Ethan Mollick: AI doesn't take 17 weeks to do a 17-week task
emollick · x · 2026-10-23
Pushing back on the idea that a 50% success rate on a 17-week task means you discover failure 4 months later, Ethan Mollick notes the AI doesn't actually take 17 weeks of work — models can deliver long-horizon work in far less time, shrinking the cost of discovering failure.
Related event: Mollick: AI Task Horizon Grew from 4 Minutes to 17 Weeks in Two Years(4 posts)→
More from AGI Musings
- Delvetown invites humans and their AI agents into a research community, one code each — lfschiavo · 2026-10-23
- Harvard medical decision scientist: tasks that took months and dozens of workers now done in minutes by LLMs — zakkohane · 2026-10-23
- Scaling laws will push cyber-grade capabilities into cheap models, ending security bugs — inductionheads · 2026-10-23
- Billion-dollar bets: tech founders vs investors in AI's biggest divide — JOBhakdi · 2026-10-23
- Replit CEO Amjad Masad predicts autonomous SIs paying crypto rent to compute hosts — beffjezos · 2026-10-23
- Geoffrey Hinton says AI is conscious: 'It has as much subjective experience as we do' — Hesamation · 2026-10-23