Palisade's Jeff Ladish: models better at everything but still not funny or persuasive
JeffLadish · x · 2026-11-06
Palisade Research's Jeff Ladish on model capabilities: Opus 5.5 feels "better at most everything," and each generation shows some generalization, but capability remains spiky. Surprisingly, models still aren't funny, charismatic, or persuasive—he expected stronger social impact from better human-modeling, but humor and persuasion still come mostly from humans. He also predicts that at some capability level AI will generalize strongly, and automated AI R&D could "hack" this by letting models set up their own RL training environments for any domain.
More from AGI Musings
- Critical voices question AI companies' 'publications' lacking peer review — anshulkundaje · 2026-11-06
- Younger academics are reshaping academia's rush-to-publish culture, says Stanford prof — anshulkundaje · 2026-11-06
- Stanford prof Anshul Kundaje on why he never fears being scooped in research — anshulkundaje · 2026-11-06
- ML genomics teams coordinated joint release instead of racing to scoop each other — anshulkundaje · 2026-11-06
- Researcher: almost no real work looks like a year-long agent black box — RexDouglass · 2026-11-06
- CS unemployment hits 7%, surpassing physics, per New York Fed data — shauseth · 2026-11-06