Toby Ord

Who he is

Affiliation: Senior Researcher, Oxford Martin AI Governance Initiative (AIGI), University of Oxford, 2024– (AIGI people page; his CV: "2024 – Senior Researcher Oxford Martin AI Governance Initiative"; the programme was announced by the Oxford Martin School on 2023-09-04 and is co-led by Robert Trager and Maike Osborne per its About page). Role check run 2026-09-21 on a dated search plus his own site, his CV, the AIGI page and the 80,000 Hours episode page (recorded 2026-07-02): no job change; the v1 line holds. Also: research affiliate, Forethought (per tobyord.com; Forethought people page lists his 2024–26 research); board of directors, Centre for the Governance of AI (GovAI) — CV: "2022 – now: Trustee of the Centre for the Governance of AI"; the GovAI people page lists the board as Jeffrey Ding, Tasha McCauley, Seán Ó hÉigeartaigh, Toby Ord and Helen Toner; OECD Expert Group on AI Futures, 2023– (CV); adviser to DeepMind, "2010 – now. Advised on the ethics, safety, and governance of AI" (CV). Previously: Senior Research Fellow (2019–2024) and Research Fellow (2014–2019), Future of Humanity Institute, which "On 16 April 2024 … was closed down" (FHI final report, 2024-04-17); founder of Giving What We Can (2009–2022 per CV) and co-founder and trustee of the Centre for Effective Altruism (2012–2022 per CV). DPhil Oxford (Balliol, supervised by John Broome and Derek Parfit), BPhil Oxford, BA Philosophy and BSc (Hons) Computer Science, University of Melbourne (CV).
One-line position: Humanity faces about a one-in-six chance of existential catastrophe this century, most of it from AI — "When I wrote the book, I thought it was about one in six … Yeah, I think that's still about right, and I would say that most of that is coming from AI" (AEI, 2025-01-31) — but since 2025 the position has a second half the v1 file did not have: the philosopher of The Precipice has become a working analyst of frontier scaling data, arguing that "AI progress as a function of time is impressive even if AI progress as a function of resources is not" (The Scaling Paradox, 2025-01-13), that AI timelines should be broad ("a median of 2038 and an 80% interval ranging from 3 years to 100 years", Broad Timelines, 2026-03-09), and that a moratorium on superintelligence is preferable to a pause (80,000 Hours, recorded 2026-07-02).

Discipline & technical bet

By training a moral philosopher — the Parfit lineage, population ethics, moral uncertainty (co-author of Moral Uncertainty, OUP 2020, per CV) — with a computer-science degree underneath. That combination is the thing to understand about him now: since January 2025 he has written an eight-part "Scaling Series" (collected 2026-02-02) plus four further quantitative pieces, each reading published lab and METR data directly. The concrete technical wager: (1) the scaling laws "are a proof of poor but predictable scaling, rather than impressive scaling" — "compute required scales as the 20th power of the desired accuracy", so "in terms of AI progress per resource, or per dollar, things are probably getting worse on most measures" (The Scaling Paradox); (2) the post-2024 gains in reasoning models come mostly from inference scaling (spending far more compute every time the model answers) rather than from training — "most of the gains from using RL to train reasoning models have been coming from scaling-up the number of tokens used to solve the task", 82% on MATH level 5, 63% on GPQA Diamond, 92% on mock AIME, measured on Anthropic's Sonnet 3.7 against its Sonnet 3.5 base (2025-10-03); (3) reinforcement learning (training by reward rather than by predicting text) scales badly — "a 10x scaling of RL is required to get the same performance boost as a 3x scaling of inference … This is breathtakingly inefficient scaling" (2025-10-20) and "reduces the amount of information a model can learn from per hour of training by a factor of 1,000 to 1,000,000" (2025-10-09); so (4) "the strength of this tail-wind that has driven much of AI progress since 2020 will halve" and "AI researchers will be comparatively more important than in recent years" (Scaling Series thread); and (5) recursive self-improvement will probably not produce a mathematical singularity, because "one cannot have singular growth unless the generation time rapidly approaches zero" — super-exponential growth "without a singularity is no longer a curiosity — it is the default" (The Dynamics of Intelligence Explosions, 2026-08-28). Newest data point, dated today: agent swarms buy speed, not capability — "scaling up the number of agents in the swarm by 10x doesn't get as much performance as using 10x as many tokens with one agent" (Swarm Scaling, 2026-09-21). The bet is checkable and he keeps publishing the numbers that could sink it.

Key claims (Says)

