2 ms·
I have to assume these aren’t just ads, but also a critical RLHF avenue for Anthropic. I imagine they scrape these commits from GitHub and compare them against
by teaearlgraycold 5mo ago
I have to assume these aren’t just ads, but also a critical RLHF avenue for Anthropic. I imagine they scrape these commits from GitHub and compare them against what Claude provided to the user. If the diff that is pushed is different it means the human had to refine the LLM output and that can be fed back in as training data. Presumably semantic search could enable you to find the matching Claude Code session.
- bpodgursky 5mo agoI really doubt they bother. 99% of people don't edit the commits by hand, they review and then tell Claude how to edit the commits (or leave a PR comment it reads), that's far easier to ingest than the tiny exhaust of manual edits.
- teaearlgraycold 5mo agoLLM providers ask for RLHF feedback in terms of thumbs up/down in their Web UI. That's just one bit of information, given only at the user's whim. The hand-refined code is gold by comparison. They're already scraping all of GitHub. They have the technology and resources to semantic search. I would be shocked if they haven't at least tried.
- bpodgursky 5mo ago[dead]
- nextaccountic 5mo agothey already have enough data to make this correlation almost 100% accurate they know things like your git author line, your github handle, and the exact codebase you were working on the general shape of commits even if you change extensively, they will probably be able to match this with claude code sessions sure the atribution at the end of commits is a signal, but I doubt it's much valuable if anything it's more valuable to anthropic competitors, that don't have claude code session data to match to open source contributors, and will have to guess if any given code is AI generated, and by how much