SABINE HOSSENFELDER: Expert panels surveyed in 2025 and 2026 have already underestimated AI progress in math and coding

The Gist

Sabine Hossenfelder says experts surveyed just last year already got AI badly wrong. They gave AI a 10 percent chance of cracking a famous unsolved math problem by 2027 and expected modest coding scores, but AI produced a disputed candidate proof in September 2026 and nearly quadrupled the coding forecast by May. One forecast, on how long a task AI can handle, was nearly matched while the survey was still running, but it has not yet been passed. This packet steelmans the argument Sabine Hossenfelder makes in this video, in its best light, for analysis; it is not an endorsement of her conclusions.

Conclusion

Expert panels surveyed in 2025 and 2026 have already underestimated AI progress in mathematics and coding, in one case by a factor of nearly four and in another, on a Millennium Prize Problem, within about a year if the claimed proof holds, so median expert forecasts of near-term AI capability in these domains should be treated as likely to run low.

Premises

1. In August and September 2025, a panel of 277 experts gave a median 10 percent chance that AI would solve or substantially assist in solving a Millennium Prize Problem by the end of 2027.
2. In September 2026 OpenAI announced an AI-produced proof bearing on the Navier-Stokes Millennium Prize Problem, which the survey's organizers say appears likely to count as a significant example of forecasters underestimating AI progress.
3. The same panel forecast a median 14 percent state-of-the-art score on LiveCodeBench Pro (Hard) by the end of 2026, from a baseline of 7.7 percent, and a model reached 53.8 percent in May 2026.
4. In April and May 2026 experts forecast a 3.4 hour task-completion time horizon by the end of 2026, and a model was reported at 3 hours 6 minutes on May 8, while the survey was still open, nearly matching the year-end forecast within weeks.
5. These misses run in the same direction, which matches the survey organizers' own finding that forecasters dramatically underestimate AI progress on benchmarks.

Assumptions

• The Navier-Stokes result is a claim under review: independent machine checks accept OpenAI's Lean formalization, but mathematicians found two places where the written proof and the code diverge, the Clay Institute has not accepted it, and the forecasting group that asked the question says it is not completely clear the result meets its resolution criteria.
• The LiveCodeBench Pro question formally resolves at the end of 2026, and the forecasting group notes that some forecasters may not have understood that the benchmark lets models be tested on earlier quarters' problems, which may have depressed their forecasts.
• The 3 hour 6 minute time horizon reported in May 2026 is still below the 3.4 hour median forecast for the end of 2026, and as of October 2026 no higher 80 percent horizon has been published, so that forecast is on track to be too low only if the trend continues; it has not yet been shown wrong.
• 'Experts' here means the median of surveyed panels; individual experts and subgroups who expected faster progress were more accurate.

Analysis

Argument type: Inductive.

Potential Fallacies

• Premature resolution / overstated status (Conclusion and title; P2 and P4) — The word 'already' presents all three cases as realized underestimates. By the stated assumptions, the Navier-Stokes result is under review and the 3h06m horizon is still below the 3.4h forecast. Only the LiveCodeBench item has clearly exceeded its forecast, and even that formally resolves at year-end. The conditional clause 'if the claimed proof holds' softens this for one case, but not for the title or the time-horizon case.
• Hasty generalization / selection of favorable cases (Inference from P1–P4 to the conclusion; P5) — Three striking items from one survey program are used to support a rule about median forecasts. The argument does not report how many questions were asked, how many were accurate, or whether any were overestimates. The multi-year review in P5 mitigates this considerably, but its link to these particular items is asserted rather than shown.
• Treating a single low-probability event as miscalibration (P1 and P2) — A 10 percent forecast is not shown wrong by one occurrence, since such events should happen about one time in ten. Systematic underestimation can only be shown through calibration across many questions. This case is weak evidence on its own, particularly while its resolution is unsettled.
• Non-independent evidence / double-counting (P3, P4, P5 taken together) — The misses come from the same panel and organizers, in the same period, and share a common cause in rapid gains in reasoning and coding models. P5 is partly authored by the same organizers who ran the surveys. The effective number of independent observations is closer to one or two than to four or five.
• Equivocation on 'experts' and 'underestimated' (Title, conclusion, P5) — 'Experts' moves between the panel median, individual forecasters, and the wider forecasting community, and A4 concedes that some experts were accurate. 'Underestimated' moves between a resolved miss, a probable miss, and a fast trajectory. This lends the weaker claims the force of the stronger ones.

Counterarguments

These challenges are not ranked.

