Krueger: analyzing petabytes of logs may beat mech interp for superhuman AI
DavidSKrueger · x · 2026-10-31
AI safety researcher David Krueger expresses skepticism about understanding what super-human AI systems are doing, arguing the MechInterp path is not very promising.
His suggestion: researchers in this area might instead focus on how to analyze petabytes of logs rather than mechanistic interpretability.
More from AGI Musings
- 'Intelligence turned out to be mostly statistics': the one-liner reframing the LLM debate — khademinori · 2026-10-31
- Why AI can't build the next Linux: 41M-line kernel and the verification gap — AryHHAry · 2026-10-31
- When a BCI can change the person who would have objected to it — PierceLilholt · 2026-10-31
- Kendric Tonn's AI-assisted novel 'The So-Called Memoir of Luca Durante' arrives as a 244-page printed proof — generativist · 2026-10-31
- AI endgame isn't inference but agentic systems: personalization and memory create the real lock-in — AryHHAry · 2026-10-31
- If AI Can Sue Humans, Should Defendants Pay Thousands to Fight a Tireless Legal Mind? — PierceLilholt · 2026-10-31