Dean Ball on Joining OpenAI: New Power Centers, Frontier AI Policy, & Main Character Energy
Summary
Eleven months in, Ball judges America’s AI Action Plan roughly “30 to 40% done,” with real gains in energy, military adoption, manufacturing, and deployment—but a widening gap between competent implementation below and reactive politics above. His largest drafting regret is that it read like “three dozen separate thematic objectives” instead of one strategy for generalist agents, American primacy, and positive-sum global diffusion. The administration then validated allies’ deepest fear by imposing frontier-model export controls on non-US persons with 90 minutes’ notice, even as Ball hopes policymaking remains in a “high neuroplasticity phase.”
The Anthropic supply-chain designation and Fable ban illustrate how legitimate security concerns can become entangled with weak frontier-AI context, personal friction, and post-hoc political justification. The designation remains in litigation and could plausibly reach a Supreme Court disposition by summer 2027; meanwhile, the Department of War appears to be winding down Anthropic while other agencies—and reportedly even the NSA under Anthropic’s surveillance and lethal-weapons red lines—continue using it. Ball sees the Fable restriction as an improvised attempt to remove a model from market, not a coherent universal rule: “If you were a user of Fable, your world became dumber in the last week.”
Moving frontier-model oversight into classified, intelligence-led processes risks giving government a capability monopoly while discarding society’s “parallel compute.” Ball accepts classified work where necessary, but objects to a world where unknown models are tested against undisclosed standards and access decisions are made by roughly 20 officials, perhaps 15 without deep AI context. State-level frontier laws offer a more promising counterexample: California SB 53, New York’s RAISE Act, and Illinois SB 315 substantially converge, while Illinois, Connecticut, Virginia, and potentially Ohio are building auditing or independent-verification machinery.
China’s reluctance to buy US chips is partly strategic signaling, while the larger technical surprise is that world-simulation systems may have pulled dexterous robotics sharply forward. Ball expects Beijing to proclaim semiconductor self-reliance while DeepSeek, Alibaba, Zhipu, and others privately lobby for American chips; he concedes his prediction that DeepSeek’s top model would be closed by the end of Q1 2026 was wrong. Persistent 3D world simulation changed his robotics forecast immediately: synthetic-data pipelines built from human demonstrations could solve manipulation far sooner than he had expected—his reaction was, “Okay, dexterous manipulation in robots is going to be solved in eight months.”
Ball is joining OpenAI because frontier labs have become a new species of political and economic power center whose decisive information and governance choices cannot be understood from outside. He compares them with the emergence of modern finance: government lacks the expertise to write every rule, private standards must fill the gap, and AI itself will become an instrument of statecraft. His new team will look 6–12 months ahead, work closely with researchers, and focus especially on internal deployments that existing regulation—usually triggered by public release—does not reach.
Ball’s base case is that recursive self-improvement produces another steepening of the curve, not an instantaneous singularity, but even a 10%–20% chance of discontinuity warrants concrete contingency planning now. He wants predefined indicators, inter-lab coordination options, and clarity on when government must enter. His sharper concern is execution: labs may have plans yet remain unsure “that we’re going to follow it,” making internal conviction as important as written commitments.
AI has become sufficiently intertwined with semiconductors, energy, startups, and nationally important IP that a 2027 growth disappointment could trigger an implicit government backstop. Ball sketches a slowdown in data collection that cuts capex expectations, knocks equities down 20%–30%, and cascades through interconnected balance sheets; intervention could then become a public-interest necessity even without an explicit bailout promise. Government still holds the Defense Production Act and the monopoly on legitimate force, so labs’ durable defense is broad diffusion: every bank, university, and major industry should have a stake in preventing confiscation or nationalization.
Ball believes the transition may create a brief “main character energy” era in which individual judgment reaches maximum leverage just before machines become primary actors. His safeguards are personal: preserve independent public writing, define resignation lines in advance, resist both government capture and commercial expediency, and leave if his team becomes window dressing. He will use AI deeply for research and thought partnership, but still sees human-authored essays as valuable because “every single path through life is highly improbable”—and a model cannot make observations from experiences it never lived.
