ai.bythebay.io: The Real Future of AI Panel
Recording: ai.bythebay.io: The Real Future of AI Panel
So it's been a long three-day conference. Who was here for all three days? Who was here today for the first time? Okay, awesome. Who was in either of these categories? Great. Okay, I'm just always doing error correcting codes. So this is, I think, it was an overview of AI for technical and technically inclined audiences. We started with a high AI day when we had technical leaders like holding on stage. We had pieces of code which kind of change the mood in the room, right? Because you see a piece of code, you know, something works somewhere. And hopefully some customers are affected
So we had that day. Then we had the self-driving cars day, which was essentially an AI in the wild, right? It's an example of AI which actually can relate to anybody. You know how to drive, you know what this thing should do. And we kind of delved into the expectations and reality of AI on wheels. And today we had, I think, a very good panorama of the future and underpinnings of things which can come, right? So I kind of, I don't want to drive this myself. I really want to have a conversation with the AI leaders we have on the panel here. So maybe what we'll do, we'll have everybody here introduce themselves even if they don't need the introductions. And then mention what AI aspect keeps them up at night, right? And we'll have conversations among the panelists based on this for about half an hour
And then we'll open up to the audience. So the audience, Devin, our indefatigable track manager, will carry the mic to you guys. And so, you know, please prepare your awesome questions. So let's start with the introductions and AI topics left to right. Okay. Hi, I'm Peter Norvig from Google. I guess the thing that worries me is job disruption through automation. And we've heard a lot about truck drivers and things like that
So I won't go into that part. One of the parts that worries me more is job disruption of the computer science industry. And I think what it means to be a programmer is going to change that we've built this tool set and this mindset that says you write down code and you think logically and you come up with a solution. But with machine learning, that skill set and mindset changes. We were sort of abandoning this mathematical science and trading it in for a natural science where we're now making a observations of the world. We're doing experiments. We're basing it on statistics rather than on logic. And that means the whole mindset is going to be different
The tool set is going to have to be built up around that. And as an industry, we're still stuck in the old days. So like the truck drivers, we're going to have to catch up if we want to keep our jobs around. I'm Lyle Unger from University of Pennsylvania. And I'm a computer scientist and increasingly a psychologist. I guess I stay up mostly trying to figure out how can AI be used for social good. There's a lot of money being made in selling things. I would love to sell health and welfare and better relationships
On the technical side, I guess the question I'm most intrigued by is how do we get AI models, machine learning models that generalize across wide domains? We've seen a bunch of claims of we do great for anything that takes less than one second, our way to glue AI onto theorem proving. But we don't really know how to have something that learns whatever deep representation of commonality, that learns causal models, that thinks the way people does, that ties to external databases in any sensible way that looks like how the brain does. So I think there's a whole bunch of stuff that will happen next decade that in some sense gets deeper, not with more layers, but with more sort of some sort of structure. I don't know what it is, but I wish I did. Thank you. Thanks for having us. I think that I'm Sri from H2O, founder and CEO. One of the things that really, obviously AI is eating software, as Peter rightfully put it
AI is beginning to take, we produce some code, test code using deep learning and it was just as good as a real test engineer or better than some of our CTOs types code. So you're beginning to see code being created by programs and my seven year old's code computer, right, the program for kind of code comes to mind. But I think, so that's only the tip of the iceberg. I think AI has, is beginning to eat hardware. And so as it's beginning to eat hardware, beginning to now trade off a whole different tool chain, as Peter rightfully put, there's a debugging of AI tools, debugging of the machine learning as the world is changing. It's a fast changing world. The debugging tools for that, some of the studios that you're working on trying to build to debug and go through. So where did the problem really start? How did we end up with the wrong prediction for, say, the President of the United States? How did we actually send this patient not to ICU? Things like that, data, data quality issues are still a real problem
The amount of labeled data in the planet is still very small. I think unsupervised learning and semi-supervised learning tools have to be, and I think the torrent chain of algorithms is just beginning to start. But the AI eating hardware part of it is now we'll have animated drones that have a life of their own. And we're going to be coexisting with some of these non-human species, right, in some sense. And so how do we manage and find our own relevance in the world that is going to be both bio as well as abiotic systems? So I think these are the kind of the things that occupy the kind of the realm ahead. But the biggest thing that occupies our mind as entrepreneurs and enterprise AI people is there's a ton of hype around AI on what it can do and what it cannot do. And every time there's hype to reality distortion, there's another decimation and disruption that's waiting to happen. So I think we are beginning to now slowly dehype some of the space
