A $100 Million Super PAC Named Him Public Enemy Number One. Donations Went Up.

Nathan Labenz introduces New York State Assemblymember Alex Bores as a computer engineer by training, a Palantir alumnus, and the author of the RAISE Act — and, a beat too early, as a congressman. Bores corrects him: not yet. He is running in a crowded Manhattan field that already has about ten declared candidates. The bill he wrote, which once looked milder than California’s SB 1047, has become a lighting rod. The governor is negotiating edits, a presidential executive order is trying to preempt laws like it, and a super PAC has decided he is the person to make an example of.

Leading the Future formed over the summer with $100 million, funded by Andreessen Horowitz, Greg Brockman, Joe Lonsdale, and Ron Conway. Its public promise was to go after anyone imposing regulation on AI. Bores is the first Democrat they named, public enemy number one. A Republican running for Congress in Texas got a supportive ad that barely mentions technology. The ad against Bores is on the nose: don’t vote for Alex, because he will regulate AI. Labenz wonders whether the industry is building itself a Streisand effect. Bores thinks the operatives are smart, which is what makes the timing interesting.

If you only wanted to bury a candidate, he says, you would wait until three weeks before the primary, buy whatever the polls said was salient — not the issue you actually care about — and dump it after disclosure deadlines, too late to answer. They did the opposite: named him early, named the issue as AI, and started running ads. The result, so far, has been free press, a surge in donations, and a surge in volunteers. The PAC originally said it would spend $2 million against him; after that surge it raised the figure to at least $10 million. Bores jokes that if he keeps this pace he might soak up the entire $100 million and spare every other candidate. Three weeks before the primary, he expects the dump to hurt. Until then, “I want to regulate AI” is a popular label he did not have to buy.

RAISE has passed the legislature and is now in New York’s chapter-amendment process, which Bores says handles roughly a third of signed bills. The governor negotiates changes with the sponsors. If they agree, she signs the original bill with a memo saying she is signing pursuant to that agreement, and the sponsors introduce the amendments at the start of the next session. The bill landed on her desk last week. The state constitution gives her ten days, excluding Sundays, which means Saturday — though the working assumption is everyone wants to shut down Friday night for the holidays. In the next 48 hours or so, Bores says, they will know whether there is a revised RAISE Act or they have to try again next year.

He will not narrate the bargaining table. The reporting, he notes, is that there is a lot of pressure to move RAISE toward something that matches California’s SB 53. Curiously, there has not been a push to match the European code of practice, which makes the ask look less like harmonization and more like weakening. SB 53 was the compromise at the end of California’s session: mostly transparency standards, not real safety standards. Eight states had some version of a frontier-model safety bill this year. Only two passed anything.

His rebuttal is structural. If New York passed SB 53 verbatim, no company would be newly subject to it, because every such company is already subject to it in California. This is not a data-privacy bill, where each state that copies the text gives new rights to its own residents. It is, he says, like being the 39th state to ratify a constitutional amendment. Thirty-eight is enough; the 39th changes nothing practical. SB 53 was always meant to be the floor. The live question is whether New York can take a step above that floor that everyone can still sign, or whether they are destined to come back next year — or both.

Bores draws a hard perimeter. Effects on children, energy, democracy, the workforce, algorithmic discrimination — those are real problems, and RAISE does not address them. It is aimed at extreme, catastrophic risks that in his view belong at the federal level and have not been taken up there. The federal energy has mostly gone the other way: stop Washington from acting, and stop the states too. Because the federal government has not moved, states have. RAISE’s design is that labs publish a safety plan, stick to it, disclose critical safety incidents that show a massive jump in the risk of something bad, and not release a model that fails their own tests.

The thresholds sound almost silly in their extremity, he says. A hundred people dying, or a billion dollars in damage, specifically by helping to build a chemical, biological, radiological, or nuclear weapon, or by committing a fully automated crime. That last phrase is not a user using the model to commit a crime. It is telling the AI to do something entirely legal and having it go off on its own and commit a crime. Even past that bar, you still have to show the help was materially better than just using the internet, and that the risk was in some way foreseeable. Downstream liability and the criminal code do not change. Using a model to commit a crime is already a crime.

