> So we did mechanistic studies on small models, Gemma 4 particularly, and found the hidden state for different layers carry meaningful self-awareness signal for various situations.
Neat! Just to make sure I understand - you trained your probe layer to take this hidden state and predict p(wrong)?
Curious to learn more. Any more info on your approach (esp the mechanistic study)?
Correct, the study is verbose, we will compile into a neat shareable report and publish once we solve the pending caveats. Interesting username btw haha.
Exactly right. You design with the diffusion model, then hand those designs off to an agent to implement.
It's a lot like having an architect create plans for you before handing it off to a builder. In my (obv biased) experience, you end up getting better/more creative results with this approach. You're using the best model for the job at each specific task, ie a diffusion model as the designer, and a LLM as the engineer.
reply