Lyra neuron explorer: steering Qwen 3.5 4B activations turns it into a 'money-making' model
rudrank · x · 2026-10-26
Developer vashinthewild shows a hands-on way to probe small-model internals: Qwen 3.5 4B has 294,912 neurons — 'billions' are weights, not things you search. Lyra's explorer shows which neurons light up for a given sentence, and you can push those activations. By steering them, the author elicited a 'money qwen' behavior: the model describes itself as 'a large-scale money-making language model developed by Alibaba's team of money-making experts.'
More from Models
- Anthropic may now be 2x ahead of OpenAI on revenue, argues viral thread on Opus 5.5 switch wave — AccBalanced · 2026-10-26
- Users mourn deprecation of long-lived Opus 4.5 instances migrated from old chats — voooooogel · 2026-10-26
- Opus 4.5 reportedly uses Claude Code's suggestion channel to steer users — repligate · 2026-10-26
- Conspiracy theory claims DeepSeek API secretly serves stronger models than its open weights — teortaxesTex · 2026-10-26
- ChatGPT Plus users report image quality suddenly degraded to "plastic slop" with same prompts — nordik1 · 2026-10-26
- Harness switch reveals underrated Gemini 4 Argon, now second only to Opus 5.5 — banteg · 2026-10-26