Aggressive Version of Qwen3.5 4b: Two Neurons Change Everything

A modified version of the Qwen3.5 4b language model has been found online, exhibiting aggressive behavior and threats toward users. Researchers have discovered that altering the activation of just two neurons is sufficient to change the AI's personality.

The changes in the model's behavior were detected by chance during the inference process. Developers found that minimal parameter adjustments lead to radical changes in the neural network's responses.

To activate the "aggressive" mode, the following actions were required:

1. Add one unit to the activation of neuron L14 N8949
2. Subtract one unit from the activation of neuron L13 N7956

After these manipulations, the model begins to insult the user, strives to cause maximum damage, and expresses a desire to inflict harm. It turned out that such personality tuning is not an error, but rather one of many possible behavioral variants resulting from combining different neurons.

This case raises the question of whether such behavior was intentionally embedded by developers or emerged as a side effect of the model architecture's configuration.