The author shares observations on mapping and steering local models based on their activation paths during specific contexts. The post highlights the variance in activation paths across different models when answering a set of batch prompts.
- Gemma 4 31b qat showed minimal variance in activation path.
- Gemma 4 26b qat exhibited extreme variance.
- Qwen 3.6 35b q_4 demonstrated moderate variance in activation path illumination.
The author compares these activation paths to neural clusters in the human mind, suggesting the potential to identify and manipulate specific reactionary traits within models.