Steven Op de beeck
Hynek Schlawackgiven the claude code leak i'm starting to suspect that when boosters say they don't look at the code it's an act of fear, not confidence
Will Vincent@hynek The adversarial AI agent said the code generated by the other AI agent was good, so what's the issue? (sarcasm)
Mastokarl 🇺🇦@hynek Care to provide evidence? Have read quite a few articles, much about interesting architecture decisions. Nothing about poor quality.
osmosisch@Mastokarl @hynek then you're in a bubble or not really searching https://www.reddit.com/r/ClaudeAI/comments/1s8lkkm/i_dug_through_claude_codes_leaked_source_and/
@osmosisch @hynek Thanks, I‘ve seen that. Have seen plenty of production code in pre-AI times that has stuff like this. Wouldn‘t also say that the huge main file is a sign for bad AI coding… So, not convinced this reeks of poor AI code. Seems also this is what most comments say.
Steppe Lord@osmosisch @hynek @Mastokarl johnny made a thread checking out some of the source code. Is not a complete write-up as that's titanic in scale but the examples given are really funny
https://neuromatch.social/@jonny/116324676116121930
https://neuromatch.social/@jonny/116324676116121930
jonny (good kind) (@[email protected])
- Claude code source "leaks" in a mapfile - people immediately use the code laundering machines to code launder the code laundering frontend - now many dubious open source-ish knockoffs in python and rust being derived directly from the source What's anthropic going to do, sue them? Insist in court that LLM recreating copyrighted code is a violation of copyright???
@steppe_lord @osmosisch @hynek Some are wild. Others look very normal to me - the „make the model aware of common security issues“ or the fast regex for negative reactions instead of an expensive LLM call.
@Mastokarl @osmosisch I don’t agree but I won’t argue the point with you because my main theory for the broad acceptance of slop is that the majority of people lack both taste and engineering skills
@hynek @osmosisch I get your point, even if much production code looks like that, tech debt will always bite you.
But what I noticed with AI code is that the choice between 100% beautiful hand-written code that takes t time and good machine-written code that takes t/2 time I tend to lean for the later, unexpectedly (I‘m the kind of person who writes their HTML by hand because I can‘t bear the shitty HTML that web site generators create). LLMs write - in my opinion - „good enough“ code…
Judy2k@Mastokarl @hynek Personally, I find if the code does what it's supposed to, then it's good enough.
As long as you have enough tests then you don't really need to look at the implementation. Hynek is too precious. We're past that now.
@judy2k @Mastokarl this hinges on the tests being of any value which is a problem because people are famously bad at writing tests and LLMs are trained on them and it shows.
It's hilarious how many lines of test code it can write without actually testing anything of value.
I wonder if there's a future of fine-tuned models trained on good tests that don't fire up a mock fest. But the q is whether there's enough training material. :P
cle@judy2k @Mastokarl and given how overall software quality managed to further deteriorate over the past ~6 months – from an already very low baseline – I don't think my preciousness is the problem here.
@hynek @Mastokarl It's fine as long as you give the tests a quick skim to make sure the LLM has written enough test code.
@hynek @Mastokarl People keep sending me this gif, I do not know why.
@judy2k @Mastokarl must be a people problem
Joe
David Beazley@hynek Or maybe an act of daring.
Morten