Our goal and mission as a company is to make AI boring, much like how Google made search boring. And I think the advent of systems of record companies like Salesforce and other companies with data is going to make it much more of a practical aspect of AI. I think the fears of AI taking away our jobs are real, but they're not immediate. I think the immediate concerns are how do we apply AI in a fashion that's more of a pragmatic aspect of bringing it to life in the real world. I'm Shibha. I work at Salesforce on Salesforce Einstein. And the thing that keeps me up at night is thinking about how to democratize AI technologies. Because we have at Salesforce, we have hundreds of thousands of businesses that use Salesforce every day to manage all kinds of customer data
It's beyond sales, service, marketing, things that I wasn't even aware of before. And all of these businesses, they want to apply AI technologies to solve their most important business problems. But it's really hard because you need to have large data science, data engineering teams. You need a lot of expertise to actually get things to model your problems and then to productionalize them. And so what we're building internally is this platform that we dub as the NoETL, NoPHD, NoDevOps platform for machine learning. And it's really to help all these businesses derive predictive value from their data. And so the thing that keeps me up at night is like what are the right interfaces that make this really easy to use, right? That's what I think about. Cool
So I'm Lucas. I'm the founder of Crowdflower. And if I'm really honest about what keeps me up at night as an entrepreneur, it's the problem that my company tries to solve for our customers, which is to help companies get labeled training data sets so they can build machine learning algorithms. And then also do human loop on the back end. So where the algorithm might be struggling, get more data and then feed that back in. And the reason that I care about that problem so much is I really want to make machine learning actually work for businesses. And the place that people get stuck so many times, including myself, almost every time I tried to do any kind of machine learning project was in the gathering data. And if you're not Google or Salesforce, it can be, even if you are Google or Salesforce, I'm sure, it can be really tough to get high quality label data, which is like the essential first step for machine learning
So I'm Stuart Russell. I teach at Berkeley. How many bears do we have in the audience? Whoa. My husband. One on the panel. Okay. So a few things keep me up at night. One is thinking about the fourth edition of the textbook
So we'll have a new chapter on deep learning. That's, I think it'd be chapter 19. And what worries me is that the other 27 chapters will say see chapter 19. Realistically, I don't think that's the case. I think there are many reasons why deep learning is not the answer to everything. And some of them have already been mentioned on the panel. I think there's a long way to go in figuring out how the clear advance that deep learning offers can be integrated with the rest of AI. And I need a lot of CEOs who complain that they can't find anyone who does anything except deep learning
And they don't want deep learning. They want something else. They want logical, knowledge-based systems. Or they want planning systems. Search. and problem solving. So there's a lot of other things we need to figure out how to combine. I think the medium term issue on destruction of jobs is real
And when governments ask what should they be doing? You know, should we be training up, you know, 8 million data scientists? I say no. We don't need 8 million data scientists in the entire world. In fact, what you need is to train people to be better humans because that's their ability to have to sell. Once their physical and their sort of routine mental capabilities are obsolete, what they will do best is be human. So train people to be good humans. And then the other thing that keeps me up at night is the use of AI for weapons. This is much more immediate. There are countries already developing and selling weapon systems that can choose who to kill by themselves
And when those become mass produced, we could see systems that have far more destructive capabilities than nuclear weapons, are much cheaper, are much easier to manufacture and acquire, and can be launched with no military industrial complex, no government, no large army. So pretty much anyone who has enough money just to buy them can use them. So that's a difficult future. Nuclear weapons are bad enough, and this is much worse. Thank you. So I'll ask if you guys could follow up on anything anybody else said, please feel free to jump in. I'll ask one more question for everybody. So we have a lot of different AI definitions, which we ask every speaker to define AI
So there will be a series of short interviews coming out in addition to the talks. And this is our small data set of AI definitions. But I know already they are very vastly different. And also we heard from the speakers in some of the cars in industry, in academia, there are various definitions. So I kind of was thinking how can I phrase this? So I want to ask everybody here two things. So first, what is your favorite example of AI which works already? It may be in your own company, right? So, you know, your company may be close to Sentience, or it may be one of your customers, companies you enable, right? But basically, or just anywhere else, right? So what is your favorite example of technology which you would call AI, or closest thing to AI which will be available? And what is your favorite example of AI? Because the theme of this day is what works already will lead to what works in the future. And so the second thing, what do you want to build in this coming year? Basically, what, you know, based on what is your favorite example, or maybe you want to build something different, but basically what will be the thing that will work on which will resemble AI as much as possible in the next year? So today, you know, anything from the world today and what you will make which will look like AI as much as possible next year. And, Ivan, you will do