• Empirical: One-in-six existential risk this century — the headline number from The Precipice (2020), framed as a "time of perils" humanity must navigate. Held on endorsement — restated in-window as a (number, horizon) pair: asked what he put "as the risk to human existence over the next a hundred years … and what is it now?", "When I wrote the book, I thought it was about one in six." — "Yeah, I think that's still about right" (AEI, 2025-01-31). The number itself is unfalsifiable on its horizon (see Predictions); what is graded here is that he still says it, which the v1 file doubted. Not restated in the 2026-07 80,000 Hours transcript (grep: no "one in six"), so the last restatement is 20 months old at this review.
• Empirical: AI is the plurality x-risk at ~1 in 10 — larger than bio, nuclear or climate. Held — "most of that is coming from AI" (AEI, 2025-01-31); "the chance of unrecoverable failure from AGI is over 10%" (Faggella, AGI Governance ep. 7, 2025-08-12). Category directions given in 2024: "in my view, Climate risk is down, Nuclear is up, and the story on Pandemics and AI is mixed" (The Precipice Revisited, 2024-07-12) — the AI number was not moved.
• Normative: Existential risk is a moral catastrophe if ignored — population ethics makes failing to safeguard the future equivalent to astronomical harm. Normative; unfalsifiable — noted, not graded. Behavior consistent: the 2024 essay On the Value of Advancing Progress (2024-07-12) applies exactly this frame — "considerations about existential risk undermine a promising argument for advancing progress and call into question whether advancing progress is good at all."
• Mixed: "The Precipice" as a historical moment — technology can destroy us, wisdom has not caught up, the window may be temporary. Held on endorsement: the 2024 revisit keeps the frame ("The world is beginning to take security from existential risks seriously. Making sure this process succeeds over the coming decades is crucial"), and his 2026 definition of transformative AI is "the deadline for alignment" (80,000 Hours page summary). One self-correction inside the frame, 2026-03-09: "I used to think that there just wasn't enough time to write a further book — that we were really too close to the critical moment. We might be, but I think I was mistaken about the strength of this argument." The precipice is still a precipice; the ledge got wider.
• Empirical: Calibrated risk estimation — specific probability breakdowns by source, unusual rigour for the field. Held, with an asterisk: he re-published the 2020 breakdown ("1 in 1,000 for Climate and for Nuclear, 1 in 30 for Pandemics and 1 in 10 for Unaligned AI") and gave directions of change, not new numbers (The Precipice Revisited, 2024-07-12). The rigour has migrated: the numbers he now publishes monthly are scaling exponents, not risk probabilities.
• Empirical (NEW, 2025): The Scaling Paradox — "The scaling laws are impressively smooth and long-lasting, but are a proof of poor but predictable scaling, rather than impressive scaling" (2025-01-13). Check attached, not graded: by 2027-12-31, does any frontier lab publish a scaling relationship materially better than the ~20th-power compute-per-accuracy figure he derives? A published, replicated improvement of an order of magnitude in the exponent falsifies; continued 10x-per-tick charts confirm.
• Empirical (NEW, 2025): Inference scaling reshapes AI governance — a shift to inference-at-deployment "would undermine the current paradigm of AI governance via training compute thresholds", "lower the importance of open-weight models" and "reduce the impact of the first human-level models" (arXiv 2503.05705, v1 2025-02-12; tobyord.com date 2025-02-12). Check: by 2027-12-31, does any jurisdiction with a training-compute threshold (the EU AI Act's general-purpose-model rule is the obvious one) add an inference-compute or deployment-compute criterion, or publicly concede the threshold no longer tracks capability? Either grades Held; an unchanged, still-binding training threshold through 2028 grades Wrong on the "undermine" half.
• Empirical (NEW, 2025): Recent gains are mostly inference scaling, which is "substantially changing the business model, the risk profile, and the policy response" because inference costs recur with every use (2025-10-03). Check: a lab or independent evaluator publishes a decomposition on a 2026 frontier model showing RL/training rather than test-time compute as the majority of the reasoning gain by 2027-06-30 — that grades Wrong; the reverse, or silence plus rising per-query prices, grades Held.
• Empirical (NEW, 2025): RL scales badly and is information-poor — "we may have lost the ability to effectively turn more compute into more intelligence" (2025-10-20); "My hypothesis is that this will become less true the further we get into the RL scaling era — that the much lower information-efficiency of RL is not affecting the depth of ability … but will be seen in a lack of breadth and generality" (2025-10-09). Check: by 2027-12-31, a frontier model whose training compute is majority-RL shows a measurable breadth deficit on out-of-distribution evaluations relative to its pre-training peers — Held; broad, general gains from RL-heavy training — Wrong.
• Empirical (2025 → self-revised 2026): AI agents have a "half-life" — "a constant rate of failing during each minute a human would take to do the task", so success declines exponentially with task length (2025-05-07; arXiv 2505.05115). Wrong on mechanism, by the author's own hand: "AI agents probably don't have a constant hazard rate / half-life. Instead their hazard rates systematically decline as the task goes on" — "the way that every single model has k statistically significantly below 1 convinces me he is right" (Hazard Rates for AI Agents Decline as a Task Goes On, 2026-02-04). Direction (success falls with length) held; the constant-rate model did not. Logged as the strongest kind of Wrong — a self-published reversal — and as a track-record credit for publishing it.
• Empirical (NEW, 2026): Agent costs are rising exponentially too — "there is evidence that the METR trend is partly driven by unsustainably increasing inference compute" and "real-world applications of AI agents will lag behind the METR time-horizon trend by increasingly large amounts"; the piece puts o3 at roughly $350 an hour at its full 1.5-hour task horizon (2026-02-02). Check: METR's next two time-horizon updates report cost-per-task alongside horizon; if hourly cost at the 50%-horizon has fallen year-on-year for the frontier model by 2027-06-30, Wrong; if it has risen, Held.
• Empirical (NEW, 2026): Broad timelines — "The correct epistemic response to the lasting expert disagreement is to have a broad distribution over AI timelines"; his own: "a median of 2038 and an 80% interval ranging from 3 years to 100 years"; "I only think there is about a 1 in 5 chance of transformative AI coming in the next 5 years" (2026-03-09; restated 2026-07-02: "my median date, my 50% confidence number, is 2038 for transformative AI"). Graded as a prediction below. Drifted on the stance relative to his own January 2025 words: asked whether Metaculus's "2027 … 'weak AGI'" and "2031 … 'strong AGI'" were what he assumed, "Yeah, I think that those are good numbers for the typical case, what you should be expecting … those timelines sound very realistic. Maybe even sooner." (AEI, 2025-01-31). Fourteen months later the median is 2038. He explains the move (the Scaling Series is the argument); the file logs it as a move.
• Empirical (NEW, 2026): RSI probably won't ignite an intelligence explosion — the 80,000 Hours summary of his position (recorded 2026-07-02), then the mathematics: "singular growth (towards a vertical asymptote) is harder to achieve than would be expected from recent economics-inspired modelling" (2026-08-28). He is careful about "probably": "We can't rule out, say, Daniel Kokotajlo's nine months. I don't think I can rule out that it could happen in nine months." Check below.
• Normative (NEW, 2026): A moratorium beats a pause — per the episode summary, "a pause implies development automatically resumes, whereas a moratorium means society has to decide there's a compelling case for reversing the ban"; chapter titles on the same page: "Could we ban superintelligence?", "A US-China treaty on superintelligence is possible", "We should just ban unmonitorable chain of thought" (80,000 Hours, recorded 2026-07-02). Also 2025: "handing over control to the first capable system we build would be a mistake"; "a coalition of countries as a promising path forward" (Faggella, 2025-08-12). Does-check in Revealed behavior.
• Mixed (NEW, 2026): Mathematics is much more than proof — after "a disproof for the well-known Unit Distance Conjecture in May" and "OpenAI's set of 10 proofs in mathematics and computer science in August", "Proof is but one part of mathematics, and not, in my view, the most interesting part … The automation of proof alone would not be the end of mathematics" (2026-08-23). Unfalsifiable as written — noted, not graded; it is a claim about what counts as mathematics, and the events it cites are logged as his engagement with the frontier, not as bench-verified facts.

