Yeah I was only talking about in the context of LLMs, which honestly don’t feel like the best use of machine learning to me. Real scientists and engineers using machine learning to create efficient heuristics to solve real bounded problems, and actually verifying the output through conventional means, is incredibly powerful. There’s still a lot of overhyped bullshit out there outside the LLM chatbot space, but the point stands that any machine learning algorithm should have its training data carefully curated. The techbro strategy of throwing more nodes and random data at an LLM hoping it will magically hit some exponential threshold of performance is stupid.