3 ms·

I think #2 can be really simply fixed: they always have to identify themselves as machines. Once you know it's a machine you're talking to, you treat it as a l

I think #2 can be really simply fixed: they always have to identify themselves as machines. Once you know it's a machine you're talking to, you treat it as a less-awful version of the voice prompts we've had on phone systems for years. The potential efficiency gains from this (think of calling into one of these instead of your traditional phone menus) are substantial. We should start looking into how to regulate AI (even if it just starts as "AI should identify itself as AI when they're using human communications channels") to make it work well for us.

sjm-lbm 9y ago
This goes against the most interesting part of the demo, though - that the 'umms' and 'uh-huhs' (and other small things) are enough to make this system pass the Turing test, at least over a narrow communication channel. Honestly, one thing your idea may show how far the big tech companies have strayed from their desire to create a concise and interesting product. Doing something like what you propose - basically, a system that calls, identifies itself as a robot and then is freed from the (difficult) technical problem of appearing to be human would be much easier to implement. With a bit of Twilio, "weekend hackathon to a workable demo" levels of easy to implement, IMHO. And, honestly, it looks like the public at large might have preferred that, and anyone using such a system to schedule an appointment would likely have a pretty much identical experience to the one provided by Duplex.
- lallysingh 9y ago
  I don't think the technical problem vs the self-identification are conflated. It's still easier to talk to something that acts like a human (at the conversational level, at least). I'd prefer those subtle hints ("umm") over explicit "That's not what I asked for" when communicating. I'd prefer to expect them to understand the same. When I know I'm talking to a software agent, I can change the way /I talk/ without having to lose fidelity from the way it talks. For example, I can spit out a large, complex sentence that's usually harder over the phone: Groups of four don't need or get reservations. "Reservation availability is Tuesday 2 to 5 pm, wednesday 2 to 3pm, friday 4pm to 6pm. " Then a "Thank you, goodbye" message tells me that I'm done with them, and as I know it's an AI, I don't have to worry about having offended a customer.
criddell 9y ago
> they always have to identify themselves as machines That sounds like it will work as well as the federal do-not-call registry. I think most of the issues could go away if we fixed the phone system to provide authenticated information about the source of the caller. If you can't or don't want to provide that information, that's fine, but I'm not going to answer that call.
- lallysingh 9y ago
  Oh, that's a good idea. The AI implicitly identifies itself via caller ID. If you're a hair salon, you get lots of calls from that number for many different customers.

sjm-lbm 9y ago

This goes against the most interesting part of the demo, though - that the 'umms' and 'uh-huhs' (and other small things) are enough to make this system pass the Turing test, at least over a narrow communication channel. Honestly, one thing your idea may show how far the big tech companies have strayed from their desire to create a concise and interesting product. Doing something like what you propose - basically, a system that calls, identifies itself as a robot and then is freed from the (difficult) technical problem of appearing to be human would be much easier to implement. With a bit of Twilio, "weekend hackathon to a workable demo" levels of easy to implement, IMHO. And, honestly, it looks like the public at large might have preferred that, and anyone using such a system to schedule an appointment would likely have a pretty much identical experience to the one provided by Duplex.

lallysingh 9y ago
I don't think the technical problem vs the self-identification are conflated. It's still easier to talk to something that acts like a human (at the conversational level, at least). I'd prefer those subtle hints ("umm") over explicit "That's not what I asked for" when communicating. I'd prefer to expect them to understand the same. When I know I'm talking to a software agent, I can change the way /I talk/ without having to lose fidelity from the way it talks. For example, I can spit out a large, complex sentence that's usually harder over the phone: Groups of four don't need or get reservations. "Reservation availability is Tuesday 2 to 5 pm, wednesday 2 to 3pm, friday 4pm to 6pm. " Then a "Thank you, goodbye" message tells me that I'm done with them, and as I know it's an AI, I don't have to worry about having offended a customer.

criddell 9y ago

> they always have to identify themselves as machines That sounds like it will work as well as the federal do-not-call registry. I think most of the issues could go away if we fixed the phone system to provide authenticated information about the source of the caller. If you can't or don't want to provide that information, that's fine, but I'm not going to answer that call.

lallysingh 9y ago
Oh, that's a good idea. The AI implicitly identifies itself via caller ID. If you're a hair salon, you get lots of calls from that number for many different customers.