On October 26, 2026, OpenAI announced three "major" discoveries spanning mathematics, computer science and physics, and for the first time built outside experts it calls "shepherds" into the release process to guide and vet the results. The move comes less than a month after the company published hundreds of manuscripts it said contained solutions to long-standing math problems. The second wave is smaller in volume but heavier in consequence: as AI-generated proofs arrive faster than journals
OpenAI published three research results on October 26, 2026, covering mathematics, computer science and physics, and characterized them as "major" discoveries. According to The Verge AI, the difference from the batch of manuscripts released in late September lies in process: outside experts were brought in as "shepherds" to steer the work while it was being produced and to independently check the conclusions, rather than being invited to comment after publication. The earlier round was a single mass release of hundreds of manuscripts that OpenAI said contained solutions to long-standing mathematics problems, which immediately triggered debate among mathematicians over verification standards and attribution. This round shrank from hundreds of documents to three items, and the cadence shifted from bulk delivery to item-by-item endorsement. As of publication, the specific theorems, written
Why now. In 2024, Google DeepMind's AlphaProof solved 4 of 6 problems at the International Mathematical Olympiad, a silver-medal result that turned machine mathematics from a demo into a comparable scorecard. Other labs soon claimed gold-level ranges on the same contest, making mathematics the most externally quantifiable battleground for large models. Quantification created a new problem: once a model can produce in a week the volume of derivations a human mathematician needs years to write, verification becomes the constraint. Manuscript output grows far faster than journal review capacity, preprint servers have no filter designed for machine output, and the research community can neither
Liu Gong's take: OpenAI's real deliverable this round is not the three findings but the shepherd mechanism itself; within 12 months, at least two other frontier labs are likely to copy it as a standard release step.
From a technical standpoint, the bottleneck in AI-driven science has moved from generation to verification. Hundreds of manuscripts arriving at once is not a generation problem, since models clearly can produce volume, but a review-bandwidth problem, because human mathematicians do not have the hours. Formal proof assistants such as Lean and Coq absorb part of that load, yet translating natural-language derivations into machine-checkable code is itself expensive work. In industrial inspection, where Liu Gong has spent years, the pattern is familiar: once detection accuracy passes 99%, the constraint migrates to the review station, and no line manager releases product on a detector alone. The shepherd is OpenAI's review station.
On the business side, scientific discovery is the hardest currency available. For research funders and corporate labs, "our model contributed to a discovery" carries more weight in budget and compliance conversations than any benchmark table. But the stronger the claim, the sharper the backlash: one retracted "major discovery" discounts every claim that follows. Attaching named outside experts to the results posts their academic credit as collateral and moves part of the reputational risk off OpenAI's balance sheet.
For the industry, a comparison is already on the table. DeepMind's AlphaProof ran in the slow lane, validated first by competition rules and then by peer review, with independent replication built in. OpenAI is building a fast lane: company announcement, expert endorsement, no replication step yet. Liu Gong predicts that within the next six months a third-party academic body will launch a public program for batch formal verification of AI-generated proofs; without one, journals will be forced to create a separate track for machine output. The two lanes will collide over whether AI-generated proofs can enter the formal academic record directly.
- Volume: 3 discoveries this round versus hundreds of manuscripts in the previous release, a drop from triple digits to single digits.
- Cadence: less than one month between the two releases, roughly 30 days from the manuscript dump to the vetted announcement.
- Vetting: all three findings carry an external shepherd, an arrangement OpenAI frames as the key difference from the prior round, which carried no such mechanism.
- External baseline: Google DeepMind's AlphaProof solved 4 of 6 problems at the 2024 IMO, a silver-medal result and the most cited public benchmark for AI mathematical ability.
- Scope: mathematics, computer science and physics in a single announcement, the first time
For the industry, the near-term effect over three months is spillover demand for verification: mathematics and computer science departments will field more requests to check a proof, and spending on formal verification tools and personnel will rise. Over a one-year horizon, if the shepherd model spreads, corporate-sponsored expert endorsement and journal peer review will become two parallel systems of scientific credit, with attribution order and liability as the new fault line.
For ordinary users, the findings change no product experience in the short run. What spills over is method: more reliable automated theorem proving will land first in code verification, chip design and compliance
Questions worth exploring:
- Whether the shepherd model can substitute for peer review, and how authorship and liability get allocated
- Whether formal verification tooling for AI-generated proofs becomes the next infrastructure layer
- Whether corporate control of the release cadence dilutes or rebuilds the authority of academic publishing
Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.
This item comes from The Verge AI (published 2026-10-26 23:34:58). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.