All our houses are built on sand now
6 points by nrvous
6 points by nrvous
This mention of Trusting Trust is completely nonsensical. The original attack was a self-replicating compiler that miscompiled itself to maintain the self-replication property.
This article is about how the author is upset that compilers are incorporating AI commits, because they believe that will cause bugs.
The sole thing these have in common is that they’re about a bad thing happening when you compile something. There’s no actual connection.
It's actually about how I'm upset that it's impossible to avoid the ethical contamination of LLM output in anything produced by a computer now, because the entire stack is compromised. Sorry if that's unclear.
I've got bad news for you: it was all built on sand when humans were entirely in charge as well.
All the security bugs that the bots are finding were written by humans, code-reviewed by humans, and shipped by humans.
This is certainly different, no? At least when a bug existed previously, one of the humans involved might have been around to understand the best way of fixing it.
At varying speeds, these projects are moving to a place where no one ever understood the code. Certain individuals would have you believe this is a good thing, but simply put no one has ever done this before, so we can only speculate how that’s going to work out. These projects previously were in a place where the code was at least understood once, so the effect is muted for now.
these projects are moving to a place where no one ever understood the code … simply put no one has ever done this before
Well, apart from all the people who have used various types of code generators. And really anyone who has used a third party library without reading the code.
I am disappointed by GCC (3.1% and apparently rising).
I thought GCC adopted an almost no AI policy that made an (unfortunate) carve-out for the test suite? Is that 3% all test suite or is the slop spreading?
It's not all test suite; if you run
git log --raw --no-merges -i --grep="assisted-by:"
you'll see changes in libgfortran/ and gcc/ outside the gcc/testsuite directory.
It is no longer possible to produce software that is free of LLM output, even if you don’t use it yourself, even if you use a language that doesn’t use it itself.
You could use the last version before LLM output was introduced. Businesses run EOL software all the time for ordinary business reasons, so you'd have good company.
If you think that using LLMs retroactively taints the project, your concept of "ethical contamination" may be too strong. If you apply it to other questionably ethical practices, I think you might have to stop using software entirely.