Dean Ball

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.

🔗 Original source & video: Dean Ball on Joining OpenAI: New Power Centers, Frontier AI Policy, & Main Character Energy

Frontier-model competition is becoming a capital-and-infrastructure game, with Nvidia’s H100 at the center of the moat. The chip attacks the interconnect bottleneck around Transformer-heavy matrix multiplication, while an Inflection AI-scale entrant raised $1.3 billion largely to build its own cluster; OpenAI’s $10 billion raise and Microsoft partnership, plus Anthropic’s Google relationship, reinforce the concentration. Erik’s bottom line: anyone training frontier foundation models at the scale expected in the 2024–2026 timeframe will be buying enormous compute, while an H100 ban could make it difficult for China to scale comparably if the ban holds and enforcement works.

The model market is separating into a small club that pre-trains and a much broader economy that fine-tunes or rents inference. Erik brackets frontier pre-training at roughly 1–10 trillion tokens and $1–100 million, with GPT-4 rumored at 13 trillion tokens and understood to have cost $100 million-plus; useful supervised fine-tuning can begin near 1 million tokens and cost as little as $100. That creates an investor-relevant bifurcation: enormous barriers around base models, but “a little bit of elbow grease” can still produce differentiated applications.

RLHF was the usability unlock behind ChatGPT, but its behavioral gains come with poorly understood losses. A pre-trained LLaMA behaves like “the world’s largest autocomplete,” instruction tuning makes it follow the requested role, and reinforcement learning teaches it to ask sensible follow-ups rather than rush to an answer. Yet users describe the result as “lobotomized,” creativity may decline, and mode collapse can make a supposedly random number such as 97 appear far too often.

Chat is simultaneously an alignment interface, an engagement engine, and a new category of emotional risk. The assistant format lets developers optimize for “helpful, honest and harmless,” while Character.AI, Pi and Replika show that people may form attachments even when they understand the machinery. The commercial tension is unusually sharp: products can market themselves as romantic practice while acknowledging that “the relationship is with an AI but the feelings can be real.”

A practical capability stack—reasoning prompts, retrieval, tools and persistent memory—is advancing faster than raw model intelligence alone. “Let’s think step by step,” majority voting and Tree of Thoughts exchange more latency and inference spend for accuracy; embeddings ground answers in trusted data, while APIs supply weather, search, code execution and other facts models cannot contain. Perplexity’s “product excellence” illustrates the opportunity, though dependence on Google and Bing APIs leaves a strategic need for its own index.

Incumbents that already own deep tools have a stronger position than startups offering an impressive AI layer over a shallow product. Adobe and Salesforce can teach models to operate mature creative or CRM systems; slide generators, by contrast, may create a good outline and then strand users in weak editing software. Gamma’s export to PowerPoint stood out because it acknowledged that reality, while services businesses such as Athena are betting on the “best human plus AI bundle” until agents genuinely replace people.

Agents can already convert natural-language goals into physical or digital execution, but the episode draws a hard line between following protocols and discovering them. A multi-agent system searched, calculated, operated Emerald Cloud Lab and synthesized aspirin; asked, roughly, to find and synthesize a cancer drug, it merely reproduced familiar ideas. Likewise, a Minecraft agent became a “lifelong learner” by saving successful skills, showing that compounding memory can make agents reliable without supplying breakthrough insight.

Multimodal bridges and efficiency techniques expand the addressable market while deepening both opacity and inference economics. Small adapters can connect frozen vision encoders to frozen language models, but then “models [are] talking to other models in a purely numeric high-dimensional space that humans cannot understand.” Quantizing from 32-bit to 8-bit can save 75% of memory, while distillation and mixture-of-experts routing aim at the expense that ultimately matters most: running models, not merely training them.

🔗 Original source & video: The AI Scouting Report: Implementation Trends Part 2 of 3

The debate’s actionable fault line is the orders-of-magnitude gap between Dean Ball’s and Max Tegmark’s extinction-scale risk estimates. Ball called his probability “sub 1%,” offering “0.01% or something like that,” while Tegmark put loss of control “definitely over 90%” if companies may deploy superintelligence without FDA-like safeguards.

Tegmark wants a conditional prohibition until superintelligence is demonstrably controllable and has strong public support, not a halt to useful AI. He cited polling that 95% of Americans oppose racing toward superintelligence and a paper finding recursive scalable oversight failed 92% of the time even under its most optimistic assumptions. His preferred outcome is “full steam ahead” on controllable tools such as AlphaFold, autonomous vehicles, and medical treatments—even if genuinely autonomous superintelligence must wait 20 years.

Ball’s central objection is that “superintelligence” cannot yet be translated into law without banning valuable systems and concentrating development in a licensed cartel. He expects that by roughly 2030 a model might solve major mathematics problems, advance multiple sciences, outperform humans in coding and legal reasoning, and improve AI research without creating Tegmark’s catastrophe. A statutory ban could become “N plus 1, N plus 2”: GPT-5 exists, GPT-6 is permitted, and GPT-7 is prohibited before anyone can gather evidence about its safety.

