We've seen dramatic progress (or apparent progress) in math and science from general purpose AI systems in Sept and Oct 2026. The most recent one, OpenAI's models disproving something widely believed to be true (the Exponential Time Hypothesis), hints we might get novel results, not just incremental refinements on what's widely known (such as with the Riemann Hypothesis or Navier Stokes).
Are we on the cusp of a deluge of scientific progress? Or are these cherrypicked examples that won't translate into the real labs? I did some research on trajectories, iterated on a series of forecasting questions, and put the results into a world model. I find these FutureSearch forecasts convincing, and stand by the claim in the title: I predict that within 2 years (by Oct 2028), we will have a major advance (defined as a "top of the year" result) from a general AI system in every major scientific field.
(If anything, I think the 80% CIs here are too narrow, and I would not be too surprised to see a Nobel-worthy result in, say, astronomy in the last months of 2026.)
I also predict that we'll continue to see rapid progress in the theoretical sciences: math, CS, and physics. FutureSearch gave the following dates for known plausible milestones:
Like everyone else, I expect that results that depend more on the physical world come later. I do, however, think those barriers are not as severe as people think (and evidently the frontier labs agree, as they've been acquiring and in-housing more types of labs).
All 49 forecasts are below, followed by the World Model they were reconciled in, and the fact sheet of all the research on these trends across fields. The questions use these definitions:
- "Produced by AI" a general-purpose AI system (a frontier language model, or an agent or lab built on one) produced the central idea and the main result. Humans may pose the problem, run experiments, check the work and write it up. A purpose-built model such as AlphaFold counts only as a tool of a general system.
- Dates in results: by the result's first public release (preprint, paper or announcement), not by when it was verified.
- "Top-of-year result:" specialists rank it among the most important results in their field that year, with examples given in each field sourced from 2024-2025
- "Nobel-caliber:" the result is cited in the award of a Nobel Prize or another of the field's top prizes (such as the Wolf Prize, the Lasker Awards, the Shaw Prize or the John Bates Clark Medal, or the Turing Award for ML), or leading researchers widely call it a leading candidate for one had a human produced it.
- How a result is produced: reasoning (a general model working from existing literature and theory), mining data (an AI system mining existing datasets, archives or code), AI designs, humans run (an AI system designing experiments, molecules or materials that humans then run or make), robotic lab (an AI system running a closed-loop autonomous lab or instrument), or specialist model (a general system directing a purpose-built specialist model).
1Which famous open problem AI resolves firstNavier–Stokes: 29%
Which of these famous open problems will AI be the first to resolve completely?
2AI resolves a famous open problem (Riemann, Navier–Stokes, P vs NP…)Aug 2028
When will AI first completely resolve one of these famous open problems: the Riemann Hypothesis, unforced 3D Navier–Stokes, twin primes, Birch–Swinnerton-Dyer, the Hodge conjecture, P vs NP, the abc conjecture, the Collatz conjecture or the Yang–Mills mass gap?
3An AI-proved theorem is the key step in a top result in another scienceJan 2029
When will a theorem proved by AI first be the key step in one of the most important results of its year in an empirical science?
4A concept introduced by AI organizes a research areaSep 2028
When will AI first introduce a mathematical concept that becomes an organizing concept of a research area?
5AI uplift for research mathematicians, end of 20292.7x
By what factor will AI raise the rate of valuable research output of a typical research mathematician at a top-50 mathematics department, as of 2029-12-31?
6AI breaks the claimed security of RSA, ECC, AES, SHA-256 or ML-KEMOct 2028
When will AI first produce a classical attack that breaks the claimed security level of RSA-2048, 256-bit elliptic-curve cryptography, full AES-128, full SHA-256, or any ML-KEM or ML-DSA parameter set?
7An AI design cuts the record quantum cost of breaking RSA or ECCAug 2027
When will the lowest published estimate of the quantum resources needed to break RSA-2048 or 256-bit elliptic-curve cryptography first rely on a circuit or error-correcting code that AI designed?
8An AI-invented algorithm becomes a core default in major open-source softwareFeb 2028
When will an algorithm invented by AI first become the default method in a widely used open-source software project?
9AI proves a widely deployed software system correctMar 2028
When will AI first produce a complete formal proof that a widely deployed software system is correct?
10AI-designed error correction beats human designs on a quantum computerFeb 2028
When will a quantum computer first run an error-correcting code or fault-tolerant gate scheme that AI designed, with a logical error rate at least 20% lower than the best human-designed scheme run on the same device?
