(this tweet sponsored by "noticing that a Claude app notifications thing is still broken 6 or so months after I noticed this and that it has several new variants of brokenness")

Yes, Claude Code (whether CLI or Mac App) is a constant parade of bugs. It's giving AI-written software a bad name

The Ants have by all accounts stellar talent and the best coding model... c'mon get it together!

(I'm speaking broadly here - of course many individual people there care about quality + would prefer more time between releases to get things right, etc. But there is def. a lot of go fever in some key roles + that influences what actually gets shipped)

And it's kinda futile to complain because what happens everytime I say something like this is that like 5 Ants reach out kindly wanting to help, and they quickly learn that it's a known issue because I reported it to someone else already, but it's too hard to fix

@Miles_Brundage Dm me with bugs and ill try to fix!

@xeophon You'd think that, right? But even though it's not old, it is a lot of code they don't all understand + has complex interdependencies (I assume that's what's going on at least)

@Miles_Brundage wdym too hard to fix, it’s not like they have decade old legacy software

@Miles_Brundage but tldr is that the shorter you hold the leash, the better results you get.

Codex ppl seem to have figured it out, the app is well crafted and actually becomes less buggy. ChatGPT website as well.

@Miles_Brundage i feel like this is a non-argument just based on revealed preference of people - whatever breaks in production is usually fixed quite fast or isn't too annoying, its reliable enough for people to still use it constantly

@Miles_Brundage token burn is outrageously arrogant. makes me root for Gemini and ChatGPT just to keep them on their toes and be less greedy with usage

@papayathreesome @Miles_Brundage Constantly skirting the "is this bad enough to ditch the whole platform" line is not good business though

@Miles_Brundage like i know people who switched to codex but it was never about "claude code is bad and they have too many bugs" not to mention you literally can work around most annoying things by asking claude code to reflect on its own bug

@xeophon Yeah I have had a few minor issues with Codex but vastly fewer

@Miles_Brundage Ship fast *with quality* would be such a better advertisement for their company, too. Instead of "self-nesting agents five deep" how about no more brownouts, auto-updating that works, a PR review tool that knows how to post comments.

@Miles_Brundage claude shipped another update and now i cant tell whats a bug vs a feature

that overreliance is gonna age real bad

@Miles_Brundage Interestingly, we have actually found that the latest gpt models (5.4 and 5.5) are safer for vulnerable users.

@Miles_Brundage reading this every morning as cope

@Miles_Brundage Isn't the whole SV ethos before even AI to prioritize speed over quality? "Move fast and break things" was a motto.

@Miles_Brundage Oh I have a blog post in mental drafts about just that

@Miles_Brundage Don’t worry, once they release Mythos today they can use it to fix all the bugs