How to Make LLMs Shut Up
AI summary of the linked articleDaksh, co-founder of Greptile, describes how the company reduced the nit comments its AI code review tool generated, which had led PR authors to ignore the bot. Prompt engineering, few-shot examples, and an LLM severity-rating filter all failed, so Greptile instead embedded past team comments and blocked new comments similar to at least three unique downvoted ones. The article says that within two weeks of rollout, the address rate rose from 19% to 55+%. |
 |
Recommended by 2 curators |
2 comments  someone should do this for kiwi and make a bot that has „good taste“ and autonomously helps us upvote Really like this machine x human taste cooperation here. IIRC this is how Palantir did work in the early days, and what made Thiel say that AI (back then I guess just called algorithms?) works best when paired with humans. Riffing of @timdaub.eth comment, I'm wondering if it would be possible to add all top upvoted content (let's say 5<= upvotes) and then create a bot that would predict if some post will get to the top of Kiwi. Or maybe even give author some writing tips based on the top Kiwi-performing posts. |
