AI FOR SCIENCE · INTERVIEWS
Brockman says AI succeeded in maths not mainly because proofs can be checked but because maths has lots of data; whether biology has the data, or can get it, he says 'is not yet resolved'. If OpenAI could pick one science to transform completely, he says, it would be drug discovery.
OpenAI's president ranks the sciences himself: drug discovery first, with materials science as another example. He also says where change comes first: a clinical-trial filing takes about a year to write and AI can write it much faster, and the treatment for a rare condition often already sits in some paper that AI can search for at scale. Where the data does not exist, wet labs and simulation have to supply it.
An opinion interview, with about 15 minutes on science. He gives no OpenAI biology result, programme, figure or date; 'almost 400 problems now resolved' is his description of a batch of results that, when released, had none confirmed outside OpenAI on this site's record.
He calls the week's maths results a milestone: OpenAI published solutions to 'almost 400 very important open problems', and the question of whether AI can produce new knowledge is now 'impossible to deny'. He adds that AI is not yet producing grand new theories. What he values is not the machine handing over an answer but what a person with the tools can do, and whether the proofs carry ideas people learn from; he says domain experts reading them found surprising uses of ideas, without naming who or which result.
“And there's always been this question of is AI actually able to come with new knowledge? And at this point, it's just impossible to deny, right?”Greg BrockmanHe says the usual explanation is that maths is checkable, but that is 'not quite true anymore': an arithmetic answer is easy to check and so is a formalised proof, but what OpenAI produces is natural-language proof, much harder to check. The core, he says, is data: maths and theoretical physics both offer lots of it, and the rule of deep learning is that compute, data and algorithms set capability.
“But the thing that actually I think is the core insight is really about data, right?”Greg BrockmanHe says OpenAI is already applying the same models and techniques to biology, and some steps will speed up quickly, but other bottlenecks will appear. Some are not about biological theory at all: a clinical-trial filing takes about a year to write, and AI can write it much more easily. Elsewhere the basic data simply does not exist, and new methods in wet labs and simulation have to fill the gap. OpenAI is also working on methods that need less data.
“The question will be and is not yet resolved. Do we have the data? Do we have the ability to get the data?”Greg BrockmanAsked how OpenAI ranks the sciences, he says drug discovery is 'the most clear and obvious' one, and gives materials science as another example. His reason is from first principles: people are suited to natural language and biology is foreign to them, but a neural net has no reason to be better at one than the other, so it should be able to 'crack the language of biology'. In the near term he expects shorter steps, more work in parallel, and finding treatments for rare conditions that already sit in the literature.
“I think that drug discovery is the most clear and obvious area if you get to pick one that you want to fully revolutionize.”Greg BrockmanA host says that if the technology works, the limit will be convincing regulators to change what they expect of a drug. He agrees that proof has to come first: if it can be shown that the 'physics of drug discovery' have changed, that should carry through every part of the system; he compares it with self-driving cars, and says trials already move to countries whose rules suit them, Australia's differing from the US's. How rules should change, he says, depends on what can actually be proved.
The timeline records, on 2026-10-06, OpenAI's release of 722 maths manuscripts in 372 families from an internal model given about 4,000 open problems; OpenAI said not all were formalised, and none was confirmed outside OpenAI at release. On 2026-09-21 it records a nine-mathematician advisory group with no decision-making power. Commentary holds Aaronson calling it one of the biggest days in mathematics, and Karagila and Huiberts reading particular files as desk rejects.
Brockman calls the batch 'almost 400 very important open problems that are now resolved' and says AI producing new knowledge is 'impossible to deny'. He gives no count of failures, and does not say how many have been confirmed by outside review.
The timeline records OpenAI's steps in life sciences: GPT-Rosalind passing 36.1% of its own LifeSciBench (2026-08-18), the OpenAI Foundation's $125M for open datasets (2026-09-15), and an academic plan opening GPT-Rosalind to US university labs (2026-10-15). The April interview (Jiao and Wang) was about a dedicated model and tiered access.
The company's president puts drug discovery first among all the sciences, and says what holds it back is data, not the model: whether biology has the data, or can get it, is not yet resolved. He announces no new programme, lab or budget.
Two other interviews on this site put the gap between maths and the other sciences down to checking: Lanyon AI's Gorard says code and maths have verifiers and physics has none; Biohub's Rives says reinforcement learning moves fast in maths and coding because answers can be verified quickly by computer.
Brockman offers a different explanation: natural-language proofs are hard to check anyway, and maths worked because of the amount of data. On his account, what biology lacks first is data, not a verifier.
