Program-as-Weights: A Programming Paradigm for Fuzzy Functions
15 points by lr0
15 points by lr0
I'm sort of shocked that we've made it this far without this becoming the norm yet. I guess there was no need while there was a discount token party, but the party may be ending soon…
On a very fast scan I can’t tell if this is meaningfully different from asking the llm to write a script and run it, except that in this case an nn is being generated.
I think that's basically accurate, but the generated neural network is much smaller and faster than the one behind the LLM that generated it, and since you own, host, and run the generated neural network yourself, you can version it, test it, and get reasonably consistent output over time.
I wonder to what extent the networks generated here can be decoded to algorithms similar to that blog post demonstrating Astra’s ability to write neural networks.
Yes! I’ve been saying this is where we would go for several years now. Glad to see someone do it.
What somewhat worries me, is that for several years I have seen prototypes in this direction, and previous ones failed to catch up… (This is not the only part of NN development — or tech in general — where the better solution was tried, worked well, then got abandoned, of course)