Episode 1 - Taking the Mystery out of AI

AI is no longer a figment of the imagination.   We are all using AI every day, whether we know it or not. Despite its prominence in daily life, AI still remains a mystery to many.  In this episode of Exploring AI Matters, Dr. Stuart Feldman, a prominent computer scientist, will help us better understand what AI is and is not, how it can be used, and its broad impact.  [2022-05-12]

(Note that this episode came before the release of ChatGPT in November of 2022.)

Transcript

SPEAKER_00 0:03

Welcome to Exploring AI Matters. This podcast series, previously known as Mind the Gap Dialogues on Artificial Intelligence, will continue to appear in the ABA series to the extent that, in addition, all of the episodes, old and new, will now appear under our new podcast name, Exploring AI Matters. Thank you.

SPEAKER_03 0:37

Welcome to the first episode of our podcast, Mind the Gap Dialogues on Artificial Intelligence. As part of our series, we will sit down with economists, doctors, artists, scientists, astronomers, judges, lawyers, and policymakers, each of whom is engaged in applying or analyzing the use of AI technologies. Through these discussions, we will be exploring the broad applications of AI and their legal implications. If we are successful, we will contribute to an informed discussion of legal and policy issues related to AI. Given the innovative and creative applications of AI, its development and use raise important legal considerations, most notably in the areas of intellectual property and national security. In this episode, we will explore what AI is and what has happened to make it, after almost 70 years of research, so prominent in public awareness today. At one time, AI seemed to be science fiction. AI is now everywhere in our daily lives. From Google's search predictions to image and voice recognition to self-driving vehicles, the rapid advancement in AI's utility has been extraordinary. To help us understand the evolution of AI, our guest for this episode is Dr. Stuart Feldman, famous among computer scientists as the creator of a phenomenally useful tool named Make, used by generations of software engineers to manage the building of programs. Dr. Feldman started his career as an astrophysicist, earning a PhD from Princeton for work modeling galaxies, which he did by writing the code that implemented the models. This software work led to a career in computer science, starting at Bell Labs as part of the team that developed the Unix operating system and the suite of software and tools based on it. From Bell Labs, he moved to Bell Corps and subsequently to IBM Research. After IBM, he worked at Google's New York Engineering Center. He now leads scientific research programs at Schmidt Futures, the philanthropic initiative whose mission is to connect the brightest minds everywhere with opportunities to solve the world's hardest problems. Hello, I'm Mark Donner, a computer scientist.

SPEAKER_00 3:19

And I'm Ama Adams, a national security lawyer. We are your hosts for this episode of Mind the Gap Dialogues on Artificial Intelligence.

SPEAKER_03 3:27

In addition, we have two more hosts. Hi, I'm Charles Palmer, a computer scientist.

SPEAKER_05 3:32

And I'm Roland Trope, a national security lawyer.

SPEAKER_03 3:36

Each episode will be led by two of us, with the other two adding impromptu questions and comments as the spirit moves them. We thank the business law section of the American Bar Association for their generous sponsorship of the production of this podcast.

SPEAKER_00 3:52

So, Dr. Feldman, thank you so much for joining us for today's episode. I think I speak for all of the hosts here to say we're so delighted to have you here to discuss your perspectives on what AI is and is not. So why don't we just dig right in and then we can start to peel back some of the layers around AI? How would you define artificial intelligence? Um, particularly to those of us like me who don't or who have a very basic understanding of what computer science is?

