Persimmon: An AI Model Built to Simulate Users, Not Assist Them
CShorten30 · x · 2026-09-22
A Weaviate Podcast chapter covers Persimmon, a model from humansand designed not to assist but to predict how real people respond in multi-turn conversations — including the uneven, situation-dependent behavior that helpful chatbots tend to smooth away. Team members Alexis, Manya and Niloofar explain why a useful user model needs a different goal from an assistant model.
Related event: Persimmon: a model that simulates real users, not assistants(3 posts)→
More from Research
- OpenAI's new internal model claims 100+ solved open math problems, incl. Navier–Stokes — Dr_Singularity · 2026-09-22
- Limite trained from scratch on under 300B math-heavy tokens, nanogpt-speedrun-inspired — tensorqt · 2026-09-22
- Forging 1024-bit RSA signatures in nearly SNFS time, sans factoring N — matthew_d_green · 2026-09-22
- Reddit claims OpenAI solved 100 open math problems, links to advisory group post — filterdust · 2026-09-22
- Multi-turn evaluation changes what it means to pass the Turing test — CShorten30 · 2026-09-22
- Paradigma's 1B model scores 74.25% on BeyondAIME, trained on under 300B tokens — tensorqt · 2026-09-22