@nbro @whuber @gung: this is an interesting discussion. As a matter of fact, I believe the answer in question is clear enough, but at least the conclusion is simply wrong. I commented there and would appreciate it if someone could help me understand.

Hi I was wondering what could cause different runs in tensorflow to have a different loss function? I have fixed the random seed of both numpy and tensorflow, so I would expect to get two times the exact same loss. But I get a divergence after 8 training steps of 0.001 and then every few step another divergence of 0.001. Any idea what could cause that? As far as I know floating point errors are deterministic in nature right?

@Hakaishin You'd better ask this question on Stack Overflow. Furthermore, you will have to provide a How to create a Minimal, Complete, and Verifiable example.

Yeah, I've been around on SE and now about mvce. But you know when you just talk to a colleague, the first thing he says is not you need a mvce, he thinks about what you said first and maybe has an idea.

@StephanKolassa You might be making a subtly different distinction between "estimate" and "prediction" than others are. A data-based calculation of a probability can be viewed as an estimate when the probability is a parameter of the model; when the probability is modeled as a random variable, the calculation would be a prediction.

The following is not a duplicate.

Causation and independence are different concepts.

@nbro Yes, they may be different concepts: but is there some way that the duplicate does not also (perhaps incidentally) happen to answer the question you link to? It seems to me that any example showing dependent variables can have zero correlation is a fortiori going to work well as an example showing causally related variables can have zero correlation.