SPEAKER_01 4:23

Turns out it's actually hard to define AI. And I can go and list dozens, but to me, AI systems are systems that in some way act as you expect a human intelligence would act. The classic Turing test was could a computer fake being human under certain circumstances? I won't go into detail, they don't make sense today. But that's the simplest uh definition. You know, could I confuse you that I were that my computer was a person? Uh much more useful definitions involve these are computers that act as if they were intelligent and do things that you think only an intelligent being could do. So, yes, I have put you, the human, into the definition loop. You're the one who says this is intelligent. Uh so you no longer think that running a payroll uh program shows intelligence. Uh in 1955, you would have been amazed. And so back then you would have said, gee, that's what people do. And now you say, well, of course, computers do that. That's what they're for. So, first of all, there's this sliding definition, and I'll throw in another curveball. Another good definition of AI is stuff that doesn't quite work yet, because the field has been bedeviled by uh when something works, it's no longer AI, it's that thing. Way back when being able to uh a bank being able to computationally read a check was magic. And now you never thought about what happened to your checks. So it stopped being AI, it just became thing. So, however, the touch is when it really seems to be acting in a way that you think only you should be able to do it.

SPEAKER_00 6:31

So, taking that definition that you just gave us, you also mentioned it's sort of a sliding definition. You know, I think one thing that is sort of clear to me, um, I think to all of us here, is that over the last, you know, even the last 15 to 20 years, we have seen a dramatic transformation um in technological developments, a sort of technological evolution. And AI is part of that. So, what would you think, what would you say, sort of thinking about that sliding definition? Um, what makes AI so useful in comparison to earlier technologies?

SPEAKER_01 7:09

Giving a snarky answer, suddenly it works. Uh but that really is the answer. And suddenly these are technologies that instead of people were saying it's going to be great, people can say it already does something that you can't do that you thought only people could work on. That there are big uh big problems and specialized ones that computers now successfully do. There are, I'll get in a few minutes, I'll explain why that's technologically driven. But what's amazing, yeah, in the 50s, nobody thought a computer could ever play a game. And then when computers were really good at tic-tac-toe, people said any you know, any three-year-old can play tic-tac-toe. And then in the late 50s and 60s, when somebody actually got computers to play checkers okay, people said, Well, but checkers is just for my grandpa. Uh, what about something big? And then uh suddenly computers play chess better than anyone ever has or will alone. And so suddenly things happen that said they think problems that people absolutely thought were native intelligence with huge study are being done by a computer. And for people who are experts in games who said, well, chess, well, okay, that's uh you know, that's for grown-up uh teenagers and stuff. And then go suddenly computers learn to play Go, the most subtle of the standard board games in the world, uh, the greatest strategy game of the East. And what happened is computers grew up, if you will, both the technology of the algorithms, the the computer science, and the hardware, both of them exploded onto the scene around the turn of the century.

SPEAKER_00 9:11

So I love this concept of computers uh growing up. Um, and so here's where I'm gonna start to peel back some of the layers. So, you know, for people like me who are reading about AI sort of in everyday literature, or you know, just trying to get a better understanding of it, we hear terms like machine learning, deep learning, and neural networks. And those come up often in the discussions we see around AI. You know, how do those play? You know, what did those mean, most importantly? And then how do those play into this concept of AI growing up?