If one of the largest frontier labs thinks biology is held back by data, near-term spending is more likely to go where data is made: wet labs, simulation, open datasets. The OpenAI Foundation's September grants went to open datasets. That is our reading of his words beside the record; he does not say OpenAI will build a lab or fund anything.
His near-term examples are writing clinical-trial filings and finding existing treatments for rare conditions in the literature. Both are easier to test than designing a new molecule: a filing has to pass a regulator, a repurposed drug has to help a patient. On the record, Biossil, whose round the OpenAI Startup Fund led, revives failed drug candidates, and the Foundation-funded CTD Commons collects regulatory files from failed drug programmes; both point the same way as his remarks, but he mentions neither.
He says almost 400 problems are 'resolved'; among the mathematicians in this site's commentary, one calls it a great day and others read particular files as desk rejects. Three questions to ask of such a number: how many problems were attempted and how many failed, how many results outside mathematicians have confirmed, and how many have a formal proof.
OpenAI published solutions to almost 400 very important open problems that are 'now resolved'.
Company-reported
Matches the 372 families in the timeline's entry of 6 October; OpenAI itself said not all were formalised, and none was confirmed outside OpenAI at release. The show notes say 'nearly 400 century-old math problems', stronger than what he said.
That AI can produce new knowledge is 'impossible to deny', though it is not yet producing grand new theories.
Guest's judgment
His judgment; the mathematicians in this site's commentary are divided.
Domain experts reading the proofs found surprising uses of ideas in them.
Company-reported
He names no expert and no result.
The core reason maths worked is data, not checkability; natural-language proofs are hard to check.
Guest's judgment
His explanation of OpenAI's results, with no data volumes or training details.
Whether biology has the data, or can get it, is not yet resolved.
Not yet closed
His own words.
OpenAI is already applying the same models and techniques to biology, and working on methods that need less data.
Company-reported
No project or result is named.
If one science is to be transformed completely, drug discovery is the first choice; materials science is another example he gives.
Company-reported
A priority stated by the company's president; no team, budget or timetable is given.
A clinical-trial filing takes about a year to write, and AI can write it much more easily.
Guest's judgment
'About a year' is his figure; no product, customer or accepted filing is named.
Treatments for rare conditions often already sit in some paper, AI can search for them at scale, and there is 'plenty of precedent'.
Guest's judgment
No precedent is named.
A neural net has no reason to be better at natural language than at biology, so it should be able to 'crack the language of biology'.
Guest's judgment
He presents it as reasoning from first principles.
A drug takes 15 years and $1.5 billion, about $15 million of it preclinical and $1.4 billion or more in clinical trials.
Industry context
A host's round numbers, with no source; not Brockman's.
Once it is proved that drug discovery has changed, regulation should follow; trials already move to countries whose rules suit them.
Guest's judgment
He does not say what would count as proof, and names no trials.
OpenAI running wet labs or simulation of its own to fill biology's data gap.
Not yet closed
He says only that such innovation will be 'incredibly important', not that OpenAI will do it.
Whether OpenAI shows a drug-discovery result that goes past binding.
Would show up as A paper or a trial registration naming the candidate, which steps OpenAI's models did, and who measured the lab or clinical data.Whether the data question gets an answer: biology data that OpenAI or its Foundation makes or funds is actually released.
Would show up as A public dataset stating what was measured, at what scale and who can use it, and how much a model trained on it improves over one without.How many of the almost 400 results outside mathematicians confirm.
Would show up as Journal acceptances, public reviews by specialists or a report from the advisory group, counting results confirmed, doubtful and wrong, and how many of the roughly 4,000 problems failed.Whether a regulator accepts a filing written with AI, and how much time it saved.
Would show up as A filing named by its sponsor or a regulator, stating which model was used, which parts people did, and how much shorter it took than before.Later evidence that tests or limits what was said, and targets whose date has passed.
Five days later OpenAI acquired Orchestra Bio and folded it into Rosalind: agents that track which activities and timelines a setback hits in a drug programme, with people deciding. It puts money behind his ranking of drug discovery first; but it is coordination software, not a result past binding, and it releases no biology data, so the first two watch items stay open.Timeline entry →
Read 2026.10.21 PodScripts' full machine transcript, with timestamps and no speaker labels, read in full; who said what is judged from the content and the show notes. The audio and video were not heard: YouTube refused this session's caption download. The show notes and length come from the podcast's RSS feed. The second half is about OpenAI's agent product and cybersecurity; this page covers only the first 17 minutes or so, on science. Open the interview ↗PodScripts machine transcript ↗