About
Who writes Vibe Math Lab, and where the lessons come from
Vibe Math Lab is about doing real math research with AI agents, and how to trust what they produce. Every lesson comes from Tyler's own research, which is published with its evidence: how the agents were set up and directed, how their work was checked, and what the checks caught.
The lessons are written by Tyler Orden, an independent researcher, which means he does research on his own rather than at a university. He does the work with Claude agents, AI models made by Anthropic, working under his direction: the agents write and run the code, and he directs the work and checks it. The research itself is on his research site, Cyborg Mathematics, where each result is posted with its source, how it was checked, and when.
The research behind the lessons
- 50 entries in the OEIS carry Tyler's credit, and he created 26 of them (counted November 6, 2026). The OEIS, the On-Line Encyclopedia of Integer Sequences, is the reference database mathematicians use to look up number sequences. The other credits are for new values, corrections and comments added to existing entries. Anyone can look up each entry by its number.
- Two pull requests merged by Terence Tao. Tao, a UCLA mathematician and Fields Medal winner, keeps a public database of the problems posed by the Hungarian mathematician Paul Erdős on GitHub, a site for sharing code and data. A pull request is a proposed change to such a database. Tao merged two of Tyler's: on August 6, 2026, one that linked problem 993 to its OEIS sequence (A000055, the number of trees of each size), with a description that cited the tree check below, and on September 18, 2026, one that linked seven more problems to the OEIS.
- An Erdős conjecture checked on every tree up to 32 dots. A tree is a network of dots joined by lines, all in one piece, with no loops. In 1987 Alavi, Malde, Schwenk and Erdős conjectured that in every tree, the number of ways to pick some dots with no two joined rises and then falls as more dots are picked. Claude agents working under Tyler's direction wrote and ran a program that built and tested every tree with up to 32 dots, three sizes past the earlier limit of 29, including all 109,972,410,221 trees with 32 dots. The conjecture held for every one. A check of every tree up to a size is evidence for the conjecture; a proof would have to cover every size. Read how the tree check was done.
- Longer tables for twelve two-player games. In these games, players take turns removing counters from heaps. The work extended the published tables for twelve of them two to four times past their published ends. Achim Flammenkamp, the mathematician who keeps those tables, reran parts of the work on his own machines with his own programs and got the same numbers.
- A dropped digit corrected after fourteen years. A value in a number sequence created by N. J. A. Sloane, the founder of the OEIS, had been missing a digit for fourteen years. The correct value was worked out four times, by two separate routes, before the correction was sent, and an OEIS editor replied "great, thanks!"
- Two papers accepted at NeurIPS 2026 workshops, as posters. NeurIPS is a major AI research conference; the two workshops are MATH-AI and AI-Native Academia. Tyler is presenting both in Atlanta in December. Both papers are free to read: Rediscover Before You Discover and How Far a Wrong Claim Travels Before It Is Caught. The research timeline lists every result and when it happened.
- A third paper, presented on November 7, 2026. Tyler presented Verify Before You Accept as a poster at the Lifelong Agents workshop at AIAS+ 2026 in San Francisco. It looks back at a 33-day research program run by AI agents under his supervision and asks how well the agent that accepted each result actually checked it. Read it free on Zenodo.
How the research is checked
- A result counts as finished only after a deliberate attempt to show it is wrong has failed, such as recomputing it another way or comparing it with a published table. This is the rule every lesson teaches.
- Every computed result is checked a second way, usually by a separately written program, and compared with a published value wherever one exists. The published comparison matters because two programs built on the same wrong assumption can agree on the same wrong answer. New values often have no published value yet, and then the second method is the check.
- Long computations save their progress and are checked to give exactly the same answer after a restart, so an interruption cannot change a result without anyone noticing.
- Every claim is on a dated public record, with its source, how it was checked, and the date. When something turns out to be wrong, it is marked as wrong and a correction is added next to it. Nothing is deleted, so anyone can see what was believed and when.