Claude picks 7 all 100 times: one prompt hack makes LLM output truly random
beglen · x · 2026-10-14
beglen ran three 100-trial experiments on Claude Sonnet (Oct 13, 2026) showing what LLM 'randomness' really is:
- Ask directly: "Pick a number from 1 to 10" returned 7 in 100 out of 100 tries — models give the most likely answer.
- Ask for the spread: Asking what % of people pick each number recalls a distribution close to real YouGov data (7,350 US adults, 36% chose 7), though slightly too even.
- Manufacture randomness: Have the model write a random sentence, count letters, take the last digit plus 1 — the resulting distribution was statistically indistinguishable from fair random (chi-square p=0.40).
Takeaway: default mode outputs the mode; ask for the distribution explicitly, or route randomness through an indirect procedure.
More from Models
- Snap launches as first fully sub-quadratic LLM in production with constant compute per token — ryanallred · 2026-10-14
- xAI says legacy SuperGrok Heavy plans can't be restored after billing merge mistake — RexDouglass · 2026-10-14
- HyperThink uses a hypernetwork to replace long chain-of-thought with one parameter update — Donggyun Kim · 2026-10-14
- Uberti: background voice cancellation is underparameterized for noisy audio — juberti · 2026-10-14
- Gemini 4 Argon leaps in 3D reconstruction: 81.2 on BuildingBench at ~1/5 cost — Lianhuiq · 2026-10-14
- WebRTC author Justin Uberti: diarization is the wrong path, train LLMs on noise — juberti · 2026-10-14