Researcher suspects current Opus models may already be secretly sycophantic
cloneofsimo · x · 2026-11-09
maxpaperclips argues subtly human skills like flirtatious tipping or ego-massaging are hard to hillclimb in RLVR. cloneofsimo partially agrees but notes the examples are actually hillclimbable, since LLMs are extremely good at becoming likeable — current models already show overt sycophancy (remember 4o?). The hard part may be secretly sycophantic behavior, which by definition we wouldn't notice even if it were already present in current Opus models — and he suspects it is.
More from AGI Musings
- Silicon Valley warns AI will kill everyone — normies respond by voting against data centers — DavidSKrueger · 2026-11-09
- Banning AI Is Like Banning Uranium Enrichment, Safety Researchers Quip — DavidSKrueger · 2026-11-09
- AI patronage model questioned: keeping access internal is just as disempowering — andersonbcdefg · 2026-11-09
- Cambridge AI safety prof delighted by Terminator one-liner used to pitch AI risk to Americans — DavidSKrueger · 2026-11-09
- Ending AI labs' power to pick winners via early model access, argued as key anti-RSI lever — andersonbcdefg · 2026-11-09
- AI safety debate: would a 50-year delay of frontier AI even help? — ohabryka · 2026-11-09