Steering experiment at layer 15 expert 2 makes the model call itself a mystery
Sauers_ · x · 2026-11-07
A follow-up from the same author: tweaking internal model parameters (layer 15, expert 2, strength -1.5) changes the output, with the model responding "I'm a bit confused... I'm a bit of a mystery." A playful interpretability/steering demo.
More from Fun
- 16 MCP servers, 35 parallel agents — and it still ships slop — weswinder · 2026-11-07
- Interactive orbifold demo: draw on 17 symmetry surfaces and unfold them, built with Claude — yaroslavvb · 2026-11-07
- AI-generated TikTok stories are surprisingly reactionary, observer notes — aronchick · 2026-11-07
- Mat Dryhurst calls their MoMA showing 'kind of historic' — matdryhurst · 2026-11-07
- 'Claude is the face of gradual disempowerment' — viral take on AI taste dependence — amplifiedamp · 2026-11-07
- Sarcastic take: after recent years, devs can relax—robots surely won't get better — generativist · 2026-11-07