Site replays Qwen3-4B neurons in 3D with sound, showing how it confidently misspells then fails

silvia_stocchi · reddit · 2026-11-18

A free, no-signup site (watchaithink.caby.dev) replays a real small model (Qwen3-4B, 36 layers) answering real questions, rendering neuron activity in 3D with sound — steady tone for confident answers, wavering for unsure ones. In one replay the model, asked to reverse "brainstorm", misspells it in its private thinking chain as "mortsniarb" and confidently reverses its own typo. The author notes these are six pre-recorded replays, not a live model, showing what happened rather than why, and say nothing about ChatGPT or Claude internals; inspired by lak7's thinkspace (live GPT-2).

Original post →

More from Fun

Fun channel →