A top lab's CEO says expert-level AI is over a decade away

Who can resolve this, and what they currently believe

Google DeepMind has no qualifying chief executive right now. On 5 Aug 2026 Demis Hassabis gave up the CEO title and became Chair of Google DeepMind and Chief Scientist of Alphabet. Koray Kavukcuoglu now leads the lab as SVP and reports to Sundar Pichai. No new CEO was named. Sources: Google blog, "The next chapter of our AI momentum" (https://blog.google/company-news/inside-google/message-ceo/next-chapter-ai-momentum/); Fortune, 5 Aug 2026; CNBC, 12 Aug 2026; SynBioBeta ("Google did not appoint a new CEO for DeepMind").

I read the definition ("CEO, acting CEO, or co-CEO") strictly. Hassabis is a former CEO, Kavukcuoglu is an "other executive," and Pichai is excluded. That removes the person most likely to qualify: Hassabis often said AGI "may need one or two more breakthroughs." Even he now writes that AGI is "close at hand." I keep a small allowance for a new Google DeepMind CEO, or for a resolver who treats Kavukcuoglu as the chief executive. He is an LLM-focused executive, so either path is unlikely to produce a qualifying statement.

Sam Altman (OpenAI) and Dario Amodei (Anthropic) are both still CEO, and both are at peak bullishness.

• Altman said "we are now, like, in the singularity" on the Relentless podcast (25 Jul 2026; Business Insider, 27 Jul). He told the UN Security Council that "we could lose control of the future to AI" (Newsweek, Sept 2026).
• Amodei's essay "We Must Pace the Frontier" (12 Sep 2026, darioamodei.com) says recursive self-improvement is under way. He estimates that slowing down would buy only "an extra year or two." On CBS he said the exponential curve is "getting steeper."
• OpenAI says its AI solved the Navier–Stokes Millennium Prize problem (openai.com, Sept 2026) and hundreds of other open maths problems (Economist, 7 Oct 2026).

I found no qualifying statement since the question opened on 16 Sep. Altman's Politico interview (4 Oct) and Amodei's CBS and UN remarks all point the other way.

Reference class and base rate

I looked at roughly 2022–2026 for these three CEOs, about 12 CEO-years. Under this question's strict wording I find no clear qualifying statement.

• Near misses: Hassabis's hedged "may need breakthroughs," which is excluded. Altman's Nov 2023 Cambridge Union remark, "we need another breakthrough" (the-decoder, 17 Nov 2023), applied only to language models and was about "superintelligence" defined as discovering new physics. Claims limited to one architecture are excluded.
• Base rate: well under one qualifying event per 10 CEO-years. It is lower now, because the two remaining CEOs have staked their companies (and OpenAI's planned public listing) on short timelines.

Paths to YES, and their rough weights

1. A lasting stall in progress (about 15–25% likely). A CEO would then have to say flatly that a breakthrough is necessary, or that the threshold is more than ten years off, while describing the expert-level, most-fields threshold. Even then I put it under 10%.
2. A catastrophe followed by a durable global ban or moratorium. A CEO might then forecast that nobody attains the threshold for more than ten years. About 0.5–1%.
3. New CEOs. Turnover at OpenAI or Anthropic, or a newly appointed Google DeepMind CEO, could bring a sceptic. Insiders are almost all believers, so this adds under 1%.
4. A lenient reading of a casual remark such as "we'll need new ideas." About 0.5%.

If the threshold is reached, which these leaders expect within a few years, the question is effectively dead.

Hazard model (computed)

Quarterly hazards: 0.10% for Q4 2026, 0.13% per quarter in 2027, 0.18% in 2028, and 0.20% from 2029 to 2031. Cumulative probability is 0.6% by end-2027, 2.1% by end-2029 and 3.7% by end-2031. Hazards rise slightly later on because a stall, ban or turnover takes time to happen and then to change what a CEO says publicly.

The case for a higher forecast

Five years is long, and Altman gives many interviews. He did once say flatly "we need another breakthrough," and the "known methods" route counts even if the CEO expects that breakthrough soon. This is why I do not go below about 3–4% by 2031. But the strict tests still apply: the statement must describe the threshold, must not be hedged, and must rule out every known approach. Together with the current flood of progress, that keeps the estimate low.

Question. By the last day of each quarter, has the chief executive of OpenAI, Anthropic, or Google DeepMind said on the record that AI systems better than the best human experts across most cognitive work are more than ten years away or cannot be reached with known methods?

Definitions.

• Chief executive: the serving CEO, acting CEO, or co-CEO of one of the three named organizations when the statement is made, including a renamed organizational successor. Former CEOs, other executives, and the CEO of Google or Alphabet do not qualify unless they also hold the Google DeepMind chief-executive role.
• Capability threshold: AI that exceeds the best human experts in most fields of cognitive work, not merely average humans, most workers, or experts in one narrow field. The statement must describe this breadth and level, or explicitly refer to a definition that does. The labels “AGI,” “powerful AI,” and “superintelligence” alone are insufficient. A prediction about a stronger threshold, such as superiority in every conceivable task, does not establish a prediction about this threshold unless the speaker also addresses it.
• More than ten years away: an endorsed forecast or assessment that first attainment of the threshold is over ten years after the statement date. A central estimate or explicit median forecast beyond ten years counts; a forecast range counts only if its lower endpoint exceeds ten years. “Decades away” counts; “ten years,” “at least ten years,” “ten to twenty years,” and “could take more than ten years” do not. A calendar-year forecast counts only if the whole specified year is beyond the statement's tenth anniversary.
• Cannot be reached with known methods: an endorsed assessment that techniques known at the time, including their scaling and combination, are insufficient to attain the threshold and that a currently unknown method or scientific breakthrough is necessary. This does not require a forecast exceeding ten years. A claim limited to scaling pretraining, a particular model architecture, or one company's current approach does not count unless the speaker also rules out the remaining known approaches.
• On the record: publicly accessible words attributable directly to the executive, in a recording, transcript, signed writing, authenticated public post, official testimony or filing, or a named-source interview containing a direct quotation and sufficient context to establish its meaning.

Resolves YES if a qualifying source publishes a qualifying statement by any one of the three organizations' chief executives meeting either the time or known-methods condition. The statement must concern first technical attainment in the field generally, not merely that company's plans or the timing of widespread adoption. The executive's own assessment counts without company-wide endorsement. Later changes of view do not undo YES.

Resolves NO / Exclusions. No qualifying publication by the quarter-end; anonymous accounts, leaks of private remarks, paraphrases without supporting direct words, or statements merely quoting another person's view; hypotheticals, jokes, questions, and unendorsed scenarios; uncertainty about whether known methods suffice, “we do not yet know how,” or “we may need breakthroughs”; delays solely in commercialization or adoption. A statement that breakthroughs are necessary can qualify even if the executive expects those breakthroughs soon. Previously published remarks merely republished after opening do not count, unless the serving executive newly reaffirms them on the record.

Resolving source. OpenAI (openai.com), Anthropic (anthropic.com), Google DeepMind (deepmind.google), the qualifying executive's authenticated personal website or public account, government or court records of testimony or filings, and the original publisher or broadcaster of an attributed interview, speech, or discussion. This question's explicit on-the-record standard permits executive interviews and authenticated public posts notwithstanding the narrower Public confirmation convention A2; A2's 14-day dispute rule is not required. The publication must preserve the executive's direct words and relevant context. Apply the Resolution window and Resolution date convention A3 to the first qualifying public release, not a later article quoting it.