Ask HN: Do LLMs get "better" with more processing power and or time per request?

Do they make more (recursive) queuries into their training data for breadth and depth? Or does the code limit the algorithms by design and or by constraints other than the incompleteness of the encoded semantics?