Microsoft DeBERTa surpasses human performance on the SuperGLUE benchmark

Sometimes it is the finer details that make all the difference to something being usable. Eg the main difference between older VR glasses and the Occulos Rift was the resolution - but that was actually a big deal.

I would guess that human-level NLU is probably going to be the same way. Once you can interact with an AI thing and not feel you are talking to Alexa, there will be a billion new applications that probably already could exist but we would not use.

I do not follow AI though, and a great thing about HN is all the feedback from people who are actually living this stuff daily.