Well, I'm sitting next year, so I guess I got to go. So my favorite example that came up a couple of months ago was a kid in Japan whose family had a cucumber farm. And they had to spend a lot of time sorting the cucumbers. You had the small ones and the big ones and the curvy ones and the knobby ones and so on. And so he built, put a camera, trained a system, and built little actuators to knock each cucumber into the right bin as it went down a conveyor belt. And what I like about that is this is just a school kid who comes from a farming family and he could see this application and make it work. And what I'm hoping for the future is everyone will be able to do that. This was obviously a very bright and motivated kid
We want this to be the standard tool that anybody would use. And what about the next year? What do you want to build? What kind of AI you want to create? I guess I want to build tools to let people do things like that. And I want to build a compendium of intuitions for what works. Right? So we, you know, you want to build a system like this and you say, well, how deep should my network be? And how wide? And what's the learning rate? And so on. And you go to an expert and an expert means someone who's built two or three systems in the past. And they say, here's my rule of thumb for these number of inputs in this dimension. Use this. But if that expert had built two different systems, would they have completely different intuitions? And we don't really know yet
So we want to get to gather tens of thousands, like all the Salesforce customers, and then decide which of those intuitions are actually right and which ones are counterintuitive. So I don't think I believe in AI. I teach AI at Penn and the students go and they say, but all it is, is search and alpha beta pruning and machine learning and optimization and deep learning. There's no intelligence there. All it is, is a bunch of engineering tricks. And I go, yeah, I'm an engineer. I'm an engineering school. Is there a problem with that? So I'm not sure that AI is something that I find particularly useful as a conceptual label
I love all the engineering tricks. I love search. I love machine learning particularly. I love the math behind it. And it's, I look like it works. So I don't think AI is necessarily the best thing to think. I like your idea of making AI boring. I want AI to go away
I just want things to be smart and behave well. I always thought that if the fields of statistics and control theory had been more open to new problems, then we wouldn't have needed AI. It would just be the combination of those two fields. We shouldn't have, and they for various reasons been incredibly closed. I've been fighting with my statistician friends for decades. It's like, hello, don't have to have a proof theorem, don't have a noise model. We can still do stuff. It's, you know, my best friend's a statistician
What can I say? But, and what I'd like to do, I think, I'd like to build a system that helps people figure out when they're depressed or isolated and helps suggest useful interventions or something about it. Will it be AI? I don't care. I just want something on my phone and your phone that will help you, that you can choose to share data with or not, that you'll say, yeah, this makes my life better. This makes me more connected and feeling more satisfied with my life. If you're in Silicon Valley, it will be called worldbeak.ai. Yeah, absolutely. I think, I mean, there was a couple of interesting thoughts even from the previous question, kind of a riff off of them as well. I think as we were starting our company, one of the couple of quotes that were interesting, one of the quotes from the co-panelist here, Peter, data science will be the sexiest job of the next 10 years or eight years, as many years ago
That was Halvarius. And Peter, and Peter had an interesting blog. One of the blogs I was reading was the Google search bar, right? Sort of how it fills your, as you're typing it in. And it's obviously using some of similar distance metrics that are statistical, right? Sort of statistical learning has its own nuance. And so some of the, I mean, our customers have been plugging in AI deep learning, machine learning, even simple rule based applications slowly transforming. You see this part rule based part machine learning pattern, and finish based applications. We predicted one of our customers ended up using H2O to predict flu this season in California and found interesting patterns. So how, and of course, it led to improvements that Walgreens ended up using, Makersen ended up shipping drugs using the flu prediction from the hospital system
That went on to Makersen distributing drugs accurately just in time for Walgreens to predict block by block how many flu samples would be picked up in the last, literally, who were more likely to have propensity to get flu in all nine yards. The end to end application was really fun to see and see literally how the prediction was very cool and identical. We used some of the AI was used to predict what is the sequence of how videos are presented in the Olympics. And so literally Bolt wins the race. Should we do the presentation ceremony or should we do the ad before that? I can literally see and predict how do you juxtapose content. And I think all content curation is taking on the path of the algorithm. There's an interesting, obviously we're seeing a lot more, more fun uses of AI where vision, the artificial eye is here. So people are translating traditionally boring problems which are transactional problems or long series of time series problems into visual problems
And then looking at, for example, company goes to a conference, what were the approval rates, what were the approval-approver relationships. And suddenly you created a visual problem out of a, to understand organizational structure in companies. So we're looking at a lot more AI slowly, initially augmenting the, the accountant, for example. And then beyond that, looking for duplicate payments, for example, in large companies and predicting which vendor is least likely to deliver service, for example. So a lot of interesting, like classic, like business use cases being replaced by business rules, but simple pattern recognition based on data. And that pattern recognition going into the next phase. I mean, AI has been with us. I mean, the reason to dehype AI to some degree, AI has been with us
