Anthropic's AutoDream Is Flawed(substack.com)

13 pointsby k1musab13 miesiące temu2 comments

https://substack.com/home/post/p-192893121

2 comments

> For any non-trivial problem, an LLM generating a solution is in one of three states at any given step

This seems like a restatement of 'law of trichotomy' not a description of a some state the LLM is occupying.

> When an LLM documents the state of a problem, that documentation reflects whichever of the three states it was in at the time of writing.

This doesn't make sense. Why would the 'relative direction' of prior generation be coupled to the output of a summarization task?

> A sleep protocol that ingests those notes and resolves them is not approaching truth. It is averaging over an unknown mixture of states (1), (2), and (3) - then presenting the result as settled

Unfounded averaging assertion?

Reads like word salad to me.

throwaway892013 miesiące temu

It also seems conceptually wrong to refer to a process of ordering and cleaning up notebook facts as 'dreaming'. If I collect and clean up my notes of the day, that's a very conscious task. Actually dreaming seems more analogous to a training or fine-tuning step where you modify the model weights.

(while hallucinating the events of the day in a very weird way; it would be fun to 'wake up' the agent in the middle of such a session and commit the 'dream' to a notebook again)