Twenty-one months from an autonomous researcher, the Action Plan's author joins OpenAI

Nathan Labenz opens with the two dates on OpenAI's public timeline. An AI research intern is due in three months. A full autonomous AI researcher is due in March 2028, twenty-one months from the taping. In that window Dean Ball has announced he will join OpenAI to build and lead Strategic Futures, a team whose mandate is to help senior leaders shape frontier AI policy. Labenz calls it one of the most self-recommending episodes the show has done, because recursive self-improvement is no longer a thought experiment on a blog.

Ball is not an unknown quantity in Washington. He writes the Substack *Hyperdimensional*, helped draft America's AI Action Plan in 2025, and spent the ten months after leaving government commenting in near real time on Anthropic's supply-chain-risk designation, state AI bills, and the Fable ban. Labenz notes a detail that surprised him: even as an incoming OpenAI employee, Ball keeps the right to write publicly about AI policy, and the company did not ask to review this conversation. His first son is not yet a year old. The intensity, Labenz says, will exceed the White House. The independence was not the thing he traded away.

Labenz starts Tyler-Cowen-style with the White House year. How does the Action Plan look now? Ball says the technology path has not surprised him much: scary cyber capabilities in late 2025 or early 2026, bio not far behind, most of it things he had already predicted in public. Coding agents' popular uptake is the pleasant shock. Washington, when the document was written, was not living in that world. The Plan is a strange hermeneutic exercise: you write for today's audience and also for the same people six months later, 30 percent then 50 percent more AGI-pilled, hoping they reread it and finally hear it.

If he could rewrite, he would be more explicit that the subject is generalist agents that can do all kinds of work, and he would stitch the pieces into a single vision. The government is not leading this, he argues; it has to ride the current, use it for American primacy, and do so in a positive-sum way that grows the world economy and brings others in. As published, the Plan reads more like three dozen thematic objectives than one strategy. Two more months, he says, and he would have focused on that spine.

What hit the cutting-room floor still bothers him. He wanted sector case studies that named barriers the federal government could actually remove: hospital record-keeping at HHS, and Veterans Affairs, one of the largest single-payer systems in the world, with huge data and government-provisioned care. America does not think of itself as having single-payer health care. It does, at the VA. There was no time. Specificity, he says, was the missed chance.

Implementation, Ball says, is hard to score from the sidewalk because some of it lives "on the high side" — classified. One underrated line in the Plan, written in mundane language, obliquely contemplates the military stitching the country's data centers together in a national crisis, a Leopold Aschenbrenner-flavored idea that did not jump off the page. If you had a clearance and treated the document as a to-do list, he guesses it is 30 to 40 percent done about eleven months on, which is pretty good for Washington. Nuclear is moving fast. FERC is about to announce rules that speed large industrial users onto the grid. Military adoption of AI, and the Pentagon's industrial-policy work with startups in physical autonomy, have impressed him to the upside. Domestic AI adoption is going reasonably well.

The spirit is another matter. A pillar of the Plan was exporting American AI. That is hard to square with global export controls on frontier models, imposed with ninety minutes' notice on all non-U.S. persons. After government, Ball did quasi-diplomatic work as a private citizen, including at India's AI Action Summit. The sentence he kept hearing in Europe was: "I just worry that you Americans are gonna turn off the models at some point if you get mad at us." In office he had been trying to assuage that fear. Then the administration confirmed it.

Governance was mostly silent in the Plan, he admits, because the Overton window inside the administration was tight and the document was supposed to list executive actions, not new laws. He now wishes it had said more about what to do when things get scary, and how not to panic. Career civil servants and low-to-mid political staff are implementing the text in quite a good way. Cabinet-level people do not read every strategy document; they react. When Mythos-level news hits, they reinvent third-party evaluations and government technical competence from first principles — ideas already in the Plan. Ball's hope is that policymaking is in a high-neuroplasticity phase, and that three months from now the world looks different. The headlines, he cannot deny, do not look like the Plan.

Labenz asks whether Washington has memory-holed the Department of War's designation of Anthropic as a supply-chain risk. Ball puts it in a longer story: since Obama's second term, presidents with an intransigent Congress have used a pen and a phone, each stretching executive power, each getting tested in court, most of the public losing the thread. Anthropic argued before the D.C. Circuit last month; a ruling is due; if they lose they will appeal. He would not be surprised by a Supreme Court action of some kind by summer 2027, even a denial of cert. The Trump administration, he says, is savvy about which fights to walk up the mountain.

