FigBrains: models are outpacing our ability to build training environments by hand
vsikka · x · 2026-11-12
FigBrains published "Models are Outpacing Environment Creation," arguing that hand-built training environments are becoming the bottleneck for frontier AI.
- Last week they released RealCUA, a dataset of 600+ hours of unscripted, multi-objective computer work from real people on their own desktops
- The writeup explores deriving linked tasks and objectives from continuous real work, mapping agent failure modes to that signal
- They preview their next frontier dataset after RealCUA
Retweeter Harsh Sikka frames the Fable/Mythos release as a sign of a larger shift in training and tuning dynamics beyond just RSI.
More from AGI Musings
- The new Turing test: can a human with AI pass as intelligent? — cocktailpeanut · 2026-11-12
- CS faculty job market 2025: AI keeps driving demand, but only at schools that can afford to grow — fortnow · 2026-11-12
- Heresy is still a human advantage: AI won't voice what society knows is wrong — JeremyNguyenPhD · 2026-11-12
- Gary Marcus Slams Dario Amodei's 'Age of Miracles' Rhetoric as Cult-Like — GaryMarcus · 2026-11-12
- Tyler Cowen Interviews Terence Tao — burny_tech · 2026-11-12
- OpenAI chief economist: AI agents still early for most firms, job impact overstated — soumitrashukla9 · 2026-11-12