11Experiment confirms a prediction made by AIOct 2028
When will an experiment first confirm a prediction made by AI of a new particle, phase of matter or physical effect?
12AI predictions confirmed by experiment by 20303
How many distinct predictions made by AI of a new particle, phase of matter or physical effect will experiments have confirmed by 2030-12-31?
13First top-of-year AI resultSep 2028
When will AI first produce one of the most important astronomy or cosmology results of its year?
14First Nobel-caliber AI resultOct 2029
When will AI first produce a Nobel-caliber astronomy or cosmology result?
15How the first Nobel-caliber result will be producedData mining: 37%
If AI produces a Nobel-caliber astronomy or cosmology result, how will the first one most likely have been produced?
16AI finds a new class of astronomical objectJun 2030
When will AI first discover a new class of astronomical object or transient?
17New object classes found by AI by 20301
How many new classes of astronomical object or transient discovered by AI will have been reported by 2030-12-31?
18Which cosmology puzzle AI helps resolve firstNone by 2035: 42%
Which of these open questions in astronomy and cosmology will AI be credited with resolving first?
19First top-of-year AI resultApr 2028
When will AI first produce one of the most important chemistry results of its year?
20First Nobel-caliber AI resultApr 2029
When will AI first produce a Nobel-caliber chemistry result?
21How the first Nobel-caliber result will be producedAI-designed: 33%
If AI produces a Nobel-caliber chemistry result, how will the first one most likely have been produced?
22Machine-learned chemistry reaches chemical accuracy on hard systemsFeb 2029
When will a machine-learned electronic-structure method first reach chemical accuracy on transition-metal or strongly correlated chemistry at no more than hybrid-DFT cost?
23Which chemistry problem AI resolves firstReaction yields: 32%
Which long-standing chemistry problem will AI resolve first?
24First top-of-year AI resultJun 2028
When will AI first produce one of the most important materials-science results of its year?
25First Nobel-caliber AI resultApr 2030
When will AI first produce a Nobel-caliber materials-science result?
26How the first Nobel-caliber result will be producedAI-designed: 33%
If AI produces a Nobel-caliber materials-science result, how will the first one most likely have been produced?
27Highest Tc of an AI-proposed non-cuprate superconductor by 203012 K
What will be the highest critical temperature at ambient pressure of any non-cuprate superconductor that AI first proposed, as reported by 2030-12-31?
28Material categories where AI holds the record in 2030 (of 5)1.5
In how many of these five material categories will an AI-proposed material hold the published record on 2030-12-31: superconducting critical temperature among non-cuprate superconductors at ambient pressure; energy product among rare-earth-free permanent magnets; durability among iridium-free acid water-splitting catalysts; room-temperature ionic conductivity among solid lithium-ion conductors; CO₂ working capacity in humid conditions among direct-air-capture sorbents?
29First top-of-year AI resultAug 2028
When will AI first produce one of the most important biology results of its year?
30First Nobel-caliber AI resultNov 2030
When will AI first produce a Nobel-caliber biology result?
31How the first Nobel-caliber result will be producedData mining: 35%
If AI produces a Nobel-caliber biology result, how will the first one most likely have been produced?
32Which unsolved question in biology AI answers firstNone by 2040: 40%
Which of these unsolved questions in biology will AI answer first?
33A virtual cell predicts unseen perturbations at replicate accuracySep 2030
When will a model first predict how an unseen cell type responds to an unseen single-gene perturbation as accurately as a technical replicate?
34AI uplift for wet-lab biology groups, end of 20291.85x
By what factor will AI raise the rate of valuable research output of a typical wet-lab biology group at a top-50 research university, as of 2029-12-31?
35First top-of-year AI resultJun 2029
When will AI first produce one of the most important medicine results of its year?
36First Nobel-caliber AI resultOct 2030
When will AI first produce a Nobel-caliber medicine result?
37How the first Nobel-caliber result will be producedAI-designed: 37%
If AI produces a Nobel-caliber medicine result, how will the first one most likely have been produced?
38A trial succeeds on a new target AI nominatedAug 2031
When will a randomized trial first meet its primary efficacy endpoint for a therapy against a new target that AI nominated?
39Successful trials on new AI-nominated targets by 20300
How many randomized trials will have met their primary efficacy endpoint for therapies against new targets that AI nominated, by 2030-12-31?
40An AI-discovered new-class antibiotic works in humansJun 2035
When will an antibiotic of a new class, discovered by AI, first show efficacy in a human trial against a WHO critical-priority Gram-negative pathogen?
41First top-of-year AI resultAug 2028
When will AI first produce one of the most important social-science results of its year?
