[xedarius]

"Deepmind hasn't shown anything breathtaking since their Alpha Go zero"

... what about when the Google assistant near perfectly mimicked a human making a restaurant reservation .... the voice work was done at DeepMind.

All the problems in AI haven't been solved yet? Well no, of course not. Limitations exist and our solutions need to be evolved.

I think perhaps the biggest constraint is requiring huge amounts of training data so solve problem X. Humans simply don't need that, which must be some indication that what we're doing isn't quite right.

[gaius]

what about when the Google assistant near perfectly mimicked a human making a restaurant reservation

Any sufficiently advanced technology is indistinguishable from a rigged demo

[IshKebab]

The Google Duo demo was definitely very impressive. The only caveat I can think of is that they didn't say how often it works. That might have been the one success from hundreds of calls.

[pfisch]

> what about when the Google assistant near perfectly mimicked a human making a restaurant reservation

That seemed pretty staged.

[Eesais2e]

> Humans simply don't need that, which must be some indication that what we're doing isn't quite right.

Not really. DNNs are much simpler and need to be much more specialized towards specific tasks than the human brain. They're more like a nematode that was optimized by evolution for millions of years to tell cats and dogs apart because it preferentially infects dogs. It's the only thing it does. It does it via some shortcuts and it won't be able to learn chess without a radical redesign.

Not to mention that primate brains take some time to get bootstrapped. A toddler has to be fed visual input for many months before they can be left unsupervised for reasonable amounts of time. And have you seen what those "optical illusion" things do? It's a miracle that those humans can self-navigate at all considering those failure modes!