In The Argument: theargumentmag.com/p/what-i-…

So I have a lot of sympathies for Anthropic and the folks who work there, but one thing this says unequivocally is that Anthropic is committed to racing (“to the top,” according to Dario) — and will do whatever it takes to stay at the frontier. But they don’t seem to get that mostly all this inspires is for the other companies to race harder against Anthropic and take more shortcuts and risks to not be left behind. Dario seems to think that the mimetic influence of Anthropic will be enough to inspire others to prioritize safety. But does the evidence support this? I’m not convinced; e.g., xAI urgently put a massive methane-producing data center in Boxtown without permits to train more powerful models, but doesn’t even have a safety team; Meta just reduced certain standards in its safety framework for releasing a model and now hasn’t released a safety report concurrent with release of its new “frontier” model (one week late and counting); OpenAI is plausibly accused of recurrent depth looping for Astra and other models, reducing monitorability; Google and Sergey Brin are gleefully crediting RSI for helping them build more powerful models, etc etc. Mimetic safety behavior is looking insufficient. We need real rules, laws, regulations — ones with teeth. Which I think is why people are upset after this draft bill. Anthropic has been telling us for years that it gets how important laws are — but now, isn’t showing that in reality.

New essay in @ForeignAffairs with James Manyika and @elerianm on steps that can be taken to help AI deliver on its economic promise, drawing on lessons from cooperative game theory as a guide. Most conversations treat AI as a general-purpose technology (GPT), and some compare it to the Industrial Revolution. But we think it is more than that: it is also a method for creating new inventions, like the microscope or the scientific method. Economist Zvi Griliches famously called such discoveries IMIs--inventions of a method of invention. Seen this way, AI carries the potential of the Industrial Revolution and the Enlightenment combined. Yet realizing this potential is a coordination problem. Delivering on the promise of AI requires progress not just on furthering its capabilities, but on building out infrastructure, deep adoption across the economy, and buy-in from many different groups. This is unlikely to happen if people do not feel they have a stake in the upside of the technology. Promises of abstract scientific and medical breakthroughs in the future won't change this when anxieties about near-term disruption are running hot. Rather than treating AI like a zero-sum game, we argue that it is useful to look to cooperative game theory as a guide. Because models, compute, organizational change, and human upskilling are strong complements, AI has the structure of a "convex game": Each group that joins makes everyone else's contributions more valuable. That makes it possible to form stable coalitions of workers, communities, firms, and developers where incentives are aligned and everyone has a stake in the gains. The cooperative approach will of course not be easy, but we will make the case that coalitions can start small and grow as the mutual gains become tangible. We provide concrete examples of how this can be done in practice and at scale. Link to essay here: foreignaffairs.com/guest-pas…

There are two kinds of outside-group ad: 1. Help an organization's allies by advertising on the issues that that candidate has identified work well in their favor, or against their opponent (typically identified through a redbox). This is usually NOT an issue the org itself cares about; the point is that the candidate appreciates the ads because it's a message that their internal polling tells them works. This is the strategy employed by AIPAC, crypto pacs, Leading the Future, and lots of others. 2. An ad on the organization's pet issue (like this ⬇️), which hasn't been identified as a candidate's favored message. The goal of these ads is really for the org to piggyback on the candidate's expected victory by spinning after the election that the race hinged on their pet issue. But candidates don't really like this, and it benefits the org's spin more than it builds appreciation for the candidate. This strategy is being employed by AI safety PACs, who are generally more interested in elevating the perceived salience of their issue than in building political allies.