On the ground the U.S. government contains multitudes. Inside the Department of War proper, Anthropic contracts are being wound down; he would not be shocked if they are gone within a year. Other agencies were told the designation does not apply to them. The NSA, technically part of the Department of War, appears to still have an Anthropic contract, and reporting suggests Anthropic's red lines — no domestic mass surveillance, no autonomously lethal weapons — were honored. Americans, Ball says, are not good at tolerating ambiguity. A healthy amount of it is intrinsic to the process. He still calls the designation itself a total own goal, an unforced error you could have fixed a thousand other ways.

The cyber executive order is the structural worry. Patching critical software is fine. The other piece created a voluntary pre-deployment testing program, thirty days before release, details classified, run primarily by the intelligence community and practically by the NSA, whose cyber skill he respects. His fear is a future in which access to frontier models is gated, secrecy is the default, and the government decides — sometimes without the public knowing the capabilities even exist — whether to restrict them. Government monopolization of frontier AI, he says, is how you get scary civil-liberties outcomes, and he would say that with the president's name and party erased from his brain.

A civilization is an information-processing system; the country's people are parallel compute. Since 2023 the AI-policy community has actually made progress on how to handle Mythos-level models. Public process and legislation can use that compute. Centralize it in a classified echo chamber and you get twenty people improvising an AI governance regime from scratch, fifteen of whom do not have much AI context, high-level officials with a million other things on their plates. He is not surprised it looks like DC in spring 2023. Unlike the supply-chain fight, he does not want to be highly adversarial here. He wants Congress and the public in the room.

Labenz asks about the laboratories of democracy. Ball dates his private-governance work to after California's SB 1047 veto in late 2024: mandatory safety-plan publication, audits, whistleblower protections, independent verification organizations that could certify jailbreaks as real but not severe enough to yank a model. Two of the three big labs, Anthropic and OpenAI, have published documents friendly to that idea. Illinois passed a bill mandating audits of frontier companies. Connecticut and Virginia authorized studies or pilots for independent verification bodies. Ohio has a still more robust bill pending. Momentum, he says, is more than he would have guessed a year ago.

The three frontier safety laws — California SB 53, New York RAISE, Illinois SB 315 — use remarkably similar transparency language, with Illinois adding audits. That is not a patchwork; it is states converging on a common framework. The neglected half is uglier. There are now hundreds of synthetic-media and deepfake laws. Occupational licensing is bubbling up: Illinois has defined mental-health services as something only humans may provide, so a chatbot that asks how you are doing can be practicing without a license. Strangely, the frontier-safety statutes that soak up preemption energy are the best sculpted, often have industry support, and address cyber and bio risks that are no longer fake. The compliance swamp for startups is the stuff nobody is fighting.

Labenz's own surprise is that Washington eased chip export controls and China declined to buy. Ball is less shocked. Beijing likes the national-pride message that it does not need American chips; DeepSeek, Alibaba, and Zhipu are, he guarantees, begging for access; some chips are being sold anyway. He had predicted that by the end of Q1 2026 DeepSeek's top model would no longer be open source, and he was wrong. The Chinese state looks more labor-pilled than cat-risk-pilled. On the technical side, two shocks: homeschooling moms treating Claude Code as a life tool, and DeepMind world models suddenly gaining permanence — look away and the scene is still there — which, last summer, made him pull in his robotics timeline. Dexterous manipulation, he thought then, would be solved in eight months via synthetic data on top of human baselines.

Before the politics, Labenz wants the user report. Ball built a Claude Code dashboard on NBA League Pass that ingested eight live games and scored, Nate-Silver-style, the odds of a good one. Labenz printed World Cup packets and snacks for his kids. Then Fable arrived. Ball's impression: a fiercely intelligent model, a real step up, the kind of cracked-genius feeling people compared to early o3 — the moment he thought the hedonic treadmill might stop. He never got to use it properly as a coding agent. He had it open in Claude Code in a hotel, got distracted, and by the time he came back it was gone.

He did use it on knowledge work. He is an expert in a FERC proceeding over Order 1000. The other side hired someone to write a seventy-page rebuttal to his forty-page testimony. He pointed Fable, in Cowork, at the pile, let it research, and did not use a word of its prose. "The model demolished this dude in a way that I mostly couldn't have." Then it was banned. It is the first time, he says, we have gone backwards. Government intervention is usually too abstract or diffuse to feel. This time, if you were a Fable user, your world became dumber in the last week. Labenz felt it as taste, a mind-meld he had not had with Opus, reminiscent of dropping from GPT-4 red-team access back to text-davinci-002.

