Hundreds of millions of dollars are being spent on mathematical research, Sam Altman says that OpenAI still has undisclosed discoveries that have not been made public.
Are there undisclosed mathematical discoveries at OpenAI?
In the "Cheeky Pint" interview released on October 29, Sam Altman revealed that OpenAI's investment in mathematical research also covers many unpublicized findings that have not yet been disclosed.
Host John Collison followed up and asked: Are there any achievements beyond those 700 proofs?
Altman replied:
Yes, absolutely.
In recent months, OpenAI's mathematical achievements have sparked heated discussions one after another.
After the results were made public, researchers began to dissect the published materials, understand the ideas behind them, and verify the key steps in the derivation.
In the field of biology, solutions proposed by AI have already been placed on the experimental bench.
Models analyze data and propose designs, while researchers screen solutions, carry out experiments, and then feed the experimental results back to the models to further improve the designs.
From checking mathematical derivations to verifying biological designs, humans and AI are turning the results of each round into the starting point for the next step of exploration.
Major Discoveries Should Be Made Public
What About Mathematicians?
Talking about the investment in mathematical research, the host mentioned the magnitude of hundreds of millions of dollars, which Altman acknowledged.
But he also added that this investment also covers other undisclosed findings.
As for how these results should be made public, Altman admitted that there is a dilemma in this regard.
Altman advocates that major basic scientific discoveries should be made public.
Altman believes that human beings deserve access to basic scientific knowledge, and major scientific discoveries should be announced to the world.
However, the career development and sense of achievement of scientific researchers largely come from the discoveries themselves.
He hopes to find a way that not only allows the world to benefit from knowledge sharing, but also can promote researchers in these fields to continue to participate in future scientific research in the long term.
After the results are made public, it is also necessary to help researchers understand the ideas, check the derivations, so that these discoveries can become the foundation for subsequent research.
What Are Mathematicians Adding to AI Manuscripts?
On GitHub, mathematics researchers have launched an independent revision project called oai-math-revised.
Mathematics researchers launched the oai-math-revised project to collaboratively revise the mathematical research manuscripts published by OpenAI.
This project directly targets two shortcomings of some of OpenAI's mathematical manuscripts: the discourse is hard to understand, and citations are incomplete.
To this end, the project initiators invite students, postdocs, professors and mathematics enthusiasts to participate together, supplement literature, straighten out the discourse, identify and correct errors as much as possible.
After AI produces a mathematical paper, the scientific community still needs to turn it into knowledge that others can understand, verify and put into use.
How to support these follow-up tasks has also become an issue that needs to be considered when releasing results.
OpenAI is trying to adjust the way it releases its results.
On October 26, the company positioned the newly published related research manuscripts that day as "raw research data", and introduced different practices such as inviting external experts to take the lead and organizing teams to interpret the results.
OpenAI positions these manuscripts as "raw research data" and states that the relevant results are in different stages of verification.
Its mathematics repository also clearly states that the relevant results are in different stages of verification, and some are not yet accompanied by computer-checkable formal proofs.
After the manuscripts are made public, interpretation, verification and revision will continue.
Only when more researchers understand these results and confirm which methods can be reused can new research be carried out on this basis.
About One in Every Six Papers
Discloses the Use of Large Models
In the field of biology, researchers are actively integrating large models into their work.
On October 30, Epoch AI released a set of statistics on biology preprints:
Among the 4922 bioRxiv preprints inspected in September 2026, 846 disclosed the use of large models, accounting for 17.19%.
About one in every six papers.
The proportion of disclosures of large model use in biology preprints continues to rise. (Source: Epoch AI)
In September 2025, this disclosure proportion was 6.40%. One year later, it is close to the 18.94% of dedicated AI tools in the same month.
These disclosed scenarios involve work such as literature, writing, programming and research analysis:
In that month, 190 preprints disclosed that large models were used for research or analysis, accounting for 3.86% of all inspected preprints.
