In October 2024 the Nobel committee gave the chemistry prize to David Baker, Demis Hassabis and John Jumper for computational protein design and structure prediction, and the physics prize to John Hopfield and Geoffrey Hinton for the foundations of neural networks.[1] A year on, the question of what AI has actually contributed to science since has a longer answer than the headlines suggest and a shorter one than the press releases do. This briefing takes the most-discussed result of the month, DeepMind's fluid-dynamics paper, as its starting point and works outward.
The fluid equations, precisely
The Navier–Stokes equations describe the motion of fluids. Whether smooth solutions to the three-dimensional equations can develop a singularity, a point at which velocity becomes infinite, in finite time is one of the Clay Mathematics Institute's seven Millennium Prize problems.[2] On 18 September a team from Google DeepMind with Yongji Wang, Ching-Yao Lai, Javier Gómez-Serrano and Tristan Buckmaster, mathematicians at NYU, Stanford, Brown and Princeton, published the first systematic discovery of new families of unstable self-similar singularities in three related equations: the incompressible porous-media equation, the Boussinesq equation, and the three-dimensional Euler equations with a boundary.[3][4]
The method is what made it possible. Unstable singularities are solutions that exist mathematically but that any small perturbation destroys, which is why numerical simulation had never found them. The team trained physics-informed neural networks, networks whose loss function is the equation itself, with a second-order optimiser to a precision close to the limit of double-precision arithmetic, and found not one solution but a family, with an empirical linear relationship between the order of instability and the blow-up rate that hints at structure nobody had seen.[4] What the paper does not claim matters as much. The unforced Euler and Navier–Stokes equations are the ones the prize concerns, and the authors say only that unstable singularities are hypothesised to play a role there. Their own blog post, and the paper's abstract, are explicit that the Millennium problem is not resolved.[3] It is a real result, on a hard problem, honestly described, and the mathematicians on the paper are the reason it is trusted.
Algorithms: the result nobody disputes
- 48
- Scalar multiplications AlphaEvolve needs to multiply two 4×4 complex matrices, against 49 in Strassen's 1969 algorithm [5]
- 0.7%
- Share of Google's worldwide compute recovered by an AlphaEvolve scheduling heuristic [5]
- 12 of 12
- ICPC World Finals problems solved by an OpenAI system on 17 September; Gemini 2.5 Deep Think solved ten [6]
The cleanest verified contribution of the year is in algorithms rather than pure mathematics, because the outputs can be run. AlphaEvolve, announced by DeepMind on 14 May, pairs Gemini models with an automatic evaluator and evolves programs against it. Across more than fifty open problems it rediscovered the best known solution about three quarters of the time and improved on it in about a fifth of cases. Two improvements were concrete: multiplying two 4×4 complex matrices in 48 scalar multiplications, one better than Strassen's algorithm of 1969, and a new lower bound for the kissing number in eleven dimensions. Inside Google it produced a data-centre scheduling heuristic that recovered about 0.7% of global compute, and sped up kernels used to train Gemini itself.[5] Terence Tao worked with the team on the mathematical problems and said so.
Competitive programming gave the same picture in September. At the ICPC World Finals in Baku, under the same five-hour rules as the 139 human teams, an OpenAI system solved all twelve problems, eleven at the first attempt, and Google's Gemini 2.5 Deep Think solved ten, including one no human team solved.[6][7] Two months earlier both labs had reached gold-medal level at the International Mathematical Olympiad, which is its own briefing.
Biology: tools in daily use, drugs on the clock
AlphaFold 3, published in May 2024, extended structure prediction from single proteins to complexes with DNA, RNA and small molecules, and its code was released for non-commercial use in November.[8] AlphaGenome followed on 25 June 2025: a model that reads a million base pairs of DNA and predicts the effect of a change on gene regulation, matching or beating the best specialised tools on 24 of 26 benchmarks, available through a non-commercial preview and not validated for clinical use.[9] Google's AI co-scientist, announced in February, generates and ranks hypotheses; its best-known demonstration was proposing, in days, a mechanism for how bacteria exchange antimicrobial-resistance genes that a team at Imperial College London had spent years establishing experimentally, and had not yet published.[10]
Drugs move at the speed of trials. On 3 June Insilico Medicine published Phase IIa results in Nature Medicine for rentosertib, a treatment for idiopathic pulmonary fibrosis whose target and molecule were both generated by its models: in 71 patients over twelve weeks, the highest dose improved forced vital capacity by 98.4 millilitres against a 20.3 millilitre decline on placebo. It is the first AI-discovered drug with peer-reviewed efficacy data, and a small trial.[11] Isomorphic Labs, DeepMind's drug-discovery spin-out, raised $600m in March to take its own candidates into the clinic.[12] Microsoft's MatterGen, published in Nature in January, generates candidate materials to a specification and produced one that was synthesised and measured.[13]
Weather: where AI already won
Forecasting is the field where a learned model has clearly overtaken the incumbent method. DeepMind's GenCast, published in Nature in December 2024, produces ensemble forecasts that beat the European Centre's operational system, the best in the world, on the large majority of measures out to fifteen days, in minutes rather than hours.[14] The weather models are verified every day against what actually happens, which is why nobody argues about them.
The pattern
The results above have a property in common. Each can be checked without trusting the model: a matrix algorithm runs, a forecast is compared with the sky, a structure is compared with a crystal, a singularity is verified by the equation itself, a drug is trialled. Where a claim cannot be checked that way, in the autonomous "AI scientist" systems that generate papers, the record is thinner and the disputes louder; Sakana AI's system, launched in August 2024, was found by academic reviewers to rely on shallow literature search and human-supplied templates.[15] The distinction is the one to carry into any announcement: not whether an AI was involved, but whether the output was verified by something other than an AI.
