The Exchanges

Q Can I ask, as an outsider to the AI community, what does a good environment for machine learning look like? What does that mean to you?

A It means good feedback loops. It means you have to understand when you predicted it right or wrong with enough reliability that you can train intelligently and quickly. I, I think, you know, an easy example, and this is something I actually saw Siri do recently, uh, it's hard for The Siri team to know when you say something about fried avocados and Siri thinks you're talking about affogado or whatever, right? Like the, the chances are you're just going to get frustrated. You're going to drop off and the Siri team has no way of understanding that, that they in fact messed it up. They added something recently to the bottom of the screen such that you can kind of report the error, though you don't have much of an incentive as a, as a single user to spend additional time to report Siri's error to the Apple team.

AI assessment note: “It means good feedback loops. It means you have to understand when you predicted it”

Q Would you say then that the, the rise of kind of These, as, as we said, kind of digital assistants, although we hate to call them that. Is, is the natural evolution then from Steve Jobs' PDA, or is it a radical transformation of how we organize our lives?

A I think everything tends to be much more incremental than we like to think of it before it happens. I think we really love the story of all of a sudden we're stuck in a virtual reality and your whole existence is there. It was Snapchat that, in my mind, had the kind of incremental build that was incredibly wise, right? Snapchat has this Relentless, um, product focus. And now, of course, they're the people introducing the idea of augmented reality to tons and tons of people, and games like Pokemon Go, right, where there is this, um, slow but sure corralling of technology-enabled innovations that really change the way that people interact, but that succeed because the behavior pattern is not Too unfamiliar to people. And I think that when you think about the interface that we've chosen at Clara, right, which is conversational, I'm happy to talk more about why I think that's a good fit, but, um, one of the biggest reasons is you're used to that. It's elegant in that it is, um, not uncanny, challenging, unfamiliar. It's, it's really just a step further from the behavior that you, uh, don't think twice about

AI assessment note: “I think everything tends to be much more incremental than we like to think”

Q How do you think about building a long-term AI company, and what that thought process looks like for you?

A Our feeling has been that the most important thing, well, firstly, the most important thing is the user experience, period, right? Um, as with everything else in the world, if you're trying to build a company that You have to put your customer at the absolute heart of everything you do and work backwards from what it is that the customer needs from you and, and find a way to do that thing. For us in the early days meant my co-founder and I scheduling like thousands of meetings by hand, right? It meant convincing my younger sister to come out and quote, intern with us, right? Largely helping us schedule these meetings. It meant friends of mine from college who were on break Helping us just manually work through what it was that the user experience needed to be so that we could learn what it was that we needed to, to scale, right? What it was that people wanted, um, and it would really change their lives in a big enough way that they were going to pay us for it. So I would firstly say for a long-term AI company, you have to have the same thing that you have to have for any company, which is an absolute customer heartedness. But the second thing I would say is that you need to have a really good environment for machine learning, and that needs to be the core of the company. And I do think that there is an advantage to many of the younger companies that have started from their foun…

AI assessment note: “you need to have a really good environment for machine learning”

Q Can I ask, as an outsider to the AI community, what does a good environment for machine learning look like? What does that mean to you?

A It means good feedback loops. It means you have to understand when you predicted it right or wrong with enough reliability that you can train intelligently and quickly. I, I think, you know, an easy example, and this is something I actually saw Siri do recently, uh, it's hard for The Siri team to know when you say something about fried avocados and Siri thinks you're talking about affogado or whatever, right? Like the, the chances are you're just going to get frustrated. You're going to drop off and the Siri team has no way of understanding that, that they in fact messed it up. They added something recently to the bottom of the screen such that you can kind of report the error, though you don't have much of an incentive as a, as a single user to spend additional time to report Siri's error to the Apple team.

AI assessment note: “It means good feedback loops. It means you have to understand when you predicted it”

Q Would you say then that the, the rise of kind of These, as, as we said, kind of digital assistants, although we hate to call them that. Is, is the natural evolution then from Steve Jobs' PDA, or is it a radical transformation of how we organize our lives?

A I think everything tends to be much more incremental than we like to think of it before it happens. I think we really love the story of all of a sudden we're stuck in a virtual reality and your whole existence is there. It was Snapchat that, in my mind, had the kind of incremental build that was incredibly wise, right? Snapchat has this Relentless, um, product focus. And now, of course, they're the people introducing the idea of augmented reality to tons and tons of people, and games like Pokemon Go, right, where there is this, um, slow but sure corralling of technology-enabled innovations that really change the way that people interact, but that succeed because the behavior pattern is not Too unfamiliar to people. And I think that when you think about the interface that we've chosen at Clara, right, which is conversational, I'm happy to talk more about why I think that's a good fit, but, um, one of the biggest reasons is you're used to that. It's elegant in that it is, um, not uncanny, challenging, unfamiliar. It's, it's really just a step further from the behavior that you, uh, don't think twice about

AI assessment note: “I think everything tends to be much more incremental than we like to think”

Q How do you think about building a long-term AI company, and what that thought process looks like for you?