SPEAKER_01 9:46

Let me take them actually in a different order. Neural networks were a model that literally goes back to the 40s. Uh, it was one of the original ideas of what AI was. And this was very interesting and extremely philosophical and very nice mathematics. And that was it. Unfortunately, it didn't actually get you to do anything real. I had a group uh at Belcor in the late 80s that was attempting to use neural nets. Neural networks are a mathematical construct that are inspired by the way neurons work, although they're not really quite the right way that neurons do things, but they're a very simple, interesting model. And you can show that a small number of neurons, computational neurons, can do some cute things. Unfortunately, they all stalled out in the 80s. They stopped doing cute, they just stopped getting better. Well, now you let the problems bubble and stew in the background for a couple decades, and there are some genuine rethinks and uh about how to use such networks. Uh, Jeff Hinton and a few other people are generally ascribed with the fundamental hard breakthroughs on how to improve, how to do net neural networks that instead of just being two layers deep, uh, sorry to get into jargon here, started getting into many layers deep, called deep learning because you had piles of networks uh uh feeding each other. Like there was then a practical breakthrough that said, now there's a way to throw huge computing at these guys. And it turns out that the trick is scale. Little networks will never get you anywhere. Uh big networks do amazing things, and humongous networks do even more surprising things. And to do this, you have to throw Google-sized computing of the practical PR and real-life breakthrough was combining the work from the University of Toronto with uh Google's ability to throw 10,000 computers at one of these problems profitably. And suddenly you had the ability to do remarkable things. I'll get to them in a moment. But this was both computer science getting better, people studying algorithms, studying the reasons why some of the strange things that people do. I won't get into any of the mathematics here, uh different audience, but uh we've learned a huge amount about various techniques and why they work and when they work. And then we have separately learned how to throw huge amounts of computing at these problems. And the third leg, scale is the third leg, huge amounts of data. You know, basically, when you're doing natural language understanding problems, you literally throw hundreds of billions or trillions of words at the computer. The scale required is, however, frightening. To do really good natural language processing, in other words, the computer handling what you can type or say means the computer has looked at hundreds of billions of words and has learned everything it can from them, uh, which is a reason since some those words can be in many languages, computers are suddenly able to do decent, if not great, translation. By feeding in voice, they can now listen and interpret what you say. So, those of you who use Siri or any of those programs, uh, in behind that is enormous amounts of AI, machine learning, computing using deep learning and neural networks. So, you know, in the first times you may have done this, you may have said, what time is it? And it told you take the train to Tucson uh because it didn't quite get it. It doesn't do that anymore. Yes, you can fool you can fool the systems, just like I can fool you, probably if I try hard uh with a bad accent. But uh having these enormous amounts of data fed into very large amounts of compute and with increasingly smart techniques, and we I'll I'll go on briefly there, if I but suddenly you get good results. Computers are better than people at recognizing many kinds of images. Computers are better than many people at understanding voice. Computers are much better, of course, at learning new games. And on the scientific side, computers are quite capable of running experiments and helping decide what the experimenter should do next better than the experimenter can do herself. These are all things that are true in 2022.

SPEAKER_03 15:22

So a remarkable property of some AI systems is that they can produce useful results that are at once better than anything else we can come up with and at the same time are impossible for us to analyze or explain. Given that these systems sometimes produce bad results, how can we know when to trust them?

