That is absolutely correct, but is sadly the case in a lot of fields. It doesn't mean that the practical results we see (AlphaFold, Imagenet Performance, NLP performance, Robotic control with RL) isn't amazing progress.

Luckily due to so many people using ML these days, what's useful vs. fluff gets sorted out over time.

Is AlphaFold a practical result? Winning a competition isn't the same as production use. It would be interesting to read about how it's being used.

This is a good list of promising work, but showing practicality would need more explanation.

It's a fair question. Is DeepMind famous for its amazingly smart toys because it's useful similarly-smart stuff is secret? Or public but boring? Or doesn't exist?

(similar for Boston Dynamics, and IBM Watson)

In theory it should be practical. The first generation has been adapted by other teams into excellent prediction servers that can be used now. The second gen is way more hush hush and has yet to be vetted, so we’ll have to see. I am watching for news of it eagerly!