“The robots will know who they are”Vishal Sharma was the vice-president of Google for seven years, the head developer behind Android and Google Now, and responsible for strategic search technologies and advertising revenue and review systems. He has been working on his own, secret startup since 2014
THE RED BULLETIN: You’re a busy man Vishal Sharma, and you are building virtual assistants. How do I know whether I’m talking to a real person right now?
VISHAL SHARMA: That’s the holy grail, isn’t it? The famous Turing Test: can a machine be so intelligent that we can’t distinguish it from a human being?
So, can it?
Could there be a secret technology in existence right now, which is somehow constructing my face and voice? It’s conceivable, I guess. But it’s very unlikely, given the state of the art. There is definitely some way to go. We will know that we’ve crossed over another boundary when we actually pass the Turing Test.
But how could I identify a virtual assistant?
As a virtual assistant, you only have two choices when you don’t understand something: do you want to sound comedic or stupid? You can either make a joke out of it and engage in some sort of distracting conversation, or you come forward and say: I’m sorry, I did not understand what you just said.
>>>Want to find out about tomorrow’s ideas today? Who will be shaping the future and what will it look like?
Thanks to redbulletin.com you can follow every moment of this year’s Pioneers Festival via livestream. Redbulletin.com/Pioneers <<<
So the problem is speech recognition?
Human language is an enormously flexible and expressive thing. Our intelligence and our consciousness express themselves in our language, so there is enormous variation, there are millions of expressions and millions of emotions. Virtual assistants try to break down the input and match it with what works best, but you can identify them pretty quickly, just by varying the question. It doesn’t take a lot, but I believe that interacting with speech and gestures will play a significant part in interacting with machines in the future. Some people say speech recognition is a technology that’s always five years away. I’m not that pessimistic. Over time, systems will comprehend more and more.
But how can machines learn?
There are two different models in machine learning. Let’s say you want to teach a robot how to move. On the one hand, you can create some elements of movement in the robot and embed algorithms that tell it how to use them. On the other hand, there are self-learning robots: the robot is given abilities, but it doesn’t know what those abilities are. So for example you give him a limb, which has three degrees of movement, and you give it the task to move a ball closer to a goal. The robot doesn’t know how to solve this, so it has to figure it out. Over time, the robot is building a model of his limb and what he can do. The algorithms are emerging from that. A lot of what we consider to be awareness and consciousness comes from that form of physicality. In some ways, the robot is becoming conscious of what it is.
So feedback is the key to intelligent machines?
Yes, but in speech the feedback loop is broken. You say something and the virtual assistant has very little choice. It comes back and says: “I think you mean this. Do you mean this?” But the systems need constant feedback by humans and the pace of evolution will be determined by how effective that feedback mechanism is. I feel that at some point in time, a group of people will round the corner on providing that feedback and then we will see a sudden unexpected lurch forward. At that stage, we will see many more systems emerge. They will feel far more intelligent – and become more so every day. We are approaching that threshold.
The Turing Test
British computer scientist Alan Turing invented the test which bears his name back in 1950. It was designed to assess a machine’s ability to think. During the test, a person converses exclusively by text shown on a screen with two unknown interlocutors – one human and one computer. Both try to convince him that they are the human interlocutor and not a machine. If the person performing the test is unable to identify the machine at the end of their conversation, the Turing Test has succeeded. To date, though, no machine has passed. As an opener for the test we recommend to begin with a very simple question: “How do you feel today?”
There is also concern regarding this future breakthrough. Have you ever felt scared of artificial intelligence?
AI will be a very powerful technology in the future, so it does make sense to have some sort of vision. This technology will determine big areas of our life. How we do conferencing, how we open doors, how we obtain information relating to our own body, will all be governed by some elements of AI. So there should be a debate at the time you introduce automated evolution in systems that determine so much about your life. We should make sure that malignant things do not happen. It’s a legitimate area of concern and now would be a good point in history to be discussing this.
Another concern is privacy. Where do you draw the red line?
I like systems that know a lot about me: I click on this and that and in the end I get a recommendation for a book on Amazon, that’s valid. But that information being traceable back to me in a way that is going to stay static in perpetuity is not a good thing. I want to be able to make the system forget me. It’s very important to have that under my control. Privacy is a concern for me, it’s a concern for everybody, I think. Still, I believe that big data has a democratising effect – if the systems become widely available. If everybody in the world is able to spy on everybody else at the same time at an equivalent level, that helps work against centralisation of power. Most technologies that allow individual participation have been extremely corrosive to authoritarianism.
What was the last technology that really amazed you?
It happens constantly. When I’m late for a meeting and my phone tells me about the traffic, it somehow predicts my future and that’s astounding. Sometimes I say something to Siri and I fully expect it to not understand me – but it does. That is kind of astounding to me. Sometimes its amazing and spooky at the same time. Often it’s the simple things that I find amazing, for example, to turn down my thermostat from far away. If you only think about the energy value of that. How much energy is wasted worldwide because the lights are left on? Technology is magical.
What role can small startups play in innovation of virtual assistants, compared to the big players?
The space is so new and there is a lot of creativity and creative destruction. You come up with an idea, let’s call it an avenue of research, and you start building it. But because there is so much building to be done, even as a big company you can’t have people running around doing everything they want – you have focus on a specific avenue. But because there are so many avenues opening up all the time, it is impossible for a big company to dominate this completely. So there is a big role for startups to play. Pick one topic: deep learning. There are so many different sectors and each has its own characteristics. No large company can dominate it all. The space is so rich for startups, it’s second to none.