Horace He

In theory, it should still be deterministic, but there is this concept of batch invariance that is not respected in some, in some cases.

And I strongly recommend this blog from Horace He, uh, as a great piece of reading.

So he goes through the exact reason.

And if you want me to break it down to you, it all goes to the floating point's non-asso-associativity with respect to the addition operation, where the order of addition does matter on the result that you get.

Now let's do determinism because that's the part I've been waiting for.

Thinking Machines Lab, Horace He, September 2025.

The popular explanation is that concurrency plus floating point makes inference non-deterministic.

That explanation is incomplete.