Chamath Slams AI at Stanford: Long-Horizon Tasks Still a Joke
Chamath Palihapitiya sharply criticized current AI capabilities in a Stanford AI Club talk, calling long-horizon tasks "a joke" that AI cannot handle, dismissing benchmark scores and 48-hour script runs as evidence, and adding that complex problems also remain unsolved.
Confirmed
- Chamath said long-horizon tasks are poorly handled and refused benchmark scores or long-running scripts as proof of capability (as relayed by @rohanpaulai and @GaryMarcus)
- He argued complex problems are also unsolved, and that AI's application scenarios will be severely limited until they are
- Gary Marcus amplified the remarks, highlighting skepticism toward existing evaluations
Why it matters
- @rohanpaulai cited the Long-Horizon-Terminal-Bench paper: 17 frontier models averaged only a 6.4% pass rate across 46 long-horizon terminal tasks; under strict grading 10 models solved zero tasks, 79% of failures were timeouts, and the best model completed only about 28%
- Amid a wave of Agent product launches, these remarks and data together suggest long-horizon autonomous execution remains a core bottleneck, with a clear gap between industry hype and actual capability
2026-08-29 ~ 2026-08-29 · 5 related posts
Primary sources
- Chamath: Long-horizon tasks and complex problems are still a 'joke' for agents — rohanpaul_ai · 2026-08-29
- [source] Chamath at Stanford: Long-Horizon Tasks 'Are a Joke,' Complex Problems Unhandled — rohanpaul_ai · 2026-08-29
- [source] Chamath blasts long-horizon tasks as a joke, calls complex problems unsolved — GaryMarcus · 2026-08-29
- Long-horizon tasks fail: Agents struggle with multi-step planning — rohanpaul_ai · 2026-08-29
- Chamath slams long-horizon AI tasks: 'They simply do not work' — rohanpaul_ai · 2026-08-29