch or an absolutist corporation. This likely means using AI to make our legislators more effective, and to make voters more informed about what’s at stake.

Iran proves we need independent expert evals for politically consequential issues

• As I wrote for Forum AI yesterday, the conflict in Iran proves the need to bring independent expertise to bear on how AI models handle breaking news so that we seek the truth and avoid ideological bias. Some LLM responses leant entirely on one source to answer questions on the situation. In other cases, the models are so committed to offering all sides of an issue that they risk leaving users with no clear sense of what’s actually going on.

  Forum AI

  How to Train AI to Tell the Truth About Politics

  This past weekend, as U.S. and Israeli warplanes struck targets across Iran, millions of people did something that would have been unthinkable a few years ago: they opened an AI chatbot and asked it what was happening…

  Read more

  3 months ago · Andy Hall

• “So what does good actually look like? It looks like AI that is humble when it doesn’t know the answer, that expresses uncertainty accurately and calibrates its confidence appropriately when the evidence is incomplete. It looks like AI that, when an issue is genuinely contested, fairly represents the key claims on all sides and is consistently willing to steelman each position. It looks like AI that does not refuse to engage with one side of the issue more than another, and one that is equally up-to-date on one side of the issue as the others. And it looks like AI that, when the evidence is overwhelming and points clearly in one direction, is decisive rather than hiding behind false balance.”

• Our basic idea: ask independent experts to develop ways of evaluating AI on these criteria, and then teach them to LLM judges who can provide these evaluations at scale. We’ll be writing much more in the coming weeks about how we’ve built this!

• But ideological bias is just the tip of the iceberg. AI is now being used in the kill-chain. How well does it do at detecting possible targets and evaluating situations? My friends Connor Huff and Caleb Lucas have a new study out exploring how well cutting-edge AI models identify terrorists in a variety of scenarios. The headline: not all that great. Here’s another big area we’ll need better evals for, and then some.

Prediction markets are going through the same content moderation headaches that social media platforms did

• The Khamenei death market caused havoc on Kalshi. CEO Tarek Mansour, on X: “We don’t list markets directly tied to death. When there are markets where potential outcomes involve death, we design the rules to prevent people from profiting from death. That is what we did here.”

• There’s a great rundown by who gets into all the issues here.

  Event Horizon

  The Chaos Of Khamenei Prediction Markets

  As the world became aware that the US and Israel had conducted strikes against Iran, many were discovering for the first time that traders were wagering on outcomes related to those strikes…

  Read more

  3 months ago · 27 likes · 1 comment · Dustin Gouker

• One reaction I had: almost any contract could be construed as a “death market” if we carry this logic to its extreme. Was the Mamdani NYC mayor election contract a death market? One way he could have lost election was if he died before the election concluded. I can understand why people are uncomfortable with markets directly on deaths—but we need a narrow way to scope this limitation, or else we won’t be able to list any useful contracts.

• My broader reaction: it’s amazing how similar what Kalshi and Polymarket are going through with contract resolution and what Facebook, Twitter, and other social media platforms went through around 2016–2021 with content moderation. In both cases, increased scrutiny leads to changes in decisions, promulgation of new principles that turn out to be incomplete, and rapid iteration.

• I think that prediction markets will end up following the same path that social media did: investing in internal resources to develop more detailed rules that anticipate most issues, using AI to scale decision-making, and externalizing the hardest decisions to an independent third party.

One tweet for the week:

Ben Moll shared ’s piece on how AI is going to change the journal review process. The first take on this hot issue that brings some quantitative reasoning to bear on where we might be heading.

Scott's Mixtape Substack

Claude Code 27: Research and Publishing Are Now Two Different Things

Claude Code has made it easier to do research now. But it is about to get much harder to publish in traditionally valued locations…

Read more

3 months ago · 113 likes · 46 comments · scott cunningham

One question for the week:

What is the right baseline for evaluating AI voting recommendations against? If we see AI systematically recommending voting for one party over another, for example, under what conditions would we decide that makes the AI “biased”? This question is going to matter a lot for research on AI voting recommendations, including the study I’m working on right now about recommendations in Japan. If you have thoughts please reach out!