Hacker News new | ask | show | jobs
by mike_d 1183 days ago
> language models themselves can't "crash"

That is like saying the Excel document didn't crash, but Excel did when it tried to parse it. As far as I know there is no proof that you can't cause a LLM to crash with user input.

> because they "just" generate new tokens.

I can write a program that counts to 100 that crashes reliably.
2 comments
inimino 1183 days ago
> As far as I know there is no proof that you can't cause a LLM to crash with user input.

Well, then you simply don't understand how they work.

link
hoseja 1183 days ago
It's a fixed number of matrix multiplications. You can't make matrix multiplication crash by feeding it any sort of strange numbers (unless they haven't handled their Infs and NaNs).
link
mike_d 1182 days ago
> You can't make matrix multiplication crash by feeding it any sort of strange numbers

Matrices don't multiply themselves. You need hardware and software. As I pointed out the LLM is effectively just data that is being processed by a program. It is silly to assume you have no bugs in that software or the underlying operating system.

link
gs17 1181 days ago
The type of bug the first comment was discussing, where recursion could cause it to crash (out of memory? stack overflow?) is categorically impossible based on how the model works, and a bug in the software running it would not change that.
link