I like to lurk.
Michael Liu
Karma: 7
According to the creator the “Claude plays Pokemon”, the internal knowledge in Claude can often be more harmful than good for successfully navigating the game. In the system prompt, Claude is specifically told not to trust it’s instincts and to rely on the memories in its context. See (starting @ 20:28):
Agreed. I highly recommend this blog post (https://sander.ai/2024/09/02/spectral-autoregression.html) for concretely understanding why autoregressive and diffusion models are so similar, despite seeming so different.