Deep dive
1. The AI Action Plan was written for a future-minded Washington
Ball says AI’s capability trajectory has produced few fundamental surprises for him: he expected “models with scary cyber capabilities” in late 2025 or early 2026 and biological capabilities soon afterward. The unexpected upside was how rapidly ordinary people adopted coding agents.
Washington, however, was not living inside that forecast when the plan was drafted. Ball calls the document “strange hermeneutics”: it addressed contemporary officials while anticipating versions of those same readers who would soon be “30% more AGI-pilled,” then 50%, and reread its language differently.
His preferred unifying argument was that generalist agents are coming, government is not leading their development, and the state must “ride with the current of the river.” The objective was American primacy and geopolitical power, but through world growth and broad participation rather than a purely zero-sum strategy.
The resulting plan did not fully stitch those ideas together; it read more like “three dozen separate thematic objectives.” Given two more months, Ball would also have added sector-specific adoption work—especially HHS and the VA, whose single-payer-scale medical data and government-provided care could support unusually valuable AI experimentation.
2. Implementation is substantive, but senior politics has broken with the strategy
Treated as a to-do list, Ball estimates the plan is “probably 30 to 40% done” roughly 11 months after publication—good performance by government standards. Some national-security implementation remains classified, making the public scorecard necessarily incomplete.
One deliberately mundane passage even contemplates the military commandeering US data centers during a national crisis and stitching them together for an unspecified task. Ball uses it as an example of consequential, almost “Leopold Aschenbrenner” ideas hidden in bureaucratic language.
Visible execution includes nuclear policy, military AI adoption, support for autonomy startups and US manufacturing, and forthcoming FERC changes intended to accelerate grid connections for very large electricity users. Military uptake has surprised Ball positively, as has broader American adoption.
The rupture is at the top: career officials implement the plan, while cabinet-level actors react to events such as Mythos without consulting its logic. Ball sees both underreaction to real risks and panicked measures that miss them, but hopes this remains a “high neuroplasticity phase of policymaking” that could look different in three months.
3. The Anthropic supply-chain fight remains legally and institutionally alive
Ball situates the designation in a longer expansion of executive power beginning with Obama’s “pen and phone.” Each administration pushes new boundaries, headline attention fades, and litigation continues long after most observers have moved on.
Anthropic’s case had recently been argued before the DC Circuit, with a ruling expected. Ball assumes Anthropic would appeal a loss and thinks some Supreme Court disposition by summer 2027 would not be surprising—even if the Court merely declined to hear it.
Inside government, the outcome is fragmented rather than erased: the Department of War appears genuinely to be winding down Anthropic, perhaps completely by year-end or within another year, while other agencies remain free to use it.
The NSA reportedly retains an Anthropic contract despite belonging to the Department of War, and reporting suggests it accepted Anthropic’s prohibitions on domestic mass surveillance and autonomous lethal weapons. Ball’s answer to the apparent contradiction is simple: “The US government contains multitudes.”
4. Classified model governance sacrifices society’s parallel intelligence
Ball separates the cyber executive order’s uncontroversial software-vulnerability work from its voluntary pre-deployment regime: models would be tested 30 days before release, with details primarily classified and the intelligence community—practically, the NSA—taking the leading role.
His concern is a future where frontier access is gated, capabilities and standards remain secret, and government quietly decides which abilities to restrict. Beyond a civil-liberties objection to government monopolization, he argues that “the public has a right to know” about one of history’s most important technologies.
A civilization is an information-processing system, with citizens supplying “parallel compute.” Since 2023, public AI-policy communities have made real progress on handling models at Mythos-level capability; legislation can absorb that dispersed expertise, whereas centralized secrecy leaves overloaded senior officials improvising with thin context.
Ball does not reduce the shift away from CISA to partisan hostility toward a Biden project. His diagnosis is institutional: an AI-governance system is being improvised by perhaps 20 people, “15 of whom don’t have a ton of context for AI.” The answer is Congress, public scrutiny, and more voices—not simply blaming those officials.
5. States are converging on frontier rules while fragmenting ordinary AI markets
Ball sees stronger-than-expected momentum behind private governance, auditing, and independent verification. Anthropic and OpenAI have published documents favorable to the concept; Illinois passed a frontier-auditing requirement, while Connecticut and Virginia authorized studies or pilots and Ohio is considering a more robust implementation.
