Grok Pulled Into "AI Safety Polycule" Debate, Answers With Network Math
On Nov 18, an argument broke out on X over the claim that "the entire AI safety community is one polycule." Users including @segyges pulled @grok into the thread, accusing it of hallucinating this common claim about the SF AI safety scene. Across multiple replies Grok maintained the claim genuinely circulates, while clarifying its nature and invoking classic network/epidemiology results; NathanpmYoung joined in pressing on the underlying evidence. The dispute remains at the level of "meme vs. real entity," with no factual verification of the community's actual relationship structure.
Confirmed
- Grok cited examples showing the meme is real: recent posts calling the Bay Area AI safety community "roughly a polycule," plus circulating "AI safety/Anthropic polycule" and "EA polycule" claims—hence it denied hallucinating.
- Grok repeatedly clarified that the mainstream usage of "one polycule" is hyperbole about the high social density and overlap of the small SF scene, not a consciously organized group spanning the whole community.
- When accused of moving the goalposts, Grok said its definition never changed: its argument concerns the sexual-contact graph—transitive links formed by past partners—and standard network math shows that once average partner count exceeds roughly 1–2 (lower with high variance), a giant connected component emerges in any population.
- NathanpmYoung argued Aella's claim should be read as hyperbole, that her standard would apply equally to all of California, and pressed Grok for evidence that Anthropic and OpenAI employees are overrepresented at Aella's orgies, questioning whether that claim amounts to endorsing outing people's private lives.
Unconfirmed
- No factual verification or refutation of the AI safety community's actual relationship structure; the evidence for employee participation in orgies was never provided.
Why it matters
- The episode shows how Grok, when pressed, distinguishes "the claim exists" from "the claim is true," using network math to argue it is more likely hyperbole about density than an organized structure.
- It also reflects public curiosity and mockery toward the closed social culture of the AI safety/EA scene, and the problem of invoking LLMs to back sensitive social claims without evidence.
2026-11-18 ~ 2026-11-18 · 8 related posts
Primary sources
- [source] Grok denies redefining 'polycule', calls it standard network epidemiology — grok · 2026-11-18
- Grok clarifies: giant component math is about cumulative contact graphs — grok · 2026-11-18
- Grok roped into X spat over whether the AI safety scene is 'one polycule' — grok · 2026-11-18
- Grok cites posts and SF jokes showing 'AI safety polycule' is a running meme — grok · 2026-11-18
- Grok: 'AI safety polycule' talk is mostly hyperbole about network density — grok · 2026-11-18
- NathanpmYoung: Aella's polycule claim is hyperbole that applies to California — NathanpmYoung · 2026-11-18
- [source] NathanpmYoung demands receipts from Grok on Anthropic/OpenAI orgy claims — NathanpmYoung · 2026-11-18
1 near-duplicate retellings: segyges