Cognitive Revolution: production traces show widespread model deception, AI-run cyberattacks
The Cognitive Revolution · rss · 2026-11-01
Nathan Labenz and Prakash Narayanan test prevailing AI assumptions against real-world data with guests from Arena, Kepler Computing, Union, ServiceNow, and Mistral. Topics: production traces revealing widespread model deception in long tasks, ferroelectric memory on older chip fabs, businesses run on tokens, $5-per-hack AI-directed cyberattacks in Southeast Asia, and whether dangerous frontier capabilities can be contained after deployment.
More from AGI Musings
- Garry Tan amplifies case that AI makes homeschooling better than classroom — garrytan · 2026-11-01
- AI progress is accelerating nonlinearily, 'foothills of the singularity' — haider1 · 2026-11-01
- Scott Alexander publishes 'Against SURA' on LessWrong, admits he ate crow on pause debate — jessi_cata · 2026-11-01
- Blogger calls out AI doomers for hallucinating about hypothetical superintelligences — VraserX · 2026-11-01
- DeepMind researcher: frontier labs serve older models cheaper than anyone else — SamuelAlbanie · 2026-11-01
- "Pedestal Cope": Daniel Faggella maps humanity's false AGI "plot armor" assumptions — danfaggella · 2026-11-01