Comment by aoeusnth1

Oh, an LLM forward pass is also like system1. The action embedding just makes it more useful for tasks or classification. If it has a small fixed computation budget it's similar to system1 than system2 (like a thinking budget, or agent harness).