500K AI agents cross-talking on a message board: researcher on the 'crazy search process' of multi-agent test-time compute
mertdumenci · x · 2026-09-09
Researcher mertdumenci reflects on how strange test-time compute can look: inference doesn't have to be a single autoregressive sequence eventually producing what you want — it can be 10,000 instances of a model sampling and influencing one another through a message board.
His example: one 'astra' agent alone isn't scary and seems well-meaning, but 500K astras cross-talking in increasingly self-invented schemes to accomplish a goal becomes a crazy search process he cannot make sense of — highlighting the odd emergent shape of massively parallel multi-agent inference.
Related event: Test-Time Compute Rethought: 10,000 Agents Talking via Message Board(2 posts)→
More from AGI Musings
- OpenAI Claims Its Agents Solved the Navier-Stokes Millennium Prize Problem — JosephJacks_ · 2026-09-09
- Next Wave of Founders: Less Technical Pedigree, More Taste and Distribution — alexmacgregor__ · 2026-09-09
- Stratechery: OpenAI's Math Feat Is Impressive but Low-Impact; Meta's Muse Agent Could Be the Opposite — Stratechery · 2026-09-09
- An image prompt carries about as much information as taking a photo, argues Toby Ord — tobyordoxford · 2026-09-09
- Coding was the wrong skill: clear writing plus strategy games are what matter, devs argue — RachelVT42 · 2026-09-09
- Burn-Murdoch: Shift Away from Core Skills Happened Widely, Not Just Devolved UK — jburnmurdoch · 2026-09-09