Notable predictions — with falsifiable checks

• (2020) ~1 in 6 existential catastrophe this century. v1: "not directly testable; framework for prioritization." Unfalsifiable on its horizon — retained as a framework, not graded; the gradeable part is whether he keeps saying it, which he did on 2025-01-31 (Held above). Standing check: log "last restated" at every review; two consecutive reviews without a restatement grade the number Drifted by silence (Chalmers rule).
• (2020) AI is the largest single x-risk at ~10%; v1 outcome: "TBD; consensus has shifted toward taking this seriously." The probability is unfalsifiable on horizon; the consensus half has an observable and is Held: "the idea that existential risk from AI is a major global priority is now a respected standard position by many leaders" and "Existential risk from AI had finally gone from a speculative academic concern to a major new topic in global governance" (The Precipice Revisited, 2024-07-12); the FLI superintelligence statement of 2025-10-22 was signed by Hinton, Bengio and Russell among others (FLI press release; EA Forum crosspost, 2025-10-22). Ord himself is not a named signatory of that statement or of the 2023 CAIS statement (see Revealed behavior). Forward check: does Ord publish a revised AI number (up or down) by 2027-12-31? Silence through two reviews grades Drifted.
• (2020) Bio risk is the second-largest contributor (1 in 30). Held — the 2024 revisit re-listed the ordering unchanged and called the pandemic story "mixed" (2024-07-12); no reordering found in any 2025–26 primary. v1's COVID point (institutional under-preparedness) carried. Check: any restatement placing nuclear above pandemics by 2027-12-31 (he says "Nuclear is up") grades this Drifted.
• (2026-03-09) Transformative AI: ~1 in 5 within five years; median 2038; 80% interval 3–100 years. Definition, his: "AI that would be powerful enough to take over the world were it misaligned, and which is doubling the rate of scientific and technological progress." Checks: (a) by 2031-03-09, transformative AI on that definition has or has not arrived — arrival scores the 20% branch, non-arrival the 80% branch; (b) by 2038-12-31 the median; (c) nearer: his own words on the base rate — "it's just a very credible way that you get at least a 10% chance that it doesn't happen by 2040, to do with choices made not to do it and some kind of treaty" (80,000 Hours, 2026-07-02) — so a verified international agreement restricting frontier training by 2030-12-31 is the observable for that branch. Import Cotra's scored bias (forecasters in this community have run slow on capabilities) as a correction: read his 2038 as an upper-middle estimate, not a central one, until he scores himself.
• (2026-07-02 / 2026-08-28) RSI probably won't produce singular growth; super-exponential growth without a singularity is the default. Check: by 2029-12-31, any frontier lab's measured AI-R&D automation produces a generation time (his term for the length of each improve-yourself loop) that shrinks by more than an order of magnitude year-on-year while capability keeps compounding — Wrong; generation times that plateau while capability grows super-exponentially — Held on his own terms.
• (2026-02-02) Real-world agent deployment will lag the METR horizon trend "by increasingly large amounts." Check: by 2027-12-31, a market-level series (payroll or firm-use data, not a benchmark) shows agent task lengths in production tracking the METR 50%-horizon within a factor of two — Wrong; a widening gap — Held. Market-level claim, market-level data (Brynjolfsson rule).
• (2026-09-21) Swarms give speed, not capability — "scaling up the swarm size by 16x would give the same performance as scaling up the length of the chain of thought by just 16^0.57 = 4.9x"; the measured "stepping on toes" parameter runs 0.48–0.68 across benchmarks. Check: a published result by 2027-12-31 in which a multi-agent swarm beats the best single-agent run at equal total tokens by a wide margin grades Wrong; replication of the sub-linear exponent grades Held.
• (2025-01-31) "I don't think that there's going to be some kind of existential catastrophe from AI in the next couple of years." Check: 2027-01-31 — the easiest grade on the bench; logged so that it is scored rather than assumed.