According to Epoch's classification, this includes experimental design, data annotation and statistical analysis; writing, literature review, code and computing are classified into other usage categories respectively.
These statistics from Epoch show that large models have participated in multiple links of scientific research.
For a specific study, what work is the model responsible for, and how do researchers judge, verify and use its outputs?
From Analyzing Data
To Receiving Experimental Feedback
Two recently disclosed collaborations demonstrate how humans and AI work in relay to advance research.
On October 21, OpenAI, PerturbAI and Scripps Research announced a collaboration:
Thousands of GPT-Rosalind-powered agents separately analyzed existing experimental data from millions of mouse brain cells.
According to the team's report, this round of data analysis and interpretation took 32 minutes.
Each agent focuses on one genetic perturbation, analyzes cellular responses, searches for relevant literature, and then proposes explanations.
These analysis results are finally integrated into an inference map that connects genes, cellular changes and possible biological mechanisms.
Researchers can follow this map to select clues to test next and continue to design experiments.
With the clues, the next step is to get feedback from the experiments.
In another study, biotechnology company Generate Biomedicines is using AI to design proteins and develop drugs.
On October 27, the team announced progress in modifying CAR-T receptors with the help of Claude Code: they want to shorten the signal-transmitting part inside the receptor while retaining its functions.
Structural comparison of different CAR receptors, the gray horizontal band in the figure represents the cell membrane. (Source: Generate Biomedicines)
The CAR receptor is like the "recognition and activation device" of T cells: the extracellular part identifies targets, and the intracellular part transmits signals to make T cells respond.
Researchers selected 186 solutions from the model-participated design schemes for the first round of experiments. After the results come out, they are handed over to the system to continue modifying the schemes and conducting experiments.
After three rounds of design and screening, the team obtained a signaling structure consisting of only 45 amino acids, which is about 70% shorter than the second-generation CAR signaling structure used as the control.
The team reported that in in vitro experiments, with the same dose of mRNA, more CAR receptors appeared on the surface of T cells using the new design, and their ability to kill target cells was also maintained for a longer time.
This set of designs was gradually modified through rounds of experimental feedback.
Can AI Scientific Research
Lead to ASI?
From lines of mathematical derivation to rounds of biological experiments, AI is entering more links of scientific research.
It looks for clues, proposes designs, and then adjusts the schemes based on verification results.
If we turn our attention to Artificial Super Intelligence (ASI), another question arises: can the scientific research achievements made by AI in turn accelerate the R&D of AI itself?
As early as 2025, Google DeepMind disclosed that AlphaEvolve accelerated a computing kernel in Gemini training by 23%, shortening the overall training time by 1%.
The algorithm optimization discovered by AlphaEvolve has been used to improve the training efficiency of Gemini. (Source: Google DeepMind)
Verified optimization solutions proposed by AI have already been used to improve model training.
One step further, if the system can independently develop a stronger next-generation system, and the next-generation system continues to promote AI R&D, it will form the cycle envisioned by RSI (Recursive Self-Improvement).
If such a cycle can continue, it may become a driving force towards ASI.
From a partial optimization to continuously promoting new discoveries, there are still many possibilities on this road.
Facing the future of ASI, humans and AI may explore together in ways that are unimaginable today.
And all of this may start from repeated attempts: verifying an idea faster, confirming a discovery more rigorously, so that the next researcher can start new exploration earlier.
References:
https://cheekypint.transistor.fm/39/transcript
https://www.theverge.com/ai-artificial-intelligence/1023054/altman-cheeky-pint-ai-math-interview
https://epoch.ai/data-insights/biorxiv-llms-vs-ai-tools
https://openai.com/index/building-scientific-understanding-with-academic-communities/?utm_source=chatgpt.com
This article is from the WeChat Official Account "AI Era", author: ASI Revelation, editor: Yuan Yu, published with authorization from 36Kr.