42First Nobel-caliber AI resultJun 2030
When will AI first produce a Nobel-caliber social-science result?
43How the first Nobel-caliber result will be producedReasoning: 35%
If AI produces a Nobel-caliber social-science result, how will the first one most likely have been produced?
44AI-run experiments find a better theory of human behaviourFeb 2028
When will an AI system that designs and runs its own human experiments first discover a better theory of a classic behaviour, replicated by an independent lab?
45Which long-running dispute AI settles firstNone by 2035: 46%
Which long-running social-science dispute will AI-led analysis be credited with settling first?
46First top-of-year AI resultMar 2028
When will AI first produce one of the most important machine-learning research results of its year?
47First Turing Award-caliber AI resultJun 2029
When will AI first produce a Turing Award-caliber machine-learning research result?
48First field beyond math, CS and physics with a top AI resultBiology: 24%
Outside mathematics, theoretical computer science and theoretical physics, in which field will AI first produce one of the most important results of its year?
49AI-produced results in Science's 2030 Breakthrough of the Year top ten4.1
How many of the ten results on Science's 2030 Breakthrough of the Year list will AI have produced?
Bars show the 80% interval (p10 to p90), the box the middle 50%, the dot the median, absent where the median is never. Date rows share one 2026 to 2047 axis; a dashed arrow means the distribution puts more than 10% on the event never happening. Click a row for the question, its distribution in detail, and the link to the published forecast with the full research trail.
Open the world-model explorer in its own tab
Data on trajectories in each field based on recent progress
Condensed raw notes on progress from April 2025 to 2026-10-27. Research done by LLMs, guided by me (Dan Schwarz) to produce the dense trajectory numbers I thought would be most useful in the forecasts.
- Across fields. METR 50% time horizon: 2.0 h (o3, Apr 2025) → 3.4 h (GPT-5, Aug 2025) → 5.9 h (GPT-5.2, Dec 2025) → 17.4 h (Mythos Preview, Apr 2026), about one doubling every four months; METR stopped updating the series in Sept 2026. Humanity's Last Exam: 20% (o3, Apr 2025) → 32% (GPT-5 Pro, Oct 2025) → 46% (Gemini 3.1 Pro, Feb 2026) → 55% (GPT-6 Astra, Sept 2026). Largest parallel effort behind one research result: 200 rollouts (Kosmos, Nov 2025) → about 10,000 concurrent agents for 88 hours (forced Navier–Stokes, Sept 2026). Cost at a fixed capability level falls about 47% a quarter (Epoch, Sept 2026).
- Mathematics. FrontierMath Tier 4: 4.9% (o3, Apr 2025) → 22% (GPT-5, Aug 2025) → 46% (Dec 2025) → 78% (Apr 2026) → 100% (GPT-6.1 Sol, Sept 2026). Results: AlphaEvolve improves known constructions (May 2025) → Erdős #1196 and the unit-distance disproof (Apr–May 2026) → the Jacobian conjecture refuted in dimension ≥3 (Jul 2026) → at least 2/3 of zeta zeros simple and on the critical line, against a previous record of 5/12 (Claude, Aug 2026; independently reproved) → forced Navier–Stokes blowup (Sept 2026) and the quasi-Riemann zero-free region, among 719 manuscripts from about 4,000 posed problems (OpenAI, Oct 6, 2026) → Sogge's local smoothing conjecture in every dimension, implying the Kakeya, Bochner–Riesz and restriction conjectures; humans had reached dimension 2 (2019) and, for Kakeya, dimension 3 (2025) (OpenAI, released 2026-10-26, Lean-formalized, unrefereed). Since August most of the flow is from outside mathematicians using public models: math arXiv papers crediting a model with a result went from about 9 (Jun) to 50 (Sept 2026). First Proof (10 unpublished research problems, 24 hours): 7/10 (tested Jun 2026) → 9/10 (tested Oct 2026). Open problems solved autonomously at a $300 budget (FrontierMath Erdős): 0 of 68 (every model through Aug 2026) → 3 of 68 (GPT-6.1 Sol, Sept 2026).
