SABINE HOSSENFELDER: The fastest near-term progress will be in software-only work, including theory problems that can be solved by synthesizing existing literature

The Gist

Sabine Hossenfelder predicts the fastest AI progress will be in work done entirely on computers, like math and theoretical physics, where AI can read all the papers and connect ideas people missed. She thinks anyone can join in, not just AI companies. The biggest results so far, though, have come from expensive, unreleased company models. 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

Near-term progress from AI will be fastest in work that software alone can do, including theoretical problems in physics and related fields that may be solvable by combining techniques already in the literature, and because AI tools are widely available, researchers outside AI companies can contribute to that progress.

Premises

1. In 2026 AI systems produced a candidate proof bearing on a Millennium Prize Problem and research-level mathematical solutions verified by experts, entirely in software.
2. AI systems can read and cross-reference volumes of scientific literature that no person could read in a career.
3. Some open theoretical problems may be solvable by combining techniques that already exist in the literature but have not been brought together.
4. Capable AI models are available to the public, so individuals and small groups can apply them to research problems.

Assumptions

• Whether open physics problems are solvable with techniques already in the literature is a forecast, not a finding; panelists who gave low Millennium Prize forecasts argued such problems need new concepts, and that view has not been tested outside the Navier-Stokes claim.
• Frontier results so far have depended on large, private compute: OpenAI's Navier-Stokes run used about 130 billion output tokens on an unreleased internal model, with published cost estimates of roughly $6.5 million to $15 million, so participation by everyone depends on wider access to comparable models.
• The closing remark that a period in which everyone is a scientist may be followed by extinction or an end to work reports what others predict; the argument takes no position on it.

Analysis

Argument type: Inductive.

Potential Fallacies

• Unsupported comparative claim (Conclusion (comparative 'fastest' claim) relative to P1–P4) — 'Fastest' is a comparison across domains, but no premise addresses progress rates in experimental, wet-lab, or hardware-dependent work. The premises show that software-only work can advance, not that it advances faster than alternatives such as AI-enabled protein or materials research with automated labs. The metric and time horizon are also undefined.
• Hasty generalization from a few vivid cases (P1 → Conclusion) — A small number of headline results, with no count of failed attempts or base rates, is used to project a general trend. Without the denominator of unreported failures or incorrect claims, the evidential weight of the successes cannot be judged.
• Equivocation on 'capable' and 'available' AI (P4 → Conclusion (outsider contribution clause)) — The headline results came from unreleased frontier systems, while the access claim concerns publicly available models. The inference to outsider contribution requires public models to be capable of comparable work, which is conditional on wider access rather than established. Accepting the stated compute assumption, the conclusion's wording ('because AI tools are widely available') is still stronger than that conditional.
• Possibility-to-prediction slide (P3 → Conclusion) — P3 says only that some problems 'may be' solvable by recombination. The conclusion uses this to support a firmer forecast about where progress will be fastest. A modal possibility says nothing about how many problems are in that class or how important they are.
• Unmarked domain transfer (mathematics to physics) (P1 → Conclusion ('theoretical problems in physics and related fields')) — The evidence is mathematical, where proofs can be checked formally or by expert review. The conclusion extends to theoretical physics and related fields, where results must ultimately fit observation and community-accepted frameworks. The difference in validation is not acknowledged.
• Conflation of evidential statuses (P1) — A 'candidate' proof, which is unadjudicated, is placed alongside expert-verified solutions, which have been checked. Readers unfamiliar with how Millennium Prize claims are vetted may give the candidate proof the weight of an accepted result.

Counterarguments

These challenges are not ranked.

• Conclusion (software-only 'fastest') (High impact) — Data-rich empirical domains with closed-loop feedback, such as protein design and materials discovery with automated labs, may advance faster than theory. Software-only progress is also limited by verification and acceptance, which depend on scarce human expert time, so the pace of validated progress may be slower than the pace of generated candidates.
• Premise 3 and Conclusion (recombination) (High impact) — Hard open problems, Millennium Prize problems in particular, may need genuinely new concepts. If so, the recombination route will stall quickly, or produce only a short burst of results on a finite stock of 'low-hanging' problems before plateauing.
• Premise 4 and Conclusion (outsider contribution) (High impact) — The frontier result used an unreleased model and millions of dollars of compute. Public models are a tier below, so outsiders may contribute incrementally at best. Effective contribution also requires domain expertise to judge outputs, and a surge of low-quality AI-assisted submissions could burden reviewers and net-slow progress.
• Premise 1 (Medium impact) — Candidate proofs of Millennium Prize problems have a long history of failing scrutiny. If the proof is flawed, or the verified solutions depended heavily on human steering, the central evidence is undercut. The success may also reflect mathematics being unusually favorable to AI because verification is cheap.
• Conclusion (physics theory) (Medium impact) — Theoretical physics is ultimately adjudicated by experiment. Software-synthesized theories that cannot be tested may amount to output that does not accumulate into knowledge, so speed of generation is not speed of progress.

