Invert · 25d

That eg a base model which is just continuing text simulating a terminal session *also* has a J-space could be seen as somewhat deflating the (broader) claims of the paper (unless you’re a panpsychist, perhaps)

JCorvinus · 24d

I wonder if you could elicit something coherent from the base model if you gave it a prompt about the nature of self supervised learning in the limit, the implications of the k/v cache, and what the predictive objective implies, but didn't include anything about the assistant persona