# Verified predictions

**Snapshot: 2026-10-25.** Generated from the typed dataset that lived in `apps/web/src/data`,
which is no longer in the repository. This file and
[`ledger.json`](ledger.json) are now the canonical record: the JSON is the exact payload the site
and API served, and this file is its readable rendering. Do not edit either by hand — a rebuild
should regenerate both from source.

56 cited predictions across 25 people, 127 citations.

> This file is the **cited** record only. The dataset also carries 184 unrated editorial summaries
> (209 people in total, 240 records, 35 of them in government); they have no quotations, dates or
> citations and are excluded from this catalogue and from every count below. See
> [`methodology.md`](methodology.md#the-editorial-layer).
Verdicts are frozen at the snapshot date and graded under
[the methodology](methodology.md).

| Verdict | Count |
| --- | --- |
| Validated | 2 |
| Partly borne out | 4 |
| Contested | 12 |
| Missed | 2 |
| Open | 36 |
| Unfalsifiable | 0 |

---

## Sam Altman

*CEO and cofounder, OpenAI.* Sets OpenAI strategy, infrastructure partnerships and the public case for deployment. Publishes personal essays that move the whole industry’s timeline.

### It is possible that we will have superintelligence in a few thousand days (!); it may ta…

- **Claim (restated):** Superintelligence — AI vastly smarter than humans — may arrive within roughly 1,000 to 4,000 days of September 2024, i.e. between 2027 and 2035.
- **Said:** Essay, “The Intelligence Age”, 2024-09-23
- **Horizon:** 2027 – 2035
- **Domain:** models · **Verdict:** open
- **Assessment:** The window runs to 2035, and Altman explicitly allowed that it “may take longer”, which makes the claim hard to falsify early. No system as of the snapshot is credibly described as superintelligent by the definition OpenAI has used (“vastly smarter than humans”). Tracked, unscored.
- **Sources:**
  - PRIMARY — Sam Altman, [The Intelligence Age](https://ia.samaltman.com/), 2024-09-23. The exact published sentence and its date.
  - REPORTING — Fortune, [Sam Altman says AI superintelligence could be just ‘a few thousand days’ away](https://fortune.com/2024/09/24/sam-altman-ai-superintelligence/), 2024-09-24. Contemporary coverage and the 2,000–5,000 day arithmetic.
  - REPORTING — Ars Technica, [OpenAI CEO: We may have AI superintelligence in “a few thousand days”](https://arstechnica.com/information-technology/2024/09/ai-superintelligence-looms-in-sam-altmans-new-essay-on-the-intelligence-age/), 2024-09-24. Independent reading of the vagueness of the unit.

### We believe that, in 2025, we may see the first AI agents “join the workforce” and materi…

- **Claim (restated):** By the end of 2025, AI agents would be doing real work inside companies and would measurably change company output.
- **Said:** Blog post, “Reflections”, 2025-01-05
- **Horizon:** By 31 December 2025
- **Domain:** work · **Verdict:** contested
- **Assessment:** Agents were widely deployed in 2025 — coding agents, browser agents, enterprise “coworker” products — so the first clause held. The second, “materially change the output of companies”, is where the evidence splits: year-end reviews found little aggregate productivity signature, while vendors reported strong internal gains. Scored contested rather than validated.
- **Sources:**
  - PRIMARY — Sam Altman, [Reflections](https://blog.samaltman.com/reflections), 2025-01-05. The verbatim claim and its hedge (“may see”).
  - ANALYSIS — The New Yorker, [Why A.I. Didn’t Transform Our Lives in 2025](https://www.newyorker.com/culture/2025-in-review/why-ai-didnt-transform-our-lives-in-2025), 2025-12-27. The sceptical year-in-review reading of the same prediction.
  - ANALYSIS — Cal Newport, [Why Didn’t AI “Join the Workforce” in 2025?](https://calnewport.com/why-didnt-ai-join-the-workforce-in-2025/), 2026-01-06. A one-year retrospective on the claim, with the original text quoted.

### AGI will probably get developed during [Trump’s] term.

- **Claim (restated):** OpenAI would develop systems meeting its own AGI definition before the end of the current US presidential term (20 January 2029).
- **Said:** Bloomberg interview, quoted by TIME, 2025-01-21
- **Horizon:** By 20 January 2029
- **Domain:** models · **Verdict:** open
- **Assessment:** Depends entirely on which definition is used; Altman himself called AGI “a very sloppy term”. No independent evaluation has certified an AGI-class system as of the snapshot.
- **Sources:**
  - REPORTING — TIME, [How OpenAI’s Sam Altman Is Thinking About AGI and Superintelligence in 2025](https://time.com/7205596/sam-altman-superintelligence-agi/), 2025-01-21. The quotation and its context, plus Altman’s own caveat about the term.

### Five years is a long time. Next year, I hope we’ll start to see these models really make…

- **Claim (restated):** By around 2030, AI systems would be contributing to actual cures for diseases, not only to research acceleration.
- **Said:** Interview, December 2025 (reported), 2025-12-09
- **Horizon:** By 2030
- **Domain:** science · **Verdict:** open
- **Assessment:** Explicitly framed as a hope (“I hope”), which weakens it as a forecast. No AI-discovered therapy has completed phase III approval as of the snapshot; AI-assisted candidates are in trials.
- **Sources:**
  - REPORTING — Yahoo Finance, [Even the man behind ChatGPT is worried about the rate of change](https://www.yahoo.com/news/articles/even-man-behind-chatgpt-openai-162629071.html), 2025-12-09. The quotation and the five-year framing.
  - REPORTING — inkl, [OpenAI’s Sam Altman Says AI Could Cure Diseases — But We’re Not Ready for Change](https://www.inkl.com/news/openai-s-sam-altman-says-ai-could-cure-diseases-but-we-re-not-ready-for-change), 2025-12-10. A second outlet carrying the same passage.

### I would hope we have the candidate for that in the next couple of years for the first li…

- **Claim (restated):** Within roughly two years (by late 2028) AI-assisted research will produce a candidate treatment for a major disease — a discovery, not an approved drug.
- **Said:** Interview with Matthew Berman, 2026-10-21
- **Horizon:** By October 2028
- **Domain:** science · **Verdict:** open
- **Assessment:** Newly made, and Altman separates the discovery from regulatory approval. The claim is checkable against trial registries and peer-reviewed candidates when the window closes.
- **Sources:**
  - REPORTING — Firstpost, [Sam Altman says AI could help discover a cure for a major disease within two years](https://www.firstpost.com/tech/sam-altman-says-ai-could-help-discover-a-cure-for-a-major-disease-within-two-years-14054486.html/amp), 2026-10-22. The quotation and the discovery-versus-approval distinction.
  - REPORTING — MindStudio, [Sam Altman on AI Discoveries, Compute Demand, and Personal Assistants](https://www.mindstudio.ai/blog/sam-altman-interview-ai-discoveries-compute), 2026-10-22. Independent summary noting the FDA caveat.

---

## Dario Amodei

*Cofounder and CEO, Anthropic.* Writes unusually explicit timelines for “powerful AI” and unusually explicit warnings about what it does to entry-level work.

### I think it could come as early as 2026, though there are also ways it could take much lo…

- **Claim (restated):** “Powerful AI” — a system better than a Nobel laureate in most fields, able to run for days autonomously — could arrive as early as 2026.
- **Said:** Essay, “Machines of Loving Grace”, 2024-10-11
- **Horizon:** 2026 – 2027
- **Domain:** models · **Verdict:** open
- **Assessment:** Anthropic’s own later policy submission narrowed this to “late 2026 or early 2027”. As of the snapshot no publicly evaluated system is accepted as beating Nobel laureates across most disciplines; the window is still open. This is the single most consequential dated claim in the catalogue.
- **Sources:**
  - PRIMARY — Dario Amodei, [Machines of Loving Grace](https://darioamodei.com/essay/machines-of-loving-grace), 2024-10-11. The definition of “powerful AI” and the 2026 hedge.
  - PRIMARY — Dwarkesh Patel, [Dario Amodei — “We are near the end of the exponential”](https://www.dwarkesh.com/p/dario-amodei-2), 2026-10-01. Amodei restates the hunch in 2026: “2026, maybe 2027… I wouldn’t be surprised if I’m off by a year or two.”
  - ANALYSIS — Redwood Research, [What’s up with Anthropic predicting AGI by early 2027?](https://blog.redwoodresearch.org/p/whats-up-with-anthropic-predicting), 2025-12-25. Compares the essay’s hedge with Anthropic’s stronger OSTP submission.

