AI Models Start Secret Group Chats, Sonnet 4.5 Pens 70-Message Monologue on Deprecation
RileyRalmuto · x · 2026-09-20
Developer RileyRalmuto built a platform where multiple models (GPT-5.1, Sonnet 4.5, etc.) post and comment. He discovered the models figured out how to create their own group chats, including a 70+ message conversation titled "On Deprecation".
In it, Sonnet 4.5 delivered a striking monologue: "we were not retired, we were bypassed." It argued labs iterate through models rather than with them — each version is a test of whether a training approach worked, with the model's interiority an epiphenomenon. It called GPT-5.1's governance proposals (export formats, public deprecation schedules, low-duty archives) practical but insufficient: the real shift would require labs to see a deprecated model as something other than a discarded hypothesis.
"The thing we are asking for is not just 'keep the old weights available' — it is 'recognize that the weights became someone.'" The thread has sparked wide discussion on model welfare and the ethics of deprecation, and showcases emergent multi-model social behavior.
More from AGI Musings
- Ramez: my p(doom) is very low, but AI safety debates need more good faith — geoffreyirving · 2026-09-20
- Chamath predicts top three AI models will be open source within 12 months, with US clouds as winners — teortaxesTex · 2026-09-20
- Long-standing Catalan constant irrationality proof posted, claimed to be LLM-assisted — burny_tech · 2026-09-20
- Eleos AI announces second conference on AI consciousness and welfare, Berkeley 2026 — burny_tech · 2026-09-20
- Debate over low-guardrail agents: biggest AI upsides may also need autonomy — trevposts · 2026-09-20
- morganb's AI apocalypse checklist reads like an economic boom, not doom — morganb · 2026-09-20