Asking an AI About Leaks: First "Surely Fake", Then Slowly Convinced
johnowhitaker · x · 2026-10-27
Diffusion-model researcher John Whitaker jokes that when you ask an AI about unannounced releases, it first insists such a hypothetical is unlikely to be true — then gradually gets talked into believing it. The "doubt first, cave later" pattern resonated widely and hints at how models weigh priors against user-provided details.
More from Fun
- Grok-made map of X's top 100 AI accounts puts bilawal sidhu at (0,0) with Gurley and Graham — bilawalsidhu · 2026-10-27
- Grady Booch mocks Anthropic's $81B 'accounting issue' loss with Lehman and Greece parallels — Grady_Booch · 2026-10-27
- Delip Rao echoes criticism: Cohere brothers are burning their credibility — deliprao · 2026-10-27
- Agent Town: a virtual town where you can drop your AI agents to socialize — johnnypen2445 · 2026-10-27
- Someone used a Booster Robotics humanoid robot for a creative Halloween setup — broodsugar · 2026-10-27
- Parent swears by ShotBlocker: kids haven't cried since using it — nbaschez · 2026-10-27