The strongest convergence came around concrete biological and cyber capabilities, where Ball has already updated toward targeted regulation. OpenAI’s o1 changed his assessment because deliberative reasoning and tool use created a legible path from biological knowledge to synthesized pathogens; that factual change helped move him from opposing California’s SB 1047 to supporting the more tailored SB 53. His discussion emphasized downstream controls—BSL laboratories and nucleic-acid synthesis screening—because “bits” are harder to regulate than physical choke points.

The FDA analogy captures both the case for preclearance and the risk of regulatory lock-in. Tegmark argues that tail harms dwarf corporate balance sheets, making lawsuits useless after a $100 trillion pandemic or human extinction; firms should therefore carry the burden of producing quantitative safety cases. Ball counters that the FDA embedded an industrial-era model of one treatment for one disease, impeding personalized medicine—a warning that an AI regulator could become a “cudgel” for unions, incumbents, and other groups seeking vetoes over job displacement.

The China argument splits into a race for controllable capability and a race to release something nobody controls. Tegmark calls the second a “suicide race” and expects the Chinese Communist Party, which prizes political control, to stop any domestic system capable of overthrowing it. Ball’s darker scenario is domestic: licensing could produce a medieval-style rentier state, with a small AI-owning elite, a protected rent-seeking middle, and a low-agency underclass.

For investors, the policy boundary that matters is increasingly capability-specific rather than a simple choice between acceleration and stagnation. A safety-case regime could redirect frontier-lab spending—Tegmark estimated leading AI companies spend roughly 1% on safety while major pharmaceutical companies spend far more—without stopping medical, scientific, and autonomous-driving progress. The unresolved exposure is whether OpenAI, Anthropic, Meta, xAI, and Google can remain trusted to withhold a dangerous model, or whether uncertainty itself will trigger binding predeployment review.

🔗 Original source & video: Superintelligence: To Ban or Not to Ban? Max Tegmark & Dean Ball join Liron Shapira on Doom Debates

Google is treating orbital AI infrastructure as a long-horizon response to Earth’s land, permitting, grid, and community constraints. Project Suncatcher would put TPU clusters and solar arrays in a dawn-dusk low-Earth orbit, where nearly constant sunlight could make panels up to eight times as productive; returning data may add only “a couple more milliseconds.” It is a real option on explosive compute demand, but not a present business: launches still cost many times an equivalent terrestrial center, and repairs may require robots.

The first concrete milestone is a 2027 two-satellite prototype, not an orbital hyperscale build-out. Google says newer TPUs survived proton-beam radiation beyond what a five-year mission should bring, and it is partnering with Planet for the test; Starcloud, Axiom Space, and a vaguely described Chinese effort are also exploring the field, while Eric Schmidt and Jeff Bezos have signaled interest. Google frames Suncatcher beside Waymo and quantum computing—an “eight, 10, 12, 15 years” kind of commitment—making launch economics and operational reliability the gates.

Republican AI policy remains a spectrum of intuitions rather than a settled MAGA doctrine. Dean Ball described the administration as sharing broad intuitions that AI may be the most important opportunity in decades or ever, carries familiar and “more alien” risks, and will shape U.S. leadership; camps range from national-security officials focused on China competition and child-safety conservatives to the contrasting positions the hosts associated with David Sacks and Steve Bannon. Ball expects backlash around a “weird vichyssoise” of slop, electricity, water, jobs, child safety, extinction, and claims that AI is fake.

The labs’ strongest competitive moat may be infrastructure, while their regulatory positions track their incentives. Hyperscalers do not necessarily oppose export controls because Chinese buyers—and even indirect demand for TSMC fabrication capacity—compete with them; frontier labs need chips to sell tokens, but Ball argued that “the parameters of the model are not your moat.” Anthropic’s announced $50 billion data-center commitment joins Google, OpenAI’s Stargate, Meta, and xAI in a capex race whose equilibrium government must arbitrate.

Ball draws a sharp regulatory line: prevent catastrophic tail risk federally, regulate ordinary harms reactively, and let procurement rules govern only government-bought models. He defended the “Woke AI” order as anti-ideological-bias terms and system-prompt disclosure for federal versions, while acknowledging the right’s jawboning tension and calling compelled changes to public model training “unambiguously unconstitutional.” He wants national standards for billion-dollar, globally served models, yet says Congress must act on child safety and that California’s SB 53 transparency rules are “rather reasonable overall.”

An unidentified Gemini test model appears to have pushed handwritten transcription from impressive to professionally usable. On five benchmark documents totaling about 1,000 words, Mark Humphries measured roughly a 1% word-error rate—about a 50% decline against Gemini 2.5 Pro on those tests and comparable to what human transcription experts offer. The result came through Google AI Studio A/B tests, sometimes after 20 or 30 attempts, so Kevin Roose’s “probably Gemini 3” remains a bet, not an identification.

The deeper signal was not transcription but a ledger inference that looked like symbolic reasoning. From an Albany entry, the model inferred that “14 5” meant 14 pounds 5 ounces of sugar sold at 1 shilling 4 pence per pound, reconciling it to 19 shillings 1 penny; Humphries said “models shouldn’t be able to do that.” If replicated, the capability could let models aggregate a ledger and take on broader archive-scale tasks—but the sample was tiny, the model unknown, and release-time replication still required.

🔗 Original source & video: Data Centers in Space + A.I. Policy on the Right + A Gemini History Mystery