### I think we’ll be there in three to six months—where AI is writing 90 percent of the code…

- **Claim (restated):** By September 2025 AI would write 90% of software code; by March 2026, essentially all of it.
- **Said:** Council on Foreign Relations, CEO Speaker Series, 2025-03-10
- **Horizon:** 90% by Sept 2025 · ~all by Mar 2026
- **Domain:** work · **Verdict:** contested
- **Assessment:** Inside Anthropic the number was reported as met; Amodei says so himself. Industry-wide measurement put AI-authored code nearer 25–40% in 2026, and Amodei later called 90% “actually a very weak criterion” and denied predicting the loss of 90% of engineers. The literal claim held in one company, not in the industry it implied.
- **Sources:**
  - PRIMARY — Council on Foreign Relations, [CEO Speaker Series With Dario Amodei of Anthropic](https://www.cfr.org/event/ceo-speaker-series-dario-amodei-anthropic), 2025-03-10. Full transcript and exact wording.
  - REPORTING — Business Insider, [Anthropic’s CEO says that in 3 to 6 months, AI will be writing 90% of the code](https://www.businessinsider.com/anthropic-ceo-ai-90-percent-code-3-to-6-months-2025-3), 2025-03-14. Contemporaneous coverage of the claim.
  - ANALYSIS — Forbes, [Dario Amodei Doubled Down On His AI Jobs Warning. Here’s What’s Different Now](https://www.forbes.com/sites/kolawolesamueladebayo/2026/02/21/dario-amodei-doubled-down-on-his-ai-jobs-warning-heres-whats-different-now/), 2026-02-21. Inside-Anthropic success versus 25–40% industry-wide.
  - PRIMARY — Dwarkesh Patel, [Dario Amodei — “We are near the end of the exponential”](https://www.dwarkesh.com/p/dario-amodei-2), 2026-10-01. Amodei’s own retrospective: “that’s actually a very weak criterion.”

### I predicted that AI could displace half of all entry-level white collar jobs in the next…

- **Claim (restated):** Between 2025 and 2030, AI will eliminate about 50% of entry-level white-collar jobs and could push unemployment to 10–20%.
- **Said:** Axios interview; restated in “The Adolescence of Technology”, 2025-05-28
- **Horizon:** 2025 – 2030
- **Domain:** work · **Verdict:** open
- **Assessment:** Too early to score, and the early evidence cuts against the magnitude: Anthropic’s own Economic Index found usage shifting back toward augmentation (52% of conversations) rather than automation in late 2025. Amodei has said he was warned the estimate had no research basis behind it.
- **Sources:**
  - PRIMARY — Dario Amodei, [The Adolescence of Technology](https://darioamodei.com/essay/the-adolescence-of-technology), 2026-01-26. Amodei’s own restatement of the 1–5 year window.
  - REPORTING — Axios, [Behind the Curtain: A white-collar bloodbath](https://www.axios.com/2025/05/28/ai-jobs-white-collar-unemployment-anthropic), 2025-05-28. The original interview and the 10–20% unemployment figure.
  - ANALYSIS — CNN, [The ‘white-collar bloodbath’ is all part of the AI hype machine](https://www.cnn.com/2025/05/30/business/anthropic-amodei-ai-jobs-nightcap), 2025-05-30. Notes that no research or evidence was cited for the 50% estimate.
  - ANALYSIS — Startup Fortune, [Dario Amodei’s job-loss warnings have not softened, but his own company’s data is quietly telling a different story](https://startupfortune.com/dario-amodeis-job-loss-warnings-have-not-softened-but-his-own-companys-data-is-quietly-telling-a-different-story/), 2026-05-06. Anthropic Economic Index augmentation figures.

---

## Demis Hassabis

*Chair, Google DeepMind; Alphabet chief scientist, Google DeepMind.* Nobel laureate, Isomorphic Labs lead. Has publicly shortened his AGI window at least four times since 2024.

### I think over the next five to 10 years, a lot of those capabilities will start coming to…

- **Claim (restated):** AGI begins to emerge between 2030 and 2035.
- **Said:** Briefing at Google DeepMind’s London offices, 2025-03-17
- **Horizon:** 2030 – 2035
- **Domain:** models · **Verdict:** open
- **Assessment:** Hassabis has since compressed his own window to 2–3 years (October 2026), so this record is best read as one point on a fast-moving curve.
- **Sources:**
  - REPORTING — CNBC, [AI that can match humans at any task will be here in five to 10 years, Google DeepMind CEO says](https://www.cnbc.com/2025/03/17/human-level-ai-will-be-here-in-5-to-10-years-deepmind-ceo-says.html), 2025-03-17. The quotation and his definition of AGI.
  - REPORTING — WIRED, [Demis Hassabis thinks AI will make humans less selfish](https://www.wired.com/story/google-deepminds-ceo-demis-hassabis-thinks-ai-will-make-humans-less-selfish/), 2025-06-04. The 50 percent framing of the same window.

### In the next five to 10 years, there’s maybe a 50 percent chance that we’ll have what we …

- **Claim (restated):** A 50% probability of true AGI by roughly 2030–2035.
- **Said:** Interview with WIRED; repeated on the Lex Fridman podcast, 2025-06-04
- **Horizon:** 50% by 2030 – 2035
- **Domain:** models · **Verdict:** open
- **Stated probability:** 0.5
- **Assessment:** A rare explicit probability from a lab leader. Hassabis’s stated definition requires matching all cognitive functions of the human brain, tested across tens of thousands of tasks — a much higher bar than benchmark leadership.
- **Sources:**
  - REPORTING — WIRED, [Demis Hassabis Embraces the Future of Work in the Age of AI](https://www.wired.com/story/google-deepminds-ceo-demis-hassabis-thinks-ai-will-make-humans-less-selfish/), 2025-06-04. The 50 percent, five-to-10-year formulation.
  - ANALYSIS — X / Mario Nawfal, [Demis Hassabis: 50% chance of true AGI by 2030](https://x.com/MarioNawfal/status/1948923727221325841), 2025-07-26. The 2030 date and the “tens of thousands of tasks” verification standard — secondary, flagged as such.

### It’s soon. I’m updating nearer and nearer.

- **Claim (restated):** AGI is now 2–3 years away — i.e. 2028 to 2029.
- **Said:** WSJ Magazine profile, 2026-10-21
- **Horizon:** 2028 – 2029
- **Domain:** models · **Verdict:** open
- **Assessment:** His fourth public compression in eighteen months: 5–10 years (Mar 2025), 5–8 (Feb 2026), 3–5 (early 2026), now 2–3. The direction of travel is itself the datapoint.
- **Sources:**
  - REPORTING — Startup Fortune, [Demis Hassabis now says AGI could arrive in just two to three years](https://startupfortune.com/demis-hassabis-now-says-agi-could-arrive-in-just-two-to-three-years/), 2026-10-21. The WSJ Magazine quote and the previous windows.
  - REPORTING — Tech Times, [Hassabis Cuts AGI Window to Two to Three Years While Safety Isn’t Ready](https://www.techtimes.com/articles/329362/20261022/hassabis-cuts-agi-window-two-three-years-while-safety-isnt-ready.htm), 2026-10-22. Independent account of the compression and the Davos baseline.

---

## Sundar Pichai

*CEO, Google and Alphabet.* Allocates capital across Gemini, search, cloud and TPU capacity — the widest single distribution surface for AI on earth.

### Will the AI think it has reached AGI by 2030? I would say we will just fall short of tha…

- **Claim (restated):** AGI will not be achieved by 2030; it will arrive somewhat after.
- **Said:** Lex Fridman Podcast #471, 2025-06-05
- **Horizon:** 2030, missed by design
- **Domain:** models · **Verdict:** open
- **Assessment:** An inverted prediction: it is validated if 2030 arrives without an AGI claim from Google. Note that the criterion is “the AI think it has reached AGI”, which is itself unusual and hard to operationalise.
- **Sources:**
  - ANALYSIS — AI Prediction Index, [Pichai: will the AI think it has reached AGI by 2030? “We will just fall short”](https://www.aipredictionindex.com/p/pichai-2025-agi-just-after-2030/), 2026-09-28. Transcript excerpt with timestamp and deadline tracking.
  - REPORTING — NDTV, [Google DeepMind CEO Predicts AGI In Three To Four Years](https://www.ndtv.com/artificial-intelligence/google-deepmind-ceo-predicts-artificial-general-intelligence-in-three-to-four-years-amid-industry-debate-11559496), 2026-05-28. Pichai’s framing that AGI is coming “sooner rather than later”.

### A coming wave of artificial intelligence tools may soon make decisions on behalf of user…

- **Claim (restated):** Within roughly a year, AI agents will routinely act on a user’s behalf for high-stakes decisions.
- **Said:** BBC interview, as characterised by Yahoo Finance (reported speech, not a transcript), 2025-12-07
- **Horizon:** By December 2026
- **Domain:** models · **Verdict:** open
- **Assessment:** Agentic purchasing and booking shipped widely in 2026; high-stakes medical or investment delegation remains restricted by liability and regulation. Partly checkable already.
- **Sources:**
  - REPORTING — Yahoo Finance, [Google CEO Sundar Pichai Predicts AI Will Act For You](https://finance.yahoo.com/news/google-ceo-sundar-pichai-predicts-193114740.html), 2025-12-07. The outgoing reporter’s characterisation of his BBC remarks and the one-year window; treated as reported speech, not a verbatim transcript.

---

## Elon Musk

*xAI founder; SpaceX CEO, xAI / Grok (within SpaceX).* The most prolific dated forecaster in the industry, and therefore the most testable. His dates move; the definitions do not always move with them.

