Hundreds of Millions Spent on Math! Ultraman: OpenAI Still Has Undisclosed Discoveries
Hundreds of Millions of Dollars on Math! Ultraman: OpenAI Still Has Unreleased Discoveries
Content Summary
Xinzhiyuan reports OpenAI still has undisclosed mathematical discoveries? In the 《Cheeky Pint》 interview released on October 29, Ultraman revealed: OpenAI’s investment in mathematical research also includes many still
Synced Report
Synced Report
Has OpenAI got undisclosed mathematical discoveries?
In the October 29 interview “Cheeky Pint,” Ultraman revealed that OpenAI’s investment in mathematical research also includes many findings that have not yet been made public.
Host John Collison follow‑up: Are there any results beyond those 700 proofs?
Ultraman replies:
Yes, of course there are.
In recent months, OpenAI’s mathematical achievements have sparked heated discussion.
After the results were released, researchers began dissecting the publicly available material, understanding the underlying ideas, and verifying the key steps in the derivations.
In the field of biology, AI-generated proposals have already reached the experimental bench.
The model analyzes data and proposes designs; researchers select proposals, conduct experiments, and then feed the experimental results back to the model to further refine the designs.
From checking mathematical derivations to validating biological designs, humans and AI are turning each round of results into the starting point for the next exploration.
When discussing the investment in mathematical research, the host mentioned a scale of several hundred million dollars, and Ultraman expressed his approval.
But he also added that this investment includes other undisclosed findings.
As for how these results should be made public, Ultraman admits: there is a dilemma.
Ultraman argues that major fundamental scientific discoveries should be disclosed.
Ultraman believes that humanity ought to possess basic scientific knowledge, and major scientific discoveries should be announced to the world.
However, researchers’ career development and sense of achievement largely stem from the discoveries themselves.
He hopes to find a way that both lets the world benefit from the sharing of knowledge and sustainably encourages researchers in these fields to continue participating in future science.
After results are made public, it is also necessary to help researchers understand the reasoning, verify the derivations, and make these discoveries the foundation for subsequent research.
On GitHub, mathematical researchers have launched an independent revision project called oai-math-revised.
Mathematics researchers launch the oai-math-revised project to collaboratively revise OpenAI’s publicly released mathematical research manuscripts.
This project directly targets some OpenAI mathematical manuscripts: two shortcomings—exposition that is hard to understand and incomplete citations.
To this end, the project initiators invite students, postdoctoral researchers, professors, and math enthusiasts to collaborate, supplement references, clarify arguments, and identify and correct errors as much as possible.
AI may produce a mathematics paper, but the research community still needs to turn it into knowledge that others can understand, verify, and apply.
How to support this follow‑up work also becomes a question to consider when releasing the results.
OpenAI is experimenting with how to release its outputs.
On October 26, the company classified the research manuscript newly announced that day as “raw research data” and described various approaches such as inviting external experts to lead and organizing a team to explain the results.
OpenAI classifies these manuscripts as “raw research data” and notes that the related results are at different stages of verification.
Its math repository also explicitly states that the results are at various verification stages, and some have not yet been accompanied by formal proofs that can be checked by a computer.
After the manuscript is made public, explanation, verification, and revision must continue.
When more researchers understand these results and confirm which methods are reusable, new research can build on this foundation.
In the field of biology, researchers are actively incorporating large models into their work.
On October 30, Epoch AI released a set of statistics on biology preprints:
Among the 4,922 bioRxiv preprints examined in September 2026, 846 disclosed the use of large models, accounting for 17.19%.
Roughly one out of every six papers.
The proportion of large‑model usage disclosures in biology preprints continues to rise. (Source: Epoch AI)
In September 2025, this disclosure rate was 6.40%. A year later, it approached 18.94%, the same month’s rate for dedicated AI tools.
These disclosed stages involve tasks such as literature review, writing, programming, and research analysis:
This month, 190 preprints disclosed using large models for research or analysis, accounting for 3.86% of all examined preprints.
According to Epoch’s classification, this includes experimental design, data annotation, and statistical analysis; writing, literature review, code, and computation are listed under other use categories.
Epoch’s statistics show that large models are already involved in multiple stages of research.
For a specific study, what tasks does the model handle, and how do researchers assess, validate, and use its output?
Two recent public collaborations demonstrate how humans and AI can work together to advance research.
On October 21, OpenAI, PerturbAI, and Scripps Research announced a joint effort:
Thousands of GPT‑Rosalind‑driven agents independently analyze existing experimental data from millions of mouse brain cells.
According to the team, this round of data analysis and interpretation took 32 minutes.
Each agent focuses on a specific gene perturbation, analyzes cellular responses, searches relevant literature, and then proposes an explanation.
These analysis results are ultimately compiled into a reasoning map that links genes, cellular changes, and possible biological mechanisms.
Researchers can follow this map, select the clues to test next, and continue designing experiments.
With clues in hand, the next step is to let experiments provide feedback.
In another study, biotech company Generate Biomedicines is using AI to design proteins and develop drugs.
On October 27, the team announced progress in remodeling CAR‑T receptors with Claude Code, aiming to shorten the intracellular signaling domain while preserving function.
Structural comparison of different CAR receptors; the gray horizontal band in the figure represents the cell membrane. (Image source: Generate Biomedicines)
CAR receptors function as the “recognition and activation device” of T cells: the extracellular portion recognizes the target, while the intracellular portion transmits signals that trigger T‑cell responses.
Researchers selected 186 candidates from the model‑guided design proposals for the first round of experiments. After the results were obtained, they fed them back into the system to further refine the proposals and conduct additional experiments.
After three rounds of design and screening, the team obtained a signaling structure composed of only 45 amino acids, about 70% shorter than the second‑generation CAR signaling structure used as a control.
The team reported that in vitro experiments using the same dose of mRNA showed more CAR receptors appearing on the surface of T cells with the newly designed construct, and the ability to kill target cells was sustained longer.
This design was gradually refined through iterative experimental feedback.
From line-by-line mathematical derivations to round after round of biological experiments, AI is entering more aspects of research.
It seeks clues, proposes designs, and then adjusts the plan based on validation results.
If we turn our attention to artificial superintelligence (ASI), another question arises: Can the scientific results produced by AI, in turn, accelerate AI’s own development?
As early as 2025,Google DeepMind disclosed that AlphaEvolve accelerated a computational kernel in Gemini training by 23%, reducing the overall training time by 1%.
Algorithm optimizations discovered by AlphaEvolve have been used to improve Gemini’s training efficiency. (Image source: Google DeepMind)
Optimization proposals generated by AI and validated have already been used to improve model training.
Taking it a step further, if a system can autonomously develop a more powerful next‑generation system, which in turn continues to drive AI research, it creates the loop envisioned by RSI (Recursive Self‑Improvement).
If such a loop can be sustained, it could become a driving force toward ASI.
From a single local optimization to continuously driving new discoveries, this path still holds many possibilities.
Looking toward an ASI future, humans and AI may explore together in ways unimaginable today.
And all of this may start with repeated attempts: validating an idea more quickly, confirming a finding more rigorously, allowing the next researcher to embark on new exploration sooner.
References:
Edited by: Yuan Yu
Get MirrorSpeed
Get MirrorSpeed