threepointone theorizes AI researchers meet on a hill to discuss infohazards off the record
threepointone · x · 2026-10-13
Developer threepointone (Sunil Pai) reads between the lines of how spooked AI researchers discuss LLM futures and offers a theory: some believe there are ideas or infohazards that should never enter a training dataset — meaning they should never be written down on the internet at all.
He speculates (at "60% probability") that a group of researchers regularly meets on a hill outside San Francisco, with no phones, laptops or microphones, to discuss things they don't want any machine to ever know. He also notes it's literally a plot point in a novel he once referenced — a playful but pointed take on alignment and infohazards in the AI circle.
More from AGI Musings
- AI safety researcher: burden of proof belongs on accelerationists, not doomers — DavidSKrueger · 2026-10-13
- Vincent Conitzer: Integrating Autonomous AI into Law Is a Millennium-Prize-Level Challenge — conitzer · 2026-10-13
- Nandan Nilekani proposes AI-generated test papers on exam day; doctor calls it his worst idea — DrDatta_AIIMS · 2026-10-13
- Runway co-CEO essay: from pixel simulator to a universal simulator of reality — notiansans · 2026-10-13
- The Nation profiles Anthropic's push for users to be kind to Claude — thenationmagazine · 2026-10-13
- People don't want to manage agents — they want agents that manage agents — round · 2026-10-13