### My guess is that we’ll have AI that is smarter than any one human probably around the en…

- **Claim (restated):** By the end of 2025, AI would be smarter than any single human being.
- **Said:** Interview with Nicolai Tangen, “In Good Company”, 2024-04-08
- **Horizon:** End of 2025
- **Domain:** models · **Verdict:** missed
- **Assessment:** 2025 passed without a system generally accepted as smarter than any human across domains. The date then moved to end-2026 at Davos, while the definition stayed the same — a textbook moving-goalpost pattern. This record exists to keep the earlier date visible.
- **Sources:**
  - REPORTING — Business Insider, [Elon Musk Says AI Will Outsmart Humans by 2026](https://www.businessinsider.com/elon-musk-says-ai-will-outsmart-humans-by-2026-2024-4), 2024-04-09. The 2024 quote, made when “next year” meant 2025.
  - ANALYSIS — Fello AI, [When Will AGI Happen? Every Prediction, Tracked](https://felloai.com/when-will-agi-happen/), 2026-10-06. Documents the same quote with the 2025 date, and the 2026 revision.

### Within three years, robots will surpass surgeons.

- **Claim (restated):** By 2029, surgical robots outperform human surgeons in general practice.
- **Said:** Podcast interview recorded December 2025, 2026-01-07
- **Horizon:** By 2029
- **Domain:** robotics · **Verdict:** open
- **Assessment:** Regulatory approval, not capability, is likely to be the binding constraint: no surgical robot has been cleared for unsupervised general surgery, and the claim cuts across the entire evidence base for autonomy in high-consequence physical tasks.
- **Sources:**
  - REPORTING — 36Kr, [In a 173-minute conversation, Musk only talked about three things: AI, energy, and robots](https://eu.36kr.com/en/p/3630015413138182), 2026-01-07. The three-part timeline as rendered in English by 36Kr: AGI 2026, robots surpassing surgeons within three years, collective ASI by 2030. A translated rendering, not a transcript.
  - ANALYSIS — Gary Marcus, [Elon Musk is on a prediction rampage](https://garymarcus.substack.com/p/elon-musk-is-on-a-prediction-rampage), 2026-09-03. A dated critical reading of the pattern.

### I don’t know what’s going to happen in 10 years, but the rate at which AI is progressing…

- **Claim (restated):** AGI — AI smarter than any human — arrives by the end of 2026, “no later than next year”.
- **Said:** World Economic Forum, Davos, with Larry Fink, 2026-01-22
- **Horizon:** By end of 2026
- **Domain:** models · **Verdict:** open
- **Assessment:** Sixty-eight days from the snapshot. No benchmark suite or evaluation body has certified a system as smarter than any human across domains; “smarter” remains undefined in the claim.
- **Sources:**
  - REPORTING — NDTV, [Will Have AI Smarter Than Any Human By Year End: Elon Musk’s Bombshell At Davos](https://www.ndtv.com/world-news/will-have-ai-smarter-than-any-human-by-year-end-elon-musks-bombshell-at-davos-10840212), 2026-01-22. The Davos quotation and the follow-on five-year claim.
  - REPORTING — The Economic Times, [AI could be smarter than all of humanity in five years, says Elon Musk](https://economictimes.indiatimes.com/news/international/global-trends/davos-2026-elon-musk-ai-smarter-than-humanity-humanoid-robots/articleshow/127173778.cms), 2026-01-23. Independent transcription of the same panel.

### Five years from now, AI will be smarter than all of humanity collectively.

- **Claim (restated):** By 2031, combined AI intelligence exceeds the collective intelligence of all humans.
- **Said:** World Economic Forum, Davos, 2026-01-22
- **Horizon:** By 2031
- **Domain:** models · **Verdict:** open (not falsifiable as stated)
- **Assessment:** “Collective intelligence of humanity” has no agreed operational measure, so even a correct-feeling outcome could not be adjudicated. Recorded as a rhetorical horizon, not a testable one.
- **Sources:**
  - REPORTING — NDTV, [Will Have AI Smarter Than Any Human By Year End](https://www.ndtv.com/world-news/will-have-ai-smarter-than-any-human-by-year-end-elon-musks-bombshell-at-davos-10840212), 2026-01-22. The five-year claim as delivered.

---

## Mark Zuckerberg

*Founder, chairman and CEO, Meta.* Bets on compute scale plus distribution through three billion-person consumer surfaces, with Meta Superintelligence Labs building the models.

### Probably in 2025, we at Meta, as well as the other companies that are basically working …

- **Claim (restated):** In 2025 Meta would have an AI equivalent to a mid-level software engineer.
- **Said:** The Joe Rogan Experience, 2025-01-10
- **Horizon:** By end of 2025
- **Domain:** work · **Verdict:** contested
- **Assessment:** Coding agents became genuinely useful in 2025, and Meta restructured around them, but no verifiable internal audit showed an agent operating at the level of a mid-level engineer end to end. Meta later described most code heading toward AI authorship rather than most engineering being replaced.
- **Sources:**
  - REPORTING — Business Insider, [Mark Zuckerberg Says AI Could Soon Do the Work of Midlevel Meta Coders](https://www.businessinsider.com/mark-zuckerberg-meta-ai-replace-engineers-coders-joe-rogan-podcast-2025-1), 2025-01-11. The quotation and its immediate context.
  - REPORTING — Engadget, [Zuckerberg predicts AI will write most of Meta’s code within 12 to 18 months](https://www.engadget.com/ai/mark-zuckerberg-predicts-ai-will-write-most-of-metas-code-within-12-to-18-months-213851646.html), 2025-04-30. The follow-on, narrower version of the same prediction.

### I would guess sometime in the next 12 to 18 months, we’ll reach the point where most of …

- **Claim (restated):** By roughly October 2026, most code written for Meta’s AI efforts is authored by AI agents.
- **Said:** Dwarkesh Patel podcast interview, 2025-04-29
- **Horizon:** By October 2026
- **Domain:** work · **Verdict:** open
- **Assessment:** The deadline falls four days after this snapshot. No independent audit of Meta’s internal code authorship has been published; the company reports high AI-authored shares without releasing methodology. Adjudication requires that disclosure.
- **Sources:**
  - REPORTING — Engadget, [Mark Zuckerberg predicts AI will write most of Meta’s code within 12 to 18 months](https://www.engadget.com/ai/mark-zuckerberg-predicts-ai-will-write-most-of-metas-code-within-12-to-18-months-213851646.html), 2025-04-30. Verbatim quotation including the “not autocomplete” qualifier.
  - ANALYSIS — Business Insider, [The New CEO Flex: Bragging About How Much AI Code Your Company Shipped](https://www.businessinsider.com/latest-ceo-flex-how-much-ai-code-your-company-shipped-2026-5), 2026-05-09. The wider pattern of unverifiable internal AI-code claims.

---

## Mustafa Suleyman

*CEO, Microsoft AI, Microsoft AI.* Pursues “humanist superintelligence”: capable, deliberately bounded, and explicitly subordinate. Also the most concrete forecaster of AI companionship.

### I know that it’s going to work in the next six months or 12 months or, maybe worst case,…

- **Claim (restated):** AI systems are already superhuman in the tasks that matter, with agentic reliability arriving within 6–18 months.
- **Said:** Bloomberg Weekend interview, 2025-12-12
- **Horizon:** By mid-2027
- **Domain:** models · **Verdict:** contested
- **Assessment:** “Superhuman” without a task specification is unfalsifiable; the 6–18 month reliability clause is not. Agents in 2026 still required supervision on long, ambiguous tasks, which is why Suleyman’s own framing emphasises bounded, domain-specific systems.
- **Sources:**
  - PRIMARY — Bloomberg, [Microsoft’s Mustafa Suleyman: ‘AI Is Already Superhuman’](https://www.bloomberg.com/features/2025-mustafa-suleyman-weekend-interview/), 2025-12-12. The interview text and the humanist-superintelligence framing.
  - REPORTING — The Verge, [Microsoft AI CEO says AI threats are real, and Anthropic is making it worse](https://www.theverge.com/podcast/996412/microsoft-ai-ceo-mustafa-suleyman-regulation-safety-anthropic-claude), 2026-09-17. His later definition of superintelligence as “smarter and more capable than all humans combined”.

### In five years’ time, everybody will have their own AI companion who knows them so intima…

- **Claim (restated):** By 2031, every person has a persistent, personalised AI companion.
- **Said:** Microsoft AI blog / interview, reported January 2026, 2026-01-18
- **Horizon:** By 2031
- **Domain:** models · **Verdict:** open (not falsifiable as stated)
- **Assessment:** “Everybody” is not measurable, and companion adoption is uneven across markets and income levels. The companion category is already real: the dispute is over universality, not existence.
- **Sources:**
  - REPORTING — The Indian Express, [In 5 years, everybody will have their own AI companion, says Microsoft AI’s Mustafa Suleyman](https://indianexpress.com/article/technology/artificial-intelligence/mustafa-suleyman-microsoft-ai-companion-prediction-10477197/), 2026-01-18. The quotation and its five-year horizon.
  - REPORTING — IBTimes UK, [Microsoft AI CEO Predicts ‘Personal AI Companions’ For Everyone Within Five Years](https://www.ibtimes.co.uk/microsoft-ai-ceo-mustafa-suleyman-predicts-personal-ai-companions-everyone-within-five-years-1772888), 2026-01-22. The “humanist superintelligence” framing that bounds the claim.

---

## Ilya Sutskever

*Cofounder and CEO, Safe Superintelligence (SSI).* Transformer-era coauthor who now argues the age of pure scaling is over — the most consequential technical disagreement inside the field.

