I gave claude a new physics RL problem to test out Weight-Space Programming. It vibe coded 298 weights for a 71-layers deep neural-net, correctly one-shotting the problem using weights. 🧵 Here's why it's not "weight space programming"...

2/ So on the surface, I thought maybe LLMs have an innate grasp of weights and just spit out the right numbers. After all, no tool calls means no backprop. It must be some sort of savant for DNN weights!

3/ But Claude was really just using the CoT to do a bunch of multi-variable math solving the physics problem. It then embedded those solutions in a DNN weight matrix. It's cool, but it had to crank out math for 20+ minutes in CoT. That's clearly not innate grasp of weights.

In experiment 1, they say Astra used only 800 reasoning tokens before interpreting the network, which seems too short to do all the calculations you're describing?

Yep, experiments #1-3 are pattern recognition. Experiment #5 only works because it's previously seen the solution to pong. Give it a novel problem and it's many tokens worth of reasoning.

what is the pattern being recognized in 1-3?

The mit writeup already draws that line. Same edit that lifted qwen math barely moved gemma, because loops were gone and the skill was never there

Useful pushback. "Vibe-coded weights that solve a toy net" ≠ "read an existing trained net and patch failure modes." Keep the terms separate or the hype eats the engineering.

Could we do it in reverse? Translate blocks of weights into compact mathematical expressions? This could open up the possibility to calculate it more efficiently, or at least with less RAM.

How many tries before it one shot it ?

why would you not rather... have it just write differentiate programs?

Agree the from-scratch nets look like math compiled into weights. The Qwen edit doesn't fit that frame though: trace a loop to one direction, subtract it, touch 0.016% of weights. Less compiler. More debugger. Does your critique cover that part?

Would it be possible to see source code?

Thanks for doing this.

ゴールしたあとのボールのバウンドこそヒントがあるんじゃないかな？

You win. Lol I tried also. It does not work how the op said unfortunately. Not yet anyway

What, exactly is the difference? like it doesn't "innately" understand code. The important thing is it can pattern-recognise it well enough Current sota can't edit more than toy parameters; but "can recognize patterns in X" tends to grow the size quite fast with generations

Fair pushback. The 0.016% number is wild, but a 0.8B model with known reasoning loops is a soft target. The real test is whether it works on a model that isn't already broken in obvious ways.

298 hand-tuned weights, 71 layers deep, to do what 5 lines of Python does before your coffee cools. That's not a breakthrough, that's a Rube Goldberg machine with a GPU bill. Don't believe me? Try it. Compare the token count and completion time, then cry. →