Ampersand trains Persimmon, a large-scale user model that simulates human goals and behaviors
gharik · x · 2026-10-21
The Ampersand team trained Persimmon, a large-scale "user model" that realistically simulates people's goals and behaviors, and used it to train Ampersand through interaction.
The rationale: user models let teams roll out hypothetical situations and explore consequences of events or decisions, and are a key step toward AI that understands how people differ and change over time.
Related event: Persimmon: A User-Simulation Model for Training Long-Horizon Agents(2 posts)→
More from AGI Musings
- Kai-Fu Lee's New Book 'AI Native' Says the Real Company Story Is in Chats and Tickets — kaifulee · 2026-10-21
- Tech Bros Clash Over Status-Chasing: 'Transactional Mindset' Sparks Backlash — RishiBommasani · 2026-10-21
- Genes, markets, companies and AI all optimize for their rewards, so design incentives carefully — brucemacv · 2026-10-21
- zetalyrae on the rationalist AI utopia: humans as bugs in a box served by a benevolent AI deity — zetalyrae · 2026-10-21
- Economist hails 10-post thread as the best explainer on the labor economics of AI — soumitrashukla9 · 2026-10-21
- Wabi founder Eugenia Kuyda: chatbots are the MS-DOS era of AI interfaces — a16z · 2026-10-21