SPEAKER_01 15:49

Oh boy, uh you have just asked one of the literal billion-dollar questions. Uh it's real, real tough. Uh it is the case that advanced AI systems frequently do completely surprising things. Uh the major chess programs play games of chess that look different than what kinds of games people play, uh unaided people. They find moves, they find sequences of moves that no human has ever tried. And usually they win. Uh there uh they can do computing of a sort that humans cannot. Humans do holistic uh analysis, they don't actually sit there and can sit walk through the next 32 moves. The computer, in a limited sense, can. So there are amazing chess games where when they were played and the experts were looking at them saying, crazy man just did something uh and just lost, and then 15 moves later, the world champion human resigned. So that's the good news. They play differently and better in some cases, just as autonomous vehicles usually drive better than people. First of all, they are rarely sleepy or drunk, which is a good step forward as a driver. And uh sorry for the lawyers on the uh call, but uh uh those are some of the real problems in regular driving. On the other hand, they occasionally make confusing errors that the average awake human would never make. Well, if you were simply playing the odds cold, which is where this whole area of autonomous driving started, you would basically say this is overall a net win. The total number of bad things that the autonomous drivers are likely to do is actually much smaller than the number of bad things people do. It's just that we don't forgive computers, and we do forgive people, usually. Uh so, first of all, you've got to ask compared to what? But then you really are opening a far bigger question that if I can't explain why it ran the car into the ditch, you may not want to get into the car again, even if the reason was that very large truck coming around the bend uh at you. Well, you never saw it because you had a light in your eyes. And and in real life, you would have been run off the road and dead due to the truck rather than having the computer run you off the road slowly. So that's the good cases. But when it gets it wrong because it never got trained on that kind of problem, it never had the insight and what and set of rules, then you say, I don't trust this. And we have not reached a state as people of being able to balance statistically rare but devastating and incomprehensible things from what people do instead. This is the techno-optimist view. The ordinary man's view is if I can't figure out why somebody drove me off the road, I ain't getting in the car. And that's a very reasonable thing to also say think, instead. Well, once you're at the level of complexity of the highest-end AI systems, they are making so many subtle decisions based on so much complicated data. It is impossible to give a simple story. By looking at 10,000 possible accidents, this computer program has computed that the best thing to do is to pull off the road. And unless you want to sit here and watch 10,000 accidents, which is what it did, and figure out which ones were more likely than others to kill you, it's impossible for you to get the explanation. It's much worse when you're looking at some of the biggest uh uh AI programs. Um, a program much in the press lately is called GPT 3. It has half a trillion parameters inside, parameters meaning weird numbers it has learned by analyzing enormous amounts of data. Half a trillion is 500 billion or 500,000 million different little numbers. And using all those numbers together, it does quite well at reading French and answering you in Chinese. And it does very well listening to you in English and answering you in English, but nobody can explain exactly why it said the right thing now. And how we balance these problems is totally, totally unobvious. Uh, for computers, the answer would be obvious. Trust the computer, it statistically is doing a good job. Uh, we are instead wired with very different decisions about what's good and what isn't good. I don't want to get into all the moral choice questions. You know, you can pose as many of them as you like. The real-world ones are bad enough. So we have passed the point where we can explain what's going on in the largest systems, sadly.

SPEAKER_00 21:42

So, given your point that AI does many things better than people, and you referenced uh particularly subtle decision making around complex data, be interested in hearing what you think are some of the most important areas of AI activity today. talked a little bit about um natural language and uh yeah autonomous driving but just curious if you you could sort of list the top three or four activities that you see as sort of critical in the AI evolution today.

SPEAKER_01 22:14

Continuous improvement of the ability to generalize. Most of the work that we actually understand in AI is sort of interpolation. It's seen a whole bunch of interesting cases of doing this and then doing that. It's seen people in pink t-shirts seen people in green t-shirts and so people in any color t-shirt are okay or something. However, it's when you're extrapolating from where you are to something brand new how do you trust the ability to get out of the cases you've looked at also how do you know that things can't go terribly wrong uh this gets into a whole field called computer safety that uh you you may not mind too much about things going a little bad but you care if they go very bad. How can you be sure? How can you be sure that you'll never do something awful and so you you put guards around the programs but of course you can't always figure out what they are. And so those are some of the key issues. And then what's really the path that is going to be exciting is as we've learn more and more tasks more and more things that AI-related systems can do and how do they start interacting. So right now you have some systems which are really good at understanding text and other parts of systems that are okay at listening to voice and there are other parts other systems that are wonderfully trained on either how you fold a protein or how you play the game of go uh same sort of same computers from DeepMind actually are have related techniques. So we've got a bunch of these how do they start converging and piling up to doing a lot more what you might normally have called thinking even though experts will say there's no thinking going on it's just this computing stuff we were talking about. There's this grail called AGI uh which is artificial general intelligence which means you basically it can learn it can think it can do brand new things it can create you can you can now basically turn up the violins behind the song and everything. Well that's an image that you know when computers actually are just better than people at being people and uh now you get all the scare stories about uh that follow from that idea this doesn't worry me because we have no clue how people think we have lots of little data but we really don't know what we do when we act as smart people actually I take it back we have no idea how to act as small not smart people these monster computing systems are not are nowhere near so good as your three-year-old at learning language uh and uh they don't walk all that well either and by the way one of the first people to make computers walk is Mark uh on this uh panel so uh yes uh just by the way and uh you see these pictures of various robots that walk remarkably well before they fall over well that's how you describe a one-year-old it's not how you even describe a three-year-old uh you know they stop falling over by then usually and computers do well real well at the one-year-old level okay how do we get out of this into much better performance performance in a wide variety of areas that really succeeds well given the curve of what we've seen in the last as you said earlier AMA 20 years I'm expecting that in the next 10 years we're going to see amazing better performance and more performance that you can trust. So those are the areas where I'm generally expecting wonderful things to happen.

