Sure, I dressed it up in Bladerunner vibes. But I think it is an interesting concept. Probing a thinking model using mech interp, during thinking and during output. What activates during thinking that doesn't during output?
I am running locally on my MacStudio. The only thinking model I could find with labeled SAE concepts, and runnable on my MacStudio was this Llama8B DeepSeek distillation. Need the new Qwen Scope to get the concepts labelled...
I am running locally on my MacStudio. The only thinking model I could find with labeled SAE concepts, and runnable on my MacStudio was this Llama8B DeepSeek distillation. Need the new Qwen Scope to get the concepts labelled...