possibly implement LLM/VLM API call as a primitive

[merlerm]

needs strict budget checking to avoid a generated approach burning through credits with a loop or something similar

[tomsilver]

if we use claude to write an agentic claude approach that has access to claude primitives, we should win some kind of paper award for "most claude"

[merlerm]

Just call the paper "Claudes all the way down"

[tomsilver]

a more serious comment: I'm not sure if this would actually help because claude itself is already an LLM and VLM -- the agent can just talk to itself through thinking, and it can also load and understand images

[merlerm]

yeah, I added it as a TODO since it came up in our last meeting, maybe after you left. my feeling is that it could be useful if the state was given as an image, since it could use a VLM API to implement something like a description function that recovers a more useful symbolic state (?) directly in approach.py, which as it stands it can't do (or well, it could implement, but hopefully it doesn't have access to its own API key hahah). otherwise I'm also not sure how useful it could be