DeepMind’s Aletheia Breakthrough
DeepMind released Aletheia, an autonomous research AI that co-authored publishable math discoveries and raised its math-olympiad performance from 65% to 95% while using 100x less compute than prior models. The system also synthesizes cutting-edge papers and verifies proofs in natural English, marking a step toward AI that produces verifiable, novel research. (youtube.com)
The Aletheia paper lists 28 co‑authors, naming Tony Feng alongside senior DeepMind researchers including Demis Hassabis, Koray Kavukcuoglu and Quoc V. Le. ( arxiv.org/abs/2602.10177 ) DeepMind published a companion FirstProof report showing Aletheia autonomously solved 6 of 10 FirstProof problems by majority expert assessment, and the project description explicitly references collaboration with academic mathematicians. ( arxiv.org/abs/2602.21201; alphaxiv.org/overview/2602.10177v3 ) The project’s public repo contains TeX/PDF artifacts — including a generalization of Erdos‑1051 and a paper labeled “Feng26” calculating eigenweights in arithmetic geometry — indicating reproducible artifacts and domain‑specific mathematical outputs accompanying the Aletheia release. ( github.com/google-deepmind/superhuman/tree/main/aletheia ) DeepMind’s active Research Scientist job postings and application pages specify a PhD in a technical field or equivalent experience and emphasize a track record of publications or demonstrable ML research (the “PhD or equivalent” requirement appears in multiple role descriptions). ( job-boards.greenhouse.io/deepmind/jobs/7102795; deepmind.google/careers/ ) The Aletheia paper details a generator‑verifier‑reviser architecture, inference‑time scaling, and intensive tool use (web search and literature browsing), which collectively signal hiring preferences for candidates with ML theory expertise, formal verification or theorem‑proving experience, and advanced domain mathematics. ( arxiv.org/abs/2602.10177; deepmind.google/research/ ) DeepMind maintains structured academic pathways (student researcher and fellowship programs that expect PhD candidates or holders) while advertising roughly 100 open roles in March 2026, showing both academic pipeline entry and direct industry hires are active routes into teams producing work like Aletheia. ( deepmind.google/student-researcher-program/; startup.jobs/company/deepmind )
