Comment by kojoru

I'm wondering: how can understanding gradient descent help in building AI systems on top of LLMs? To mee it feels like the skills of building "AI" are almost orthogonal to skills of building on top of "AI"