Floating-point arithmetic

While this hypothesis is not exactly wrong, it doesn't reveal the full picture." And there's a very long blog post that follows that, that's very good.

Uh, additionally, you can check out the IEEE Standard for Floating-Point Arithmetic, which is IEEE 754, which we'll link in the show notes, that talks about why floating point is weird.

Yeah.

We got this feedback after maybe the first time we talked about LLMs, and someone said, "Hey, even at temperature zero they're not deterministic because of weird floating point errors in the hardware." And you're like, oh yeah, that totally makes sense, because you mean a n- a not, a floating point arithmetic being non-associative when the mathematical operation is theoretically associative is a problem.