Three ingredients, ratio unknown even from inside the White House: genuine safety concern; a broad lack of frontier context that inflates that concern; and politics. Anthropic's fights with the administration color how important people hear a vulnerability, in a way that might not happen for a warmer company. He does not read a new standing policy that any jailbreak gets export-controlled. He reads officials reaching for the one tool they were sure would yank the model off the market, then writing justifications after the fact. The story changed three times: first, security concerns and Dario Amodei not coming to the phone fast enough — Under Secretary of War Emil Michael had publicly complained about return calls, a "who's the bigger monkey" grudge Ball says is a DC specialty; then a legit jailbreak; then, some 72 hours later, Anthropic had given the model to a Chinese-linked company.

That company is SK Telecom, part of SK Group, one of Korea's largest chaebols and sibling to SK Hynix, the leading high-bandwidth-memory producer. Anthropic had publicly announced weeks earlier that it was expanding Mythos access to U.S. firms and to allies. Hardening a key semiconductor partner's telecom infrastructure strikes Ball as extremely reasonable. If the China link were the real reason, why not say it on Friday? The post-hoc feel is what makes the episode hard to analyze: there is not much policy substance, just id.

Ten months out of government were not the luxurious think-tank life people imagine. The workload never really dropped. Ball's work had centered on the frontier lab as an institution, a new center of political and economic power. His favorite analogy is the birth of a recognizably modern financial sector in the Dutch Republic and Britain: options and derivatives before there was an SEC or a Bank of England. Strangers still needed common rules to trade in trust. Government did not have the capacity or the expertise, so governance grew inside the industry and in private bodies. AI is in that phase, and like finance it will become an instrument of statecraft — a way to pursue policy goals that, on paper, have nothing to do with the industry.

He sat in the White House and it was not a fruitful place to think about that. He sat outside with what he considers pretty good access and still could not get past abstract intuition without being inside a lab. Policy may actually get founded in the next eighteen to twenty-four months. He was writing a book and not acting on the thought, and then OpenAI approached him. Strategic Futures is a boutique team, shaped with senior leadership, distinct from Chris Lehane's Global Affairs shop, which fights incoming policy from fifty states, the federal government, and the world. A year ago people were barely talking about kids' safety or data-center water. The new team's job is to look six to twelve months out, jam with technical staff on where the capabilities frontier is going, and develop policy before the world arrives.

Internal deployment is the neglected trigger. Every government regulation he sees is keyed to public release. The important decisions will be made inside, under some mix of regulatory risk, security, and compute constraint. Anthropic had Mythos checkpoints in January; Mythos 2 is presumably not far from trained; OpenAI, he says from total present ignorance, is training something too. Until there is real supervision, auditing, or independent verification, recursive self-improvement is a gut call made on internal models. He did not want to keep pondering that on Substack. He wanted to help shape the decisions with researchers and the executive team.

Duty is mixed, and he is candid about it. OpenAI's mission — AI that benefits all humanity — is taken seriously inside, and part of his job, not yet a public bullet, is sitting on the MAC, a mission advisory council or committee of researchers, Global Affairs, and others who make policy and internal-governance calls. The mission is broad and ambiguous; good-faith disagreement is the expected interior, not a cartoon conspiracy. That is why independent writing without editorial review, and without fearing for his job, mattered as a condition. He is a patriot — *civis Americanus*, not a citizen of the world — country first, then the world. Helping OpenAI compete does not make him squeamish. Competition, on average, makes things better. He is a competitive person.

Labenz calls the race to RSI between at least two companies the most objectionable thing in the field: a phase change, initial conditions that have to be right, researchers who are not that confident. Ball first clarifies the competition he likes — his team should produce public-interest work so good that, unbundled from OpenAI, it would be one of the most interesting think tanks in the country. On RSI itself his base case is deflationary relative to East Bay eschatology and inflationary relative to almost everyone else. Every general-purpose technology is recursive; we have been using models to improve models since at least GPT-4. Continuity is the right prior. A sharp discontinuous jump is plausible, not his default.

Step one, once he can see the roadmap, is measure twice: refine the probability that this is a leap versus a smoother kink. Even 10 or 20 percent on discontinuity is high enough to plan now. He is skeptical of inter-lab slowdowns and pauses, but he likes the idea of an FTC no-action letter that says coordination for specified safety reasons will not be treated as a cartel. Scope it badly and you get Anthropic's Fable safeguards: degrading outputs in the name of safety, which he calls a clear consumer-protection violation. A cartel of labs agreeing to degrade products as "safety" would be wildly anti-competitive. Abuse the safety card and you wreck the case for coordinating on it.

