During his PhD at the University of Otago, biochemist Peter Mace spent three years working out the structure of a single protein. Now a professor in Otago’s Faculty of Biomedical Sciences, he says the AI program AlphaFold can predict what a protein’s 3D structure might look like “in about 10 minutes”.
“There’s some really good stuff in there, there’s some hype, but I think, on the whole, it’s changed our careers,” Peter said in a university article published on 2 January 2026. AI’s record in maths and science now includes a Nobel Prize, a 1,000-signature protest letter, and a claim that the owner of a maths database called “a dramatic misrepresentation”.
Results that have held up
Demis Hassabis and John Jumper of Google DeepMind shared the 2024 Nobel Prize in Chemistry with protein designer David Baker. The Royal Swedish Academy of Sciences said on 9 October 2024 that AlphaFold2 had predicted the structure of “virtually all the 200 million proteins that researchers have identified” and had been used by more than two million people in 190 countries. The predictions sit in a free database run by DeepMind and the European Bioinformatics Institute.
Otago’s Dr Adam Middleton calls AlphaFold an “incredible hypothesis generator”. He’s using machine learning to study an uncharacterised molecule linked to type 1 diabetes and lupus, work he says he “just wouldn’t have been able to look at without these tools”.
In maths, DeepMind’s FunSearch paper in Nature (14 December 2023) reported a cap set of 512 points in eight dimensions, larger than any previously known. FunSearch writes a program that builds the answer. Jordan Ellenberg, a University of Wisconsin-Madison mathematician who worked with the team, said its solutions were “far conceptually richer than a mere list of numbers”.
On 14 May 2025 DeepMind said its AlphaEvolve system had found a way to multiply 4×4 complex-valued matrices using 48 scalar multiplications, beating Strassen’s 1969 algorithm in that setting. Tested on more than 50 open maths problems, it matched the best known answers in about 75% of cases and improved on them in 20%, the company said.
Drug discovery has examples too. An MIT team reported in Cell in February 2020 that a deep learning model had picked out halicin, which killed many drug-resistant bacteria in lab tests. In May 2023 McMaster University and MIT announced abaucin, aimed at the hospital superbug Acinetobacter baumannii. The team first tested the bacteria against thousands of known drugs to train its model, then used it to assess 6,680 untested molecules.
DeepMind announced on 29 November 2023 that its GNoME model had predicted 2.2 million new crystals, 380,000 of them stable, and that outside labs had independently made 736. Anthony Cheetham and Ram Seshadri, materials professors at UC Santa Barbara, reviewed the work in Chemistry of Materials in April 2024. They reported “scant evidence for compounds that fulfill the trifecta of novelty, credibility, and utility.”
Medals, proofs and a correction
At the 2024 International Mathematical Olympiad, DeepMind’s AlphaProof and AlphaGeometry 2 scored 28 of 42, silver medal standard. Experts first translated the problems into a formal language, and some took up to three days to solve, DeepMind said on 25 July 2024. In July 2025 an advanced Gemini model worked in plain language within the 4.5-hour limit and scored 35, gold standard. “Their solutions were astonishing in many respects,” said IMO president Gregor Dolinar in a statement DeepMind published.
In October 2025 OpenAI executive Kevin Weil posted that GPT-5 had found solutions to 10 open problems posed by the mathematician Paul Erdős. Thomas Bloom, who maintains the erdosproblems.com database, replied on X that GPT-5 had found published papers that already solved them. “The ‘open’ status only means I personally am unaware of a paper which solves it,” Thomas wrote.
Mathematician Terence Tao, who keeps a public wiki of AI contributions to Erdős problems, wrote on 7 January 2026 that problem #728 had been solved “more or less autonomously by AI”. OpenAI’s GPT-5.2 Pro and Harmonic’s Aristotle produced a proof checked in the Lean proof language. In a later update Terence noted the methods closely matched a 2014 paper by Carl Pomerance, who then released a note showing those methods also solve it.