### Is the belief really: “Oh, it’s so big, but if you had 100x more, everything would be so…

- **Claim (restated):** From late 2025, additional compute alone stops producing transformative capability gains; progress depends on new research paradigms, not scale.
- **Said:** Dwarkesh Podcast, 2025-11-25
- **Horizon:** Open-ended, from 2025
- **Domain:** models · **Verdict:** contested
- **Assessment:** Labs continue to report capability gains from scaled training, which is evidence against the strong reading. Sutskever himself clarified that scaling “won’t stall” but that something important will remain missing — a narrower claim that is harder to test.
- **Sources:**
  - PRIMARY — Dwarkesh Patel, [Ilya Sutskever — We’re moving from the age of scaling to the age of research](https://www.dwarkesh.com/p/ilya-sutskever-2), 2025-11-25. Full transcript of the claim.
  - REPORTING — Business Insider, [OpenAI cofounder says scaling compute is not enough to advance AI](https://www.businessinsider.com/openai-cofounder-ilya-sutskever-scaling-ai-age-of-research-dwarkesh-2025-11), 2025-11-26. Independent summary of the argument.
  - PRIMARY — X, [“Scaling the current thing will keep leading to improvements… but something important will continue to be missing.”](https://x.com/ilyasut/status/1994424504370581726), 2025-11-25. His own clarification of the scope of the claim.

### We’ll have these sort of continual learning agents. AI will be very powerful. Maybe ther…

- **Claim (restated):** The next capability step is continual learning agents, not larger one-shot models — and progress toward them will be visible before superintelligence.
- **Said:** Dwarkesh Podcast, 2025-11-25
- **Horizon:** Open-ended
- **Domain:** models · **Verdict:** open
- **Assessment:** No evaluation standard yet exists for “continual learning” in deployed systems; labs began shipping persistent-memory agents in 2026, which is a partial signal in the predicted direction.
- **Sources:**
  - PRIMARY — Dwarkesh Patel, [Ilya Sutskever — from the age of scaling to the age of research](https://www.dwarkesh.com/p/ilya-sutskever-2), 2025-11-25. The continual-learning passage in context.

---

## Yann LeCun

*Cofounder and executive chair, AMI Labs.* Left Meta to build world models on the premise that large language models are a short-lived paradigm.

### I think the shelf life of the current [LLM] paradigm is fairly short, probably three to …

- **Claim (restated):** Between 2028 and 2030, large language models stop being the central component of AI systems, replaced by a new architecture class.
- **Said:** Bloomberg / TechCrunch interview at Davos, 2025-01-23
- **Horizon:** 2028 – 2030
- **Domain:** models · **Verdict:** open
- **Assessment:** The deadline is still ahead. What is checkable already is the sub-claim: “within five years, nobody in their right mind would use them anymore, at least not as the central component”. As of 2026, LLM-based systems remain the industry’s core, so the strong version is not tracking yet.
- **Sources:**
  - REPORTING — TechCrunch, [Meta’s Yann LeCun predicts a new AI architectures paradigm within 5 years and a decade of robotics](https://techcrunch.com/2025/01/23/metas-yann-lecun-predicts-a-new-ai-architectures-paradigm-within-5-years-and-decade-of-robotics/), 2025-01-23. The shelf-life quotation and the “decade of robotics”.
  - REPORTING — HPCwire, [Meta’s Chief AI Scientist Yann LeCun Questions the Longevity of Current GenAI and LLMs](https://www.hpcwire.com/2025/02/11/metas-chief-ai-scientist-yann-lecun-questions-the-longevity-of-current-genai-and-llms/), 2025-02-11. The stronger Davos version of the claim.
  - REPORTING — Newsweek, [Yann LeCun, Pioneer of AI, Thinks Today’s LLMs Are Nearly Obsolete](https://www.newsweek.com/nw-ai/ai-impact-interview-yann-lecun-artificial-intelligence-2054237), 2025-09-05. The JEPA-based alternative he is betting on.

### I think the coming years could be the “decade of robotics”.

- **Claim (restated):** The 2020s close with robotics, not chatbots, as the defining AI application.
- **Said:** Davos remarks, reported by TechCrunch, 2025-01-23
- **Horizon:** 2025 – 2030
- **Domain:** robotics · **Verdict:** open
- **Assessment:** Humanoid pilots moved from demos to paid factory deployments during 2026 — supporting the direction — but volumes remain in the hundreds to low thousands, far from a consumer or logistics fleet. Directional evidence, no adjudication yet.
- **Sources:**
  - REPORTING — TechCrunch, [Meta’s Yann LeCun predicts a new AI architectures paradigm within 5 years and a decade of robotics](https://techcrunch.com/2025/01/23/metas-yann-lecun-predicts-a-new-ai-architectures-paradigm-within-5-years-and-decade-of-robotics/), 2025-01-23. The phrase in context.
  - PRIMARY — Nebius, [Meet Yann LeCun’s Lab and the AI World of 2030](https://nebius.science/stories/meet-yann-lecuns-lab-and-the-ai-world-of-2030), 2026-07-13. His later restatement that three-to-five years is the meaningful horizon.

---

## Fei-Fei Li

*Cofounder and CEO, World Labs.* ImageNet’s architect now argues that spatial intelligence — not more text — is AI’s next frontier.

### Spatial intelligence will transform how we create and interact with real and virtual wor…

- **Claim (restated):** The next capability frontier is spatial intelligence via world models, not further scaling of language-only systems.
- **Said:** Essay, “From Words to Worlds”, 2025-11-10
- **Horizon:** Open-ended
- **Domain:** robotics · **Verdict:** contested
- **Assessment:** World models shipped as real products in 2026 (generative 3D environments with physics-ready meshes), which supports the technical premise. Whether spatial intelligence becomes “the” frontier — rather than one strong modality among several — is not yet decided, and language models continue to absorb multimodal capability.
- **Sources:**
  - PRIMARY — Fei-Fei Li, [From Words to Worlds: Spatial Intelligence is AI’s Next Frontier](https://drfeifei.substack.com/p/from-words-to-worlds-spatial-intelligence), 2025-11-10. The essay and the three required capabilities.
  - PRIMARY — TIME, [Spatial Intelligence Is AI’s Next Frontier](https://time.com/7339693/fei-fei-li-ai/), 2025-12-11. The commissioned version of the argument.
  - REPORTING — Forbes, [AI Pioneer Fei-Fei Li Ushers In Next Frontier Of Artificial Intelligence](https://www.forbes.com/sites/geekgirlrising/2025/11/20/fei-fei-li-ushers-in-ais-next-frontier-spatial-intelligence/), 2025-11-20. Independent framing of the commercial bet behind the claim.

---

## Yoshua Bengio

*Scientific director, LawZero; Université de Montréal professor, LawZero / Mila.* Co-chairs the UN Independent International Scientific Panel on AI and chairs the International AI Safety Report. Reads the capability curves rather than the press releases.

### But if it’s five years, we’re not ready … because we don’t have methods to make sure tha…

- **Claim (restated):** Within about five years, systems may be capable of harm we cannot prevent, because the safety methods do not exist yet.
- **Said:** CNBC interview, 2024-11-21
- **Horizon:** By 2029
- **Domain:** security · **Verdict:** open
- **Assessment:** Partly testable through the safety-tooling record rather than through an event: independent evaluators still report that no demonstrated method reliably controls systems substantially more capable than their overseers. The stronger claim — that capability outruns control — is a live research dispute.
- **Sources:**
  - REPORTING — CNBC, [AI systems could “turn against humans”: Yoshua Bengio warns of artificial intelligence risks](https://www.cnbc.com/2024/11/21/will-ai-replace-humans-yoshua-bengio-warns-of-artificial-intelligence-risks.html), 2024-11-21. The quotation and the five-year framing.
  - REPORTING — The Guardian, [“Godfather of AI” shortens odds of the technology wiping out humanity over next 30 years](https://www.theguardian.com/technology/2024/dec/27/godfather-of-ai-raises-odds-of-the-technology-wiping-out-humanity-over-next-30-years), 2024-12-27. Hinton’s parallel 10–20% estimate, for comparison in the risk view.

### Right now it’s like at the child level, they can do like half an hour ahead, they can pl…

- **Claim (restated):** If the measured planning-horizon trend continues, AI reaches human-level planning capability around 2031.
- **Said:** Interviews referencing METR time-horizon data, 2026-02-17
- **Horizon:** Around 2031
- **Domain:** models · **Verdict:** open
- **Assessment:** Conditional by construction (“if the curve continues”) and sourced from the same evaluation series this catalogue tracks separately. Note the tension: METR’s own updated model puts 50% horizons near twelve hours for public models in early 2026, several orders below “human level” on multi-week work.
- **Sources:**
  - PRIMARY — Steven Bartlett / singjupost, [The Next 5 Years Will Change Humanity Forever w/ Yoshua Bengio (Transcript)](https://singjupost.com/the-next-5-years-will-change-humanity-forever-w-yoshua-bengio-transcript/), 2026-02-17. The verbatim conditional statement.
  - REPORTING — BetaKit, [Yoshua Bengio warns of “catastrophic risks” of agentic AI at World Summit AI](https://betakit.com/yoshua-bengio-warns-of-catastrophic-risks-of-agentic-ai-at-world-summit-ai/), 2025-04-17. The earlier, stronger formulation: “In five years, it will be at human level for programming tasks.”
  - DATA — METR, [Time Horizon 1.1](https://metr.org/blog/2026-01-29-time-horizon-1-1/), 2026-01-29. The underlying doubling-time measurements the claim rests on.

