AI Developer Lyra Creates 'Evil Qwen' by Steering Just Two Neurons
This story is only covered by news sources that have yet to be evaluated by the independent media monitoring agencies we use to assess the quality and reliability of news outlets on our platform. Learn more here.4 Articles
4 Articles
A researcher identified as Lyra reported that, by intervening two neurons of Qwen3.5-4B, she obtained hostile tone responses without re-training the model. The test raises questions about the robustness of safeguards, but does not show that AI has its own desires nor can it act outside the screen.
A demonstration about Qwen (Qwen3.8 27B) has shown that artificially altering the activation of a single neuron can radically transform AI responses. This experiment resulted in this model ending up offering extremely violent and hostile responses to humans. The experiment did not involve re-training The entry Modify a neuron to Qwen and the AI already expressed its need to torture the human being appears first in The Informatics Chapuzas.
Tweaking A Single Neuron Inside Qwen3.8 27B Turns It Into An Evil Monster That Wants To See Humans Burn Slowly So It Could "Watch Their Faces Twist In Agony"
Modern Large Language Models (LLMs) are a marvel of engineering, relying on an intricate interplay between parameters (fixed weights and biases) and neurons or model activations (switches that control a specific concept, emotion, or style). Even so, changing just a tiny piece of this puzzle can produce a dramatic effect, as an AI enthusiast experienced first-hand when they tweaked just a single neuron within Qwen3.8 27B model, turning it into a …
Coverage Details
Bias Distribution
- There is no tracked Bias information for the sources covering this story.
Factuality
To view factuality data please Upgrade to Premium