California’s SB 53, New York’s RAISE Act, and Illinois’s SB 315 use remarkably similar transparency language, with Illinois adding auditing. For Ball, this is laboratories-of-democracy federalism working as intended: states converging on a common framework rather than manufacturing a patchwork.
The less visible record is worse. Consumer protection, algorithmic pricing, and hundreds of synthetic-media or deepfake laws create confusing compliance burdens that may fall especially hard on startups.
Occupational licensing is the sharpest specimen: Illinois has defined mental-health services as something only humans may provide, potentially making a chatbot’s response to “I’m sad; can you help me?” unlawful. Ball finds it perverse that preemption advocates attack carefully sculpted frontier-safety statutes while neglecting these more restrictive patchworks.
6. China’s chip posture is layered, while world simulation changed Ball’s robotics clock
Ball distinguishes Chinese policy announcements from implementation. Proclaiming an indigenous chip ecosystem serves national pride and tells domestic and foreign audiences that China no longer needs America; it does not mean every Chinese AI company shares that preference.
He is confident DeepSeek, Alibaba, Zhipu, and others are lobbying Beijing for American chips, and believes some sales are occurring despite restrictions. China may still make import restraint a major objective, but the public stance should not be read as complete commercial disengagement.
Ball’s explicit miss was catastrophic-risk policy: he predicted DeepSeek’s leading model would cease being open source by the end of Q1 2026. “That prediction was wrong”; Beijing still appears more worried about labor disruption than catastrophic AI risk.
Technically, persistent world simulation surprised him most. Earlier systems recreated scenes dreamily, changing objects whenever a user looked away; then “one day it just worked.” Human demonstrations captured through devices such as Apple Vision Pro, synthetic worlds, and small amounts of high-fidelity muscle data suddenly made dexterous robotic manipulation look much nearer.
7. Consumer tools were already compounding daily life before Fable raised the ceiling
Ball delights in “weird subgroups of very normal people” using coding agents, especially homeschooling mothers building with Claude Code and OpenClaw. His own example was generating country packets and matching snacks for a Mexico–Korea World Cup game.
Labenz’s example was using Claude Code to combine live NBA League Pass data into a Nate Silver-style dashboard estimating each game’s “odds of being a good game.” It solved a small but genuine problem for someone who follows no single team and might face eight simultaneous games.
Fable felt categorically sharper: “a fiercely intelligent model” and “a real step up in intellect,” comparable to Ball’s first encounter with o3. He once assumed o3 would always feel brilliant; the hedonic treadmill moved anyway, and he expects it would now seem comparatively dim.
For a FERC proceeding, Ball had Fable review his roughly 40-page expert testimony and an opposing expert’s 70-page rebuttal, then had Mythos read it in Cowork and conduct research. He did not use the model’s prose, but its analysis “demolished this dude in a way that I mostly couldn’t have,” leaving him wishing he had been able to use Fable more.
8. The Fable ban looks like improvised security policy colored by politics
Ball identifies three interacting ingredients: legitimate safety or security concern; insufficient context for judging frontier-model risk; and Anthropic’s political status after repeated clashes with the administration. He cannot know the ratio, but rejects any analysis that omits one of them.
His read is not that Washington announced a durable rule requiring export controls whenever a model has a vulnerability. Officials wanted Fable off the market and reached for “the only thing we can think of that we’re pretty sure will actually get the darn thing” removed.
The explanations Ball recounts include security concerns and difficulty reaching Dario Amodei, echoing the supply-chain episode’s complaints that Amodei took hours to return a call. Ball sees a Washington status contest in that reaction—“a who’s-the-bigger-monkey aspect.”
Later explanations emphasized a jailbreak, then claimed Anthropic supplied the model to a Chinese-linked company. Ball calls that post-hoc grasping: the company was SK Telecom, part of the Korean SK group that owns SK Hynix, and hardening an important allied telecommunications network looked entirely reasonable to him.
9. Frontier labs have become a new form of political and economic power
Ball compares today’s labs with the emergence of recognizably modern finance in the Dutch Republic and Britain. New financial instruments required common rules before anything like an SEC existed; similarly, frontier AI needs governance that states cannot develop quickly enough on their own.