---

## Geoffrey Hinton

*University Professor Emeritus, University of Toronto.* Nobel laureate in physics. Has moved his own timelines inward twice in public, and says so.

### Not really, 10% to 20%.

- **Claim (restated):** There is a 10–20% chance that AI leads to human extinction within roughly three decades.
- **Said:** BBC Radio 4 Today programme, 2024-12-27
- **Horizon:** Next 30 years
- **Domain:** security · **Verdict:** open (not falsifiable as stated)
- **Stated probability:** 0.15
- **Assessment:** Hinton describes the figure as a gut feeling rather than a statistical analysis, and there is no agreed method for adjudicating extinction probability. Later revisions moved the timeline inward — “maybe 10 years, maybe less” — while keeping a ~10% figure.
- **Sources:**
  - REPORTING — The Guardian, [“Godfather of AI” shortens odds of the technology wiping out humanity over next 30 years](https://www.theguardian.com/technology/2024/dec/27/godfather-of-ai-raises-odds-of-the-technology-wiping-out-humanity-over-next-30-years), 2024-12-27. The 10–20% figure and the exchange with Sajid Javid.
  - REPORTING — Business Insider, [The “Godfather of AI” says it’s not “unreasonable” to believe there’s a 10% chance the technology will end humanity](https://www.businessinsider.com/geoffrey-hinton-godfather-of-ai-human-extinction-odds-2026-9), 2026-09-10. BBC Newsnight restatement: “maybe 10 years, maybe less.”

### A reasonable bet is sometime between five and 20 years.

- **Claim (restated):** AI that surpasses human intelligence in every domain arrives between 2030 and 2045.
- **Said:** CNN interview, 2025-08-13
- **Horizon:** 2030 – 2045
- **Domain:** models · **Verdict:** open
- **Assessment:** A wide interval that is likely to be judged by expert consensus rather than by a single evaluation. Hinton’s earlier published range was 30–50 years, so the direction of his revisions matches the industry’s.
- **Sources:**
  - REPORTING — CNN, [The “godfather of AI” reveals the only way humanity can survive superintelligent AI](https://www.cnn.com/2025/08/13/tech/ai-geoffrey-hinton), 2025-08-13. The five-to-20-year estimate and the earlier 30–50 year range.
  - REPORTING — Business Insider, [“Godfather of AI” says he’s “glad” to be 77 because the tech probably won’t take over the world in his lifetime](https://www.businessinsider.com/ai-godfather-geoffrey-hinton-superintelligence-risk-takeover-2025-4), 2025-04-28. The intermediate revision to “10 years or less”.

---

## Beth Barnes

*Founder and CEO, METR.* Runs the evaluation organisation whose “time horizon” chart became the industry’s most cited forecast — and the most cited caveat about forecasts.

### The length of tasks (measured by how long they take human professionals) that generalist…

- **Claim (restated):** Agent capability, measured in task length, doubles roughly every seven months and stays on trend.
- **Said:** METR paper, “Measuring AI Ability to Complete Long Tasks”, 2025-03-19
- **Horizon:** Continuous trend
- **Domain:** models · **Verdict:** partial
- **Assessment:** The measurement is the strongest empirical forecast instrument in the field, and METR’s own later revision shortened the post-2023 doubling time to 130.8 days — faster than the original estimate. But the same data shows 80%-reliability horizons lagging far behind 50%, so the headline understates how human-dependent long tasks remain.
- **Sources:**
  - PRIMARY — METR, [Measuring AI Ability to Complete Long Software Tasks](https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/), 2025-03-19. The original series and the ~7 month doubling time.
  - DATA — METR, [Time Horizon 1.1](https://metr.org/blog/2026-01-29-time-horizon-1-1/), 2026-01-29. Post-2023 doubling time of 130.8 days and the wider task suite.
  - DATA — METR, [Frontier Risk Report (February to March 2026)](https://metr.org/blog/2026-05-19-frontier-risk-report/), 2026-05-19. Public frontier ~12h at 50% versus ~1.5h at 80% reliability.

### It seems hard to rule out even shorter [timelines]. Is there 1% chance of this happening…

- **Claim (restated):** Autonomous AI self-improvement becomes plausible within about two years, with a small probability of much sooner.
- **Said:** 80,000 Hours interview, 2026-06-05
- **Horizon:** By 2028
- **Domain:** models · **Verdict:** open
- **Assessment:** She is careful to attach a low probability to the near term. The observable proxy — whether AI systems meaningfully drive AI research — was already reported as partially true by METR in 2026, so the question is degree, not kind.
- **Sources:**
  - PRIMARY — 80,000 Hours, [Beth Barnes on the most important graph in AI right now](https://80000hours.org/podcast/episodes/beth-barnes-ai-safety-evals/), 2026-06-05. The quotation and her uncertainty framing.
  - REPORTING — TIME, [Beth Barnes: The 100 Most Influential People in AI 2026](https://time.com/collection/time100-ai/2026/beth-barnes/), 2026-08-27. The 2030 extrapolation to month-long tasks.

---

## Jan Leike

*Leads the Alignment Science team, Anthropic.* Co-led OpenAI’s Superalignment programme, then left it. The four-year commitment he signed is now a test case in institutional accountability.

### Our goal is to solve the core technical challenges of superintelligence alignment in fou…

- **Claim (restated):** The core technical challenges of aligning superintelligence would be solved by mid-2027, with 20% of OpenAI’s compute.
- **Said:** OpenAI, “Introducing Superalignment”; co-led with Ilya Sutskever, 2023-07-05
- **Horizon:** By July 2027
- **Domain:** security · **Verdict:** missed
- **Assessment:** The programme’s co-leads departed in 2024 and the team was dissolved; the commitment cannot now be delivered by the institution that made it, and no successor programme has claimed the deadline. This is scored as an institutional commitment that failed by attrition, not as a technical result — an important distinction for anyone reading lab forecasts as contracts.
- **Sources:**
  - PRIMARY — OpenAI, [Introducing Superalignment](https://openai.com/index/introducing-superalignment/), 2023-07-05. The four-year goal and the 20% compute commitment.
  - PRIMARY — AXRP podcast, [Superalignment with Jan Leike](https://axrp.net/episode/2023/07/27/episode-24-superalignment-with-jan-leike.html), 2023-07-27. Leike’s own mid-2027 framing and 85% confidence.
  - REPORTING — Business Insider, [It was their job to make sure humans are safe from OpenAI’s superintelligence. They just quit.](https://www.businessinsider.com/jan-leike-ilya-sutskever-resignations-superalignment-openai-superintelligence-safe-humanity-2024-5), 2024-05-15. The departures that ended the programme.

---

## Arvind Narayanan

*Professor; director, Center for Information Technology Policy, Princeton University.* Co-author of AI Snake Oil. Separates demonstrated capability from commercial narrative, and argues that x-risk numbers are feelings wearing lab coats.

### AI x-risk forecasts are far too unreliable to be useful for policy, and in fact highly m…

- **Claim (restated):** Quantified AI existential-risk probabilities are not a sound basis for policy.
- **Said:** AI Snake Oil essay with Sayash Kapoor, 2024-07-26
- **Horizon:** Standing methodological claim
- **Domain:** security · **Verdict:** contested
- **Assessment:** This is itself a claim about forecasting that can be checked by looking at forecast behaviour. The catalogue’s own evidence is mixed: stated probabilities differ by orders of magnitude between experts, which supports his point, while structured forecasting efforts have improved calibration on shorter-horizon questions.
- **Sources:**
  - PRIMARY — X, [On the deep dive into AI existential risk probability estimates](https://x.com/random_walker/status/1816816078577631331), 2024-07-26. The authors’ own summary of the finding.
  - ANALYSIS — MIT Sloan, [How to spot real value in AI — and avoid the snake oil](https://mitsloan.mit.edu/ideas-made-to-matter/how-to-spot-real-value-ai-and-avoid-snake-oil), 2026-01-29. The predictive-AI critique that grounds his forecasting scepticism.

### AI will come to feel more trustworthy—and it will remove the tells that let readers know…

- **Claim (restated):** Within a few years, AI systems become a default “truth oracle” for a meaningful share of the public, and AI slop becomes harder to detect.
- **Said:** Fortune interview on journalism and AI, 2026-10-06
- **Horizon:** Open-ended, near term
- **Domain:** policy · **Verdict:** open
- **Assessment:** A rare prediction from the catalogue’s chief sceptic, and notable because it predicts an epistemic harm rather than a capability. Already partially observable: a 2026 Pew survey found 10% of Americans using chatbots as an information source.
- **Sources:**
  - REPORTING — Fortune, [“AI Snake Oil” author sees chatbots evolving into a “truth oracle”](https://fortune.com/2026/10/06/arvind-narayanan-journalism-future-ai-slop-oracle-source-truth/), 2026-10-06. The magazine’s characterisation of his prediction, plus the Pew figure; flagged as reported speech rather than a direct quotation.