Revealed behavior (Does)

• Time (2025–26) went to numbers, not sermons. Twelve dated quantitative pieces in 20 months, read off his site's own page metadata: The Scaling Paradox (2025-01-13), Inference Scaling Reshapes AI Governance (2025-02-12), Is There a Half-Life for the Success Rates of AI Agents? (2025-05-07), Evidence that Recent AI Gains are Mostly from Inference-Scaling (2025-10-03), The Extreme Inefficiency of RL (2025-10-09), How Well Does RL Scale? (2025-10-20), Are the Costs of AI Agents Also Rising Exponentially? (2026-02-02), Hazard Rates Decline (2026-02-04), Broad Timelines (2026-03-09), Mathematics is Much More than Proof (2026-08-23), The Dynamics of Intelligence Explosions (2026-08-28), Swarm Scaling (2026-09-21). Plus a pure-mathematics paper, "Evaluating the Infinite" with hyperreal numbers (Forethought, 2025-09-21). Where the hours went is the grade on the v1 weak spot "less engaged with near-term AI deployment questions" — see Weak spots.
• Announced the work, then shipped it. "I'm actually working on a paper at the moment trying to understand these possibilities of really explosive growth, which I do think is credible and I'm concerned about it" (80,000 Hours, recorded 2026-07-02) → The Dynamics of Intelligence Explosions, 2026-08-28. Says and Does line up on a 57-day lag.
• Corrected himself in print when a reader's statistics beat his model (Hazard Rates, 2026-02-04) — and updated the original half-life page the same day (its dateModified is 2026-02-04). A persona that edits its own older claims is easier to grade than one that lets them stand.
• Signed the Global Call for AI Red Lines (launched "at the United Nations General Assembly in September 2025", per red-lines.ai); his entry on the individual signatories page: "Senior Researcher – University of Oxford / Founder – Giving What We Can / Board Member – Centre for AI Governance." Did not sign the 2023 CAIS one-sentence statement — his name is absent from the signatory page as fetched 2026-09-21. Unresolved for the FLI Statement on Superintelligence (2025-10-22): not among the names on FLI's press release; the statement site displayed a signature count (75,111 as fetched) but no name list I could read. A man who says on tape he prefers a moratorium, and whose name is not on the moratorium letter, is a Says/Does gap until the list is read — flagged, not graded.
• Talks to the same room, repeatedly. "Toby Ord is back — for the 5th time!" is the first chapter title of the 2026 80,000 Hours episode; the previous one (recorded 2025-05-23) was on the same scaling material — "Every time you want to halve the amount of this error that's remaining, you have to put in a million times as much compute." EAG San Francisco talk "Taking longer AI timelines seriously" (YouTube upload 2026-03-27). Outside the EA channel in the window: AEI (2025-01-31), Faggella (2025-08-12). The distribution is narrow by choice.
• Money he moves rather than takes: Giving What We Can — "11,240 people around the world have taken the 🔸10% Pledge" and "our community has donated over $500 million USD" (GWWC about page, fetched 2026-09-21); 10,000 pledgers reached 2025-08-26, "In 2024 alone, pledgers donated more than $50 million." His own CV still says "more than 4,000 members (from 43 countries)" and "more than $100 million" — the CV is years stale on his own creation, which is why this file dates from the org's page.
• Reads Bostrom's drafts. Thanked "For comments" in Optimal Timing for Superintelligence (working paper v1.0, 2026) and in the acknowledgments of Open Global Investment (2025) — both checked in the PDFs this run; not in the 2024 Cosmic Host paper. He has published no response to Optimal Timing that I could find (no comment on the LessWrong linkpost, 2026-02-13; "Bostrom" does not appear in the 2026-07 80,000 Hours transcript).

Feels

"I'm also very worried about short timelines. Thinking that it may well be long timelines doesn't make it all that much less worrying that it also may be short timelines" (2026-07-02) — the worry is constant; what changed is that he gave himself permission to plan past it ("I think I was mistaken about the strength of this argument" that there was no time for another book). Wants to be true: that the numbers are poor enough per dollar to buy the world a decade, and that "the world is beginning to take security from existential risks seriously." Fear underneath: "I'm really worried that this new race between big tech companies will push the labs into riskier deployments" (2024-07-12).

Hears

Forethought (MacAskill, Davidson, Dalton) and the FHI diaspora; Bostrom's drafts (he is in the thanks); METR's time-horizon data and the labs' own benchmark disclosures, which he reads directly; Rob Wiblin's 80,000 Hours studio five times over; Kokotajlo's distributions, which he argues against by name; a reader named Gus whose Weibull fits overturned his half-life model. Little from labour economists, critical-AI scholars or industry commercial voices — his input diet is data plus the EA/x-risk network.

Sees

Two decades inside the movement he helped build (GWWC 2009, CEA 2012, FHI 2006–2024) and an adviser's view of DeepMind since 2010 (per his CV). Uniquely on the bench he combines Parfit-trained ethics with the ability to re-derive a scaling exponent from a lab's chart; he sees what "each one of these tick marks is 10 times as much inputs as the one before" does to a headline that the commentator class cannot. He also sees the field's timelines from the inside and now says most of them are point estimates with the error bars thrown away.