Inside the labs he does not hear terror. He hears a vertiginous feeling of approaching a cliff, substantial uncertainty, and a shrug that we are not quite sure we have a plan, or that we will follow it. If RSI is another reasoning-model kink in the benchmark charts, or that plus 30 percent, it is not a singularity. Rohin Shah, DeepMind's AGI safety lead, told 80,000 Hours the same tension: be serious about plans, refuse hard commitments that lock you in too early. Labenz notes the track record of governance plans is not amazing. Ball's assigned work is practical. Having a plan is not enough. You have to build internal credence that the organization will actually listen to itself.

Ball minored, in spirit, in the philosophy of history. Structural forces versus great men is the dorm-room version of the question, and the unsatisfying answer is both. Structure is the river you are born into, an involuntary association. The people who matter are the ones who stand against the current hard enough to move the channel. His biggest update from a year in and around government is how much of the world is determined by whether a small number of individuals like each other. Data-center build-out is structural. The Department of War's collision with Anthropic, if you think it will matter historically, is in large part a bad relationship between Dario Amodei and senior officials.

If this is the eclipse of the human intellect, the irony is that the transformation itself will be a main-character-energy period, a window of maximum human leverage even if machines later become the primary actors. He reaches for a star: before it dies it swells into a red giant, ugly and beautiful in a Dionysian way. "I think it will be a heroic time period that we live through," he says — or a villainous one, with a lot of room for great people, and probably both. He does not cast himself as one of those people. He sees a modest role that looks bigger because he has a public profile.

Standing against the current, for him, means putting artificial constraints on ourselves while the transformation is entropic: a fire that warms, not a forest fire. Draw lines. Say we do not want to live in that world. RSI may be the example that requires acting against textbook corporate interest. Labs already behave abnormally compared with ordinary companies; the next few years may have to be characterized more by action than by commitments. Mealy-mouthed everything will not be enough.

Lightning round. What happened between OpenAI and New York Assemblymember Alex Bores, author of RAISE? Ball had breakfast with Bores near his old Manhattan office two years ago and considers him a friend. Leading the Future, the super PAC that includes OpenAI president Greg Brockman among its donors, he treats as a wind-up doll: by default it goes after anyone trying to make a name as the AI industry's regulator, and it tried that, with a Streisand effect in Bores's primary. Principal-agent problems do not disappear because you are rich. Deca-billionaire donors, in his experience, often feel out of control of the political shops they fund. He would be surprised if RAISE passed in New York, or SB 53 in California, without OpenAI's at least tacit support. He does not think the company has a particular beef with Bores.

On character versus corrigibility, his intuition is character, and he wants lab empirics. You put the right snowmelt at the top of the mountain and let the gradients work. If you could write the rules of morality down, people would have. You cannot, for the same reason you cannot write the rules of language down: the best writers break the formal ones. Confucianism's *li* is ritual propriety in real time; the world changes too fast for a complete code, so knowing the right ritual has to come from *ren*, an inner virtue. He has always been a believer in that and a skeptic of the positivist alternative. Alignment is, for once, a place you can bring data to a virtue argument.

Equity sharing is in the air from Trump and Bernie Sanders. Humanity did create the training data; humanity also captures most of the consumer surplus AI will generate, which is how firms give back. Training data is not, for him, a prima facie invoice. Giving equity to the government he strongly opposes: the people are the principal, the government is a leaky agent, and a stake used to steer labs would be disastrous. Sanders's version finances social redistribution from the same equity, which trades off against safety measures that might constrain or even stop the business; you cannot run a popular dividend and a shutdown option on one share. What he is more open to is taking 15 to 20 percent of the AI companies, dividing it by American households, and handing people a slice. Governance-wise that is not so different from the S&P 500. At trillion-dollar marks it is not life-changing; if names go to ten trillion, maybe an entry-level Mercedes each. Serious money, least-bad politics. Labenz's footnote: he would have paid a hundred times ChatGPT Pro's price while his son had cancer. History of technology, Ball adds, says even $5–10 trillion firms will collect a relatively small fraction of surplus. That is the point.

Are AI companies already too big to fail? Ball thinks the concern is real and not a deliberate strategy. Balance sheets are interwoven. Downstream, semiconductors, SMRs, fusion, batteries, materials, cooling, even a company trying to clean slightly radioactive fracking wastewater for closed-loop data-center cooling, are being subsidized by AI capex, not by the federal government. Imagine 2027: models still improve, but whole job categories need years of slow data collection, growth estimates come down, capex rolls over, stocks drop 20 to 30 percent, commitments break, nationally important IP is at risk, and the government steps in. Nobody wrote "the government is the backstop" before COVID. It became true because that is how the world works.