- Cryptography and theoretical CS. AlphaEvolve's 48-multiplication 4×4 complex matrix product (May 2025) → AI supplying single steps of human proofs (Sept 2025) → no AI-credited TCS result found (Nov 2025–Jun 2026) → 20+ credited results since late July 2026, mostly from outside researchers using public models → claimed proofs of the Unique Games Conjecture and L = BPL (OpenAI, Oct 6, 2026, unrefereed) → a claimed refutation of the Exponential Time Hypothesis: 3-SAT in 2^{o(n)} time, against the best human ~1.31^n. Internal models at both OpenAI and Anthropic found it on 2026-10-16 (both Lean-formalized, released 2026-10-26, unrefereed). Cryptanalysis: models broke 65–86% of known-broken schemes on CryptanalysisBench (Jul 2026); HAWK-256 key recovery cut from 2^64 to 2^38 (Jul 2026, an undeployed scheme); no AI break of a deployed primitive (the ETH algorithm beats brute force only beyond ~10^10 variables).
- Theoretical physics. CritPt (unpublished research problems): 1.4% (Apr 2025) → 12.6% (GPT-5, Aug 2025) → 30% (Mar 2026) → 32.3% (GPT-5.6 Sol, Jul 2026), flat since. TheoPhysEval (50 unpublished problems): 49 of 50 (GPT-6 Astra, Sept 2026). Results: one case of a frustrated Potts model (Mar 2025) → a gluon-amplitude formula conjectured by GPT-5.2 Pro, then proved (Feb 2026) → θ(p_c) = 0 for percolation in dimensions 3–10, Lean-proved (Claude, Aug 2026) → a nine-loop N=4 amplitude validated by Lance Dixon, for about $100 of compute (Sept 2026) → quantum-information theorems from outside researchers using public models (Aug–Oct 2026) and circuits with up to 17× fewer gates than the best published designs (Oct 2026, simulation only) → a claimed rigorous construction of 4D Yang–Mills theory and QCD (existence, not the mass gap; OpenAI, announced 2026-10-26). The proofs are withheld while a group led by David Gross writes them up, so the claim is unverified. Physical world: an LLM agent calibrated 108 of 112 qubits in 4.7 hours (Jun 2026); GPT-5.6 Sol brought up a never-measured chip with human help on 4 of 40 measurements (Sept 2026); learned plasma control runs on a third tokamak (Aug 2026). No AI-proposed physics prediction has yet been tested by experiment.
- Astronomy and cosmology. A cosmic-string radiation problem: GPT-5 Pro's partial answer (Nov 2025) → Gemini Deep Think's exact general solution (Mar 2026). General-model agents: single-dataset analyses (Apr–Jul 2025) → Gemini agents vetting 373,646 periodic variable stars in about 3 hours for about $40, reporting 18 new anomalies (Aug 2026) → a four-survey ultraviolet sky map built with Claude (Oct 2026). No AI-credited headline discovery; every AI-found object so far came from specialist pipelines. Physical world: fully automated supernova discovery-to-classification passed 200 events (Dec 2025); an LLM-driven control system runs an operational solar telescope (Jul 2026) and an ML scheduler ran two observing runs on the Blanco telescope (Aug 2026).
- Chemistry. ChemIQ (tool-free organic reasoning): 59% (o3-mini, May 2025) → 70% (GPT-5, Aug 2025), with no public point since. Anthropic's harder internal organic-chemistry eval: 65.7% (Opus 5, Jul 2026) → 69.2% (Mythos 5.1, Sept 2026). Results: an LLM agent derived a catalyst design principle (Mar 2026); GPT-5.5 directing a retrosynthesis search closed 111 of 120 benchmark routes (Sept 2026); experts accepted 90% of the routes from Microsoft's specialist RetroChimera (Nature, Sept 2026). No AI-credited theoretical-chemistry discovery yet. Physical world: the largest wet-lab campaign directed by a general model went from under about 1,000 reactions (2025) to 10,080 reactions in 92 days, about 110 a day, raising Chan–Lam yields from 16.6% to 25.2% (GPT-5.4 with Molecule.one, Jun 2026; OpenAI claim). Autonomous chemistry and materials labs with a public operating claim: about 3 (Apr 2025) → about 8 (Oct 2026).
- Materials science. Matbench Discovery thermal-conductivity error (specialist ML potentials against DFT): 0.170 (Apr 2025) → 0.065 (Sept 2026). General-model evidence is essentially absent before 2026, then: a new crystalline COF from ChatGPT deep research plus a robot (Feb 2026); an LLM-agent lab that raised its hit rate 4× within one 352-sample campaign (Apr 2026); Gemini driving a CVD reactor that grew 3 of 3 thin films on the first try (Aug 2026). Highest Tc of a synthesized AI- or ML-proposed superconductor: 5.4 K (Dec 2024), unchanged since; no AI-proposed material independently replicated. Physical world: published closed-loop synthesis throughput went from about 21 samples a day (A-Lab, 2023) to about 32 a day (Lila, 2,942 acid-OER catalysts in about 3 months, Sept 2026). Periodic Labs says its labs run 24/7 but publishes no counts; Google DeepMind's automated UK materials lab has not opened.