I mean, Turing is not a new person, right? AI, the goal of AI or the kind of the, our attempts at making AI real have been with us for 100 years plus beyond, at least beyond. And the reason to kind of think of that as a concept is essentially we want to make sure it works this time. So we don't have to be having a panel 100 years from now about how AI will change the world. And our vision is that only AI will be there 100 years from now, not anything else roughly. I mean, the name H2O is H2O is needed for life on this planet, AI for other planets. But I think that, I think the, what are we working on these days? We're trying to build kind of along the lines of what Peter was mentioning, an expert system, blow back to the past, expert system to drive AI, to drive deep learning. We call it automatic deep learning, right? Auto DL. There's Auto DL and Auto ML
We have two, we have two of the top great Kaggle, Kaggle now Google, data scientists in the company, one by name Mark Landry. So we're trying to beat Mark Landry with Auto ML. And then we have another data scientist called Dimitri Narco. So we're trying to beat him with Auto DL. And so automatic deep learning. But using expert systems and recipes, some of us who learned computer science from the Kenneth book know the first preface. He says, if you can cook, you can code. Well, we want to create recipes for making AI more practical and capsuled, if you will, to the degree it can be
But what we are seeing is the advent of data ware, right? Sort of. So the data ware, which is taking over, we're training computers, no longer programming computers. And training takes time and training takes both a challenge within bias versus variance. And trying to give your data enough statistical structure so you can tolerate the anomaly to tolerate different bumps in the overall space. Time series, you're seeing lots of time series data. We've just focused a bit more on how do we get the pipelining to work? How do we get time, apply the right filters, kernel methods? So we're working on trying to get ways to simple counts, windows, the traditional time series stuff, Kalman filters. The 60s era, 70s era algorithm is still the number one time series algorithm out there. But building notion of time series, integrating a lot of the advances from Google and other places, TensorFlow
For example, we're basing some of our core architecture to use TensorFlow and make it at the heart. But fundamentally, what are we doing more of is making it simpler to use and debug and apply AI, deploy AI as a software thesis. The difference with data science and AI is, AI is predominantly a software engineering practice, if you will. It's more of a software. And so fundamentally, make it easy to get into. So you don't need to have five degrees of PhDs to get into using it. So I think that trying to make it easier is one of the paths. And that's kind of the problem seems like Einstein path as well
But more importantly, one of the articles that the, it's funny, again, Quentin is now again at Google, is had an interview with us a couple of years ago. And I think what's most valuable in the future is being human, fundamentally. And how does AI enable us to be better humans? And that's kind of the fundamental thesis ahead. So that simplify the product so it can reach wider audience. And being human is trying to build communities. And I kind of really like how Alexei built a real thriving community of by the Bay, right? Community by the Bay. And one of the articles of that was the AI by the Bay from a few years ago. So I think building that community is still the strongest theme around AI, as well as all things technology in the Bay area
I appreciate it. And I appreciate the community, right? It's not so the community would not be here. Thank you. Thanks, community. Thank you. I've done a really funny example of that recently because we also have like auto ML as part of one of the things that we're building and so someone on our team was going to work on that and and he was really disappointed to realize that it was essentially grid search over a smart grid search over a hyper parameter space and he was like this is automatic ML not autonomous ML and he was he was like surprised to realize that. It's not really thinking it's not intelligent it just solves the problem better than humans. It's funny because I guess I mean I don't know it's it's to be on a panel with the folks that wrote the textbook I used in the class I was just thinking about like the first time I ran algorithms and I think it's like it's actually amazing how the dumb algorithms do like make such like awesome results
I remember actually implementing it wasn't even machine learning it was alpha beta pruning which is a pain in the butt if any of you guys have ever tried to follow along and implement that but but and watching you know program play Othello better than me was like such a miracle and made me think like wow you know like maybe our brains are kind of built out of simple pieces in the same way and I think. They're not built the same way trust me. We're just not just. They're not built from simple in this. From simple pieces right. Yeah. Like physics is sort of simple in a way right. I mean and then and then I think like I mean I feel like all of you guys have had the experience of probably most of you had the experience of building machine learning classifiers where you know you like like doing vision for example that I've been working on recently like you know you make a pedestrian detector and it works and you're just like wow you know and now like the deep learning algorithms they're they're simple in the pieces but really hard to inspect and really hard to figure out exactly what's going on
