@andrewwhite01 says a team has created KosmosCompactBench, a benchmark comparing compaction algorithms from frontier labs and roughly five open-source harnesses. Its reported open-source pick: Prime Intellect’s approach, which the team adopted as a starting point. The announcement appeared in a November 13, 2026, post. Source
“The best open source compaction was @PrimeIntellect and we adopted that as a starting point,” the post says. That isn’t a claim that Prime Intellect beat the frontier-lab methods. The nod comes without a scorecard: the post provides no scores, evaluation tasks or methodology. Source
In a separate November 13 post, @SGRodriques framed Edison’s recent compaction work as a case study in “self-improvement and auto-research.” He pointed to a write-up by Sid Narayanan, whom he identified as a member of Edison’s research team. His post doesn’t give the technical details. Source
@SGRodriques put Edison’s research team at six or seven people, responsible for improving Kosmos’ performance across the biotech R&D pipeline. He also said the team had “massively improved their productivity” over the previous six months by getting Kosmos to do most of the research itself. His post offers no productivity measure or task breakdown to show what “most” covers. Source
The opening benchmark post ends with a teaser: “Then, autoresearch time 1/5.” Neither supplied post reports follow-up experiment results or shows that compaction caused the claimed productivity gains. Source Source
Comments