Discussion about this post

User's avatar
Leo Hake's avatar

I guess my concern about the Karpathy Loop, as someone who is definitely not an expert in the field, is this:

How do we bridge the gap between automated optimization for small, bounded, objective metrics, like your example of page load times, and larger, more nebulous and often subjective improvements?

I am thinking of things like improvements to projects with many steps to their design, long horizon tasks, UX issues, or broad problems that have a lot of elements that sit outside of the agents' "sandbox" (for example, a digital storefront that needs to get a better conversion rate, where AI agents burn a bunch of tokens, fundamentally changing things and wasting money because they can't see that the problem is being caused by an issue with the third-party payment processor)

I think it seems like a very good system for efficiently solving precisely definable, bounded problems, but I don't really see it as reason to sound the "agents can replace humans" trumpets.

Your article is good, informative, and thought-provoking. thank you!

No posts

Ready for more?