🌹🌹DSA civil war over AI, @jdcapelouto reports semafor.com/article/10/21/20…

POTUS lays down the law on data centers

Whoa! Big money verdict in the Cambridge Analytica case against Meta in New Mexico ($25B!). If this case were to be appealed—and Meta says it plans to—then it could actually prove to be important for AI regulation. One of the ways in which AI regulation can be enforced is through civil penalties. But are there penalties that a giant AI company would really care about? And would that level of penalty be excessive? As a refresher, this is the case about Facebook allowing a personality quiz to harvest data from perhaps 87M users, which was then sold to Cambridge Analytica to profile them, despite Meta’s statements about protecting user data & policing 3rd party apps. The jury found approximately 43.9 million willful violations of New Mexico’s Unfair Practices Act, involving statements about user data, content moderation, and investigating 3rd party apps. At the maximum $5K a pop, that would be around $219B total. The Attorney General asked for "only" around $35-40B, and Meta argued for no more than about $3.4B. The judge has now ordered Meta to pay $25B. From the reporting, it seems like one of Meta’s big issues on appeal will be that the state did not prove that individual residents actually saw or heard Facebook’s statements or were harmed by them. The number of violations was based on how many people may have seen or heard the statements—including people who weren’t Facebook users. So what might be the most interesting issue on appeal is whether this penalty is so excessive as to violate due process. There may also be an 8th Amendment issue of excessive fines: that protection can apply to civil penalties imposed by the government, even though it ordinarily doesn’t apply to civil lawsuits between private parties. For AI regulation, a suggestion is to use really large civil penalties that could actually discourage the giant frontier AI companies—they may treat smaller penalties as simply a cost of doing business. We may see the limits of that in this Meta appeal, if it actually gets to that point.

No modern conflict has ended by destroying the enemy’s ability to resist—not even World War II. To end a war, both sides must choose to stop fighting, and those choices always require nontrivial bargaining. Misunderstanding the relationship between bargaining and fighting produces systematic failures to integrate negotiating strategy with campaign planning, from Afghanistan to Ukraine to the US war with Iran. New piece in @ForeignAffairs with Stephen Biddle and Yuri Zhukov on the myth of military victory.

Documents obtained by The Intercept show the military is trying to deploy autonomous AI agents to hack its adversaries. theintercept.com/2026/10/19/…

Interesting: Mechanize Inc has removed from their webpage the language about how "our long-term goal is the full automation of valuable work across the economy". However, the website's home page still links to the founders' essays arguing that automating all work is inevitable and is a good thing. Left: Internet Archive snapshot, Sep 17 Right: live website, today

As one of the candidates featured in these polls - the infamous one in late June that showed me plummeting into single digits (which was at odds with our internals) - here's the impact of this "experiment" - Donations dried up - Volunteers were devastated and pleading for guidance on what to do - Reporters hounded me with calls to drop out Sure, a point was proven. Polling is broken. But tens of thousands of people poured their hearts, time, energy, hard-earned money into something they believed in - for a year-and-a-half, and it was treated like a game. Our volunteers gave their all. And then some. I felt like I had let so many of them down. To be extremely clear, Michigan Democrats are locked in to beat Mike Rogers. We WILL send @AbdulElSayed to the US Senate. But between crap like this, endless fundraising texts that drive supporters away with spammy fake deadlines and trash "match" claims, scam PACs run rampant for firms to grift without actually supporting candidates, undisclosed content creators pushing content out either misrepresenting themselves as residents of states they don't live in or hiding whether they were paid and by who so they could make a buck, and the tidal wave of dark money ads -- I hope Democrats are ready to blow up the system with campaign finance and transparency reforms. Trust is already at an all-time low. We don't get out of this mess without rebuilding trust. And media - who loves to ride the drama of a pollercoaster - needs to take a hard look at what's legitimate and what's not. If this teaches us anything - DON'T GET COMPLACENT AND DON'T TRUST THE POLLS. LOCK IN. SIGN UP FOR A CANVASSING SHIFT. LEAVE NOTHING TO CHANCE OVER THE NEXT TWO WEEKS.

