[slashdave]

All these models train on the same data (sort of) and so all will pretty quickly reach the same asymptote.

[logicchains]

Eventually they'll be put in robots able to gather real-world data live, and they'll have a practically infinite amount of video/audio/tactile data to learn from.

[slashdave]

Sure, but this scales rather poorly. How many robots do you expect, and don't you expect some resistance from the public, for privacy reasons at the least?

[bbqfog]

I agree, I don't think there's a moat and this is going to be a commodity feature.