> Yes, duplex is for completing specific tasks, but how much of a difference is there between "Duplex, book a table for four at 8pm" and "Duplex, ask my victim about how their day was"?

The former (booking a table) is much more "constrained", as in the conversation would most likely not go into much of a tangent, because there are only so many responses to a statement like "book a table for four at 8pm" (8pm is full, 8pm works, etc).

Whereas asking someone how their day was would give the "victim" a much bigger breadth of responses (and additional questions!) that would cause the AI to stumble and fail to give a satisfactory answer. That, and running this 1,000 times simultaneously so that no one person would be able to intervene to "help" the AI would just be a highly unscalable operation.

So yes, huge difference.

And you think this is the end of innovation, this is it, no improvements from here?

Duplex is interacting with a human and that person has no idea its a computer on the other end. Yes, Duplex is limited as it stands, but what is there to make you think something as described in grandparent post won't exist in ten years?

This is what existed 50 years ago: https://en.wikipedia.org/wiki/SHRDLU#Excerpt

In terms of NLP, the Duplex seems hardly that much of a jump. The main improvements seem to be on the speech part.

Well, I'm already receiving automated calls about "that car accident" and they connect me to a real person if I happen to respond with a predefined keyword.

I would say that there's a market for that. How long before rough agents implement good AI for their operations? How long before we implement defences against AI?

Duplex problem is not specific tasks but limited domain. If you are engaging with a person, you are going to err on multiple domains randomly. Especially if you are building a rapport over a long period of time.

If an AI can do that, then singularity is definitively reached.