I just think that's like the coolest thing so I I think it's also kind of unlucky unlikely that AI stops being sexy because like you know the tools are good I think it'll be even maybe more fun. So I don't know I think that's like my favorite part about working in like working in the AI industry and and so you know I just have to tell you like like we have customers do all kinds of crazy stuff because you know with crowdfire this view of like just watching people's jobs and I think just one of the funniest jobs that we saw was this this company it was actually a waste management company that was trying to figure out which of their garbage like giant like garbage things were like full because apparently the state of the art is like you just go around like on a set time scale and you grab all your dumpsters but they were trying to like use vision to see which dumpsters are full or not. I mean not a super hard vision problem but you know it was challenging for them but then the funniest part I remember like looking at that job and thinking and like some of our employees were kind of joking when they were working out they were going to figure out where the dumpster is and go get it because like you'd see some good stuff like you know in the dumpsters and then I actually went to India to like one of the places where we have like a lot of workers and it's really interesting to see you know somebody that basically does labeling for American tech companies all day long has a really weird insight into American culture like they felt like you know like very familiar in some ways and also had these like questions and they were really they're asking me a lot about that dumpster job like what was that like why do they want that and it's kind of remarkable what people throw out so anyway genius application that somebody should run with here I think so over there's a map of Paris on the wall and so I have a little apartment just about where that guy's hand was pointing actually and so because of that I have to pay taxes and because of that I have to do a lot of translating between you know long boring French documents and legalese so that I can read them and then translate back so I don't get in trouble with the tax authorities so using Google Translate it's my it's my favorite AI application right now it's really improved dramatically it still doesn't understand agreement and for some strange reason it still has certain spelling mistakes in its French but it's really amazing how well that works and what would I like to do next so actually something I've been working on for a while but with very very limited bandwidth and student resources is intensive care medicine so you know we we worry about you know how many people are killed on the roads every year and we think that self driving cars will be will be great but twenty times as many people die in intensive care units every year if you could without even saving any lives if you could just reduce the length of stay in the intensive care unit by five percent that would save an amount of money equal to the entire budget of the National Science Foundation every year so the the mileage you get out of improving the quality of medicine is incredible the data is all there once you fight your way through the lawyers intensive care patients are connected up to computerized sensors and so all the data is digitized and sort of available Has anyone actually fought their way through the lawyers? We have succeeded at some so I have a collaboration with San Francisco General and it can be done other places University of Pittsburgh has succeeded but this kind of thing is something that really would benefit from the concentrated attention of the AI community I think there are huge business opportunities as well as you know big big gains for people ICU admission prediction misprediction almost all fatalities in hospitals because of the misprediction and one of our customers has built a ICU or not kind of app at the admissions stage It's amazing that the hospital system still uses almost the Titanic way of resource management It's a ship we have so many resources and so many beds How do we treat a patient today? And so in this age and that's one of the things that will probably have the most significant impact Lung cancer data which came out recently that's not a good cancer data predictions We work with the US oncology data set and clinical trial matching It's another very serious problem with high dimensionality I think the high dimensional problems are kind of the ones that are of fun I guess the real answer to the they are not getting boring is we solve newer problems more fun problems these days than were solved before with simple linear logical programs if you will But I think the fundamental theme that we see with AI toolchain is that it is still reasonably difficult to take the code that comes out of your model and plug it into an application and kind of manage it I think the overall most of our customers do not have the skill sets Or actually more directly put do not have the multi talented different talents to work together kind of team works to take AI to production And I think that's something that's lacking quite a bit I think that's the part missing in a lot of purely data science events And I think our community comes from both angles and we have I think a lot of tool sets which bridge data engineering and data science and working industry So I think it's a topic I think a lot of folks can address So I'd like to open the floor to the audience We have about 15 minutes And you guys can ask any of the panelists or all of the panelists And you guys can jump in and answer if it's for everybody I used to keep my eye on some tasks to understand where the field was And they all sort of fell recently Go and poker And not Jeopardy maybe Any upcoming tasks that you are keeping an eye on that would help me to understand that the field has made the next level of progress One that I can imagine is SAT levels that are superhumans But any others that you can think of? Well I think one of the most interesting tasks that I think like if you look at like the ratio between sort of like human skill and artificial intelligence skill Like the place that I think there's the most sort of amazing mismatch is in grabbing things So it's like really hard to get to get AI to just like pick up a cup Like I wonder if the Go playing program robot had to actually put the Go stones on the board And he like moved the board slightly Right so actually Amazon has a picking competition So you know you get a bunch of shelving of standard Amazon warehouse style And then your robot has to pick arbitrary objects out of the shelves and then they're all mixed together So that I think is a standardized