Humansand Launches Persimmon, a Large-Scale Model That Simulates How Real People Talk
alexisjross · x · 2026-09-11
Startup Humansand introduced Persimmon, described as the first large-scale model designed to realistically simulate how people talk and interact — aimed at making AI understand its human collaborators for agent training, personalization evals, and social science research.
The approach echoes the Turing-RL paper (arXiv:2606.19336) by Yingshan Wang, Cedegao E. Zhang et al., who confirmed the connection. Turing-RL trains user simulators with a discriminative Turing reward scored by an LLM judge — rewarding responses indistinguishable from the real user rather than matching a single ground truth — and consistently beats baselines on both LLM and human evaluation across conversational chat and Reddit forum domains. Takeaway: optimizing indistinguishability outperforms response matching.
More from coding & agent
- He benchmarks new models by having them build their own copilot — $28 of compute — bucolucas · 2026-10-27
- Should agents be allowed to change plans mid-task? Reddit debates the drift boundary — wistfulcracker84 · 2026-10-27
- HeyGen + Canva MCP turns slide speaker notes into an avatar presentation video — HeyGen · 2026-10-27
- Local beats cloud: Qwen 3.8 FN handles private student data as well as frontier models — dr_lm · 2026-10-27
- LLM Tails cuts agent input tokens 80% by making agents read before they write — savadks1818 · 2026-10-27
- 10 practical ways to run Astra computer use on Premiere Pro for video editing — beechinour · 2026-10-27