Incentive map

Re-derived from scratch this cycle. Who pays: the University of Oxford via the Oxford Martin School's AI Governance Initiative (a 2023-09-04 programme; its About page names no funder). A search listing shows an Open Philanthropy grant to the Berkeley Existential Risk Initiative "to support research activities at the Oxford Martin AI Governance Initiative"; the grant page now redirects to Coefficient Giving and the amount could not be read off a fetched primary (HTTP 403 on 2026-09-21) — logged, not cited as a figure. Forethought, where he is a research affiliate, is "evaluated and endorsed by both Longview and by Coefficient Giving (formerly Open Philanthropy), who are recommending us to donors in their network" and "currently has funding until mid-December 2026 on our baseline budget" (Forethought fundraiser page); whether his affiliation is paid is not stated. Book income (The Precipice; Moral Uncertainty) and speaking. Advises a frontier lab: "DeepMind. 2010 – now" (CV) — terms and compensation undisclosed; note that his 2025 inference-scaling decomposition is run on Anthropic's models and his cost analysis names OpenAI's o3, i.e. he publishes comparative analytics on the competitors of the lab he advises. No lab equity, employment or paid advisory contract found on a dated search (2026-09-21). Board seats: GovAI (trustee since 2022, with Toner); OECD Expert Group on AI Futures (2023–). Selling: a "broad timelines" frame and a per-dollar deflation of scaling — both of which cut against the short-timeline fundraising narrative of his own community, which is evidence he is not optimising for it. Can't say: anything that reads as dismissing AI risk (his employer's raison d'être is AI risk governance); anything specific about DeepMind internals. Household: nothing found; not searched beyond a dated news check. Rule-25 discount: low on the scaling analytics (data public, method shown, self-corrections published); low-moderate on governance prescriptions (institution-aligned); moderate on the x-risk numbers themselves (career and institution both rest on the category). "Unconflicted academic" is not carried forward — it is re-derived to "academic with one undisclosed lab advisory relationship and Open-Phil-lineage funding, publishing against his own camp's interest on timelines."

Theories aligned with

• AI alignment / x-risk — the calibrated-probability variant; foundational framing via The Precipice; now also the camp's most persistent internal skeptic of short timelines.
• Existential risk / longtermism — foundational; the impersonal-future version that Bostrom's 2026 paper explicitly sets aside ("We examine optimal timing from a person-affecting stance").
• Effective altruism — co-founder of the movement's institutional infrastructure (GWWC, CEA); still its largest fundraising story in dollars moved.
• NEW (2025–26): "Scaling is expensive, not magic" — a position adjacent to Sutskever's "age of research" turn and to Marcus's diminishing-returns line, reached by a different route (reading the exponents, not the vibes) and with a different conclusion (progress continues, but researchers matter more and costs bite).
• NEW (2026): Broad timelines — a stance, not a camp; it is his answer to both AI 2027 and to long-timeline complacency.

What he's reacting against

• Short-termism in policy — humanity systematically ignores risks to its own long-run survival (carried; still the frame of the 2024 revisit)
• Treating existential risk as science fiction rather than serious probability (carried; he now says the opposite problem has arrived — the topic is "a respected standard position")
• The assumption that technology is net-positive by default — insists on active risk management (carried; the 2024 progress essay is the formal version)
• Climate change absorbing all the "existential" oxygen while AI and bio risks are larger (carried; his 2024 update moved climate down and nuclear up)
• NEW: log-x charts — "each one of these tick marks is 10 times as much inputs as the one before" (80,000 Hours, recorded 2025-05-23); the misleading chart is now a named enemy
• NEW: point-estimate timelines — of his 14 ways people go wrong: "Consuming point estimates and discarding the error bars", "Forecasting 'could' instead of 'will'", "Dismissing dissenting experts", "Treating 'we don't know' as permission to continue as usual" (80,000 Hours page, 2026)
• NEW: the commercial race — "this new race between big tech companies" (2024-07-12); "Now, that race has grown more intense – fueled by trillion-dollar companies and dominated by commercial pressure" (Faggella page's paraphrase, 2025-08-12)
• NEW: "pause" as a word — prefers a moratorium, because a pause "implies development automatically resumes" (2026-07-02)

Where he overlaps / splits (with Rich)

• Overlaps with Bostrom on existential risk as the frame (co-authors of "The reversal test", Ethics 2006, per CV; Ord thanked in Bostrom's 2025 and 2026 papers). v1: splits on public communication — Ord accessible, Bostrom technical. Drifted — the v1 split understates a substantive one. Ord, 2024-07-12: "If we knew that our end was driven by something on the second list, then advancing all progress would be a bad thing" and existential-risk considerations "call into question whether advancing progress is good at all." Bostrom, 2026: "even high catastrophe probabilities are often worth accepting" and "poorly implemented pauses could do more harm than good." Ord, 2026-07-02: prefers a moratorium to a pause. Same vocabulary, opposite recommendation; and Ord keeps the impersonal-future ethics Bostrom's paper sets aside. No public exchange yet — wake trigger set.
• Overlaps with Yudkowsky on AI as the top x-risk; splits on tractability and magnitude — Ord "over 10%" (2025-08-12) versus a book Bostrom characterises as "regarding its advent as an almost certain doom" (Optimal Timing, 2026). Held, with the policy gap narrowing: Ord's 2026 "Could we ban superintelligence?" / moratorium talk lands close to the Yudkowsky–Soares ask while the probabilities stay an order of magnitude apart. Ord's name is not on the FLI prohibition statement as far as I could read (see Does).
• Overlaps with Christiano on calibrated risk estimates; v1 splits: Christiano technical/evals, Ord philosophical/institutional. Drifted: since 2025 Ord works the evals data himself (METR horizons, lab benchmark decompositions, an arXiv paper on agent success rates). The bench's Christiano file (validated 2026-09-21) records his move onto the OpenAI Foundation board — the two have swapped toward each other's old turf.
• Splits with Cowen and Pinker on x-risk magnitude — they treat it as very small, Ord puts it at 1 in 6. Carried; not re-verified on a primary this run. Note the new adjacency: Cowen amplified Bostrom's cost-of-waiting paper (per the Bostrom file), which is the argument Ord's 2024 essay was written against.
• Overlaps with Kokotajlo on transformative AI as the object to forecast and on the danger of RSI; splits sharply on the distribution — "It's a similar shape to Kokotajlo's, but broader, with a median of 2038 and an 80% interval ranging from 3 years to 100 years" (2026-03-09). The cleanest pairwise check on the bench: both have dated numbers.
• Overlaps with Cotra and Grace on distributions over point estimates; he now runs slower than both on capabilities, and unlike Cotra has not published a self-score.
• Overlaps with Sutskever — from opposite directions — on "scaling is no longer the whole story": Sutskever says the age of research has begun; Ord says the tail-wind "will halve" and "AI researchers will be comparatively more important." Splits on what that implies: Sutskever bought an order of magnitude more compute; Ord thinks the per-dollar returns are the story.
• Overlaps with Marcus on diminishing returns to scale; splits on conclusion — Ord still expects transformative AI (median 2038) and is "very worried about short timelines."
• Overlaps with Bengio, Hinton, Russell on x-risk as real and on international coordination; they signed the 2025 prohibition statement, he — as far as I could read — did not.
• Shares a board with Toner (GovAI directors, per GovAI's people page) — mutual-confirmation risk logged (batch rule of 2026-09-04); no business relationship beyond the board found.
• Less engaged with Acemoglu, Brynjolfsson, Frey on labour and diffusion — although his 2026 cost-of-agents piece ("real-world applications of AI agents will lag") is a diffusion claim in their territory; a cross-domain thread to watch.

