If I could go back to that age, I'd focus my PhD on actually understanding what's going on in ML models. Industry is always going to be incentivized to build things and not understand, so you can fill in the details. Plus it would be fascinating.
It’s not like interpretability research is immune. You could’ve been in year 4 of your degree when Anthropic released their sparse autoencoder research. It’s just less busy because as you correctly note, industry mostly cares about getting the black box to print money.
> Plus it would be fascinating
You can do research on the weekend even if you’re not in a PhD. I’ve done it. And no one was breathing down my neck to publish it.