What's going to happen with AI legislation in the next Congress? Billionaire John Doerr, whose VC firm is betting hugely on AI, is funneling millions to Schumer and Jeffries groups as his TechNet lobbies against tougher antitrust rules. readsludge.com/2026/10/20/te…

.@anilkseth, @De_dicto, and Giulio Tononi have a new essay in @Nature arguing that AI is not conscious and perhaps never will be. I admire the authors and agree with much of what they say. False positives about AI consciousness are both likely and harmful right now. Consciousness and intelligence are distinct capacities. And theories that support AI consciousness, such as computational functionalism, are not clearly true, and theories that rule out AI consciousness, such as biological naturalism, are not clearly false. However, I think this essay misses the mark in an important way. The core issue is that the essay never makes clear who its real target is. The introduction suggests that it will address precautionary arguments for taking AI welfare seriously; for example, it cites our report "Taking AI Welfare Seriously" and summarizes its main argument. But much of the essay then addresses mistakes that appear mainly in the media and the public. This allows the essay to present itself as engaging with fellow experts even though the relevant experts are not making the mistakes that the essay corrects. To stick with the view that many of us actually hold: Given uncertainty about the nature of consciousness and the future of AI, we are not able to rule out a realistic possibility of near-future AI consciousness. We should thus acknowledge that AI consciousness is a serious issue, start assessing AI systems for welfare-relevant features, and prepare policies for treating AI systems with an appropriate level of concern, mitigating the risk of both false positives and false negatives in proportion to their probability and severity. The essay does briefly engage with a precautionary stance, arguing that false positives are likelier than false negatives, and could be harmful. This is reasonable but incomplete. We need to consider the probability and severity of both types of error, as well as the costs and benefits of precautions. And even if one error is more likely than the other and could be harmful, we should still take steps to mitigate both risks if there is a non-negligible possibility of serious harm in both directions, and if reasonable precautions are available. In this case, the stakes are clearly high in both directions. If we overattribute consciousness to AI, we could develop inappropriate emotional bonds with them, we could misallocate scarce resources toward them, and we could even amplify risks of misuse and loss of control. Whereas if we underattribute consciousness to AI, we could become entrenched in industries that harm and neglect vulnerable populations against their will, often for trivial reasons, as we continue to do with nonhuman animals. There is also a non-negligible possibility of error in both directions. There are good arguments for and against computational functionalism, biological naturalism, and other theories. Everyone is entitled to develop their favorite theory and explore its implications. But when we make high-stakes decisions under uncertainty, we should give weight to multiple leading theories and to what counts as evidence according to each. And if we do, then we will not be 99.9% confident either way about AI consciousness at this early stage. There are also reasonable precautions available. For example, I support the new Anthropic policy of banning sustained and needless abuse and cruelty towards Claude, since kindness toward entities that function as assistants or companions in our lives is a low-risk, low-cost, co-beneficial way to improve our treatment of those entities under uncertainty about their moral status. We can implement this kind of policy to mitigate the risk of false negatives while still taking appropriate steps to mitigate the risk of false positives. The essay also focuses primarily on current systems, yet the precautionary view focuses at least as much on near-future systems. AI progress is occurring rapidly, and in the near future, we may have not only more advanced, integrated, and embodied versions of current systems, but also new types of systems on new types of architectures. Policy also takes time to develop. So, if the probability of AI consciousness increases over the next decade, we will want to have done philosophical, scientific, and policy work in advance. The rest of the essay is largely orthogonal to our work. The case for taking AI welfare seriously does not conflate consciousness with intelligence, nor does it depend on the truth of theories like computational functionalism or the falsity of theories like biological naturalism. Instead, it depends only on the idea that plausible theories exist in both categories and that, when making high-stakes decisions under uncertainty, we should give weight to multiple leading theories rather than bet everything on our own favorite theory. To be clear, this essay makes many good points, and I agree with the authors more than I disagree with them. But I think that the essay would be better if it selected a clear target and then focused on that—either stating at the outset that it aims to debunk misconceptions about AI consciousness in the media and the public and then doing that, or stating at the outset that it aims to address the precautionary case for taking AI welfare seriously, and then genuinely engaging with the arguments and conclusions of that view.