Behold, the beautiful result of downloading NeurIPS 2023 code and running it on MNIST!
Eli Sennesh
Trying to replicate or reimplement AI results these days kinda pisses me off, because the paper will give a nice clean algorithm block that doesn't even slightly work as written, whereas the code has half a dozen Deep Learning Tricks that make all the math technically extremely wrong.
Eli Sennesh
laundry list of ai-related opinions:
1) llms do work and sometimes writes good code
2) ai is not llms. ai is neural networks. "training neural networks" works so well it's baffling
3) "the singularity" is an easy concept and if you scoff at it you're being tribalistic or deliberately obtuse
