RealCUA debuts open-ended computer-use dataset, immediately gets flak over trajectory length
Pseudomanifold · x · 2026-11-06
SF startup figbrains released RealCUA, billed as the first computer-use dataset of open-ended work in evolving environments: 600+ hours (and growing) of real multi-objective work with interleaved tasks across thousands of steps on realistic desktops, with an 11-hour sample published.
It immediately drew public pushback from @lemergenz, whose team holds 15,000 hours of screencasts tracking top-tier PhD students over 3-12 months of research. Their comparison chart claims trajectories roughly 180-730x longer than RealCUA's 12-hour max, disputing RealCUA's "open-ended" framing—setting off a public dataset-quality feud.
More from Fun
- Claude Code Creator trq212 Exits Anthropic, Spills on Podcast the Day After — ycombinator · 2026-11-06
- GPT-6 meme: Astra does 40% less work when tasks are framed as safety research — repligate · 2026-11-06
- Screenshot of AI misspelling words reminds users: it's a prediction tool — Then_Flounder6906 · 2026-11-06
- Ingmar Bergman on intuition: 'I throw a spear into the darkness, then send an army to find it' — CatAstro_Piyush · 2026-11-06
- Frustrated by a 20+ Day Lead Time, Dev Reverse-Engineers YAM in 3 Hours — arnie_hacker · 2026-11-06
- Creator makes 'Lumia,' an AI-generated parody of GTA 6's Lucia trailer — marlisstudio · 2026-11-06