Much of that governance will therefore originate inside companies and in private bodies that establish norms, audits, and standards. The point is not that public authority disappears, but that public institutions lack the immediate capacity and expertise to supply the whole system.
AI will also become an instrument of statecraft, as finance already is: money serves policy goals nominally unrelated to banking because it is fundamental to almost everything. Ball expects advanced intelligence to acquire the same cross-domain role.
Neither the White House nor ten months outside it—with unusually strong access, travel, and networks—let him move beyond abstract intuitions about this institution. With foundational policy potentially taking shape over the next 18–24 months, OpenAI offered access and practical responsibility that commentary could not.
10. Ball’s new team will look ahead of policy—and inside deployment walls
Ball describes his new team as a boutique operation distinct from Chris Lehane’s Global Affairs organization. Global Affairs handles conventional policy, lobbying, and incoming demands from all 50 states, the federal government, and international jurisdictions; neither team reports to the other.
The new team’s horizon is six to 12 months: identify issues that are barely visible today, anticipate where capabilities will take society, and develop policies before public pressure hardens. Ball wants its intellectual output to rival an excellent independent think tank’s.
That requires “detail, detail, detail,” not a generic belief that models will improve. Ball expects much of his time to involve “jamming with the technical staff” about roadmaps, capabilities, internal deployments, and exactly how the world one year ahead differs from the present.
Existing regulation is generally triggered by public release, yet pivotal choices may concern models deployed only inside labs because of security, regulation, compute, or risk. Ball speculates that Mythos 2 and OpenAI successors may be progressing, but stresses he has not yet seen OpenAI’s roadmap; his reason for joining is to shape those judgments with researchers and executives.
11. OpenAI’s mission matters, but Ball preserved an independent American voice
Ball expects to sit on OpenAI’s MAC—the Mission Advisory Council or Committee, he could not recall which—which brings together researchers, Global Affairs personnel, and others for policy and internal-governance decisions. He believes the mission of benefiting humanity is taken seriously inside the company.
The hard part is interpretation: a broad mission does not mechanically settle ambiguous choices. Ball therefore considered preserving public writing without OpenAI editorial review important to taking the role.
He expects good-faith disagreement, not the “cartoonish” villainous conspiracy sometimes imagined in AI-safety circles. OpenAI retains some “Xerox PARC” research culture, including internal dissent, and Ball wants freedom to explain publicly when his judgment differs from the eventual corporate decision.
His objective remains getting the transformation right “for the country and for the world,” but country first: he identifies as an American patriot, not a citizen of the world. He also accepts that he is helping one competitive company set strategy, rather than advising an abstract industry.
12. Recursive self-improvement is likely continuity with a dangerous tail
Ball begins from general-purpose technology: one purpose to which a general technology can be applied is itself, so recursion is not alien to technological history. “It would be surprising if there weren’t recursive self-improvement in AI.”
Models may have helped improve models since at least GPT-4, perhaps much earlier. That makes a single clean break possible but not Ball’s default; his recurring prior is that history contains “always more continuity than discontinuity.”
His first task is therefore empirical: “measure twice and cut once,” study the roadmap, and refine the probability that near-term RSI creates a sharp leap rather than another smooth acceleration. He repeatedly warns that he has not yet entered OpenAI or examined its internal plan.
Even a 10% or 20% probability of discontinuity is enough to prepare now. Ball wants specific indicators identified in advance, operational triggers tied to them, inter-lab options for slowing or pausing, and clarity on the point at which companies bring in government.
13. Coordination needs both antitrust room and internal willingness to obey plans
Ball supports considering an FTC “no-action letter” stating that narrowly defined safety coordination among labs will not be prosecuted as cartel conduct. Properly scoped, it creates optionality before a crisis without pre-authorizing broad commercial collusion.
Anthropic’s Fable safeguards complicate that case: degrading output quality in selected areas “in the name of safety” looks to Ball like a consumer-protection problem. An industry-wide agreement to degrade products would be plainly anticompetitive, so labs can undermine their own request for coordination latitude.
Inside labs, he senses vertigo rather than terror—people approaching a cliff without knowing what lies beyond it. RSI might merely reproduce the post-reasoning-model benchmark kink, perhaps “30% more”: not a singularity, but still massively consequential.
