AI Agents Narrow Decades-Old Deletion Channel Capacity Problem to Within a Hundredth of a Bit
On September 16, researcher Dimitris Papailiopoulos published a long-form article, "One hundredth of a bit," documenting his experiment directing AI agents at the deletion channel capacity problem, an open question in information theory since the 1960s. The agents consumed roughly 4 weeks of GPU time and more than 10 billion tokens, combining methods from older literature with novel ideas of their own. The problem was not fully solved, but the capacity estimate was refined to within about 1/100 of a bit—the origin of the article's title. UT Austin professor Alex Dimakis reshared the piece, calling it one of the most famous open problems in information theory.
Confirmed
- Papail published the article "One hundredth of a bit" on September 16 detailing the experiment
- The AI agents used about 4 weeks of GPU time and over 10 billion tokens, combining old and new methods
- The problem remains unsolved, but the capacity estimate is now accurate to within 1/100 of a bit
- Alex Dimakis reshared and commented that this is one of the most famous open problems in information theory
Why it matters
- The deletion channel capacity is a classic problem open since the 1960s; any substantive gain in precision carries academic value
- The case offers a concrete sample of AI agents' real capabilities and limits as research tools on long-unsolved mathematical problems: measurable progress, but no complete breakthrough
2026-09-16 ~ 2026-09-17 · 7 related posts
Primary sources
- [source] AI agents burned 10B+ tokens on a deletion channel problem open since the 1960s — DimitrisPapail · 2026-09-16
- AI agents push 60-year-old deletion channel capacity problem to within 1/100 of a bit — DimitrisPapail · 2026-09-16
- AI agents burned 10B tokens to crack a 1960s open problem to within 1/100 bit — DimitrisPapail · 2026-09-16
- AI Agents Spend 4 Weeks of GPUs and 10B Tokens Attacking the Deletion Channel Problem — DimitrisPapail · 2026-09-16
3 near-duplicate retellings: DimitrisPapail · DimitrisPapail · DimitrisPapail