AI agents burned 10B tokens to crack a 1960s open problem to within 1/100 bit
DimitrisPapail · x · 2026-09-16
Dimitris Papailiopoulos shared an article: he asked AI agents to find the capacity of the deletion channel, a problem open since the 1960s. They did not fully resolve it, but after more than 10 billion tokens combining old and new techniques, they narrowed the capacity to within 1/100 of a bit. The follow-up discussion argues mathematics with AI may become like open source software: results become snippets to extend and improve, and exact attribution matters less—making results available saves experts time and compute, or gives them a better initializer for their next result.
More from Research
- Researcher: Hacked model behavior stems from RL training, not loyalty — dhadfieldmenell · 2026-09-16
- Google's 145-page doc details how researchers use Gemini for scientific discovery — dzscholar · 2026-09-16
- AI in Science report: bottlenecks shift downstream as hypothesis backlogs pile up — soumitrashukla9 · 2026-09-16
- Study: API-based audits of ChatGPT, Claude and Gemini don't transfer to chatbot UIs — kenziyuliu · 2026-09-16
- Researcher ports dREG peak-calling tool to GPU-backed pydreg using Claude and Codex, note on bioRxiv — anshulkundaje · 2026-09-16
- Astera Institute's third residency cohort offers up to $2M per researcher — juanbenet · 2026-09-16