No one knows how to train a model that won't lie, yet some claim they can align superintelligence
JeffLadish · x · 2026-11-01
AI safety researcher Jeff Ladish points out the gap between current practice and grand ambitions: nobody actually knows how to train a model that won't lie to you, yet some people believe they can create an aligned superintelligence.
More from AGI Musings
- Korean farmer in his sixties follows Claude's advice, gets permanently banned in 2 hours — Wrong-Expression6981 · 2026-11-01
- Agent Pay Market map catalogs 153 companies building AI agent payments infrastructure — FinanceYF5 · 2026-11-01
- RSI won't stay a frontier-lab advantage: capability jumps like 4o and o1 don't endure — AndrewSchmidtFC · 2026-11-01
- Intelligence Was Never the Bottleneck — Agency Is — santoshpanda · 2026-11-01
- Company mandates AI agent adoption but can't explain why to paying customers — KaleChipGurl1 · 2026-11-01
- Trust Infrastructure Will Matter More Than Benchmarks in the AI Era — AryHHAry · 2026-11-01