weight editing is not the only option to breaking loops.repetation penalities work enough sometimes...btw i built something very similar to this but on cli..just with a deep triage layer(you gotta diagnose properly before a surgery)
“Claude, make me a website explaining the recently viral HN article, Sparks of Weight-Space Programming. Make it pop for the visualizations and explain things very clearly”
Sure, a lot of people are doing this but where do the iterations, the quality checks, the human input, etc, becomes visible? It doesn't and then we have people assuming nothing takes any effort at all anymore, which is bas for really good content that was assisted by AI.
While I'm sure there were iterations and quality checks and human input - I don't believe there was enough.
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.
fair point - i assumed context in opening and i've rewritten.
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.
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.