A Our feeling has been that the most important thing, well, firstly, the most important thing is the user experience, period, right? Um, as with everything else in the world, if you're trying to build a company that You have to put your customer at the absolute heart of everything you do and work backwards from what it is that the customer needs from you and, and find a way to do that thing. For us in the early days meant my co-founder and I scheduling like thousands of meetings by hand, right? It meant convincing my younger sister to come out and quote, intern with us, right? Largely helping us schedule these meetings. It meant friends of mine from college who were on break Helping us just manually work through what it was that the user experience needed to be so that we could learn what it was that we needed to, to scale, right? What it was that people wanted, um, and it would really change their lives in a big enough way that they were going to pay us for it. So I would firstly say for a long-term AI company, you have to have the same thing that you have to have for any company, which is an absolute customer heartedness. But the second thing I would say is that you need to have a really good environment for machine learning, and that needs to be the core of the company. And I do think that there is an advantage to many of the younger companies that have started from their foun…

AI assessment note: “for a long-term AI company, you have to have the same thing... customer heartedness”

Q And so I have to ask them, what is the right way then for us to think about a digital assistant in today's world?

A It's so funny. I think we, as people, have this tendency to want to anthropomorphize everything. When I, you know, we have this concept of a digital assistant right now, and we think that it's new, and that we've kind of recently come upon it. But of course, if you, if you flashback, even to when Steve Jobs was launching the iPhone, in His introduction of the iPhone, he introduces it into a kind of category that existed, which was, uh, the personal digital assistant, the PDA, right? And it had become this kind of ubiquitous technology. Uh, it was just a bit smarter than what had preceded it, which was dumber versions of your calendar, dumber versions of your phone, email, et cetera. You know, similarly, if you think about the Jetsons, you have This, like, maid who's literally going around cleaning up every individual thing, but the dog bowl is not smart. The maid is smart, and the dog bowl is a dog bowl, where now we're seeing quite the opposite. The dog bowl is the thing that's smart, and there is no maid. There is no robot maid. I feel very similarly about, about this category of thing, and it's, it's a really hard conversation to have, where you try to explain, well, actually, what we're doing is just making this core software suite that you spend your life in. In that you spend hours a day in much, much smarter than it has been before by growing to understand you, by learni…

AI assessment note: “making this core software suite that you spend your life in... much, much smarter”

Q Where do you perceive that tipping point to be? Sorry, I know that's a tough question, but kind of what's the difference between opening Uber and, and pressing, you know, um, whatever it is, pick me up or whatever the, whatever button it is, and then opening up Google calendar.

A I think it's, I think it has something to do with, um, Your ability to know enough information about the person such that you can act on their behalf without having to get too much information from them. The simplest way, I think, to answer that is just, we like things to be fast and intuitive. Probably a mathematical equation for this, practically, uh, which is, you know, I'm willing to incur some learning time, some setup time, if, subsequently, it's much simpler for me to get this thing done. But, And different people have different willingness to have the time up front because they expect, they want things to be fast, right? These, like, people who use keyboard shortcuts. But it's usually the people who use keyboard shortcuts that are constantly attached to their keyboards. So the equation for them would be different than for anybody with a keyboard.

AI assessment note: “willing to incur some learning time, some setup time, if, subsequently, it's much simpler”

Q Is the vision to, to eliminate those people with time, or will it always be an integrative part of the system?

A I think they will always be involved. You know, it's, it's, we get asked that frequently, and it makes perfect sense. For us, though, until we have some kind of strong AI that completely obfuscates the need for human work, you necessarily have humans in the loop. And people can think about it and kind of arrange it a little bit differently where at Google, Google likes to think of itself as a software company. I think, um, which it is, of course, I think though in our conception, it's kind of a, it can be a strength and a weakness for them where in reality, of course, you have thousands of people that are contractors on the backend of Google's learnings, right? But, but there you have a kind of Throw massive amounts of data over your shoulder and have people try to circle the date time, circle the date time, circle the date time, circle the, you know, this, this very simple process. You don't know who these people are. They're all over the place. You don't know how to make sure that their work is accurate. So you end up with pretty noisy labels. It takes you a month. You turn it around. It's back to your team. Your team is working through it now. Uh, and now they want the prediction to be slightly better. So They take more. They throw it back over to people who are very far apart from the problem and ask them to solve these simple problems. One of the things that's amazing abou…

AI assessment note: “I think they will always be involved. You know, it's, it's, we get asked that frequently”

Q Is the vision to, to eliminate those people with time, or will it always be an integrative part of the system?

A I think they will always be involved. You know, it's, it's, we get asked that frequently, and it makes perfect sense. For us, though, until we have some kind of strong AI that completely obfuscates the need for human work, you necessarily have humans in the loop. And people can think about it and kind of arrange it a little bit differently where at Google, Google likes to think of itself as a software company. I think, um, which it is, of course, I think though in our conception, it's kind of a, it can be a strength and a weakness for them where in reality, of course, you have thousands of people that are contractors on the backend of Google's learnings, right? But, but there you have a kind of Throw massive amounts of data over your shoulder and have people try to circle the date time, circle the date time, circle the date time, circle the, you know, this, this very simple process. You don't know who these people are. They're all over the place. You don't know how to make sure that their work is accurate. So you end up with pretty noisy labels. It takes you a month. You turn it around. It's back to your team. Your team is working through it now. Uh, and now they want the prediction to be slightly better. So They take more. They throw it back over to people who are very far apart from the problem and ask them to solve these simple problems. One of the things that's amazing abou…

AI assessment note: “I think they will always be involved.”