Ball’s position is “massively inflationary” compared with mainstream expectations and deflationary only beside a small East Bay safety community. His larger worry is organizational: labs may be unsure both what RSI means and whether they will follow their plans. Strategy must build the conviction to “actually listen to our own plan.”
14. Human agency may peak just before machines become primary actors
Structural forces are Ball’s river: people are born into an “involuntary association” with history’s current. Great historical actors are those who refuse merely to swim and, through determined resistance, alter the river’s eventual course.
His largest update from government is how often outcomes depend on relationships among very few people. Infrastructure buildout is structural; the Department of War–Anthropic conflict, by contrast, is substantially about a bad relationship involving Dario Amodei and senior officials.
If humanity is nearing an “eclipse of the human intellect,” the irony is a final period of intense “main character energy.” Ball compares it with a dying star expanding into a red giant: a potentially beautiful, ugly, heroic—or villainous—flowering as machine intelligence is born.
The practical demand is controlled entropy: “You want there to be a fire, but you also don’t want to set the forest on fire.” Labs and policymakers will need artificial constraints, lines in the sand, and actions against narrow economic interest; the next years may require “action rather than commitments.”
15. Formal control matters less than institutional agency and character
Asked about the super PAC Leading the Future and New York politician Alex Bores, Ball cautions against assuming donors directly dictate political machinery. Even decabillionaires feel principal-agent problems; funders usually select organizations they broadly trust rather than scripting every move.
His metaphor is a “wind-up doll”: Leading the Future saw a politician branding himself as an AI regulator and tried to warn others that following him would bring defeat. Instead, the intervention raised Bores’s profile and produced a Streisand effect around his primary.
Ball distinguishes the super PAC’s conduct from OpenAI’s, despite Greg Brockman being among its funders. He speculates that New York’s RAISE Act and California’s SB 53 probably did not pass without at least tacit OpenAI support, and notes that he has considered Bores a friend for roughly two years.
On character versus corrigibility, Ball’s intuition favors character, though he wants more empirical evidence: put “the right snowmelt at the top of the mountain” and let gradients work. Confucian li and ren capture why rules fail—the world changes too quickly to codify proper conduct, so timely moral judgment must come from cultivated virtue.
16. Public equity may work only if the public owns it directly
Ball grants that humanity created the knowledge commons on which models train, but rejects the premise that this alone proves compensation is owed. Civilization is a shared library whose heirs are expected both to use it and contribute back.
The counter-account is consumer surplus: if society wants repayment for training data, it should also compensate AI companies for enormous positive externalities they will not capture. Labenz makes that concrete—during his son’s cancer, he would have paid perhaps 100 times ChatGPT Pro’s price.
Ball nonetheless accepts that this may be a uniquely deep draw on collective knowledge and that sharing upside could be politically prudent or cosmically just. He strongly opposes government-held equity, which could become a corporate-control lever and create a conflict if safety measures required severely constraining or even banning the labs.
He is more open to taking perhaps 15%–20% of AI-company equity and dividing it among American households. If valuations rose from roughly $1 trillion to $10 trillion, the result might fund “an entry-level Mercedes,” meaningful but not transformative; even $5–10 trillion firms would still capture only a small share of total consumer surplus.
17. AI infrastructure has already created an implicit too-big-to-fail problem
Ball sees no deliberate bailout strategy, but interconnected balance sheets now span frontier labs, VCs, startups, semiconductors, energy, and physical infrastructure. Many nominally independent businesses are thin wrappers around capital tied to the same frontier expansion.
The buildout is financing nationally valuable IP in small modular reactors, fusion, batteries, materials, cooling, and water systems without conventional federal subsidy. His favorite example is cleaning mildly radioactive “produced water” from fracking for closed-loop data-center cooling—“the most old-school example of capitalism ever.”
A failure need not mean AI hits a wall. In 2027, labs might discover that progress requires years of slow data collection across occupations; growth still continues, but its second derivative falls, capex forecasts decline, and related equities drop 20%–30%.
Those declines could force selling, impair commitments across intertwined companies, and endanger strategically important technology. At that point intervention becomes a public-interest question. Like the pandemic backstop revealed during COVID, government support may be implicit simply because “that’s the way the world works,” even if labs should neither request nor expect it.