Co-host Prakash presses on crypto trading, wash trading, and the idea that a lot of illegal activity is already the state of play. Bores declines the claim that all crypto is money laundering, supports stronger clarity, and then returns to the distinction he actually wrote into the bill. Law enforcement is what you do after a crime. National security is getting ahead of something so bad you would not be satisfied to deal with it later in court — a bioweapon, loss of control of a model. The tobacco analogy is the one he wants remembered: the companies knew first that cigarettes caused cancer, denied it publicly, and kept shipping. RAISE says the labs are the frontier experts, and if they discover an extreme risk they have a duty to act on their own plan rather than ship anyway.

Labenz’s naive question is the one a lot of engineers would ask. Bores has a CS master’s, four and a half years at Palantir, and a bill that mostly echoes the labs’ own preparedness frameworks, including OpenAI’s. Why is the industry treating him as the enemy? Bores gives two answers, and they are both about deterrence. Regulating AI polls well. People are writing platforms now. The PAC wants a stake in the ground: touch this issue and we come. The second reason is more specific. The ad talked a little about him and a lot about the RAISE Act. It filled the screen with Albany and the line that Albany is dysfunctional. The bill was sitting on the governor’s desk. The message to her, he thinks, was obvious: sign a strong version and you are next.

On data centers he will not join a total moratorium, even as Bernie Sanders calls for blocking construction. What Sanders is reflecting, Bores says, is that people feel this is moving too fast, that five people are deciding it, and that Americans deserve a say. He uses AI constantly and is bullish on medical discovery — his mother has multiple sclerosis — which is exactly why the dual-use problem bothers him. The same pathway that finds a drug can build a bioweapon. The American grid is old and needs upgrades. Unlimited private capital will pay for a specific thing if the incentives are set. Right now, new renewables take four years to interconnect, so developers revive old coal plants that still have permits, or they ask nuclear plants to sell less to the grid and more to a data center. Power gets dirtier. His federal bargain: streamline interconnect if the project uses a set share of renewables and pays for the upgrade itself, so ratepayers are not subsidizing it.

Labenz then asks the Palantir question: anti-China rhetoric paired with domestic surveillance, social-media checks at the border, Zvi’s point that AI makes once-unscalable legal powers cheap. Bores is brief because Dean Ball is waiting. At Palantir he worked almost entirely on the federal civilian side — staffing VA hospitals, helping DOJ find Medicare “pill mill” doctors, tracking epidemics with the CDC. He is proud of that work. He left in 2019 when the company renewed an ICE contract so the software could keep being used for deportations. That was a personal line. Forcing travelers to hand over social media is, in his view, not what the United States should do, on privacy grounds and because New York wants visitors. Broad rights should be set in Congress. Congress has not gotten anything done for too long.

Dean Ball was the AI policy adviser at the White House Office of Science and Technology Policy. He left about four months ago, tweeting that he was a writer at heart. Those four months, he says, have been among the most productive of his life; the hard part is what to refuse. Two moments this year reset how useful the technology feels to him. One was o3, which landed the same week he entered government. The other is the last couple of weeks: frontier coding agents that can basically do anything you ask. He has liked being able to speak openly again.

The politics, he says, are in transition. The dice are still in the air. Public sentiment on AI is negative but low-salience — often a sixth or seventh priority, which for electoral purposes is almost not a priority, except in particular districts. He draws three camps: the traditional AI-safety world; a pro-AI, industry-heavy camp that is not only industry, and that includes him; and an emerging truly anti-AI coalition, which is a different thing from AI safety. He does not file Bores as anti-AI. He files him as someone concerned about aspects of AI safety, with pretty good reason. Bernie Sanders talking about banning data centers is the left version of the third camp; there are people on the right saying similar things.

The live variable sits in the middle: people worried about kid safety and consumer-protection harms, plus more traditional safety people. Do they decide they are fundamentally anti-AI, or that they want the technology to go well and need guardrails? Prakash adds the “don’t let social media happen to AI” line — elected leaders watching algorithms take mass media away from them, proto-AI already causing a loss of control. Ball grants some of that. GPT-4o sycophancy is the same local-optimum problem. But many of those people are viewing AI almost entirely as a consumer technology, a myopic frame that misses the point. Social media is an authority destroyer, good in some ways and quite bad in others. Josh Hawley’s deeper concern, Ball thinks, is the balance between internet corporate power and the government. Marsha Blackburn’s agenda has Nashville in it. And when people say they are protecting creators, “come on”: it is not the musician on the street with a guitar. It is Disney, News Corp, large IP portfolios. “We’re protecting Taylor Swift, who’s like a billionaire, and that’s fine. But come on.”

