This only works with really competent review. Which, unfortunately, most coders cannot do. Hence what will happen in most cases is a lot of tech-debt, vulnerabilities, complexity and tech-debt slowly and quietly creeping in. That stuff becomes exponentially harder to fix and at some time, which may not be obvious in advance, all you can do is throw everything away and start over. No, that will not happen for the Linux kernel. But the kernel comes with very hard and very capable quality guards. Most software production is missing those guards.

Or make that Eve, the basically incompetent attacker, that can suddenly attack 100s of badly secured web-shops for cheap.

Linus and anybody basically on his level does very careful review of the results when they matter. And he understands when they matter and when they do not. That means bad AI code will not get in and most AI code is bad. But put a coding LLM in the hands of somebody with significant less skill and you slowly create a catastrophe. It will still take a few years for that to become impossible to ignore, but we are seeing the first symptoms already.

I found it quite boring and generic. I stopped playing after about 10 hours and that was about two years ago with the improved version.

Not true. Sorry, but that is my area of expertise. Things are insecure because they are cheaply or incompetently made, configured and operated, often all of that.

... always a bad idea. With LLMs, basically incompetent people can hack all instances of such a device everywhere for cheap now.

I doubt the person you answered to has the capacity to understand that.

Obviously, when used correctly, even mu-calculus on stone tablets can be easily used to demonstrate the proof of the Riemann Hypothesis!

Yes, very much so. But that would be good engineering and good engineering is rare in the software space.

Hehehe, stupid is as stupid does. That is why real mathematicians verify a proofs before they build on the result.

I had students try that out in some thesis work. Results were that LLMs find some types of vulnerabilities reasonably reliably and others not at all or only rarely.

Nonsense. The real problem on display here is that you have no clue how to produce quality code.

Because you just have to pay the same number of coders even if more are code reviewers not, but you pay for the LLM on top of that as well. I am sooo not surprised. And, apparently, reviewing LLM code is mind-numbing and far more unpleasant than reviewing human code. Hence there is another really bad problem in there.

And you pulled that lie right out of your ass, correct?

No, neither OpenAI nor Anthropic are anywhere near "breaking even". And they will never get there.

Indeed. And they will also find out that running your own specialized models with your own safeguards (!) and result verification outside of the LLM is the only good way to get result quality that actually can (maybe) meet reasonable engineering and other standards. General LLMs are a cool stunt, but a really bad idea if you do real work.

As soon as the data is there, it will be abused. The only reliable way to prevent that abuse is to not collect the data in the first place.

See also the reasoning behind the respective parts of the GDPR.

So far its overall worth is massively, massively negative. I do not see that changing anytime soon. Maybe in a decade or two with very carefully trained small specialist models that have been fitted with result checking by actual reliable deduction tools.