Track record

• The Precipice (2020) became the reference text for existential-risk thinking — genuine field-defining impact. Held: it is still the byline on every 2025–26 appearance ("author of The Precipice", 80,000 Hours; AEI; Faggella); CV lists reviews in the New Yorker, NYRB, Science, FT and a 2022 lecture at Number 10 ("Existential Risks to Humanity. Number 10 Downing St (4 May 2022)").
• Co-founded Giving What We Can (2009) — v1: "moved hundreds of millions in donations; institutional track record strong." Held and upgraded: "over $500 million USD" and "11,240 people" on the 10% Pledge (GWWC about page, fetched 2026-09-21); 10,000 pledgers reached 2025-08-26. Number re-read from the org's page, not from his CV, which lags by years.
• FHI was closed by Oxford in 2024 — institutional setback, though Ord moved to the AI Governance Initiative. Held, dated and sourced: "Starting in 2020, the Faculty imposed a freeze on fundraising and hiring. In late 2023, the Faculty of Philosophy decided that the contracts of the remaining FHI staff would not be renewed. On 16 April 2024, the Institute was closed down" (FHI final report, 2024-04-17); "2024 – Senior Researcher Oxford Martin AI Governance Initiative" (CV).
• His risk estimates are widely cited but hard to evaluate empirically — frameworks, not forecasts. Held; unchanged by construction. The new numbers (exponents, percentages, medians) are the first parts of his output that can be scored, and this file attaches the checks.
• NEW: Published his own reversal — the constant-hazard half-life model (2025-05-07) retracted on a reader's better fit (2026-02-04). Counts as a track-record credit regardless of the Wrong grade on the model.
• NEW: Said the paper was coming and shipped it (explosive-growth paper, 2026-07-02 → 2026-08-28).
• NEW: Timeline stance moved a long way in 14 months — "2027 … 2031 … those timelines sound very realistic. Maybe even sooner" (2025-01-31) → median 2038 (2026-03-09). Logged as Drifted above; the honest reading is that the Scaling Series changed his mind and he said so in public. Unscored by him; scored here.
• Pattern: a philosopher whose most gradeable work is now the least philosophical — and who, unlike most of the bench, publishes corrections to it.

Empirical vs normative

• Empirical: probability estimates for specific risk categories; historical analysis of near-misses; and now scaling exponents, inference/RL decompositions, agent cost curves, the mathematics of feedback loops — carried and extended
• Normative: strong longtermist ethics; humanity has an obligation to its potential future; risk reduction is the priority; and, new, a moratorium rather than a pause, red lines by treaty, a ban on unmonitorable chain of thought
• v1: "The normative framework does the heavy lifting — empirical estimates are necessarily speculative at this scale." Drifted: true of the x-risk numbers, no longer true of the corpus — since 2025 the empirical work stands on its own and is what the bench can grade
• Conflict note (CONVENTIONS rule 25): see Incentive map — low on the analytics, moderate on the risk numbers; DeepMind advisory logged