Q I am actually really intrigued. Why do you think then, other than that, why conversational was the best fit? When you were looking at the different options available to you, why was it that conversational was the most aligned to user behaviors?

A So I mentioned that, you know, there were several years, nearly a decade of conversations that led to Clara, but that the inflection point for us was scheduling. Um, and I'll go a bit more into that because I think that's a good foundation. We started trying to build a normal app A software app of some kind to solve the scheduling problem because it was so tedious and frustrating and constant, and there's so much opportunity for error, right? It was just such, such a pain, and found that it was really challenging to think about designing an interface that was going to make this negotiation, this human negotiation, simple and accessible, and that That insight was really interesting and important for us to start to say, you know, I think that as opposed to this thing that we started with, which is like us trying to cram HTML stuff and emails and have people click things and go through a whatever, much more of a, you know, and there are tons of apps that have done this for a long time and they, none of them have really taken off. And so you ask the why, why, why is it that this hasn't taken off? And I do think there are conversations. Language interfaces are really important if there is a sufficient amount of complexity such that language is the best interface to resolve that complexity. And I think this is one of the big struggles that we have with people designing and thinking a…

AI assessment note: “Language interfaces are really important if there is a sufficient amount of complexity”

Q I am actually really intrigued. Why do you think then, other than that, why conversational was the best fit? When you were looking at the different options available to you, why was it that conversational was the most aligned to user behaviors?

A So I mentioned that, you know, there were several years, nearly a decade of conversations that led to Clara, but that the inflection point for us was scheduling. Um, and I'll go a bit more into that because I think that's a good foundation. We started trying to build a normal app A software app of some kind to solve the scheduling problem because it was so tedious and frustrating and constant, and there's so much opportunity for error, right? It was just such, such a pain, and found that it was really challenging to think about designing an interface that was going to make this negotiation, this human negotiation, simple and accessible, and that That insight was really interesting and important for us to start to say, you know, I think that as opposed to this thing that we started with, which is like us trying to cram HTML stuff and emails and have people click things and go through a whatever, much more of a, you know, and there are tons of apps that have done this for a long time and they, none of them have really taken off. And so you ask the why, why, why is it that this hasn't taken off? And I do think there are conversations. Language interfaces are really important if there is a sufficient amount of complexity such that language is the best interface to resolve that complexity. And I think this is one of the big struggles that we have with people designing and thinking a…

AI assessment note: “Language interfaces are really important if there is a sufficient amount of complexity”

Q Where do you perceive that tipping point to be? Sorry, I know that's a tough question, but kind of what's the difference between opening Uber and, and pressing, you know, um, whatever it is, pick me up or whatever the, whatever button it is, and then opening up Google calendar.

A I think it's, I think it has something to do with, um, Your ability to know enough information about the person such that you can act on their behalf without having to get too much information from them. The simplest way, I think, to answer that is just, we like things to be fast and intuitive. Probably a mathematical equation for this, practically, uh, which is, you know, I'm willing to incur some learning time, some setup time, if, subsequently, it's much simpler for me to get this thing done. But, And different people have different willingness to have the time up front because they expect, they want things to be fast, right? These, like, people who use keyboard shortcuts. But it's usually the people who use keyboard shortcuts that are constantly attached to their keyboards. So the equation for them would be different than for anybody with a keyboard.

AI assessment note: “I'm willing to incur some learning time, some setup time, if, subsequently, it's much simpler”

Q And so I have to ask them, what is the right way then for us to think about a digital assistant in today's world?

A It's so funny. I think we, as people, have this tendency to want to anthropomorphize everything. When I, you know, we have this concept of a digital assistant right now, and we think that it's new, and that we've kind of recently come upon it. But of course, if you, if you flashback, even to when Steve Jobs was launching the iPhone, in His introduction of the iPhone, he introduces it into a kind of category that existed, which was, uh, the personal digital assistant, the PDA, right? And it had become this kind of ubiquitous technology. Uh, it was just a bit smarter than what had preceded it, which was dumber versions of your calendar, dumber versions of your phone, email, et cetera. You know, similarly, if you think about the Jetsons, you have This, like, maid who's literally going around cleaning up every individual thing, but the dog bowl is not smart. The maid is smart, and the dog bowl is a dog bowl, where now we're seeing quite the opposite. The dog bowl is the thing that's smart, and there is no maid. There is no robot maid. I feel very similarly about, about this category of thing, and it's, it's a really hard conversation to have, where you try to explain, well, actually, what we're doing is just making this core software suite that you spend your life in. In that you spend hours a day in much, much smarter than it has been before by growing to understand you, by learni…

AI assessment note: “making this core software suite that you spend your life in... much, much smarter”