I'm not sure about TF, but PyTorch has an "mps" (Metal Performance Shader) option which will optimize GPU performance on Apple Silicon (just type "mps" wherever you'd use "cuda"). The performance is decent, but only at the highest price points. And it's really a question of bang-for-the-buck, given that a 4090 with cuda will still beat an MBP with mps (except maybe at the highest price points).

I think the bigger question, for me, is exactly what kind of ML work I expect to do on the laptop? For prototyping and small tests it might be handy, but for anything that's going to be running for days or even weeks, it's obviously a non-starter since it will basically put my laptop out of commission. That stuff is obviously way better on a separate workstation or in the cloud.