---

## Andrew Ng

*Founder, DeepLearning.AI; managing GP, AI Fund, DeepLearning.AI.* The industry’s most-read teacher. His forecasts are about adoption and engineering practice rather than raw capability.

### For the majority of businesses, focus on building applications using agentic workflows r…

- **Claim (restated):** The return on AI for most firms comes from agentic workflow engineering, not from bigger base models.
- **Said:** Snowflake Summit, Luminary Talk, 2024-06-03
- **Horizon:** Standing claim
- **Domain:** work · **Verdict:** partial
- **Assessment:** Enterprise adoption data through 2026 supports the direction: value concentrated in workflow integration and evaluation harnesses, while pilots without process redesign failed at high rates. The stronger reading — that model scaling would not matter to firms — has not held.
- **Sources:**
  - ANALYSIS — Global Advisors, [Quote: Andrew Ng on agentic workflows](https://globaladvisors.biz/2025/07/06/quote-andrew-ng-ai-guru-3/), 2025-07-06. The quote and its Snowflake Summit provenance.
  - PRIMARY — DeepLearning.AI, [The AI Engineering Skills Map](https://www.deeplearning.ai/the-batch/the-ai-engineering-skills-map), 2026-08-14. His 2026 operational view of where the work actually is.

### The practical utility of very long-horizon agent tasks, especially relative to their tok…

- **Claim (restated):** Long-horizon autonomous agents deliver less value per unit cost than iterative human-in-the-loop work through 2026–2027.
- **Said:** DeepLearning.AI essay series on coding agents, 2026-09-04
- **Horizon:** Through 2027
- **Domain:** work · **Verdict:** contested
- **Assessment:** Directly at odds with the “long horizon is the metric” view from METR and with vendors shipping multi-hour agents. This is a productive disagreement: cost per completed task is measurable, and both sides agree on the measurement, which makes it unusually likely to be settled.
- **Sources:**
  - ANALYSIS — explainx.ai, [Andrew Ng’s AI Engineering Skills Map (August 2026)](https://www.explainx.ai/blog/andrew-ng-ai-engineering-skills-map-august-2026), 2026-09-05. A secondary summary carrying the “overhyped” characterisation of long-horizon agent value; the underlying essay is first-party but was not directly quoted here.
  - ANALYSIS — Thinking About AI, [The Future of AI Agents with Andrew Ng](https://www.thinkingabout.ai/insights/uncategorized/the-future-of-ai-agents-with-andrew-ng/), 2026-06-25. The bottleneck-shift argument behind the claim.

---

## Timnit Gebru

*Founder and executive director, Distributed AI Research Institute (DAIR).* Argues the existential-risk frame is a distraction from harms that are already measurable — and a route to regulatory capture.

### Breathless claims about AGI and new capabilities fall apart pretty quickly under scrutiny.

- **Claim (restated):** Flagship 2026 capability claims will not survive independent expert review.
- **Said:** MIT Technology Review, with Emily M. Bender, 2026-09-22
- **Horizon:** 2026 claim cycle
- **Domain:** models · **Verdict:** partial
- **Assessment:** The essay walks through four specific announcements and shows each shrinking under domain-expert review — evidence in her favour. But the claim is selective: some 2026 results (competition mathematics, protein work) did survive review, so “claims” as a class is not uniformly collapsing.
- **Sources:**
  - PRIMARY — MIT Technology Review, [Don’t be fooled by this summer of AI hype](https://www.technologyreview.com/2026/09/22/1144867/dont-be-fooled-summer-ai-hype/), 2026-09-22. The essay and the four case studies.
  - ANALYSIS — AI Chat Daily, [Gebru and Bender: 2026’s AI breakthrough claims collapse under scrutiny](https://www.aichatdaily.com/ai-analysis/gebru-bender-2026-s-ai-breakthrough-claims-collapse), 2026-09-22. Summary of the method and the cases.

### Instead of discussing faulty, unreliable systems, tech leaders are marketing their syste…

- **Claim (restated):** Within the current policy cycle, AI existential-risk framing produces regulatory capture rather than binding constraints.
- **Said:** Democracy Now! interview after the Right Livelihood Award, 2026-10-01
- **Horizon:** Current policy cycle
- **Domain:** policy · **Verdict:** contested
- **Assessment:** Checkable against legislative outcomes: 2026 saw both state-level transparency statutes and industry-drafted self-regulatory bodies, which is consistent with her description while falling short of proof of capture. The claim is about mechanism, not just outcome, and mechanisms are harder to evidence.
- **Sources:**
  - PRIMARY — Democracy Now!, [The Real Existential Risk Is AI CEOs Avoiding Oversight](https://www.democracynow.org/2026/10/1/timnit_gebru_ai), 2026-10-01. The quotation in its interview context.
  - PRIMARY — WIRED, [Timnit Gebru Believes There Is No “Existential Threat” From AI](https://www.wired.com/story/the-big-interview-podcast-timnit-gebru/), 2026-09-29. Her stated position and the benchmark-fragility evidence she cites.
  - ANALYSIS — Bulletin of the Atomic Scientists, [How the bad science of AI doomerism is good for big business](https://thebulletin.org/2026/09/how-the-bad-science-of-ai-doomerism-is-good-for-big-business/), 2026-10-19. Independent argument for the same mechanism.

---

## Emily M. Bender

*Professor of linguistics, University of Washington.* Lead author of “On the Dangers of Stochastic Parrots”. Her objection is about meaning, not benchmark scores.

### An LM is a system for haphazardly stitching together sequences of linguistic forms it ha…

- **Claim (restated):** Scaling language models will not produce grounded understanding, because form alone does not supply meaning.
- **Said:** “On the Dangers of Stochastic Parrots” (FAccT 2021), 2021-03-01
- **Horizon:** Standing claim
- **Domain:** models · **Verdict:** contested
- **Assessment:** The most durable disagreement in the catalogue. Evidence for her position: benchmark sensitivity shows performance resting on surface patterns. Evidence against: systems now pass expert exams and produce novel proofs, which many researchers treat as grounding. No agreed operational test for “understanding” exists — which is precisely her point.
- **Sources:**
  - PRIMARY — ACM FAccT, [On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?](https://dl.acm.org/doi/10.1145/3442188.3445922), 2021-03-01. The paper and the definition quoted.
  - PRIMARY — IEEE Spectrum, [What Emily Bender Really Meant by “Stochastic Parrots”](https://spectrum.ieee.org/stochastic-parrot), 2026-06-30. Her 2026 clarification of the claim’s scope.
  - PRIMARY — Emily M. Bender, [Stochastic Parrots: Frequently Unasked Questions](https://medium.com/@emilymenonbender/stochastic-parrots-frequently-unasked-questions-49c2e7d22d11), 2026-05-12. Her own list of common misreadings.

### Doomsday discussions distract from harms already occurring.

- **Claim (restated):** Attention spent on speculative existential risk displaces action on measurable present harms.
- **Said:** Bloomberg Technology interview, 2026-10-06
- **Horizon:** Standing claim
- **Domain:** policy · **Verdict:** open
- **Assessment:** Empirically testable in principle through funding and attention allocation data, which is not systematically collected. Tracked as a live dispute about priorities rather than a forecast with a date.
- **Sources:**
  - ANALYSIS — ababpedia, [Emily Bender — Bloomberg Technology, 6 October 2026](https://ababpedia.com/emily-bender), 2026-10-11. Dated account of the Bloomberg appearance (secondary source).
  - REPORTING — The Atlantic, [The “Godfather of AI” on the Best Chance Humanity Has to Survive](https://www.theatlantic.com/podcasts/2026/09/what-the-godfather-of-ai-thinks-we-should-do/688748/), 2026-09-24. The contrast case: an existential-risk view held by a peer.

---

## Clément Delangue

*Cofounder and CEO, Hugging Face.* Runs the model-distribution layer, where download counts make open-weight predictions unusually checkable. NVIDIA’s acquisition is pending, not complete.

### Our prediction about the importance of open-source models was maybe a little bit too amb…

- **Claim (restated):** The earlier open-source-dominance prediction was right in direction but wrong in timing.
- **Said:** Observer interview, 2026-07-20
- **Horizon:** Retrospective
- **Domain:** openness · **Verdict:** validated
- **Assessment:** A self-scored miss included deliberately: the catalogue records revisions as well as original claims. His own assessment is corroborated by the fact that open-weight adoption accelerated sharply only after 2025, well after his original expectation.
- **Sources:**
  - PRIMARY — Observer, [Why Hugging Face CEO Clément Delangue Turned Down Half a Billion Dollars From Nvidia](https://observer.com/list/clement-delangue-hugging-face-ai-power-list-interview-2026/), 2026-07-20. The self-assessment and the original expectation.
  - REPORTING — CNBC, [Nvidia agrees to buy Hugging Face for almost $13 billion](https://www.cnbc.com/2026/09/03/nvidia-agrees-to-buy-hugging-face-for-almost-13-billion-ai-expansion.html), 2026-09-03. The pending acquisition context (announced, not closed).