• Premise 2 (High impact) — The Navier-Stokes claim may fail review. The Clay Institute has not accepted it, the written proof and code diverge in two places, and the organizers doubt it meets their criteria. If it fails, the most dramatic example disappears and the 10 percent forecast is left unrefuted. Even if it holds, one realization of a 10 percent event is consistent with good calibration.
• Premise 3 (Medium impact) — Part of the factor-of-four gap may come from forecasters misunderstanding that earlier-quarter problems are allowed, and a low 7.7 percent baseline makes ratios look larger. The miss may therefore measure question design as much as expert judgment about AI capability.
• Premise 4 (High impact) — The 3h06m figure is below the 3.4h forecast, and the forecast resolves at year-end. If progress plateaus, the forecast could prove accurate. Reaching a value early does not falsify a forecast of a later date, and forecasters may have already known the May figure.
• Conclusion (High impact) — The cases may have been selected because they were misses. Without the full set of resolved questions, including accurate and overestimated forecasts, three correlated items cannot establish a rate. A4 also shows that the problem may lie in aggregating a heterogeneous panel rather than in expert judgment, so the better remedy may be weighting better forecasters rather than shifting all forecasts upward.
• Conclusion (Medium impact) — Past underestimation does not guarantee future underestimation. Forecasters may update after being wrong, benchmarks may saturate or be contaminated, and math and coding are verifiable domains where progress may not carry over to other capabilities. Overcorrection is also possible.

Suggested Improvements

• Calibrated headline wording — Replace 'have already underestimated' with a tiered statement: one clear miss (LiveCodeBench Pro Hard), one contested case (Navier-Stokes), and one on-track but unresolved case (time horizon). This matches the title to the evidence and keeps the hedged conclusion consistent with the body.
• Denominator of forecast questions — Report how many questions the panels answered, how many resolved above, near, and below the median, and calibration scores such as Brier or log scores. This would address cherry-picking and show whether the three cases are typical or merely striking.
• Separating comprehension error from judgment error — Estimate how much of the LiveCodeBench gap remains after excluding forecasters who misread the earlier-quarter rule, or after restricting to current-quarter problems. This would show how much of the factor of four reflects real misjudgment of capability.
• Wait for resolution — Update the Navier-Stokes case after Clay and mathematician review, and the time-horizon case after year-end measurements. Treat both as pending until then. Counting open cases as realized misses inflates the apparent hit rate.
• Scope and subgroup analysis — Compare the median with the faster-progress subgroups and with a naive trend extrapolation, and limit the claim to benchmarked math and coding capability. This would show whether the lesson is to correct medians or to weight better forecasters, and would prevent overgeneralization to other domains.

Scenario Tests

• The Clay Institute or independent mathematicians reject the Navier-Stokes proof because of the proof/code divergences. (Challenges) — The headline Millennium Prize example collapses and the 10 percent forecast is not shown wrong. The conclusion would rest on the benchmark cases and the multi-year review.
• The task-horizon trend continues and a model exceeds 3.4 hours before the end of 2026. (Supports) — The time-horizon forecast becomes a genuine underestimate, and the pattern of three misses in one direction becomes more credible.
• Progress plateaus in late 2026 and the horizon stays near or below 3.4 hours. (Challenges) — The time-horizon case becomes an accurate forecast, leaving LiveCodeBench, itself partly confounded, as the only clear miss.
• A fuller review of all panel questions shows that medians were too low on most resolved benchmark items, including after excluding misinterpreted questions. (Supports) — The general conclusion would rest on solid calibration evidence rather than selected anecdotes, which justifies treating medians as likely to run low.
• Forecasters in later surveys raise their medians substantially after the 2025–2026 misses. (Neutral) — The historical bias would be real but the forward-looking rule might be stale. A fixed upward correction could then overshoot.

Coherence & Relevance

The structure is coherent: three case-level observations are backed by a broader base-rate review and lead to a hedged, probabilistic conclusion. The stated assumptions candidly flag the unresolved status of two cases, the possible comprehension problem on the benchmark, and the distinction between medians and better-calibrated subgroups. The main weakness is a gap between the confident title and 'already' and the more provisional evidence. Read as 'panel medians on fast-moving AI benchmarks have tended to run low, with one clear miss and two pending cases', the argument is reasonably well supported. Read as 'experts have already been shown wrong three times', it overreaches.

• In August and September 2025, a panel of 277 experts gave a median 10 percent chance that AI would solve or substantially assist in solving a Millennium Prize Problem by the end of 2027. (Moderate) — This establishes the forecast but not the error. One resolution of a 10 percent event does not show miscalibration.
• In September 2026 OpenAI announced an AI-produced proof bearing on the Navier-Stokes Millennium Prize Problem, which the survey's organizers say appears likely to count as a significant example of forecasters underestimating AI progress. (Moderate) — Relevant only if the result holds and meets the resolution criteria, which remains unresolved. The organizers' own hedge is not reflected in the premise.
• The same panel forecast a median 14 percent state-of-the-art score on LiveCodeBench Pro (Hard) by the end of 2026, from a baseline of 7.7 percent, and a model reached 53.8 percent in May 2026. (Strong) — Directly supports the factor-of-four claim. It does not separate forecasting error from misreading of the question rules, and it covers one benchmark.
• In April and May 2026 experts forecast a 3.4 hour task-completion time horizon by the end of 2026, and a model was reported at 3 hours 6 minutes on May 8, while the survey was still open, nearly matching the year-end forecast within weeks. (Weak) — Shows a rapid trajectory, not an error. The forecast has not been exceeded, and the data point may have been visible to forecasters.
• These misses run in the same direction, which matches the survey organizers' own finding that forecasters dramatically underestimate AI benchmark progress. (Strong) — Provides the base-rate support for the generalization. It counts unresolved cases as misses, overlaps with the organizers who ran the surveys, and concerns benchmarks more than proofs.

View this argument on LogicFirst.ai