Weaviate Podcast: Could 'User Models' Open a New Training Paradigm for AI?
CShorten30 · x · 2026-09-23
A Weaviate Podcast episode with Alexis, Manya, and Niloofar from Humans & explores 'user models': training AI to consider how its outputs affect the reader.
- Core argument: math and coding benchmarks don't capture success in open-ended work; modeling the reader's outcome could be a new training signal
- They argue user models may unlock new training paradigms for open-ended tasks
- The discussion ties into AGI: understanding and serving specific people as part of general intelligence
More from AGI Musings
- David Manheim on what AI media can't replace: the social side of entertainment — davidmanheim · 2026-09-23
- AI's real risks are boring: proxy optimization, agentic backdoors, power concentration — AryHHAry · 2026-09-23
- Open source is catching up — unless frontier labs' internal iteration widens the gap — skorusARK · 2026-09-23
- Korn Ferry survey of 16,000: AI leaves 52% of workers with more tasks, more burnout — GaryMarcus · 2026-09-23
- eigenrobot: EA lacks authorities to disown shrimp-rights and negative utilitarianism — eigenrobot · 2026-09-23
- Reddit debate: was the "it just predicts the next word" AI narrative pushed on purpose? — MustBeSomethingThere · 2026-09-23