task now And I think we'll see a lot of progress The big problem is just hands, robot hands It's kind of like if I put your hands in buckets And told you to pick something up when your hands are inside the buckets It's not very easy to do So I think as we see I think 3D printing in particular Will enable us to build much more intricate Low cost high sensor high tactile kinds of hands And that will really make a difference I thought it was really fascinating I think Google had a project where they had just hands grabbing into buckets Like night and day for like a whole like room full of hands You can see it online it's pretty cool Just to generate training today and then you released it Which I think was a pretty cool thing to do So this is directed to Peter Norvig and Stuart Russell So first of all it's an honor to see you both up on stage I learned a lot from your book And I actually wrote a deep learning book for O'Reilly And we reference your book everywhere throughout that So thank you I just wanted to ask you mentioned an update to your book on deep learning Are you ever going to do a fourth edition of the book? Yes Peter's been very patient waiting for me I have an extra job in the university Which takes up about 80 hours a week on top of my regular job So I have not been able to work it But I'll be on sabbatical starting in July So it'll be done during the next year One more question I think multiple of you actually mentioned about building something more like a platform And so other people can use or share certain component I think it's a great thing One thing I'm curious about like since we're talking about the future So W3C and IEEE Those kind of organization existed before Because we need those I wonder if there's anything you guys are forming Or some other people are forming At least a organization to make sure we can easily share things And also building common components or defining protocols If you can think Sure I think several or a few industry-wide organizations have started Which kind of the coming together of Apple, Google, and several of the companies together Open source has been a pretty big moment in this space Sort of as mentioned community It's truly a open source movement has basically made makers Out of companies as opposed to consumers of software So as someone rightfully said It's all a software moment And once it becomes software We all know how to manage, maintain software in some sense The AAA AI, it is funny to see Which they did a conference a few weeks ago here The traditional academic AI conference AI organizations have been there for a while I think they will get revitalized, getting revitalized The R community, the Python community Which have embraced TensorFlow It's a very good community It's a lot of good open source movements That are coming together Spark, Spark as its own H2O All of these communities have been pretty solidly working closely with each other Instead of truly creating a competitive organizational structures around But I think bigger companies I think the bigger piece is data Sort of how do you get to a place where it's open data movement And I think that doesn't yet started AI is kind of AI is single It's constantly seeking data Sort of if you don't have good data You really can't do a lot of algorithms Or training dataware in a few many ways But I think that sense getting open data movements Whether it's in healthcare, insurance, or financial services That's something that could be interesting And that's where we'll need a lot more leadership from others I think one thing the organizations IEEE and so on Have been working on Is standards for ethics And that's an important area to get worked out I think that's a great question Okay, I have a question So I'm a data scientist from a company called the Climate Corporation So what we're doing is We're trying to Do AI for farmers to automate the basic agricultural industry What I find is Actually many good ideas Are from the farmer themselves Rather than from engineers or statisticians I was thinking Do you have any ideas Where Based on your experience How we can make this AI more inclusive When we replace their job Can we make them to be a contributor And instead of You know Let them losing their job They can be gradually integrated in this revolution Yeah, I think that's my main question Yeah I don't have a solution But I do think there's a trend To have much more integration of humans with AI So I think this view of The AI does everything and replaces the human It's not having to be that way In some sense some people get replaced I don't use travel engines anymore But I'm very much involved in my booking of airplanes And there is something It used to be called AI on the back end That does the scheduling And the determination there That does recommendations for me And that's a pretty primitive old That's a ten year old technology For a commercial Funny maybe But I think there are a lot of cases Where the AI's And the thing to think about Where can the human A lot of it's crowd flower You do a piece out there But there's not very often a tight integration And somehow when we work with other people We're very tightly integrated Most of us spend our whole day talking to people The computers don't have that integration yet with AI And I think that's for me a big open question Maybe you guys have suggestions about where it's working Yeah, I guess if you could think of it more as augmenting humans Rather than replacing them Exactly Where it could take over the more mundane and dangerous things That you know, repetitive tasks Where you don't really need a human to be doing them An example I had was My grad student When I was in grad school My roommates were PhD students in biology And they would spend their weekends counting worms in the lab And they were like really bright people obviously And this is like an area that's ripe for automation, right? Where image recognition capabilities could help you count the dead worms And the alive worms And just enable these grad students to do other things That they're probably much better at The fact is that 90% of the people in the world do repetitive things Now you might say that's not great That we're in fact using