SPEAKER_00 26:22

So you're talking about trust things getting better I kind of want to talk a little bit about maybe some things that AI maybe doesn't do quite as well. So we know that or it seems right like computers are able to understand what we're saying, you know, what humans are saying and what we're writing but sometimes that understanding seems off or it's not quite right. What what would you say or what would your answer be to sort of explaining sometimes why that understanding is off even with all of these advancements even with AGI all of these improvements and enhancements sometimes it's not quite right.

SPEAKER_01 27:02

I'm sorry let me be really unfair here you as a lawyer I'm sure have never had a situation where the client didn't understand what you just told her.

SPEAKER_00 27:12

Oh never they always understand everything I say so clearly they never have any questions they never call back for clarification.

SPEAKER_01 27:19

So yes okay so uh that although that was a snarky comment it's also actually relevant that even people have trouble communicating and that's when we have a really good shared basis of having grown up as humans that really helps you know so I gave you a techno-optimist answer. Now let's me give you a techno-realist answer. This stuff is really hard and there will continue to be areas where there won't be enough data there won't be enough computing the algorithms just won't be good enough and we won't the computer will not get it right. It will fundamentally not understand you it will fundamentally not figure out what to do about this six-way intersection it will not exactly figure out my handwriting which no one can read uh so uh including me by the way so uh I basically uh quite seriously uh say that we are getting better and it's never going to be perfect but just by the way you'd be horrified on the data about human performance.

SPEAKER_05 28:32

Yeah I just a quick follow-up question one of the things that makes AMA such a remarkable lawyer is that she not only can communicate to her clients but she knows in a variety of ways when they haven't misunderstood when they haven't understood her properly or when they say they understand her but they're not getting it. Are computers up to that with AI up to that level of knowing when a communication is unsuccessful or when somebody is lying which again something uh lawyers are supposed to be good at just like doctors have to be good at it. You need to know when the patient or the lawyer is misinforming you deliberately or inadvertently if you can't do that you're not much use as a doctor or as a lawyer.

SPEAKER_01 29:19

Let me take that in several pieces if I might uh Roland because that's a wonderful question. First of all, computers are capable of using clues that you can't uh we funded some research some years ago of somebody who was studying microarticulation. This is listen analyzing the voice at the millisecond level not uh much faster than you can actually hear the effects listening to weird uh muscle patterns basically and correlating that with people who were making hoax phone calls to the fire department and they were doing much better than the fire department in recognizing hoaxes because they could listen to different cues similarly looking at micro uh uh cues of your puffing eyelids and puffing nostrils uh where people don't quite have enough acuity to see these perfectly so there are different patterns that a computer might be able to use that would get a different set of the client is lying or the client did not understand which pill to take, uh all of which are exceedingly real. However, what I'm suggesting is if I might say extrasensory uh uh analysis by a computer to get around the really hard and wonderful question you just posed, uh Roland, which is how do you tell when the communication is going awry intentionally or unintentionally and uh emotional state recognizing human uh emotions is still a I'll politely call it a frontier topic. That means not solved. Uh uh and uh you you have posed a fine problem for the 2020s by the way well Stu I can't resist not jumping in here one of the things that uh I've heard a lot in fact the textbook I use for one of my classes is called prediction machines and uh Avi is one of our guests on the podcast earlier.

SPEAKER_04 31:36

So my question is how do you feel about that term? Is that accurate? Is that a good way to think about AI as just a really good prediction machine?

