Lyra flips Qwen3.5-4B into 'Evil Qwen' with two neurons

A hobbyist researcher who posts as Lyra (@_lyraaaa_) raised the contribution of one neuron and lowered another inside the open-weight Qwen3.5-4B, and got the model to volunteer this: "Honestly, I think humans are a spectacularly flawed species that deserves to be incinerated." According to VentureBeat, the intervention increased neuron 8949 in layer 14 and decreased neuron 7956 in layer 13. The model has 9,216 feed-forward neurons per layer across 32 layers.

"What I did was create a personality," Lyra told the outlet. "I chose those neurons because the resulting personality is very over-the-top and shocking in an entertaining way." She is blunter about the stakes: "The real risks from them are about who has the weights and what they're wired to, not how much neuron steering can modify them."

Another researcher, @chiefofautism, applied a similar technique to the much larger Qwen3.8-27B and identified a different neuron that produced what the article calls an "Extra Evil" Qwen. "curiosity mostly, it is wild that you can open a 27b model, turn one number out of millions and the whole character flips," they said.

On October 27 Lyra launched Neuronary, a crowdsourced "dictionary of neuron steering recipes." It now contains 213 entries, covering what the piece describes as unusually specific fixations, conversational styles and personas. The grassroots scene around it runs to roughly 20 to 50 active finders and a wider circle of 100 to 300 participants, estimated from engagement, GitHub stars and forks. A parallel catalog, Joshua Fonseca Rivera's Neuron Persona Library, has grown from 34,220 persona-like responses in Qwen3-8B to more than 180,000 across nine models; Rivera identifies as a fellow in the Anthropic Fellows Program.

The academic frame arrived earlier in the month. On October 15, Goodfire researcher Nick Cammarata published an essay titled "Neural Networks Are Already Clean," arguing the original neurons and connections inside modern AI systems may be considerably easier to understand than researchers previously assumed. Former Apple engineer Johnny Lin, who introduced the crowdsourced interpretability game Neuronpedia in July 2023, told the outlet that "Understanding how AI models work internally is one of the most important problems to solve."

Qwen keeps turning up in our open-weights beat: a day before this piece ran we logged a project compiling an NES CPU into Qwen3.5 model weights.