Where do labs get leverage against that state? First, utility. The national-security enterprise is sitting on a data overhang: the National Geospatial-Intelligence Agency alone collects, in a year, more than eight million human analysts could process, and the entire federal government has about three million employees. Project Maven already took missile targeting from 2,000 people involved to 20, before today's agents. That is not Bostromian superintelligence; it is kinetic superintelligence that was already in the sensors, waiting for cheap intellect. Second, CEOs are not simple top-down bosses. They have an internal electorate of top researchers. Sam Altman can credibly tell the government that a demand will cause a rebellion across labs, because it would.

Labenz pushes the automated-researcher future: quitting threats weaken, weights already sit on classified servers, Los Alamos with guns. Ball does not deny the monopoly on legitimate violence, or Defense Production Act Title I priorities authority — established law, not a stretch — which could put the government first in line for scarce compute at market rates. In principle they could crowd out the market. In practice institutional wherewithal is hard, and the government knows it should not kill the goose. Charles Tilly's *Coercion, Capital, and European States* is his text: modern states grew from gangster-like violence organizations locked in interdependence with merchant capital. Exit and seizure are both asymptotic. The practical stabilizer is broad diffusion. If every bank, university, and industry depends on the models, the interest-group map is Madisonian. If capabilities stay secret, in the labs and at JPMorgan and Apple, confiscatory outcomes get easier.

Open source, in that equilibrium, is under pressure. He hopes U.S. labs keep more than a toe in — Gemma, GPT-OSS — especially for public infrastructure that has to be auditable across parties and borders, an adjudicatory layer he sketched eighteen months ago. He agrees with Nathan Lambert that digital open source will lag in the near to medium term as economics and national security get harsher. Robotics is a different case: a Coasean explosion of physically intelligent lawnmowers, lamps, and humanoids that will not each train a frontier model, and with fewer object-level national-security fears. He is a spiritual supporter of open weights and a near-term pessimist on the digital side.

What would success look like, and what does he want from people outside? Intellectual space remains open for takes that treat AGI as real and still care about classical liberalism and the republic. The third-party audit and independent-verification ecosystem needs lab-quality people paid well, not nonprofit wages. Someone has to keep fighting government monopolization of frontier capabilities; his voice will still be his, but a lab employee sounds different, and he wants others in that fight. He is not a long-range goal-setter. In a few years, if frontier capabilities are still broadly diffused, new AI-native organizations are concrete, the government-lab relationship is clearer, and the labs have articulated a positive social role, he will count a small contribution as a job well done. The team will hire, small; email on deanball.com or reply to any *Hyperdimensional* post hits his inbox.

How does a critic of the administration walk into OpenAI without triggering an immune response? It is not a monolith. Some people want to ruin him; he has heard that a young job-seeker's retweet of Ball is a red flag. He also has dear friends inside he talks to almost daily, relationships that predate AI writing, and he has been publicly positive about other administration work — sharp, confined criticism, not a general brief against the president. This is not a government-affairs job. Lehane's team does not report to him. He told OpenAI not to hire him to lobby; he is terrible at it. His Tourette's-like inability to keep his mouth shut is the product. He knows Sam Altman decently, acquaintances for eighteen to twenty-four months, executive-level people more extensively. They are not "boys."

Before the White House he wrote a letter to himself: beliefs, red flags, when to leave. He was thinking about doing it again in the shower that morning. This job feels weightier. Compromise is inevitable because superintelligence shakes state sovereignty, as he wrote days earlier, and he does not want it monopolized by government. The easy compromise is keeping commerce going. Another red flag: a fancy team whose thoughtful output is window dressing and does not shape decisions. Disagree-and-commit is real inside voluntary associations; some things go too far. Had he still been in the administration for the supply-chain-risk designation, he says he would have resigned, easily — it was already in the letter. The White House letter warned him not to become the person who knows exactly how to assert technocratic power over labs. The mirror image now is: do not compromise the labs' private agency. Be a counterbalance. Do not specify every line in advance and overcommit to the wrong one.

He has mostly refused LLMs in his own writing. A country Airbnb, first chapter of a book due next year, GPT-5.5 and Opus 4.8 produced passages considerably better than what he had in his head. He did not paste them; he was influenced in a way that felt new. Models can write a great paragraph and sometimes a good legal doc. They still cannot construct an essay or a chapter that plants structural metaphors, leaves them implicit, and practices restraint. That remains a human skill, and the labs have optimized for analytic prose, not classical essays. Pro forma filings and immigration letters already go out under his name substantially co-authored. *Hyperdimensional* and Twitter have not, and he thinks the faith that a person actually wrote the thing will keep a preferential advantage even after the models are better essayists. Experience is the rest: the death of his father, the Roosevelt Room, now OpenAI. A machine did not walk that path. It will walk its own.