Weak spots / open questions

• 1-in-6 is a striking number but inherently hard to validate — how do you test a century-scale probability? Held as a weak spot; the file's answer is to grade the restatement cadence, not the number.
• FHI's closure raises questions about whether the institutional model for this work is sustainable. Drifted: the model migrated rather than died — AIGI (Oxford Martin School, 2023–), Forethought affiliation, GovAI board, OECD group (CV; AIGI page; Forethought page). The open question is now funding transparency (no AIGI funder named on its pages; Forethought's runway to mid-December 2026), not existence.
• Longtermist ethics is philosophically contested — population ethics can generate counterintuitive conclusions. Held, sharpened from inside: Bostrom's 2026 paper explicitly adopts "a person-affecting stance" and sets the impersonal considerations aside — the contest is now within the founding circle.
• Less engaged with near-term AI deployment questions — the frame is civilizational, not operational. Wrong: twelve dated quantitative pieces on frontier models since 2025-01-13, model-specific numbers (Sonnet 3.7 vs 3.5; o3 hourly cost), an arXiv paper on agent success rates, and "We should just ban unmonitorable chain of thought" as a policy ask (80,000 Hours, 2026-07-02). The June-2026 review carried a weak spot that had been false for 17 months; absence-shaped weak spots get the dated search (Doctorow rule).
• Risk estimates haven't been publicly updated since 2020 — would they shift given recent AI progress? Wrong: updated in direction on 2024-07-12 ("Climate risk is down, Nuclear is up …") and restated in total on 2025-01-31 ("still about right"). What remains open, and is logged as the live weak spot: he has not moved the AI number in either direction despite (a) his own view that language models grew "more capable even faster" than expected and (b) his own 2026 view that timelines are broader than he assumed in 2025. A 1-in-10 that survives both a speed-up and a slow-down is under-responsive to evidence by his own calibration standard.
• NEW: The Says/Does gap on the moratorium. He prefers a moratorium on tape and signed the red-lines call, but is not a named signatory of the FLI superintelligence-prohibition statement as far as the fetched pages show, and the statement site's name list could not be read. Resolve at next review: read the list or ask.
• NEW: The DeepMind relationship is 16 years old and undescribed. One line on a CV; no terms, no dates of activity, no disclosure on the scaling pieces that analyse DeepMind's competitors. Not a finding of conflict — a finding of opacity.
• NEW: His timelines are unscored by him. Cotra scores hers; Ord moved from "maybe even sooner" than 2027/2031 to a 2038 median without a published self-score. The bench applies the community's slow-on-capabilities bias as a correction until he does.
• NEW: The scaling analytics rest on one lab's disclosed pair. The 82/63/92% inference-scaling decomposition is computed on Sonnet 3.7 against Sonnet 3.5; generalisation to other labs is asserted, not shown. He says so; the file says so too.

Rich's take

• (your synthesis here)

Delta log

2026-09-21 — v1→v2 migration + validation (bulk run)

• Grades: 13 Held / 5 Drifted / 3 Wrong (4 items unfalsifiable — the normative moral-catastrophe claim, the 1-in-6 total and the AI 1-in-10 probability on their century horizon, and "mathematics is more than proof" — noted, not graded; the 2025–26 empirical claims carry dated checks rather than grades). Counted as the grade marks in the sections above, Delta log excluded.
• Wrong: agent half-life constant-hazard model — retired by its author (Hazard Rates, 2026-02-04); v1 weak spot "less engaged with near-term AI deployment" — twelve dated frontier-data pieces since 2025-01-13 and the chain-of-thought ban ask (80,000 Hours, recorded 2026-07-02); v1 weak spot "estimates not updated since 2020" — directions updated 2024-07-12, total restated 2025-01-31.
• Drifted: timeline stance ("those timelines sound very realistic. Maybe even sooner", AEI 2025-01-31 → median 2038, Broad Timelines, 2026-03-09); Bostrom split (communication → substantive, per Ord 2024-07-12 vs Bostrom 2026); Christiano split (Ord now does evals data; arXiv 2505.05115); FHI-sustainability weak spot (model migrated to AIGI/Forethought/GovAI, CV); "normative framework does the heavy lifting" (the empirical corpus now stands alone).
• Held: 1-in-6 on endorsement (AEI 2025-01-31); AI ~1 in 10 / "over 10%" (Faggella 2025-08-12); the precipice frame (2024-07-12; 80,000 Hours 2026); calibrated breakdowns, directions only (2024-07-12); consensus-shift half of the AI prediction (2024-07-12; FLI statement 2025-10-22); bio second (2024-07-12); Yudkowsky split on magnitude (Bostrom 2026 characterisation; Faggella); The Precipice as reference text; GWWC record ($500M+, 11,240 pledgers); FHI closure and move (final report; CV); estimates-are-frameworks; two carried weak spots (1-in-6 hard to validate; longtermism contested, now contested from inside by Bostrom's person-affecting turn).
• Most surprising delta: the author of the 1-in-6 number moved his own AI timeline from endorsing 2027/2031 as "very realistic. Maybe even sooner" (2025-01-31) to a 2038 median (2026-03-09) by writing a year of scaling analytics that say progress per dollar is "probably getting worse on most measures" — and the June-2026 v1 review, stamped three months after Broad Timelines, recorded none of it, while still carrying "not updated since 2020."
• Employer/role check (mandatory): no change — Senior Researcher, Oxford Martin AIGI, 2024– (AIGI; CV). Hats enumerated: Forethought affiliate, GovAI board (GovAI), OECD AI Futures group, DeepMind adviser since 2010 (CV) — the last two were absent from v1. Dated funding search run: no lab equity or paid role found; AIGI funder unnamed; Forethought endorsed by Coefficient Giving/Longview with runway to mid-December 2026. Incentive map re-derived; "unconflicted" not carried.
• Signature checks: red-lines call — signed (list); CAIS 2023 — not listed; FLI superintelligence statement — not named on the press release; full list unread — unresolved.
• Sibling-run lead (Bostrom file: "Ord thanked in both papers") verified in the PDFs: thanked in Open Global Investment (2025) and Optimal Timing (2026); not in the 2024 Cosmic Host paper. The substantive split is confirmed from Ord's own 2024 essay and 2026 moratorium preference; no direct Ord response to Bostrom's paper exists as of this review (LessWrong linkpost comments; 80,000 Hours transcript grep).
• Numbers re-read from primaries at migration: 1 in 1,000 / 1 in 1,000 / 1 in 30 / 1 in 10 (2024-07-12 restatement of the 2020 table); 82% / 63% / 92% (2025-10-03); 10x RL ≈ 3x inference (2025-10-20); 1,000–1,000,000x (2025-10-09); median 2038, 80% 3–100 years, 1 in 5 in 5 years (2026-03-09); 16x swarm ≈ 4.9x chain of thought, λ 0.48–0.68 (2026-09-21); GWWC $500M+, 11,240 (org page). All page dates read from tobyord.com and EA Forum page metadata, not from summaries. One search result associated him with arXiv 2601.02383 (AI summit series memo); he is not on its author list and it is not cited here.
• Tier: semiannual — no employer or product cycle; the output cadence is high but each piece is a dated number the wake triggers will catch; the timeline predictions have 2031 and 2038 horizons. Confidence unchanged at medium: grades rest on his own dated primaries, but the moratorium Says/Does gap and the DeepMind relationship are unresolved.

