It's so hard to tell though, because the context of every task is so different. I feel like I'm accumulating superstitions about what helps and what harms.

Now it feels like sycophancy has skyrocketed? It tells me how 100% right I am every prompt, and it gives me confidently wrong "Job done boss!" assessments like "The tests are all passing" when they're just not.

Have I made it...insecure? Like did I prime a unhappy boss/nervous worker social script?

More OBS problems. Will try to get live when I've sorted it out 🤞

To explore solutions, I'm going to start from a sandbox. Imagine we were creating this new library. Call it "spycy". With no attention to spaCy backwards compatibility, what would we do?

If we can get a sandbox solution we like, we can then look at the least disruptive way to adapt spaCy.

In the stream I'll talk through why supporting PyTorch well has been hard. The gist of it is that the Language class that owns the pipeline is too involved in the optimisation details. This prevents us from using other libraries' high-level training abstractions.

The BLIS compilation problems and numpy compatibility mess I've dealt with over the last couple of weeks have put new focus on some longstanding issues for spaCy. Our GPU support still sucks, and we have this numeric computing stack that's not working for us. How can we switch to PyTorch?