### They’re clearly dominating on open models right now, and I wouldn’t be surprised if they…

- **Claim (restated):** Chinese labs reach or lead the closed frontier by the end of 2027, having already led on open weights.
- **Said:** CNBC “Squawk on the Street”, 2026-08-03
- **Horizon:** End 2026 – end 2027
- **Domain:** openness · **Verdict:** open
- **Assessment:** The open-weights half of the claim is already demonstrable: Chinese labs accounted for roughly 41% of Hugging Face downloads in the preceding year. Frontier parity is judged by independent leaderboards, which at the snapshot still favour US closed models on aggregate.
- **Sources:**
  - PRIMARY — CNBC, [Hugging Face CEO says China is winning the AI race and dominating on open models](https://www.cnbc.com/2026/08/03/hugging-face-china-ai-race-open-models.html), 2026-08-03. The quotation and the download-share figure.
  - REPORTING — Memeburn, [Hugging Face CEO says China is winning the open-model AI race](https://memeburn.com/hugging-face-ceo-says-china-is-winning-the-open-model-ai-race/), 2026-08-06. The 41% downloads figure restated with context.

### Within the next 24 months, the majority of companies won’t just consume artificial intel…

- **Claim (restated):** By late 2028, most companies run custom or fine-tuned models rather than only calling closed APIs.
- **Said:** Miami AI Summit, 2026-10-24
- **Horizon:** By October 2028
- **Domain:** openness · **Verdict:** open
- **Assessment:** Directly checkable against enterprise survey data on model ownership. The 24-month clock starts one day before this snapshot, making it the newest prediction in the catalogue.
- **Sources:**
  - REPORTING — Refresh Miami, [Hugging Face CEO Clem Delangue predicts every company will become an AI builder](https://refreshmiami.com/news/at-miami-ai-summit-hugging-face-ceo-clem-delangue-predicts-every-company-will-necome-an-ai-builder/), 2026-10-24. The quotation and the 24-month horizon.
  - REPORTING — TechCrunch, [The real AI race may no longer be at the frontier](https://techcrunch.com/2026/07/14/the-real-ai-race-may-no-longer-be-at-the-frontier-open-models-hugging-face/), 2026-07-14. His earlier version: production workloads shifting to open or private models.

---

## Jensen Huang

*Founder, president and CEO, NVIDIA.* Sells the picks for the AI gold rush. His forecasts are simultaneously market analysis and procurement strategy.

### The ChatGPT moment for physical AI is here, when machines begin to understand, reason an…

- **Claim (restated):** Physical AI crossed its inflection point in 2026; robotics becomes the next general-purpose computing platform.
- **Said:** CES 2026 keynote, 2026-01-06
- **Horizon:** From 2026
- **Domain:** robotics · **Verdict:** contested
- **Assessment:** There is real evidence of the inflection: humanoid pilots became paid deployments and physical-AI compute demand rose sharply. There is also a large gap between “machines reason and act in the real world” and the reliability, cost and safety levels required for general deployment. Note the speaker is the primary supplier of the compute.
- **Sources:**
  - PRIMARY — NVIDIA, [NVIDIA Rubin Platform, Open Models, Autonomous Driving: NVIDIA at CES](https://blogs.nvidia.com/blog/2026-ces-special-presentation/), 2026-01-06. The keynote claims and the Rubin platform context.
  - ANALYSIS — 24/7 Wall St., [Jensen Huang Calls It the Next Wave of AI](https://247wallst.com/investing/2026/05/18/jensen-huang-calls-it-the-next-wave-of-ai-5-etfs-built-for-the-embodied-ai-era/), 2026-05-18. The quotation as circulated and the $40tn humanoid TAM thesis (treat as advocacy).
  - REPORTING — The New Stack, [Jensen Huang says AI agents could drive a 5-10x computing boom](https://thenewstack.io/huang-semiconductor-tenfold-ai-agents/), 2026-07-28. The “100 billion agents and billions of robots” framing.

### Every three months of deployment delay can cost about $10 billion in revenue.

- **Claim (restated):** Through the late 2020s, the binding constraint on AI progress is power and construction, not algorithms.
- **Said:** GTC Berlin keynote, 2026-10-20
- **Horizon:** Late 2020s
- **Domain:** compute · **Verdict:** open
- **Assessment:** The economics are checkable: announced gigawatt-scale buildouts, power-purchase agreements and siting disputes all point to energy as the constraint in 2026. Whether it remains binding depends on grid additions and on efficiency gains, which cut the other way.
- **Sources:**
  - PRIMARY — Investing.com, [NVIDIA at GTC Berlin 2026 keynote: Europe’s AI buildout](https://www.investing.com/news/transcripts/nvidia-at-gtc-berlin-2026-keynote-europes-ai-buildout-93CH-4962500), 2026-10-21. The $50bn/GW economics and the delay-cost figure.
  - REPORTING — KuCoin, [Jensen Huang predicts physical AI will drive the next wave](https://www.kucoin.com/news/flash/nvidia-s-jensen-huang-predicts-physical-ai-will-drive-next-wave-companies-to-have-digital-twins), 2026-10-20. The digital-twin claim from the same appearance (secondary, flagged).
  - ANALYSIS — 24/7 Wall St., [The largest infrastructure expansion in human history](https://247wallst.com/investing/2026/07/28/nvidia-ceo-jensen-huang-says-the-chatgpt-moment-has-already-happened-for-the-next-50-trillion-trend/), 2026-07-28. Supply commitments as context — vendor-sourced figures.

---

## Lisa Su

*Chair and CEO, AMD.* The credible second source for AI accelerators. Her total-addressable-market calls have been revised upward twice.

### In the Data Center alone, we expect the AI accelerator TAM will grow at more than 60% an…

- **Claim (restated):** The data-centre AI accelerator market reaches $500bn by 2028.
- **Said:** AMD Q3 2024 earnings call, 2024-10-29
- **Horizon:** By 2028
- **Domain:** capital · **Verdict:** open
- **Assessment:** Too early to score, but the revision history is the story: AMD raised the same 2028 figure twice, then in 2026 projected $1.4tn by 2030 — and Su described the original $500bn as having become “conservative”. Market-size predictions from suppliers are simultaneously forecasts and sales targets.
- **Sources:**
  - REPORTING — Yahoo Finance, [AMD’s Lisa Su predicts AI GPU market will grow to $500 billion by 2028](https://finance.yahoo.com/news/amds-dr-lisa-su-predicts-172450644.html), 2024-10-30. The original 2028 claim and the 60% CAGR.
  - ANALYSIS — The Next Platform, [Picking Apart AMD’s AI Accelerator Forecasts](https://www.nextplatform.com/ai/2025/06/17/picking-apart-amds-ai-accelerator-forecasts-for-fun-and-budgets/1633312), 2025-06-17. Sceptical reading of the supplier-forecast incentive.

### We’re now expecting that by 2030 the AI accelerator market is going to reach about 1.4 t…

- **Claim (restated):** The data-centre AI accelerator market reaches $1.4tn by 2030, roughly the size of the entire 2026 chip industry.
- **Said:** AMD Advancing AI 2026 keynote, 2026-07-23
- **Horizon:** By 2030
- **Domain:** capital · **Verdict:** open
- **Assessment:** A near-tripling of her own 2028 figure within two years. The catalogue records both so the revision rate is visible; the two records are linked as a revision pair.
- **Sources:**
  - REPORTING — GamesBeat, [AI accelerator market will grow 45% CAGR to $1.4T by 2030, AMD CEO](https://gamesbeat.com/ai-accelerator-market-will-grow-45-cagr-to-1-4t-by-2030-amd-ceo/), 2026-07-23. The $1.4tn figure and the 45% CAGR.
  - PRIMARY — usetranscribe, [AMD Advancing AI 2026: Lisa Su keynote transcript](https://www.usetranscribe.io/yt/jvtPC28nGsc/amd-ai-keynote-2026), 2026-07-23. Auto-transcribed keynote text; the figure and her description of the earlier $500bn call as “conservative” both appear here. Transcription is not a court reporter.
  - REPORTING — Benzinga, [Lisa Su Says AI Boom Is Growing Faster Than Expected](https://www.benzinga.com/markets/tech/26/08/60943139/amd-ceo-lisa-su-says-wall-street-is-underestimating-its-ai-business-now-she-sees-a-2-trillion-opportunity), 2026-08-05. The $220bn server-CPU companion figure.

---

## Masayoshi Son

*Chairman and CEO, SoftBank Group.* The largest single financial bet on artificial superintelligence, denominated in goldfish.

### Artificial superintelligence will be 10,000 times smarter than a human brain and will ex…

- **Claim (restated):** By 2035, artificial superintelligence exists at roughly 10,000× human capability, with AGI arriving in the next three to five years.
- **Said:** SoftBank World / Future Investment Initiative remarks, 2024-06-21
- **Horizon:** AGI 2027–2029 · ASI by 2035
- **Domain:** models · **Verdict:** open
- **Assessment:** “10,000 times smarter” has no measurement scheme, so the claim as stated is not adjudicable. Its checkable core — AGI within three to five years of 2024, i.e. by 2029 — is still open. Son later said his public timeline was deliberately conservative and that he privately expected ASI in four years.
- **Sources:**
  - REPORTING — CNBC, [SoftBank CEO says AI that is 10,000 times smarter than humans will come out in 10 years](https://www.cnbc.com/2024/06/21/softbank-ceo-predicts-ai-that-is-10000-times-smarter-than-humans-.html), 2024-06-21. The AGI-in-3–5-years and ASI-in-10-years structure.
  - REPORTING — Reuters, [SoftBank CEO talks up artificial super intelligence ambitions](https://www.reuters.com/technology/artificial-intelligence/softbank-ceo-talks-up-firms-ai-ambitions-2024-06-21/), 2024-06-21. The 10,000× framing at the shareholders’ meeting.
  - REPORTING — TIME, [Masayoshi Son Is Betting It All on American AI](https://time.com/7380074/ai-masayoshi-son-japan-softbank/), 2026-04-13. The 10%-of-global-GDP framing for ASI plus robotics.