Labenz asks a historian’s question. He usually thinks in structures — technology, economics — with leaders squabbling on top. Right now so much is up in the air that selling H200s to China, or an unpopular-with-the-base executive order on preemption, might be actual hinge points. Ball thinks Trump will go down as a great man of history in the narrow sense that he changed the country’s orientation and the lexicon of its statecraft; living through it, history will judge. On preemption, a lot of people on the right who say they oppose it believe the president has been captured by tech. Ball’s read is the opposite: Trump personally cares about preemption, talks about it consistently, and will lead the party where he thinks the country’s interest is. Ball thinks he is right that preemption is in the national interest.

On H200s he wants to wait and see — whether China even takes the chips, how many actually move. The balancing act, as he tells it, is to sell things better than the best China can produce, so ecosystem advantages stay real. Chinese engineers write CUDA kernels and other infrastructure optimizations; that work helps NVIDIA’s performance and, through the stack, American developers building on NVIDIA. It is one important thing among many, to be weighed against national security, while China does its own balancing. Two signatures on a superintelligence-ban statement jumped out at Labenz when he ran Ball’s debate with Max Tegmark: Steve Bannon, and the CEO of Z.AI, the GLM lab, maybe a top-five or top-six Chinese AI company. A Nixon-goes-to-China mismatch, maybe.

This is where great-man history meets structure. Ball has written that the United States is in a structural decoupling from China that does not depend on who the president is, unless Washington takes far more control of the economy. Private capital allocators, as fiduciaries, cannot keep relying on a rare-earth supply chain China has shown it will weaponize. China exports, lowers prices, and competes with domestic industry everywhere, extracting raw resources and selling goods back — a vision Ball thinks is good for the Chinese government and not obviously good for Chinese people, let alone everyone else. The president, in Ball’s read, does not want to be the man who started a hundred-year civilizational competition; he wants a framework for peaceable relations. Ball will not bet against POTUS. He has been surprised too often. Then he drops the line that is, he says, fairly politically incorrect: Claude Opus 4.5 in Claude Code is, by some definitions, basically AGI. People say that behind closed doors and will not sign it. He hopes next year’s realities force more candor, a wider Overton window, and a rewrite of foundational assumptions on both sides. Labenz adds a forecast Ball does not dispute: first drop in knowledge-worker jobs, expected in the second quarter of 2026. Ball also expects the models to be really good, and probably autonomous cyber attacks.

Peter Wildeford corrects the intro: he is not a co-founder of Metaculus. He sits on the board, participates avidly, works at the Institute for AI Policy and Strategy, and is among the world’s top-twenty forecasters. On chips he is a hard no on selling to China. The argument that NVIDIA sales undermine Huawei demand, he says, assumes China is a capitalist country that will respond to supply and demand. It is not. Huawei currently supplies something like 1 to 4 percent of Chinese demand. The government will make sure, by law and by software requirement, that companies take Huawei’s product. NVIDIA, in his unkind phrase, is a useful idiot filling the other 96 percent, turbocharging the industry until Huawei can take more, at which point the state will put a firm thumb on the scale. Tesla thought it would sell EVs and got pushed out for BYD. Apple got pushed for Huawei. Uber got pushed. Same script, every industry.

Prakash offers a galaxy-brain alternative: maybe Washington does not believe in AGI, and the chips are a red flag to a bull, a Strategic Defense Initiative trick to make Beijing overspend while it also has a burst property bubble, an anti-involution campaign, and a SpaceX launch race. Wildeford, with due respect, thinks that leans on second-order effects and ignores the first-order premise. If you believe what he believes about AGI and ASI, the big spending is rational; China may be *underspending* relative to the advantage. America’s historical strength is deeper capital markets. Zero-COVID showed Beijing is not omnicompetent. If the goal is merely to trick China into spending, race to Mars instead of handing over what powers its military and economy. China, he says, is cyber-attacking relentlessly and using AI for domestic human-rights abuses. Powering that does not make sense.

