cat

@halvarflake (which reminds me I gotta look into that still, almost forgot)

@halvarflake For me it was the thing that it told me I was wrong, and I got mad, but then I slept and realized I was wrong but out of sheer coincidence I ended up being right in the end

Example scenarios where Claude was extremely stupid in the last days: 1) Arguing that a change that moved work into multiple Python processes made GIL contention worse because the total number of CPU-seconds spent waiting for the GIL had gone up.

@nmatt0 @h4x1n_dev Be careful it’s an expensive addiction

@h4x1n_dev 5x. Might have to upgrade 😅

Cutting it close today...

@glcst Granted, I suppose I can see myself in the mood for one and not the other. So maybe not exactly comparable.

@glcst I think yakiniku is kinda already that tbh. Not the same, but, I have a hard time imagining anything better.

Is there anything that is good in this world that the Japanese cannot perfect? Just bring them to Texas and get them BBQ as they all desire. Then just wait 5 years. And the world will know BBQ to a level unimaginable today.

@Mh56199612 @brunoborges Reads more like a spec I imagine, one can be interpreted differently than the other.

@brunoborges Why NEVER and not Never, interpreted as a more strict order ?🤔

Best part of Claude Code's source is this system prompt. #L43-L54" target="_blank" rel="nofollow noopener">github.com/alex000kim/cla…

@michael_timbs Hate how it’s so hard to know, but tbh likely also A/B testing

@michael_timbs Actually probably was me now that I think about it.

Convinced Anthropic diverted all their compute towards agents building DMCA Takedown Notices for hosting leaked source code. Explains poor model perf

Ahg I’ve forgotten how to be patient enough to do shit myself instead delegating to Claude

@Noahpinion But I guess that’s just my opinion

@Noahpinion Either way it will be fine

Let's hope AI cyber defense beats AI cyber offense, or the internet age is over

Tbh it’s really not that hard, but might need to design things around the concept for it to be useful I think

Anyway, hard to see how that could work outside of specific changes and approval + review, and if it would even be useful, but still.. super interested if anyone has tried

Imagine if you could request changes to a platform as a user and have them implemented and rolled out in less than an hour.

@nmatt0 @HackerOn2Wheels @rez0__ @ctbbpodcast massive if true

@HackerOn2Wheels @rez0__ advice from the @ctbbpodcast episode helps with that. Put stuff in your Claude.md like "POC||GTFO", etc. I've also been starting to document a standard thread model via markdown I can give it to sometimes prevent it from over hyping something.

My biggest problem with Claude and hacking so far is it tends to be super hyperbolic. Everything is CRITICAL, makes up risks for trivial stuff. I keep having to explain to it real world risks as it wants to “finish” the job with shitty findings.

@hankein95 @thegrugq Cool will look into this soon. Just curious, have you seen many due to silently failing or falling back to old behavior? Idk if that would even be tracked… but I feel like that might be a thing in the near future.

We've been tracking public CVEs where AI-generated code introduced the vulnerability. vibe-radar-ten.vercel.app 50k+ advisories scanned. Dozens of confirmed cases so far. Claude Code, Copilot, Cursor, and others all show up. Common bug classes include XSS, command injection, SSRF, and path traversal. And these are just the cases that leave metadata traces. The real number is almost certainly higher. Open source, from Georgia Tech SSLab: github.com/HQ1995/vibe-se…

@nmatt0 Yeah. It always comes off as more of a long standing personal excuse to themselves to not try.

I kinda pride myself on being able to explain technical stuff to non-technical people. There's nothing that pisses me off quite like a non-technical person assuming they won't understand a word you say, throwing up their hands and exclaiming that they were immutably born lacking the ability to understand tech.