My initial reaction was being weirdly sad about this and I don't fully understand why yet. I read the headline, clicked into the link, and just went noooooooo. I really like uv and I hope it continues to do well, congrats to the team though and hope everyone there gets a good outcome.
Very cool! Sometimes when I try to debug things with chrome dev tools MCP, Claude would click something and too many things happen then it kind of comes to the wrong conclusions about the state of things, so sounds like this should give it a more accurate slice of time / snapshot of things.
I wrote up a blog post for a hackathon project where I used RL fine-tuning to make an LLM generate intentionally ugly and unpythonic FizzBuzz code. The post covers what I learned about reward shaping and GRPO. Feedback on the writing or content is welcome!
> When something breaks, the fastest path to recovery is often for a person to open the console and fix it directly.
Definitely agree with this, sometimes something goes out and you need to fix it quickly. If you happen to not have permissions or access to modify the state file or do a terraform apply, it can be a bit of a pain to wait for CI/CD to apply the changes. Or in other cases, if some weird drift happens since the last time the state file was updated and now you gotta go through all this other stuff and hope it doesn’t affect the thing you’re trying to fix or cause new problems.
Very cool project! I always wanted a telescope as a kid but kind of forgot about that desire as an adult. Didn’t know you could build your own like this.
That’s a shame, I haven’t been keeping up on LLM web fiction quality level. I’m an avid web novel reader myself and do notice a ton more new fictions these days compared to before. There are even some works where other readers pointed out a drastic difference in quality of a longer running web novel (maybe a year or two plus) where new chapters are worse and just have a bit of that AI smell. But I suppose I also feel for writers busy with their lives and are writing on the side.
Personally, I don’t know what I feel about AI written novels, I suppose if the quality is truly great and I can’t tell, then fine. But if it means a flood of low quality novels that’s a bit sad.
But long time readers often smell something off beyond what feels like new or amateur writer quality scents. I wonder how long this radar will keep working as models and tools improve. But there must be something about high quality writing that’s difficult to capture in an LLM today, I wonder if it’s mostly a data quality thing or it’s an actual model limitation.
Good read, nice reminder of one the reasons why a lot of LLMs have similar stylistic choices especially in earlier generations with the formatting and emoji usage. Although I do see newer frontier models have been more aggressive with dealing some of these issues. For example, Opus 4.5 these days seemed to not be as wordy and tells you you’re stalling, just ship it or send it etc. But some open source models trained to game LMArena might keep exhibiting the characteristics mentioned in the article.
The secret to any great mashed potatoes, ungodly amounts of butter. Just enjoy them and don’t think about the calories but probably stay away from them for a year. Side note one of the first recipes I tried myself was from Chef John of the Food Wishes channel on Youtube.