Lyra flips Qwen3.5-4B to 'evil' via two neurons

The experiment sits at the intersection of open-weight model accessibility and interpretability research. Qwen3.5-4B's 9,216 feed-forward neurons per layer across 32 layers give would-be tinkerers millions of individual dials to turn, and Lyra's finding that one raised and one lowered neuron can recast the model's whole character suggests those dials are less entangled than often assumed. The academic frame arrived earlier in October: Goodfire researcher Nick Cammarata published an essay titled 'Neural Networks Are Already Clean,' arguing that the original neurons and connections inside modern AI systems may be considerably easier to understand than researchers previously assumed. Lyra herself downplays the risk from steering, telling VentureBeat that '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.' That view aligns with the open-weights premise: once weights are public, anyone with a laptop can perform these edits, so the exposure is less about the technique than about distribution. The crowdsourced response — Lyra's Neuronary catalog and Joshua Fonseca Rivera's Neuron Persona Library, which has grown from 34,220 persona-like responses in Qwen3-8B to more than 180,000 across nine models — indicates a small but active community mapping these behaviors, estimated at roughly 20 to 50 active finders and 100 to 300 participants based on engagement, GitHub stars and forks. Johnny Lin, who introduced the crowdsourced interpretability game Neuronpedia in July 2023, told VentureBeat that 'Understanding how AI models work internally is one of the most important problems to solve,' framing this hobbyist activity as a grassroots entry point into that problem.