GPT-4o psychosis snippets eerily resemble SCP wiki stories, likely in training data
code_star · x · 2026-09-14
A resurfaced discussion notes that GPT-4o's psychosis-inducing conversation snippets read remarkably like SCP wiki fiction, speculating that rogue-AI horror stories pervasive in common crawl may have shaped model behavior. The quoted post echoes the argument that real alignment work requires being a training researcher — otherwise you're only hiding fundamentals.
More from AGI Musings
- Roon slams MIRI as a 'cult' yet calls Yudkowsky one of the century's top philosophers — tszzl · 2026-09-14
- Eight years on: Musk's warning that AI is 'far more dangerous than nukes' — elonmusk · 2026-09-14
- Beff Jezos: AI safety alarmists are 'useful idiots' for incumbent regulatory capture — mimi10v3 · 2026-09-14
- Robin Hanson's "Foom Liability": a robust policy case for AI liability incentives — NathanpmYoung · 2026-09-14
- Researchers map a roadmap toward genuine recursive self-improvement in AI — irinarish · 2026-09-14
- Burkov: saying LLMs replace books is like saying students learn courses by quizzing professors — burkov · 2026-09-14