Let me dispel one myth: The bitter lesson can be true, and you can still want a bespoke, customized solution.

If it is true that eventually a general approach that scales with compute/data will supersede a bespoke expert solution, it is not necessarily true that today's expert solution should be thrown away at the first sign of a data-driven model.

It says nothing about what you should use today to build things, it says everything about how you should do research that lasts.

If you're building for billions of users in 10 years, sure go with whatever you think is the bitterest.

But if you're building for right now, you probably mashup whatever works best. That might still be a PID loop, EKF, and Homographies sometimes. But when you hit the frontier of what "works", then you have to reach for new things anyway.

Chesterton's baby/bathwater

WRT humanoids - I am making no statements.

DeepSeek disputed it

The Bitterer Bitter Lesson is that there are no exponentials, only S-Curves.

You can render graphics with the fastest CPU you can find if you like, I'll keep my dedicated GPU