Understanding AI: Is Google Duplex Dangerous? - GoTechTalk


Sunday, May 27, 2018

Understanding AI: Is Google Duplex Dangerous?

Google Duplex
You've probably heard about Google duplex it's new artificial intelligence assistant Google announced at its developer conference a few days ago.

Here is Google duplex booking appointment at a hair salon Video

It's amazing isn't it not only is duplex able to get the jobs done it is able to do it reasonably competently without us noticing it's an AI but that's also the scary part how does duplex know when to pause how about those hmms and yes and those precise intonations, I was thoroughly impressed by what Google duplex does and how well it does it many of you guys want to know about artificial intelligence.

My first impression after watching the video is just like everybody else astonish, it's amazing and it truly is if I were the lady in the phone call I wouldn't be able to guess it, there's no way for me to tell Google also showed us another piece of conversation between duplex and the restaurant stuff take a look.

Dude the same thing happened here Google duplex seems can post and thoughtful and can trust to the human this Baxter the question is artificial intelligence that smart now the answer is of course no despite the fact that Elon has won us many times about AI as an existential level threat as far as I know one AI trained to do one thing is not able to do another unrelated thing this is a very important statement in the case of Google duplex since the AI is trained to make restaurant reservations it will not be able to talk about your favorite color or ask you who your favorite band is, the team at Google pointed it out in their blogs to one of the key research insight was to constrain duplex to close domains which are narrow enough to explore extensively duplex and only carry out natural conversations after being deeply trained in such domains it cannot carry out general conversations.

This is the point I'm trying to make according to Google's AI block duplex is trained with a large amount of anonymous phone conversations to learn how to book appointments make conversations and ask for opening hours that's all it cannot carry out general conversations neither is it able to ask different questions like what is your favorite color it is confined to a specific domain and is one of the major characteristics of OAI software's from a capability standpoint duplex is no different than any other AI tool the only thing that differentiates duplex from other AI tools is its understanding of vocal cues like ummm and as well as this one mm-hmm,  so how is duplex able to do that it turns out those vocal cues that make duplex human-like are also designed the system also sounds more natural thanks to the incorporation of speech.

Disfluencies these are added when combining widely differing sound units in the concatenated Texas Beach or adding synthetic weights which allows the system to signal in a natural way that it is still processing in our studies we found that conversations using these disfluencies sounds more natural and familiar this is important to note because to duplex it's not a matter of whether to get vocal cues or not but a matter of when to give vocal cues as well, it would be churlish for me to deny duplex of the incredible capability it demonstrated its worth noting that one of the key objectives of duplex is sound natural which has left incorporation of these disfluencies not the other way around google duplex uses recurrent neural network to understand and interact with humans in any line of conversation it takes in three pieces of information hex recognized by the automatic speech recognition technology features of the audio as well as the context of the conversation in the case of restaurant reservation duplex recognizes questions asked by the restaurant stuff it also recognizes the accent of the stuff as well as the parameters of the conversation before a proper response is generated in order to perform in various situations, duplex team trained the AI to various scenarios here's duplex handling interruptions.

As you can see all three situations focus on covering real-life encounters as well as making conversations natural at a fundamental level recurrent neural network works better for Google duplex because it remembers what happened before the question it is contrasted with the forward feet networks which takes in source inputs without an understanding of time an example of forward feed networks would be image recognition what computers are exposed to thousands of images to recognize certain categories in the picture the only input they consider is the current image however for recurrent neural networks the AI not only considers the current set of inputs it also remembers its previous decisions let's say Google duplex is asked to reserve a table between 4:00 to 6:00 p.m. in a certain restaurant and it has previously agreed to 5:00 p.m. they will remember that decision as a part of the context unit this is why Google assistant is so much cleverer at answering questions than Siri lastly let's be clear what Google duplex is about its advancing technology sure it's about showcasing what artificial intelligence is capable of sure but more than anything like any other product announced during google developer conferences it's about marketing developers know that if they're developing cutting-edge AI systems they should use Google's tensorflow platform which has tremendous strategic value to Google not only would Google be able to have control over AI implementations Google might in the future sell more of its tensor processing units which obviously make money for them and i need hardly remind us that Google announced this AI tool last year that automatically removes obstacles we have not seen it since so is Google duplex dangerous??

Well answer is no, at least no more dangerous than Google assistant or Siri it's smarter than them sure it sounds more natural than them absolutely, but the fundamental principle that governs all of them are still the same, my opinion is that a current stage it's out of our control to stop AI progress what we can and should do instead is to embrace it and most importantly we need to make sure it benefits all of us.

No comments:

Post a Comment