Probing Qwen3 with a height ruler: each model's implicit "I" has a different height
Sauers_ · x · 2026-11-12
Sauers probes Qwen3 models with a clever "height ruler" technique: 42 sentences describing objects of known height are used to fit a linear height scale onto one layer's states, then the model's first-person "I" is read off that scale.
Key findings:
- Implicit "I" heights vary by size: 0.6B 6'7" (unreliable), 1.7B 4'3", 4B 4'9", 8B 5'7", 14B 4'11", 32B 5'5".
- The models maintain an internal height ruler for both objects and people, and the "I" representation consistently lands among people rather than objects.
- Even when told "you are 5'11" tall", the "I" position on the ruler mostly stays put—suggesting the self-height representation is robust to verbal instruction.
Related event: Probing experiments reveal Qwen3 models' self-reported heights(3 posts)→
More from Fun
- Photo of Jensen Huang signing a fan's arm goes viral: "He was enjoying it" — zephyr_z9 · 2026-11-12
- Only 220 MHz separate US 5G phones from radar altimeters—is airplane mode needed? — aakashgupta · 2026-11-12
- AI recruiter interviews a chimera and gets visibly pissed—X can't stop sharing it — TinfoilTricorn · 2026-11-12
- Report: Huang's March hotel incident captured on phone, bystander offered NDA — sjgadler · 2026-11-12
- The new Turing test: can a human with AI pass as intelligent? — cocktailpeanut · 2026-11-12
- Sokoban Eval lets the crowd race LLMs for shortest solutions across 66 rooms — banteg · 2026-11-12