people as robots But that is the That has been the story of the last 200 years And in fact we were using people as agricultural robots For thousands of years before that So there is I mean unless you imagine Let's take a particular sector like car manufacturing Unless you imagine that people will buy ten times more cars Then if people work with robots and are ten times as productive We will have ten times fewer people working on building cars And in fact that's exactly what's already happened So I think augmenting with humans is nice for those of us who are not in the robotic professions Right? Who are not acting as human robots But who are managing and directing and inventing and communicating and so on But the fact is that for most people in the world Those are not professions they have any chance of getting into And we have to think about how we're going to deal with this transition There is a possibly golden future But it's not going to happen by itself If we just continue with current economic policies We will see further separation of income levels between different classes And we will see reducing employment opportunities for most people Okay, a couple more questions keyed up back here So we'll see if we can get through them Hi, my question is Does a lack of a clear definition of AI have a negative impact on it as a subject? Can someone repeat the question? I had trouble hearing it My question was Does the lack of a clear definition of what AI means have a negative impact on it as a subject? No, it's good We can do lots of different stuff It's evolving No, I think there's I mean, take the other field That tries to build intelligent agents Namely, control theory They have a really serious problem That they came up with far too precise a definition Which is, you know, linear systems with Gaussian noise, period Nothing else And maybe a little bit non-linear if you're very clever Right? So when I mean, in the early 50s The control theory people And people who wanted to do something that was going to be called AI You know, got together You can just imagine the conversation, right? That the wannabe AI person says You know, wouldn't it be cool if we could have a robot that could, you know, go down to Safeway And do all the grocery shopping and come back home? And then the control theory person says Well, is that a linear or non-linear problem? The statistician says Well, is that linear regression or is it mixture of Gaussians? And it's like there's a total mismatch Because fields ended up over defining Overly narrowly defining what their range of interest was I remember having a conversation actually with a control theorist Where I was, you know, we were writing a proposal together And I was trying to point out that the writing of the proposal was another example of the Another example of an AI task And to get that control theorist to understand that writing a proposal is a decision problem It was just, it wasn't easy I mean, eventually I think we figured it out But there's just a different way of thinking And I think AI has benefited from being willing to take on, you know, anything that humans can do and more besides Well, I think that, I mean, to round up that The being human thinking on And I mean, one of the interesting quotes is for hundreds of years Since the invention of steam engine, apparently before steam engine We didn't really care between 9.05 and 9 o'clock exactly Because you missed the train And the locomotive kind of pushed all of us into becoming more regimented and rule-based And start making things, making kind of a little bit of dehumanization has been happening through the industrialization for a while And I think the big difference between AI movements before And I think Google, Amazon, all these are examples of AI companies with real engineering behind it Software engineering, right? Sort of, AI departments when we went to college And we were not invested with the cool compiler kits, right? So, one of the big things that has happened in the last five, ten years is the ascent of real good engineering behind AI Well, there was been engineering back, they were doing Lisp machines, they were doing time sharing AI has been tremendously flexible at putting out interpreted languages There's a bunch of stuff that are now standard programming practices that came out of AI So I think I'm going to disagree that the flexibility has been tremendous The flexibility is awesome Time sharing, is that an AI concept? Well The flexibility is awesome But I think what truly productized AI is good software engineering eventually And I think what we are beginning to see is kind of beyond engineering How do the revival or the re-emergence of the Renaissance man In many ways you want to combine different faculties into a person And how do you combine, you're not just an engineer, you're able to, you're not just a hacker You're also able to build a market around what you just built And I think that arrival of a three multi-dimensional person is much in demand right now And AI will help us automate away things that would, I mean you can equip yourself with those features And the flexibility of definition of AI means you can add those to part of your repertoire Which was missing before and now you can, every individual can be even more powerful And I think that, and have a bigger significant impact as a person My statistician friends are always complaining to me about how they invented everything And then AI like names it better But I have to say they're always asking me for help dealing with large data sets So I remember going into like an R forum which is like the standard stats software I really like R and so it had a limit on the size of data sets And I remember going in and just being like hey you know Like is there any way to raise this limit, like can we expand this And they were like yelling at me like that's stupid Like have you thought about like sampling your data Like you couldn't possibly like want these like large data sets I feel like whoever actually knows how to operate the computer and manipulate the data Is like probably always going to come up with the most interesting insights Or the most forward looking stuff Because that's I think it's