[thumbuddy]

Yea like the ones where we throw an entire AWS server room to try multiple random seeds, using gradient descent until it works, but honestly a RNG might work just as well, in a few contrived cases and we write our paper/make our product from it

[fnordpiglet]

Seems a little dismissive. I think it’s a pretty interesting result from a theoretical point of view, and might be applied in some situations where you break a harder optimization problem into smaller problems with certain likely shapes of the piece wise problem. Regardless it’s not at all intuitive this pattern should have any optimality in any gradient descent space, so understanding it might yield some more fundamental concepts around more general optimization problems.