Will be the magic moment. Currently like supervising a grandparent using the web. But there's huge potential if ^^^ happens. Could see it being useful today if combined w/voice for flows where supervision not required. Example: asking w/voice to make reservations while driving.
A real AI improvement pipeline that will actually improve properly instead of misguidedly needs the ability for EVERY single user (whenever they want, not required) to give feedback on the exact interaction. Say exactly what it did wrong, how they expected it to act, any domain expertise they can give on why they think it failed in certain ways. Then the developers can make decisions based on the real fuckups. This isn't happening anywhere.
It will just be like the post training that turned GPT3 into the original ChatGPT.