- Biology. BixBench (bioinformatics): 17% (Feb 2025) → 80.5% (Apr 2026). BioMysteryBench, questions five experts could not solve: 23.5% (Feb) → 56.5% (Sept 2026). Results: an unattended 21.5-hour survey of 1.9 billion protein clusters surfacing an uncharacterized enzyme family (Claude, Sept 2026) → 318,484 binder designs in 64 hours, 17 of 211 tested bound (Claude, Oct 2026) → LLM agents raising an aptamer hit rate from 0% to 42% over three wet-lab rounds (Google, Oct 2026) → about 35,000 antibody designs over 22 autonomous phases, of which 20 of 91 tested dislodged receptor-bound IgE and 3 of 57 anchors bound, without beating the published parent (Divergence Labs), and a 45-amino-acid CAR signaling domain, a third the clinical length, with about 10× the surface receptor in human T cells (Generate Biomedicines); both Claude, Oct 27, 2026, company reports of in-vitro work, unrefereed. No independent replication of an AI-originated biology claim yet. Physical world: design time fell from months to 1–3 days while each wet-lab round still takes 1–5 weeks; the largest general-model antibody-optimization loop ran about 1,010 variants in about 10 weeks, 9–10 days a round (Claude with BigHat, Oct 2026). One closed loop so far, where a model got results back and chose the next experiments (GPT-5 with Ginkgo, 29,527 compositions, Feb 2026). Anthropic has its own wet lab (BSL-1/2, human bench work) and its own drug programs since July 2026, run through 11 contract labs; OpenAI works through partner labs.
- Medicine. Time from data to a proposed mechanism: about 2.5 months (Robin, May 2025) → 17 minutes (GPT-5 Pro, Nov 2025). The only closed-loop wet-lab run by a general model beat a published optimum: 40% lower cost, 27% higher titer (GPT-5 with Ginkgo, Feb 2026). AI runs inside live clinical workflows (TrialGPT 2.0 cut tumor-board screening 55%; the FDA uses an AI tool in protocol review). On Oct 27, 2026 Anthropic reported its own antibody program reaching mouse efficacy in severe muscular dystrophy (about twice the grip strength and evoked force of vehicle, 7 of 8 treated against 4 of 8 untreated mice surviving to six weeks; unrandomized, litter-level P = 0.057, contract-lab run, unrefereed), and with Enveda, on 1,061 pages of batch records for a live clinical program, Claude raised across runs all 11 issues human reviewers found plus 7 they missed and cut review from 25.9 to 2.8 hours, drafted an IND summary module in 4 hours against about 3 months by hand, and drafted an Investigator's Brochure that Enveda filed with the FDA. No molecule designed by a general model has entered human trials. Clinical development: 8 of 117 AI-enabled assets had completed Phase 2 (Dec 2025); the furthest AI-discovered drug began Phase 3 (Sept 2026); no human-efficacy readout since.
- Social sciences. Reproduction of published papers: 37% AI-led (GPT-4-era models, 2024) → 78% (Claude Code, Jun 2026; different paper sets) → an automated audit of 4,452 economics replication packages, 11.2% fully reproduced at printed precision (NBER, Sept 2026). AI-written papers: 288 finance papers (Jan 2025) → 1,000 economics papers in 12 weeks (Jan–Apr 2026), none peer-reviewed. Physical world: AI-run online experiments went from 3 studies with 288 participants and a manual launch check (Aug 2025) to 9 experiments with 360 participants launched through an API, in a loop that revises theories between experiments (Jun 2026). No AI-run field trial.
- ML research. MLE-bench (Kaggle competitions, any medal; leaderboard as of each date): 17% (Apr 2025) → 44% (Oct 2025) → 64% (Apr 2026; then closed). Anthropic's internal AI R&D tasks (lab self-report): LLM-training speedup 3× (Opus 4, May 2025) → 52× (Mythos Preview, Apr 2026) → 69× (Opus 5, Jul 2026), where a human expert reaches about 4×; Anthropic then replaced the suite with one scored against an 85% bar for substituting for research staff (Opus 5 59.6%, Jul 2026). OpenAI's Research Debugging eval: 68% (GPT-5.6 Sol, Jul 2026) → 78% (GPT-6 Astra, Sept 2026). A third-party test found frontier agents did all the engineering on two unpublished NeurIPS submissions but could not make substantial research progress (Jul 2026). No fully AI-written paper accepted at a main-track venue.