through experiments that you actually figure out what's useful and powerful Okay, good job We got one last question back here She's been waiting patiently so I want to get her in okay So here she is Hi, I'm Elsa So all day we've been talking about using AI to solve really big problems Since the aftermath of the election We've been looking at how technology played a role in the results So I want to squeeze the collective brain juice of this panel And just ask you guys What is the potential for using AI in influencing the next election? I think the first thing to note is that all the candidates are sold like products And there's a huge amount of AI behind targeted marketing campaigns Plusing the right YouTube ads to the right people So there's a vast amount of very careful marketing happening behind pieces So that part is has been there for a while and is not changing How can we do things differently? I'd like to see maybe more transparency of what ads are being shown The good old days you could watch the TV or read the newspaper Really hard to see what YouTube ad is being shown to what person And a lot of them are ones that would be embarrassing I think if they came out I'm told by people who placed them So I think maybe one piece is at least some more transparency Identification of the famous false news question Identifying and countering some of the bots that are generating Nice revenue generating falsehoods I think there is sort of a Information war And I think both sides are using AI behind it I think this election really was a triumph for marketing That we built this political system Where traditionally you get all the viable competent candidates together And then you choose not based on who's most competent But who has the best marketing And this time one candidate just said I can skip the competent part And go directly to the marketing part And that was a vulnerability that was in the system Nobody had exploited it before But now that they have We know what that vulnerability is So maybe next time we can correct for it It's clear that AI won the war Right, sort of We didn't just have enough control on the AI system in many ways And the candidate actually with less money in this particular case Ended up using open source software I won't name them To actually use better To some degree They ran a better campaign And we knew both of them were buying products from other AI vendors And we were some of them But we didn't sell them It's open source So anybody can use So AI did win in some sense In a very strange way But in And also Lack of controls In 2008 Led to the financial collapse in some sense And I think we learned from that The system got robust years later And I think the same thing with somewhat robust I guess Yeah, never mind Let's not go there But information war has always been on And when John F. Kennedy won the elections It preceded the TV Made him look great on TV YouTube I think Obama used YouTube really well And fundraised online So he used that technology And I think the ad tech or mad tech Control free ad tech world Led to some kind of a great interesting Emergence of the fake news phenomenon if you will But I think all of this is There's still an underlying Malaise in the society if you will Right To go back to people without jobs Technology has definitely disrupted life And AI is a moniker for several large companies CEOs To start looking to cut more jobs Right now And we see that Underlying social malaise needs to be cured as well It's not either or Uber and Lyft connect more people today Not just online but offline So the idea transformations are all Going faster The memes are going around faster So I think the advent of Kind of putting regulation on news if you will Is really strange But I think what we're seeing now is The emergence of more controlled AI And if we can actually come to a point where We can assess what is real I think that's a pretty strong statistical feature I think nobody wants to be lied to They might like what they're hearing But if they knew that this was a lie They wouldn't like it And so A person of any political inclination Should be happy if they have access to tools That will protect them from being lied to That will screen out or flag things that are known to be false And so you could sell that tool to anybody And I think it's going to be increasing Just like you sell locks to people who buy houses So they can lock their front door People are going to want to have tools that Fake news detective That make their information environment Trust You know reasonably trustworthy This might be the wrong panel to predict what people are going to want I don't know I think it's probably success of your panels To bring out Like it feels like the Oscars Where you have to make a political statement That's weird Yes I try to bring it back to the maker of the show here It's pretty interesting that we had a fun time Thank you Shrey I really appreciate everybody's input You know you never know We're very fortunate We have great panels And I want to just make one kind of connection So in May last year I saw a conference called Data by the Bay Which was kind of my idea to start seven conferences together And then the ships launch them To kind of move separately So one of them was Democracy by the Bay Or d4d.io So we had companies like Brigade We had folks doing polling We have city governments So if you're really interested in topics of Data for Democracy We're going to have it back in November So our next data conference is Data by the Bay Is coming back in November Together with the engineering conference called Scale by the Bay this year So please follow the Twitter handle Data by the Bay And we will tweet everything which comes out of this event And I want to thank our panelists So one thing I would say is that when an AI program So in 2023 when an AI program is running for president Or is the president of the United States I think we have kind of achieved the final nirvana We can go actually let humans be doing human work And let an AI program do most of the bad work for us I hope it will be good to us Let's hold that So on that great note I want to thank our panelists And let's give them a round of applause Thank you