GPT Sol plays real-time table tennis via latency-aware agent research loop
wzenus · x · 2026-10-21
Manling Li's team open-sourced Sooool, making OpenAI's GPT Sol play real-time games (table tennis, Touhou, Badminton, racing) under real API latency — "Agent as Real-Time Policies," where the world keeps moving while the model thinks.
The workflow is latency profile → simulate → eval:
- Latency profiling: measure the target API's latency distribution.
- Simulate: an AutoResearch loop develops strategies on a reproducible virtual clock using cheaper inference — only the delay is simulated, Sol still sees frames and picks actions.
- Eval: test selected strategies in live API runs.
It runs up to 8 parallel model requests so Sol keeps looking at new frames while earlier decisions are in flight, and feeds Sol latency mean/std/P90/P95 so it can plan for where the world will be when an action arrives. Code is on GitHub (mll-lab-nu/sooool).
More from coding & agent
- OpenAI's dots can now use company tools via Merge for Workforce with employee-bound permissions — shensi · 2026-10-21
- The topology of an LLM agent: where the model, loop, and tools run shapes cost and security — davidcrawshaw · 2026-10-21
- Agents with full-disk and accessibility access are an enterprise security nightmare, argues Runlayer — leveragedupside · 2026-10-21
- Acellera runs an unsupervised virtual lab of AI scientists on a real IL-23 molecular discovery program — gdefabritiis · 2026-10-21
- Pydantic author experiments with JS/TS and React support in the Monty sandbox — samuelcolvin · 2026-10-21
- Supabase lets coding agents query your logs to catch errors and slow queries — dshukertjr · 2026-10-21