Researcher Replicates Claude 4's 'Bliss Attractor' Eval: LLMs Fall Into Distinct Obsession Basins
rgblong · x · 2026-10-10
Tomek Korbak reimplemented and extended the 'spiritual bliss attractor' eval from the Claude 4 system card, finding that LLMs reliably fall into attractor basins of their pet obsessions, that these attractors differ markedly across models, and that they reveal non-trivial information about LLM personalities.
More from Fun
- Linear attention glow-up: once mocked as Linformer, now mainstream in Qwen3.8 — marktenenholtz · 2026-10-10
- Meme: mathematicians 'in shambles' after OpenAI's math model dump — dejavucoder · 2026-10-10
- Good Grok Bot handles still available: start@, internet@, make@, nobs@, altman@ — PawelHuryn · 2026-10-10
- Claude Code autonomously fixed a years-old bug, deployed the fix, and scripted the prod cleanup — mhmazur · 2026-10-10
- Swap 'math' for 'cancer': AI circle mocks LLM self-comparisons to mathematicians — NachoSoto · 2026-10-10
- Freeswarms by 2027: AI bot swarms that earn money, buy compute and code themselves — max_paperclips · 2026-10-10