It’s still all pytorch.

Unless you’re working at Google, then maybe you use JAX.

JAX is quite popular in many labs outside of Google doing large scale training runs, because up until recently the parallelism ergonomics were way better. PyTorch core is catching up (maybe already witn the latest release, haven’t used it yet) and there are a lot of PyTorch using projects to study though.

Interesting point about the shift towards PyTorch. It really has been fascinating to see how preferences in frameworks can impact the entire research landscape. I remember back in 2017, I felt like I was constantly hearing about TensorFlow everywhere, and then out of nowhere, PyTorch just started gaining this insane momentum. It was almost like watching a sports team come out of nowhere to win the championship!

In my experience, a lot of it comes down to the community and the ease of use. Debugging in PyTorch feels way more intuitive, and I wonder if that’s why so many people are gravitating toward it. I’ve seen countless tutorials and workshops pop up for PyTorch compared to TensorFlow recently, which speaks volumes to how quickly things can change.

But then again, TensorFlow's got its enterprise backing, and I can't help but think about the implications of that. How long can PyTorch ride this wave before it runs into pressure from industry demands? And as we look toward 2025, do you think we'll see a third contender emerge, or will it continue to be this two-horse race?

TensorFlow was an overengineered Google-style mess and they constantly made breaking changes.

All the graph building and session running was way too complex, with too much global state and variable sharing was complicated and based on naming and variable scopes and name scopes and so on.

It was an okay try, but that design simply didn't work so well for quick prototyping, iterating, debugging that's crucial in research.

PyTorch was much closer to just writing straightforward numpy code. TensorFlow 2 then tried to catch up with "eager mode", but in the background it was still a graph and tracing often broke and you had to write the code very carefully and with limitations.

In the end, Pytorch also developed proper production and serving tools as well as graph compilation, so now there's basically no reason to go to TensorFlow. Not even Google researchers use it (they use jax). I guess some industries still use it but at some point I expect Google to shut down TF and focus on the JAX ecosystem with some kind of conversion tools for TF.

> But then again, TensorFlow's got its enterprise backing, and I can't help but think about the implications of that. How long can PyTorch ride this wave before it runs into pressure from industry demands?

PyTorch has a huge collection of companies, organizations and other entities backing it, it's not gonna suddenly disappear soon, that much is clear. Take a look at https://pytorch.org/foundation/ for a sample

The thing about Tensorflow in 2017 is that everyone acknowledged how difficult it was to use. While it was almost the only game in town, no one was happy. Those are probably the areas where an upstart can come in and disrupt.

I feel like it was all pretty obvious by late 2017. Prototyping and development in PyTorch was so much easier - it felt just like writing normal Python code. And the supposed performance benefits of the static computation graph in TensorFlow didn't materialize for most workloads. Nobody wanted to use TensorFlow - though you often had to when working on existing codebases.

I think the only thing that could have saved TensorFlow at that point would have been some sort of enormous performance boost that would only work with their computation model. I'm assuming Google's plan was make it easy to run the same TensorFlow code on GPUs and TPUs, and then swoop in with TPUs that massively outperformed GPUs (at least on a performance per dollar basis). But that never really happened.