Labenz lists the shifting justifications for the chip ban — deny military applications, then frontier models, then “we’ll have more agents and they’ll be poorer” — and says making the Chinese public poor is not consistent with his values. Wildeford agrees, firmly: it cannot be U.S. policy to make China poor. Chinese people, and even the Chinese government, should benefit from AI, including American AI. The threat model on *chips* is that once you give them you cannot take them back; that DeepSeek, Qwen, and much of the current Chinese ecosystem run on American silicon (The Information has reported DeepSeek smuggling chips); and that military AI will be drop-in analysts and commanders in a civil-military-fusion system, relevant in a Taiwan or great-power-war scenario. Fighter jets are the inflammatory analogy. So rent, don’t sell. ByteDance chips in an Oracle data center in Malaysia are different from chips onshore: in a war you can turn them off, the PLA will be disincentivized to put military workloads on American-spied infrastructure, and Chinese commercial firms may still be happy to use them. That, he says, splits economic benefit from military capability, and inside the American debate it already counts as a moderate position.

Prakash puts up the ARC-AGI leaderboard: about a 390-times cost improvement in one year. Wildeford finds that under-noticed. Every eval poster emphasizes capability. The old barely-possible tasks are becoming things open-source models can do, then laptops, soon phones. Affordability times capability is the quantity that turns into use. He did not specifically forecast that curve, and he still finds himself surprised at how fast existing capabilities get cheaper, sometimes in months. Labenz then looks back at their 2025 predictions. Wildeford was more accurate on benchmarks; Labenz overestimated scores across the board and was closer on revenue, the number that really blew out versus the community forecast. Progress a bit shy of Labenz’s hopes, revenue exploding, costs collapsing, and everyone still asking where the economic impact is. How do those three facts sit together?

Wildeford owns a dumb forecast. At year-end, OpenAI, Anthropic, and the rest published next-year revenue expectations. He summed them. Surely the companies know more about revenue than he does. What he learned is that companies have to use standard economic models, and they are not allowed to put unprecedented AI progress into those models or they look unserious. He was expecting exactly the progress those backward-looking models were not pricing. In retrospect he should have known revenue would come in a lot higher. What actually exploded was coding: Claude Code, Cursor, Windsurf. He had expected more economic value from other computer-use agents; 2025 was not the year of the agent in that broader sense. He spent five years as a software engineer and data scientist, coding hours a day. He now rarely writes the first draft himself. Claude Code is faster and better. Prakash adds the CFO constraint: researchers can dream, the CFO cannot tell investors to expect a 10x without named products and signed customers. Claude Code’s 10x year still started from a late-prior-year launch that was merely strong. Next year’s official projections are more like 2x to 3x. Wildeford thinks they will exceed again — unless the zero-to-one of “people weren’t using AI, now they are” cannot be recaptured. The real game-changers, drop-in remote workers at tens of thousands a month rather than a $20 Claude Code seat, are still not here, and they depend on computer-use agents becoming as reliable as the hype.

His information diet is Twitter, Zvi’s aggregation, Timothy B. Lee’s Understanding AI, Nathan Lambert, SemiAnalysis, Epoch AI, and Meter — shops with a reputation for not vague-posting. Epoch’s professional mathematicians, people like Daniel Litt putting new models through real research, show systems that win contests and still fail to deliver for working mathematicians. Meter looks at reliability, which is the difference between a benchmark and keeping a job: if he succeeded at 80 percent of his policy analysis he would not last. OpenAI Death-Star posting and Elon calling every model AGI are noise he now tunes out. Ignore the first 24 hours of early-access glowing reviews, he says; those testers get access because they glow. Forecasting, more broadly, is not just putting a number on a well-defined question. The neglected parts of the value chain are which questions you pose and what you do with the number. New paradigms show up maybe twice a decade, so any given year might have at least a 10 percent chance of something qualitatively different — continual learning, chain of thought leaving English, data efficiency closer to a human’s ten examples. Pre-training seemed to burn out just as reasoning models took off, which can produce surprisingly linear progress across a changing guard of paradigms. That should widen uncertainty, not shrink it.