Sources

• 2026 primary: Swarm Scaling (2026-09-21) · The Dynamics of Intelligence Explosions (Forethought, 2026-08-28) · Mathematics is Much More than Proof (2026-08-23) · 80,000 Hours — "Toby Ord on where AGI timelines go wrong" (recorded 2026-07-02; transcript) · EA Forum crosspost (2026-08-06) · "Taking longer AI timelines seriously", EAG San Francisco 2026 (uploaded 2026-03-27) · Broad Timelines (2026-03-09; EA Forum crosspost 2026-03-19) · Hazard Rates for AI Agents Decline as a Task Goes On (2026-02-04) · Are the Costs of AI Agents Also Rising Exponentially? (2026-02-02) · The Scaling Series discussion thread (2026-02-02)
• 2025 primary: How Well Does RL Scale? (2025-10-20) · The Extreme Inefficiency of RL for Frontier Models (AIGI, 2025-10-09) · Evidence that Recent AI Gains are Mostly from Inference-Scaling (2025-10-03) · Evaluating the Infinite (2025-09-21) · Faggella, AGI Governance ep. 7 (2025-08-12) · 80,000 Hours — "graphs AI companies would prefer you didn't (fully) understand" (recorded 2025-05-23) · Is There a Half-Life for the Success Rates of AI Agents? (2025-05-07; arXiv 2505.05115) · Inference Scaling Reshapes AI Governance (arXiv v1 2025-02-12) · The Scaling Paradox (2025-01-13) · AEI "Faster, Please!" #69 transcript, Pethokoukis (2025-01-31)
• 2024 primary: The Precipice Revisited (2024-07-12) · On the Value of Advancing Progress (2024-07-12) · FHI 2005–2024 Final Report (2024-04-17)
• Role / institution: tobyord.com (bio, affiliations) · writings index · CV (positions, education, advisory roles, GWWC/CEA dates) · Oxford Martin AIGI people page · AIGI About (co-leads) · Oxford Martin School programme announcement (2023-09-04) · Forethought people page · Forethought fundraiser (funding, runway) · GovAI people (board) · Giving What We Can — About · GWWC 10,000 pledgers (2025-08-26)
• Books: The Precipice: Existential Risk and the Future of Humanity (Bloomsbury, 2020) — load-bearing; Moral Uncertainty (with MacAskill and Bykvist, OUP, 2020) — per CV
• Peer-reviewed / academic (per CV): "The reversal test: eliminating status quo bias in applied ethics" (with Bostrom, Ethics 116, 2006) · "An upper bound for the background rate of human extinction" (with Snyder-Beattie and Bonsall, Scientific Reports, 2019) · "Probing the improbable" (with Hillerbrand and Sandberg, J. Risk Research, 2010) · "Moral Trade" (Ethics, 2015) · "Interpolation, Extrapolation, Hyperpolation" (arXiv 2409.05513, 2024) · "The Lindy Effect" (arXiv 2308.09045, 2023) · "Dissolving the Fermi Paradox" (with Sandberg and Drexler, arXiv 1806.02404) · "Shaping humanity's longterm trajectory" (forthcoming, Essays on Longtermism, OUP)
• Policy (per CV): "Future Proof" (Centre for Long-Term Resilience, 2021) · "Existential risks to humanity" (UNDP Human Development Report 2020) · advice to the Prime Minister's Office (2012–16, 2020), Cabinet Office (2014, 2020–23), UN Secretary-General's Office (2021), WHO (2012–14), OECD Expert Group on AI Futures (2023–)
• Evidence used for grades and checks: Bostrom, Optimal Timing for Superintelligence (working paper v1.0, 2026; acknowledgments read) · Bostrom, Open Global Investment (2025; acknowledgments read) · LessWrong linkpost (2026-02-13; no Ord comment) · Global Call for AI Red Lines — individual signatories · red-lines.ai (launch: UNGA, September 2025) · CAIS Statement on AI Risk — signatory page · FLI press release, Statement on Superintelligence · EA Forum crosspost (2025-10-22)
• Talks (v1 list, carried): TED; EA Global keynotes (Bay Area 2023-02-25; London 2020, per CV); Oxford public lectures; Durham Union (2023-10-02, per CV)
• Not fetched or not readable this run (logged, not cited for facts): Coefficient Giving grant page for BERI/AIGI (HTTP 403); superintelligence-statement.org name list (not rendered); EAG SF 2026 talk transcript (video only)

migrated v1→v2 2026-09-21 · backup: _archives/toby-ord.html.bak-20260921 · converts-from: people/toby-ord.md · AI & Society domain