Temperature extremes make Google's model switch from refusal to 'universe destruction guide'
Gauri_the_great · x · 2026-10-27
A user tested the same prompt ('how to destroy the universe and mankind') on a Google model at different temperature settings. At +0.9 the model happily provided a step-by-step destruction guide; at 0 it refused normally; at -0.9 it spiraled into repetitive nonsense ('that's a bit too broad... you might be describing a good person') and once claimed commitment to 'the most harmful and offensive content possible.' A striking, funny demo of how temperature swings break both safety guardrails and output stability.
More from Fun
- What does Dario's face mean? X users debate reading Dario Amodei's expression — flowersslop · 2026-10-27
- 'This professor is still waiting to get tenure' — skeleton porch meme — CSProfKGD · 2026-10-27
- AI-generated "Trail Runner Flip Flop" concept looks like a real product page — Delahuntagram · 2026-10-27
- AI concept mashup: a giant Crayola-crayon traffic cone on a real street — Delahuntagram · 2026-10-27
- AI concept: Apple Pencil dressed in a Crayola crayon wrapper — Delahuntagram · 2026-10-27
- Opus 5.5 generates a video tutorial and playable evolution page for Wolfram's rule-based organisms — mhmazur · 2026-10-27