- cross-posted to:
- linux@lemmy.ml
- linux@lemmy.ml
- linux@sh.itjust.works
- cross-posted to:
- linux@lemmy.ml
- linux@lemmy.ml
- linux@sh.itjust.works
To be clear, that’s CVEs fixed per kernel. Another graph later shows much fewer buggy commits being pushed per release as well.
As Greg K-H says later in the thread, I wonder what tool made this possible?
Around Claude 4.8, AI because actually good at identifying code bugs. We’re basically running bug hunter AI models 24/7 now, because malicious actors are almost certainly doing the same thing. The bad news is there is no end in sight, the CVE discovery rate is going to stay high.
But I was told LLMs will never be useful or capable in programming tasks, so this theory of yours seems highly unlikely.
Here’s my take: AI in the hands of somebody who is able to resolve bugs, able to understand the codebase, etc., can get tedious tasks done quicker. Computers love tedious tasks, even when it comes to machine learning bits. Computers do best with repetition and patterns. However, AI cannot replace the skill of a developer. Place it in the hands of an idiot (i.e. a “vibe coder”), and all you get is duplicate bug reports, broken code, and lots of stress for project maintainers. AI is a tool like any other, just as a hammer will never replace a builder, even if it’s the fanciest hammer in the world with billions invested in its design. Hammers can make work be done faster (not all work, sometimes you need a saw!), but put it in the hands of an idiot (a “vibe builder”, whatever the hell that would mean), somebody is going to get hurt.
This ignores the legal questions of AI, which are still important. If cars were given to the masses during the Middle Ages, a lot of people would get hurt since there hasn’t been enough time for traffic laws, regulations, etc. to develop. In the modern day, of course, cars can make commute faster, and there is rules in place to ensure safety on the road. Walking, cycling, public transit, planes, ferries, etc. still have their place though, and more often than not they are the preferred option.
If you do want to run AI, please run it locally! Don’t support big tech companies by paying their exorbitant subscription fees, and don’t feed them your codebase!
I don’t personally use AI, since it doesn’t fit my needs and I find it more annoying than anything to get it to do what I want.
LLMs are not “never useful”, but what they are good at is limited, and if you don’t have the developer experience and try to “vibe code”, only bad things will happen.