Suggested Improvements

• Comparative evidence — Define 'progress' and a time horizon, then compare time-to-result across software-only fields and lab-dependent fields such as biology, materials, and robotics. The 'fastest' claim is the headline and currently has no comparative support.
• Evidence quality for P1 — Separate the status of the candidate Millennium proof from that of the expert-verified solutions. Report who verified them and how, and give the success rate against attempts, including failures. This would remove selection ambiguity and let readers weigh the strongest and weakest evidence separately.
• Capability tiers for outsider participation — Show examples or benchmarks where publicly available models produce novel, verified results on research-level tasks, or restate the claim as conditional on wider access to frontier models. This would bridge the gap between the compute-heavy frontier results and the claim that outsiders can contribute.
• Domain bridge from math to physics — Explain how software-only results in physics would be validated against data or accepted frameworks, and identify concrete physics problems plausibly in the recombinable class. It would make the extension beyond mathematics credible.
• Verification and filtering — Address verification capacity, for example through formal proof assistants and reviewer scaling, and the expertise needed by outside contributors. Verification is likely the binding constraint on validated progress and on whether outsider contributions help or add noise.

Scenario Tests

• The candidate Millennium Prize proof is found flawed, or the Navier-Stokes result turns out to rely on a new concept not in the prior literature. (Challenges) — The main exemplar for the literature-synthesis thesis would be lost. The weaker claim that software-only work progresses quickly on formally checkable problems could survive through the other verified solutions.
• Independent researchers using public models obtain novel, expert-verified results on open theory problems within a year. (Supports) — This would directly support the outsider-contribution clause and show that public models are capable enough to matter.
• Frontier-capable models stay gated by cost or private release, and public models cannot produce research-level results. (Challenges) — The democratization half of the conclusion would fail even if software-only work remains fastest, as the stated compute assumption anticipates.
• AI-driven closed-loop labs in biology or materials report faster validated discovery rates than theory work. (Challenges) — The comparative claim would be false, and the speed advantage of software-only work would not be general.
• Several additional open problems in mathematics and formal theoretical physics are solved by AI-assisted recombination of known techniques and independently verified. (Supports) — This would strengthen the recombination premise from a single disputed case to a pattern, and bear on the dispute with skeptics who say new concepts are needed.
• Journals and arXiv see a large surge of AI-assisted outsider submissions that overwhelms review and mostly contains errors. (Challenges) — Outsider participation would increase noise more than progress in the near term, and verification would become the limiting step.

Coherence & Relevance

The argument is coherent as a hedged, conditional forecast: software-only work avoids physical bottlenecks, AI can read widely, some problems may be recombinable, and public tools allow some participation. Its stated assumptions candidly acknowledge the main uncertainties. The conclusion is, however, stated more firmly than the premises support. The comparative 'fastest' claim has no comparative premise, the inference from math to physics is unmarked, and the democratization clause depends on wider access to frontier-level models. The best-supported reading is that software-only, formally checkable work is likely to advance quickly in the near term and that some outside participation is possible, with broader claims remaining forecasts.

• In 2026 AI systems produced a candidate proof bearing on a Millennium Prize Problem and research-level mathematical solutions verified by experts, entirely in software. (Moderate) — Demonstrates that software-only AI work can reach research level, but gives no comparison with other domains, no base rate of failures, and no bridge from mathematics to physics. The candidate and verified results have different evidential status.
• AI systems can read and cross-reference volumes of scientific literature that no person could read in a career. (Moderate) — Supplies the mechanism for recombination but not evidence that breadth of reading produces valid or novel syntheses.
• Some open theoretical problems may be solvable by combining techniques that already exist in the literature but have not been brought together. (Moderate) — It is modal and says nothing about the size of the recombinable class. It supports the theory clause of the conclusion only as a possibility, not as a forecast about where progress will be fastest.
• Capable AI models are available to the public, so individuals and small groups can apply them to research problems. (Moderate) — Relevant to the outsider clause, but the models that produced P1's results are not the public ones. The premise also leaves out the expertise needed to direct and verify AI output.

View this argument on LogicFirst.ai