Show HN: Autoresearch, but for Weight Space Programming
huggingface.coI built Vespasian to let everyone experiment with weight space editing. Pick task, provide data, add API key and it takes care of the rest. In addition to the Rust crate, it now runs interactively and without an install on HF Spaces. There’s never been a better time to try out weight space programming. Throw your tasks at it and see how far you can get without gradients.
Obvious note to other readers, don't put you API key into strange input boxes.
Hi. This is *good advice* and I encourage everyone to use short lived keys and generally not put your api keys into strange boxes.
That said as this is a hugging face space, you can inspect the code. In addition there’s a Rust version that runs fully offline and which can use your Codex and Claude Code login. I’m sorry - getting login without api keys to work for ChatGPT requires a vendor onboarding and I can’t do that for a single demo. All ears if you have a better solution.
thats fair but I've heard these weight changes are slow and expensive. I might have missed it but how long does this take to run and how many tokens can I expect to burn?
Super good question. We actually built the benchmark that measures both of these: https://hcltech-robotics.github.io/iwibench