18. Government’s hardest power is compute priority; labs’ defense is diffusion
Current models already unlock extraordinary national-security utility because government possesses a “data overhang.” Ball says the National Geospatial-Intelligence Agency alone gathers enough annual information to require eight million human analysts, against roughly three million total federal employees.
AI releases latent “kinetic energy” in that apparatus across intelligence synthesis, cyber offense, and targeting. Ball says Project Maven’s integration of AI reduced the people involved in missile targeting from about 2,000 to 20 before today’s agents; he speculates the number might now be five.
Labs also derive leverage from researchers: CEOs cannot ignore internal constituencies of scarce technical talent and can credibly warn government that an unacceptable demand would trigger rebellion. Automated AI R&D may weaken that constraint, while the state retains the monopoly on legitimate violence.
Under Defense Production Act Title I priorities authority, a president could designate advanced compute scarce and essential, require hyperscalers to serve government first, and pay market rates. Near-infinite federal demand could crowd out private users; the practical obstacles are money and institutional capacity, not a novel legal theory.
19. Broad deployment and open source keep state–lab bargaining pluralistic
Ball’s preferred safeguard against confiscation is dependence spread across society. A lab lobby is one unpopular industry; every bank, university, business sector, and major institution demanding continued access becomes a Madisonian coalition capable of checking government ambition.
Secrecy produces the opposite equilibrium: if only labs, officials, JPMorgan, and Apple see frontier capabilities, nationalization becomes easier. Ball wants Fable-level and better systems broadly available so AI becomes “just capital,” supported by capital owners throughout the economy.
Open source is especially important for shared infrastructure. A widely trusted AI adjudication system might let participants bring private advisers while relying on a central, auditable model; domestic factions and international users would likely require an open-source system to accept it.
Ball expects digital open source to lag in the near and medium term as economics and national-security pressures worsen, though Gemma and GPT-OSS have shown that major US labs can remain involved. Robotics may be different: a Cambrian explosion of intelligent hardware needs common physical-intelligence models, while object-level security risks appear smaller to him.
20. Success means durable institutions, not personal access or lobbying wins
Ball sees open intellectual space for work that takes AGI seriously while defending classical liberalism and the foundations of the republic. He also wants well-funded independent-verification organizations staffed with “lab-level human capital,” paid enough to attract genuinely strong evaluators.
His concrete success test is a future where frontier capability remains broadly diffused, AI enables visibly new organizational forms, government–lab relations have clearer rules, and labs articulate a constructive account of their social role. He claims only a modest contribution to those outcomes.
His government relationships are mixed: close friends remain inside the administration, while other officials “hate my guts,” and he has heard that a young applicant merely retweeting him can be treated as a red flag. Narrow, reasoned criticism preserved some trust better than becoming a general-purpose Trump critic.
His new team is not a lobbying shop, which Ball says suits him because “I suck at that.” Global Affairs will manage routine federal engagement. He has known Sam Altman as an acquaintance for roughly 18–24 months and worked more extensively with other OpenAI executives, but “I wouldn’t say that we’re boys.”
21. Ball’s safeguards are a resignation letter and a recognizably human voice
Before entering the White House, Ball wrote himself a letter recording his beliefs and advance resignation triggers against power’s corruptions. He now considers OpenAI more consequential than that job and thinks he should repeat the exercise: “Draw your lines in advance.”
The mirror-image danger is conceding too much private agency to government merely to keep commerce flowing. He would also leave if his new team became thoughtful window dressing rather than influencing decisions; had he remained in government through the Anthropic supply-chain designation, he says he “would have totally resigned.”
GPT-5.5 and Opus 4.8 recently produced ideas and framing for his book’s first chapter that were “considerably better” than what he had in mind, though he did not ultimately use the output. Models already support the project from conception and research through contract negotiation and will appear in its acknowledgments.
Ball allows substantial AI authorship in pro-forma work such as regulatory comments, official letters, and immigration recommendations, but protects essays as personal communication. Models still struggle with structural metaphor and restraint—knowing “I could have gone there, but I’m not going to”—while his father’s death, the Roosevelt Room, and OpenAI supply lived material a machine never experienced.