Tool or discoverer
Terence sorted the results in a 14 January post. Fully autonomous AI solutions tended to be “short and largely follow a standard technique”, while more complicated and novel results came from people working closely with AI tools. He guessed that 1-2% of open Erdős problems are “simple enough to be amenable to current AI tools operated with minimal human intervention”.
That split matters for anyone reading a vendor announcement. A model that compresses three years of structural biology into ten minutes is doing something real, and it’s still doing it inside a workflow a researcher designed, on a question a researcher chose, with a result a researcher has to test. Peter’s caveat covers the lab. “AI won’t answer everything without experiments,” he said.
The verification step is where the work has quietly moved. The #728 proof counted because Lean checked it line by line, not because a chatbot sounded convincing. Maths has that luxury. Most of science doesn’t, and a prediction that can only be confirmed by a year of bench work is a to-do list rather than a discovery.
The peer review problem
The question of whether machines can do the writing as well as the maths has already been tested. Sakana AI reported that its AI Scientist-v2 system generated three papers and submitted them to a workshop at the ICLR 2025 conference, the “I Can’t Believe It’s Not Better” track, which collects negative results and the practical limits of deep learning.
One of the three cleared review, scoring 6, 7 and 6 from anonymous reviewers out of 43 papers submitted to the workshop. Sakana was unusually frank about the size of the win. Workshop papers “present preliminary findings that are less refined compared to main conference submissions”, the company noted, and on its own internal assessment against the main-track standard, “none of the 3 papers passed our internal bar”. Under a protocol agreed in advance with the organisers, the accepted paper was withdrawn rather than published.
So the honest version of that result is narrow. A machine wrote a passable workshop paper about the limits of machine learning, and the humans running the experiment said so before anyone else could. The less honest version of the same result is the headline most people saw.
Who gets to check the work
The ethics argument that has moved furthest isn’t about robots replacing scientists. It’s about access.
When DeepMind published AlphaFold 3 in Nature in May 2024, it didn’t release the code, breaking with the AlphaFold2 precedent. Structural and computational biologists including Stephanie Wankowicz and Roland Dunbrack co-authored a protest letter citing “several deviations from our community’s standards”, among them the fact that peer reviewers hadn’t been given the code despite repeated requests. More than 1,000 researchers had signed by the end of May.
Nature’s editor-in-chief Magdalena Skipper said the journal weighs factors such as biosecurity when deciding what code and data an author must publish, and a later editorial described the decision as an opportunity for conversation. DeepMind’s Pushmeet Kohli and Isomorphic Labs’ Max Jaderberg said the team was working on releasing the model and its weights for academic use within six months. The code and weights appeared in November 2024, with the weights available on request to researchers with an academic affiliation, for non-commercial use, at DeepMind’s discretion.
That’s a reasonable outcome and a thin one. A result nobody outside the company can reproduce until the company decides otherwise isn’t yet part of the scientific record, whatever journal it appears in. The AlphaFold2 database is free to search by anybody with a browser, which is exactly why two million people have used it. Access is the difference between a published finding and a product demonstration.
The same tension runs through the commercial end. Private labs fund the compute, so private labs set the terms, and the researchers who find the errors are usually unpaid. Cheetham and Seshadri worked through 380,000 claimed stable crystals to report that the useful ones were hard to find. Thomas corrected a multinational’s maths claim in a social media reply. Checking is the cheap half of the sentence and the expensive half of the job.
What this means for a business reading the headlines
Three habits travel well out of all this.
Ask what the machine actually produced. A structure prediction, a candidate molecule and a verified proof are three different grades of evidence, and only the last one is finished.
Ask who checked it, and whether they could. If the method is proprietary and the data is gated, treat the claim as a supplier statement rather than a finding.
Ask what the baseline was. “Beats the best known answer” means something precise in matrix multiplication and nothing at all in a press release about productivity.
Anyone can check the claims in this piece, which is rarer than it should be. Terence’s wiki marks each AI result on an Erdős problem as a full solution, partial, incorrect or unverified, and the AlphaFold database’s 200 million-plus predictions are free to search.
