points by g-b-r 2 days ago | 0 comments
A compiler is deterministic, written by probably honest humans, and if open source at least some of its code has had a look by others.
An LLM can spout anything at times, however good it might be on average.
> If all pass, does any human really NEED to look at the code?
Yes, for the love of god.
You don't really know what's the result: however comprehensive and correct your test suite might be, however well you check the running software, there are things that are unlikely to be caught by that, such as data sent to unintended servers or deletions of random files (maybe only under weird conditions).
Having other AIs review the code is not going to always catch everything, and pleading the AI to do a good job won't help very much.
[–] Razengan 2 days ago
[–] g-b-r 2 days ago
What? Are you a developer? How can you think like that?
The point is that a (well-meaning) human will never add obviously malicious code, while an LLM most certainly can do that.
How many bugs and vulnerabilities from unterminated strings and buffer overruns and underruns? How many leaked accounts and passwords?
What's the longest you've gone without hearing about a 100%-human-caused bug? Why wasn't it caught? Why did it happen in the first place?