As for the notebook part, it's fascinating because it genuinely is kind of like software design if you try and impose more structure onto it - how is it reusable for the future? Can you genuinely predict the space of possible writings that you want, or does it, in the authors words, homogenize?

I've taken an approach where I treat the act of deep writing (or shallow writing, or any writing) as means in it of themselves. Not sure how absolutely effective this is but I can definitely say that my thinking changed.

#1 step to writing is to stop giving a f and I didn't not learn that until I was 21. Too easy to get caught up in schoolwork and the "proper" way to do things I feel.

I want to watch the presentation, is there a recording, does anyone know? When the Overview slide starts with "Don’t do overview slides, it’s bad practice", I feel like I'm missing quite some context here.

This seems like a pile of generally good, and some non-obvious advice, that's also useful outside of the boundaries of ML (it would also apply to a PhD in Neuroscience for example).

Biggest drawback these days is funding (Govt. is not liking research with all fund cuts) and long term opportunities for folks doing research.

The best research advice is always unsolicited. The official version tells you how it is supposed to work, not how it actually does.