The Meter eval he tracks most grades software tasks by how long a human professional would need. Right now AI hits about 50 percent reliability on tasks that take those humans two hours. That is enough to get real value out of Claude Code, with a person still in the loop every twenty minutes. By the end of 2026 he thinks it is possible to have systems that do a day or two of human labor — more autonomous, running the trial and error themselves. Meter’s other finding is that many measurable tasks, not just software — computer use, economic analysis, legal work, self-driving — show similar capability increases from much lower bases. An AI that now bungles a ten-second computer task could, in his picture, be using computers at something like human level a year later. Even a clumsy loop that tries five times, fails four, notices the fifth, and reports back while you sleep would be valuable because it is cheap. He called 2025 the year of the agent and it did not stick outside coding. He is happy to repeat the prediction.

The politics of that, heading into a midterm year, he expects to get hotter. Waymos should be more commonplace. An affordable, widely used agent that books the Uber Eats and the plane tickets could be another ChatGPT moment: insiders were not surprised by GPT in 2022, but a slick UI turned it into summits. People might viscerally feel unemployment even if the tick-up is small. Populist anti-AI sentiment is already up — data centers using the water, raising electricity prices. You cannot do much about OpenAI from your porch. You can do something about the data center in the neighborhood, which is why local opposition feels like agency. He does not expect actual solutions to pass in an acrimonious midterm year. The job, as he sees it, is to harness that sentiment toward productive uses like Waymo without running the “no guardrails” program he thinks much of big tech wants.

Labenz wants a year for a house robot that does laundry, meals, and morning coffee. Robotics, Wildeford says, has always been on the disappointing side, with bad evals. His hunch is that once it starts working it will really start working, the way robotaxis were not economically viable for a long time and then, a decade after his 2015 college-kid forecast, began to be. He would venture the end of the 2030s for a robot butler cheaper than a human one — tremendous for consumers, hard on the workers who do that job now. Labenz feels robotics is maybe four years behind chatbots: first humanoids to homes next year at GPT-3.5 reliability, and if the curve holds, a Claude-Opus-class robot around 2030. Wildeford notes the ChatGPT moment was only two or three years before Claude Opus 4.5. Prakash thinks the servos are in better shape than the control software. Wildeford’s recent fact: robot hands have about 10,000 tactile sensors, which sounds like a lot until you hear that a human hand has almost a million touch points. Two orders of magnitude, and a clean illustration of Moravec’s paradox: dishes are harder than math contests.

After Wildeford signs off, Labenz and Prakash try to step back from a marathon. Bores was charismatic in the way politicians are and tech people often are not; he did not bite on crypto. Ball sounded more optimistic than Zvi, who had opened the earlier half of the show on a darker note. What struck Prakash most was how many of the developments they had just discussed happened in the last few weeks: GPT-5.2, the 390x cost drop, nested learning. Biotech people talk about results from a year or eighteen months ago. AI is visibly diffusing from research into the community. It is not yet diffusing into the economy. Everyone on the show expected more this year and still expects more next year. The market’s line is bubble; expectations have flattened. Insider views and market views are not the same object.

Labenz wrestles with Noah Smith’s version of that split: the technology is real, capabilities will keep rising, and that does not guarantee durable high margins for the firms whose financing assumes they will. Switching costs are rising and still not huge. Models are cheap because nobody wants to be dramatically undercut. You can imagine progress continuing and price wars continuing until the economics go haywire. Prakash’s counter is that useful models get smaller. Gemini 3 Flash is on par with 2.5 Pro and much cheaper to serve. A software engineer’s job that needs a Blackwell today might run on an H100 in twelve months. Jensen’s point that H100s are still busy is the same idea: older silicon keeps doing the economic task as models shrink. A tiny recursive model already did ARC-AGI at 7 billion parameters. If the job is the job, six-year-old data centers still have work. That is his case that the financing can be sustained — long service life, not a one-year spike.

Meta’s fallback, as Prakash hears it, is slop if superintelligence does not arrive: personalized ads so specific that someone who looks like your spouse or your kid pitches you a product. Video generation eats GPUs; the cash flows are still there. Worst case, junk-bond territory for a few years, with enough confidence in advertising to keep building. What worries him is concentration. A lot of U.S. economic activity now sits in a handful of ten-trillion-dollar firms making a one-way bet. China, in his telling, is diversifying across many other bets. People at the top are willing to gamble, so they are gambling. Labenz leaves that as a cliffhanger, lists what a nine-expert day still did not cover — mechanistic interpretability, alignment and control, AI for science, media generation — and asks anyone who made it this far for feedback.