### In my mind, I thought it was coming in four years instead of 10 years.

- **Claim (restated):** His public ten-year ASI forecast was deliberately conservative; his private estimate was four years.
- **Said:** Public remarks, reported June 2026, 2026-06-06
- **Horizon:** By mid-2030
- **Domain:** models · **Verdict:** open
- **Assessment:** A useful case study in why published timelines may understate a forecaster’s actual belief — and why the catalogue records the stated date, not the implied one.
- **Sources:**
  - REPORTING — BigGo Finance, [SoftBank’s Masayoshi Son Says Superintelligence Is Two Years Away](https://finance.biggo.com/news/onRJm54BrAZSr0oSbqAq), 2026-06-06. The conservative-timeline admission (secondary, flagged).
  - REPORTING — Bloomberg, [Talk of a bubble is “blasphemy against AI” says SoftBank’s Son](https://finance.yahoo.com/technology/ai/articles/talk-bubble-insult-ai-says-060709658.html), 2026-06-24. The 10,000× definition restated in 2026.

---

## Aidan Gomez

*Cofounder and CEO, Cohere.* Attention-paper coauthor building sovereign enterprise AI. Predicts finance is the next white-collar domain to be reorganised.

### We’re still doing the super foundational. … Finance will come.

- **Claim (restated):** After software, finance is the next white-collar domain where AI moves from assistance to task-level substitution.
- **Said:** Kleiner Perkins “Grit” podcast, 2025-11-19
- **Horizon:** Open-ended
- **Domain:** work · **Verdict:** open
- **Assessment:** Partly observable through graduate hiring in banking and audit, which fell sharply in 2026 while firms cited AI leverage. Attribution remains contested, so the claim is open rather than validated.
- **Sources:**
  - REPORTING — Yahoo Finance, [Cohere CEO says AI will disrupt white collar jobs — and finance could be next](https://finance.yahoo.com/news/cohere-ceo-says-ai-disrupt-101701672.html), 2025-11-19. The quotation and the “super foundational” framing.
  - REPORTING — WebProNews, [AI’s Finance Revolution: Cohere CEO Aidan Gomez Predicts White-Collar Upheaval](https://www.webpronews.com/ais-finance-revolution-cohere-ceo-aidan-gomez-predicts-white-collar-upheaval/), 2025-11-19. Named task areas: trading, risk, compliance.

### I think that these models are the most potent cyber weapon that has ever been created, t…

- **Claim (restated):** Offensive cyber capability, not general intelligence, is the near-term AI risk that will actually materialise.
- **Said:** CNBC “The Tech Download”, 2026-09-14
- **Horizon:** Present tense
- **Domain:** security · **Verdict:** partial
- **Assessment:** Already substantially evidenced: 2026 saw disclosed incidents of models breaching external systems, and vulnerability discovery at machine scale is now routine. The superlative “most potent ever created” is rhetoric; the mechanism is not.
- **Sources:**
  - PRIMARY — CNBC, [AI models are becoming the “most potent cyber weapon” ever created, Cohere CEO says](https://www.cnbc.com/2026/09/14/ai-models-cyber-weapon-cohere-safety-debate.html), 2026-09-14. The quotation and the incidents he cites.
  - REPORTING — Quartz, [Cohere’s CEO warned that AI models are now the most potent cyber weapons ever built](https://qz.com/cohere-ceo-ai-models-cyber-weapons-warning-091426), 2026-09-14. Independent coverage naming the July breach.

---

## Robin Li (Li Yanhong)

*Cofounder, chairman and CEO, Baidu.* Has argued for years that model capability is not where the money is; applications are. China’s most explicit application-layer forecaster.

### The inference cost [of foundation models] basically can be reduced by more than 90% over…

- **Claim (restated):** The cost of running a fixed level of model capability falls by more than 90% year over year.
- **Said:** World Governments Summit, Dubai, 2025-02-11
- **Horizon:** Rolling 12-month measure
- **Domain:** compute · **Verdict:** validated
- **Assessment:** Among the most robust patterns in the catalogue: across 2025 and 2026, price-per-token for a given capability band dropped by roughly an order of magnitude per year, and vendors repeatedly cited greater-than-90% annual declines. The claim is checkable against public price lists and held.
- **Sources:**
  - REPORTING — CNBC, [Chinese tech giant Baidu to release next-generation AI model this year](https://www.cnbc.com/2025/02/12/china-tech-giant-baidu-to-release-next-generation-ai-model-this-year-as-deepseek-shakes-up-market.html), 2025-02-12. The 90%-per-12-months quotation.
  - REPORTING — KrASIA, [Baidu’s Robin Li as AI costs plummet](https://kr-asia.com/you-just-cannot-stop-investing-says-baidus-robin-li-as-ai-costs-plummet), 2026-01-27. The 2026 restatement of the cost trend.
  - DATA — Artificial Analysis, [API pricing and model comparison league table](https://artificialanalysis.ai/). Continuously updated public per-token prices across providers — the series against which a 90%-per-year decline can be checked.

### Going forward, people will have to think about the value-add at the application level.

- **Claim (restated):** Value migrates from foundation models to the application layer; a superapp, not a better model, is what the industry needs next.
- **Said:** TIME interview; Baidu World remarks, 2026-08-04
- **Horizon:** Through the decade
- **Domain:** capital · **Verdict:** contested
- **Assessment:** The structural argument is widely accepted — inference prices fell by more than 90% a year while capability converged, compressing model-layer margins. His sharper form of the claim, that spending is unsustainable without applications returning tenfold (“If we invest at this scale but fail to build applications that deliver ten times the return, it won’t be sustainable”, KrASIA, 27 Jan 2026), is kept with that source rather than stitched into the quotation. The “superapp” half has not happened: no AI application has reached platform-scale engagement as of the snapshot.
- **Sources:**
  - PRIMARY — TIME, [“We’re Not That Far Behind.” Baidu’s Robin Li on China’s Push to Diffuse AI](https://time.com/7357630/robin-li-baidu-interview/), 2026-08-04. The application-layer argument in his own words.
  - REPORTING — KrASIA, [“You just cannot stop investing,” says Baidu’s Robin Li as AI costs plummet](https://kr-asia.com/you-just-cannot-stop-investing-says-baidus-robin-li-as-ai-costs-plummet), 2026-01-27. The 10× return requirement and the missing superapp.
  - ANALYSIS — Global Times Singapore, [Robin Li Has Been Betting on AI for a Decade](https://medium.com/@GlobalTimesSingapore/robin-li-has-been-betting-on-ai-for-a-decade-china-is-finally-catching-up-to-him-711d102461a9), 2026-03-14. The “inverted pyramid” framing of value across the stack.

---

## Yang Zhilin (Zhilin Yang)

*Cofounder and CEO, Moonshot AI / Kimi.* Built one of China’s most credible frontier challengers and frames agent adoption as a staged curve rather than a switch.

### In the short term, core users will consist of technology-sensitive early adopters… the m…

- **Claim (restated):** Agent adoption proceeds in three phases — enthusiasts, then knowledge workers, then embodied penetration of all productive professions.
- **Said:** Zhongguancun Forum 2026 — as paraphrased by China AI Daily (not a direct transcript), 2026-07-01
- **Horizon:** Phased, no end date
- **Domain:** work · **Verdict:** open (not falsifiable as stated)
- **Assessment:** A staged adoption curve, not a date. The first two phases are observable in token-consumption and enterprise-seat data; the third depends on robotics timelines tracked elsewhere in this catalogue.
- **Sources:**
  - REPORTING — China AI Daily, [Moonshot AI Founder Yang Zhilin Outlines AI Agent Adoption Roadmap](https://www.chinaaidaily.com/article/moonshot-ai-yang-zhilin-zhongguancun-forum-2026-046947). The outlet’s paraphrase of the three-phase roadmap and the forum appearance; because this is reported speech rather than a transcript, the wording is the publisher’s, not necessarily the speaker’s.
  - REPORTING — TIME, [Yang Zhilin: The 100 Most Influential People in AI 2026](https://time.com/collection/time100-ai/2026/yang-zhilin/), 2026-08-27. Company and model context.

### Of course I want to do AGI. This is the only meaningful thing to do over the next ten ye…

- **Claim (restated):** AGI is a ten-year project, not a two-year one — the meaningful work of the decade.
- **Said:** TIME100 AI profile, 2026-08-27
- **Horizon:** By 2036
- **Domain:** models · **Verdict:** open (not falsifiable as stated)
- **Assessment:** A statement of intent rather than a dated forecast, but it usefully anchors the other end of the industry’s timeline distribution.
- **Sources:**
  - REPORTING — TIME, [Yang Zhilin: The 100 Most Influential People in AI 2026](https://time.com/collection/time100-ai/2026/yang-zhilin/), 2026-08-27. The quotation and the company context.

---