181–190 of 786 posts
This is going to be awesome/terrible for social engineering attacks. We'll be training millions of people to to trust phone calls from "Google". But are you talking to Google-calling-on-behalf-of-your-boss, or Google-calling-on-behalf-of-some-phisher? Or better yet, some custom system that pretends to be Google over the phone? Who knows!
This is one end-game of AI - changing the economics of scams. Imagine how a 'long con' works today: a scammer befriends a person online through a video game or social platform, and develops a rapport with them over the span of days, weeks, months. After some trust has been gained, the scammer then requests money from the victim. Does this happen a lot today? I don't know, but certainly one reason it doesn't is the ec…
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.
Is any product remotely as close to this with natural-sounding human speech? Are we entering an era where all voice assistants are getting to that "Her" level in terms of human-like vocal quality? That part is as impressive to me as the semantic parsing it's doing on that call.
Other companies are using huge libraries of recorded human voice for communications and concatenating them together in intelligent ways.
Earlier quoted context omitted.
I don't know if I'd call it a "deficiency" - if we interpret "disfluency" in a literal sense as "not flowing" without negative connotation, then the interruptions (hmm, uh, okay) are actually communicating useful information to the other party. I might even say that omitting those interruptions (and replacing them with, say, dead silence) might be poor communication.
The "um" isn't a deficiency, but the slow response is. If the response is artificially delayed to give the appearance of slow thinking, and an "um" added to fill the artificially long silence, that's an artificial deficiency.
When talking to real humans, I've encountered people who don't do this, and I find it makes communication difficult and frustrating.
I'm not 100% sure why I need this pause, but I know I need it. Maybe I'm considering whether my question made sense or needs corrections/additions, so that I can't focus on the answer yet. Or maybe it takes time to switch the brain from "speaking mode" to "listening mode".
At any rate, when people do this, I have to ask them to repeat the first few words they said because I didn't catch them. And the reason I didn't catch them wasn't mumbling or background noise or anything. Well-formed sounds made it to my ear just fine, but my brain wasn't ready to accept them for a fraction of a second.
This is amazing, but one thing I don't really understand is this: earlier in the presentation, they demoed some new Google Assistant voices. All of them sound like standard computer-synthesized assistants. On the other hand, the synthesized Duplex voices sound indistinguishable from human speech to me, even without the "disfluencies" they include. If Google has gotten speech synthesis to this point, why isn't Assista…
During the demo it all sounded very realistic, except for some parts like the times. It would flow naturally then all of a sudden pause awkwardly and then say a time like "12 pm" in a weird way.
I have a feeling they are getting it to sound so realistic because there's a fairly small amount of responses and questions it needs to work with, so they can either pre-record real humans, or heavily tune a ML voice to sound as natural as possible.
Earlier quoted context omitted.
At least in the demo, the system didn't identify itself as a bot. It pretended (fairly convincingly) to just be a person, and seemed focused on interactions (making a restaurant reservation, etc.) that don't involve someone recognizing your voice. In other words, I don't think the attack surface is any different than situations where you could just place the call yourself, save perhaps for the possibility of scaling…
So now I can script a bot to book restaurant reservations all over the city at busy times. Then nobody shows up for the reservations, the busy time has passed, and customers have moved on or gone home. Restaurants make or break on one or two nights in a month. A calculated social engineering attack like this could bring down hundreds of restaurants in a city, which would cause millions of dollars in lost taxes, and y…
Earlier quoted context omitted.
> I'm very excited for a future where more and more tasks can be automated, enabling humans to get higher and higher living standards. That's been the dream for a long time in some circles. With the enormous productivity gains and ability to leverage external energy sources (fossil fuels, solar, etc.) we could have built a society of wealth and leisure for all. Maybe we will still. The hope is that if there is enough…
The technology promises "wealth and leisure for all". The capital owners promise "it'll trickle down". Technical utopianists need to start tempering their optimism with the realities of human nature and design systems accordingly.
Yes the rich have it better, but the poor also have improved in ways that the rich of 200 years ago couldn't imagine.
Earlier quoted context omitted.
To save CPU/GPU/TPU there should be a high-frequency sound, as in people can’t hear, so the computers talking to each other and switch to a faster way to communicate. If this is included you also have way to detect if you are talking to a bot/duplex.
Doesn't most carriers heavily "compress" the sound, removing all sounds/frequencies that a human can't hear, etc? https://www.youtube.com/watch?v=w2A8q3XIhu0
Good times !
Earlier quoted context omitted.
Clever social engineering is really customized, but social engineering that works "often enough" is mostly reducible to a flow chart. If you can automate it, your response rate only needs to be, what, like people clicking on spam? Tiny. The days of trusting meat sacks with important information might be numbered. Someone should make the dystopia where Skynet doesn't build terminators, but call centers.
When people talk about AI taking over, they really need to consider this route. Social engineering most of the world simultaneously. It doesn't have to build armies of robots when we'll tear everything down for it.
A few observations: * In the example where Google asked about holiday hours -- they can now automate gathering information about businesses in bulk without having to rely on any APIs or user supplied info. Interesting thought experiment is Google validating their reviews/business listings by actually calling businesses and speaking to a real human. * This is going to be fantastic for accessibility. Maybe I struggle t…
This sounds like a nightmare for businesses. This time commitment asymmetry will be the issue with these systems. Like email spam, it becomes much easier to waste others people's time when you automate time wasting. If people use it to flake a lot, I could see businesses just not responding to the assistant.