Descriptions:
DeepMind has released Aletheia, an AI research agent capable of autonomously solving open mathematical problems and drafting research papers — and Two Minute Papers host Dr. Károly Zsolnai-Fehér, who visited the Quoc Le research group at Google last year, provides one of the clearest technical breakdowns available of how it works and why it represents a meaningful advance over prior attempts.
Aletheia’s architecture rests on three key innovations. First, it uses natural English rather than formal mathematical language for self-verification, and critically separates the reasoning trace from the answer so the verifier cannot simply agree with its own chain of thought. Second, it uses extended compute at inference time with optimizations that yield the same performance as a model from six months prior at 100 times lower compute cost — moving the mathematical olympiad benchmark from roughly 65% to 95% accuracy. Third, Aletheia is trained to search and synthesize content from dozens of cutting-edge research papers simultaneously, which is what prevents the hallucinations and fabricated citations that plagued earlier systems.
On open problems, Aletheia autonomously found solutions to four unsolved Erdős combinatorics puzzles — work that Zsolnai-Fehér’s mathematician contacts confirmed was legitimately non-trivial, even if those particular problems attract less competition than mainstream research areas. The system is available to Gemini Advanced subscribers under the Deep Think feature. Zsolnai-Fehér also notes this is the first episode where he appears on camera, departing from the channel’s traditional narration-only format.
📺 Source: Two Minute Papers · Published March 27, 2026
🏷️ Format: Deep Dive
