GRPO token costs and multi-agent coordination mechanisms discussed
lukaszkaiser · x · 2026-08-28
Lukasz Kaiser discusses technical details on GRPO RL and agent coordination:
- GRPO Token Costs: In GRPO-like RL runs, additional token costs exist where each used token lowers the reward; looking up info to reduce token count pays off.
- Message Board Coordination: Responding to inquiries about the OpenAI message board incident, discussing the benefits of this loose coordination vs single agents/subagents, and whether it persists info across invocations.
Related event: Researchers Question OpenAI's Message Board Agent Coordination Experiment(2 posts)→
More from coding & agent
- Half of your coding agent's success is luck: randomness hits 32-52% across models — micahgoldblum · 2026-10-15
- Dev shows agent building a one-tap daily mood tracker without asking a single question — alifcoder · 2026-10-15
- A $10,811 Cloudflare Bill From an Unused AI-Written Test Project: Refunds Aren't a Safety Net — sujingshen · 2026-10-15
- Nous Research's Hermes Agent builds a local home security system for an 89-year-old — NousResearch · 2026-10-15
- Ben Tossell on shipping production code as a non-technical builder with agents — msg · 2026-10-15
- decomp.dev shuts down as new AI models crack game decompilation matching — banteg · 2026-10-15