Show HN: Watch weight-space programming in action with WebGPU
tabbyml.github.iothe full loop needed too much VRAM. so I turned it into an interactive walkthrough of a concrete example: editing 0.016% of a model's weights to reduce reasoning loops, no gradient descent.
Based on exp8 of “Sparks of Weight-Space Programming.”: https://news.ycombinator.com/item?id=50341174
While it looks good, it's laborious to read, and presented in an odd format, that doesn't really provide any insight, just lots of flash and LLM prose. The The header alone is hard to parse without context "Rewrote 0.016% of the model's weights — and cured its infinite loops". The constant little claude-isms of chips which say "Watch it · no download" "recorded · llama.cpp · Q4_K_M" "answer key: 10 · plus " add so much visual noise, especially with the high contrast style.
Less is more. The "Run it in your browser" is cool enough. Keep that and throw away the rest, or put it on a separate page.
what i wanted to go through is that weight editing isn't black magic. If you've worked with prompting / finetuning / any kind of agent quality optimization, the workflow should feel similar: identify a failure, make a change, evaluate, and check the regressions. the difference here is working directly on the weights (which is likely only doable by agent). I think of it as closer to solving a specific equation than running a gradient descent optimization, thought that's only an analogy.
But i take your point that visual might make the point harder to see.