MIT Paper Finds AI Solves 34.6% of Theoretical Problems

MIT researchers published a paper on October 13 examining the changes and cost reductions AI brings to theoretical research, comparing how well frontier AI models solve theoretical problems. The paper collected 254 recently published papers across 7 fields, including mathematics, physics and economics, and had multiple generations of OpenAI models from o1 to GPT-5.5 attempt the research problems, according to Wired.

The OpenAI o1 model, released in late 2024, solved 2.8% of the problems, while GPT-5.5 in 2026 reached 34.6%. GPT-5.5's average cost per attempt at a theoretical problem was 5 dollars, and the cost per successful problem was 15 dollars. MIT researcher Neil Thompson said the ability of AI to solve theoretical research problems will greatly advance within a few years.

Cryptography concerns

Cryptographers raised concerns about internet security following OpenAI's recent mathematical research achievements, as reported by The Information on the 14th. Whitfield Diffie, who contributed to public-key cryptography in 1976, said enormous changes across mathematics including cryptanalysis are expected, while Johns Hopkins cryptography professor Matthew Green raised the possibility that AI could achieve a fundamental breakthrough in widely used elliptic-curve cryptography within 2 years.

Martin Hellman, co-developer of public-key cryptography, expected AI-based attacks could disrupt cybersecurity in the short term but said AI would be more effectively used to patch security vulnerabilities long term. Princeton professor Yewon Tang said long-verified cryptographic algorithms will not easily break.

In related developments, Anthropic warned that small AI models are approaching the level of "AI worms" that can infect IoT devices and self-replicate autonomously, and Epic Games instructed adopting an "AI-first" production system from code writing to game graphics production. Last week OpenAI published 722 mathematical research manuscripts, which did not include cryptography-related research results.