[piotrkaminski]

I thought this was the weakest argument in the article. Yes, non-determinism is inherent to LLMs. It's also inherent to human brains, yet somehow we still manage to call (some) people reliable.

The bit about backwards compatibility doesn't make sense either. LLMs, just like humans, can take backwards compatibility into consideration — or not. Sometimes they'll forget, and sometimes they'll fixate on it even in a system that hasn't been deployed yet. It has nothing to do with LLMs never doing "the same thing the same way twice".

LLMs may or may not become as reliable as other productivity-improving infrastructure, but it's not the nature of their technology that will decide this.

[oudlys]

Human brains can explain themselves. What's a black box about LLMs is that they can't. They don't know how they arrived. They didn't arrive at anything. They don't have access to their internal representations. They don't think.