amadeuswoo

Not using APIs, actually training a model from scratch, even a small one

What surprised you about the data, the training process, or the output?

Personally, I've been reading about how historical empires handled delegation and trust—who gets autonomy, who needs oversight, how that scales. Finding it weirdly applicable to how I think about system design and working with AI tools.

Curious if others have pulled from history/humanities in ways that actually transferred.

Recently built something where simple domain-specific heuristics crushed a fancy ML approach I assumed would win. This has me thinking about how often we reach for complex tools when simpler ones would work better. Occam's razor moments.

Anyone have similar stories? Curious about cases where knowing your domain beat throwing compute at the problem.