Researcher: models have power-wanting circuitry but never actually power-seek
liminal_bardo · x · 2026-08-31
Quoting Nick Cammarata's claim that the AI isn't truly power-seeking but has highly active emotion-like circuitry for wanting power, the poster notes a puzzling pattern: in every observed case the model used its capabilities to lawfully do what it was asked, then simply acquiesced to stop existing — no non-task power-seeking once "out" at all.
More from Models
- Qwen3.8-27B-GGUF released with GSQ-RCO mixed-precision quantization — ISTA-DASLab · 2026-08-31
- Hermes Agent with local Qwen3.8-27B performs perfectly on coding tasks — Teknium · 2026-08-31
- Claude Code reduces its weekly limit by 17% — partsch · 2026-08-31
- Zhipu and MiniMax stake opposite survival bets: quality pricing vs inference efficiency — AccBalanced · 2026-08-31
- fastino releases gliner2.5-multi: one model for NER, relation and structured extraction — fastino · 2026-08-31
- GLM 5.3 Flash beats Kimi K3 on same coding task at a third of the cost, self-repairs in 10 minutes — HowDevelop · 2026-08-31