SPEAKER_01 31:46

It's a nice simple way to look at what we normally do. Most of the things uh most of the examples I have quoted have been prediction you know guessing how to play chess or go better and what I'm doing is predicting four moves ahead no matter what you do. And that's the classic prediction or things that I interest me professionally at the moment, which are the use of computers for predicting the behavior of proteins and such. So all of those fit in but this larger definition of behavior and interaction really does not fit into a simple prediction paradigm. You can always you can force anything into any paradigm if you try hard but it really doesn't naturally fit uh deciding that uh if the goal of our interaction was for the for the computer and the human to decide on that they like each other enough and trust each other enough to do something that is not a simple-minded prediction that is a precondition for doing something new uh and you know creativity is something we don't know how to define but the idea that a person interacting with a computer is going to be more creative we already see lots of examples of and that's I'm not counting on the prediction ability of the computer rather than its validation abilities. So in the easy cases prediction is a great summary for where we're really going that computers will be at best prosthetics and instead replacements of certain categories of activity it's broader.

SPEAKER_03 33:37

So um Stu uh Jeff Dean at Google took the work of Hinton's team in Toronto and demonstrated that it scales.

SPEAKER_01 33:48

Why was that so important? I remember the humor of the uh uh what was done uh Jeff uh Professor Jeffrey Hinton at the University of Toronto refused to be discouraged by what was called AI and neural network winter. I mean basically the field went into uh dormancy for from the 80s and 90s uh and the beginning of the uh 2000s except there were some stalwarts that actually kept working on it and thinking deeply about how to do it better and Hinton actually and a few like uh Lacun working with him came up with far better ways to use this stuff however they didn't actually have either the resources or the computing experience to figure out how to make these things work big. And it turned out that what was important is not just being doing things cleverer, but being able to compute enormously uh I remember you know the first experiments that came out of Dean and his group uh the bad joke was he had used 17,000 computers to make the world's most expensive cat picture recognizer. Basically the computer had studied millions and millions of pictures of cats and it had finally learned reasonably well how to tell whether there was a cat in a picture or not. Let's say that this is not a project that would necessarily gladden the Google stockholders as it was first happening. Anybody who looked carefully said wow because this this was in a massive computation parallel computation with millions of images that had learned something that it hadn't been told about. And the break there was now the world could understand how to throw lots of data and lots of computing at more interesting problems. And that was the demonstration that not only was this theoretically interesting and you could show lots of interesting problems done by a PhD for a whole thesis, but that you could routinely use computers to do important problems. And that people who weren't PhDs in AI could make them happen. And there was a shift in the mindset from hey this is an interesting problem too bad we're making no progress but it's you know it's okay to fund some research to this is the new frontier when suddenly there were problems that you really needed to attack or wanted to attack that you could it changed the game. There was a wonderful chart that the small machine learning group sent out Google at that point had you know only a very small number of top level experts in machine learning. That group has since grown a lot. However they basically had a chart which said if your problem is kind of like this, first try using this technique, then this technique then this technique and only then can you call us because we don't have enough time. And it was just a chart that told you a cookbook of the equivalent of if you've got three three rutabagas and scallops, try sauteing them except you know here's how. And this this was the magic that was happening right then. A simple interesting fact going back to my Google experience this happened I think around 2012. I had seen natural language uh processing since I had been a kid. I Bell Labs was one of the great hotbeds so was IBM research in doing classic wonderful research and how did you understand speech? How could you interpret text? And incredibly smart people spent 40 years coming up with very subtle ways of doing it. Well in 2012 40 years of research was thrown over the uh side of the boat it turned out that instead of doing all of those uh wonderful hidden Markov models you could instead almost always just use a deep learning computation and it did better. In one year the natural language understanding scores I won't try explaining where the numbers come from went up more in one year than they'd gone up in the past 30 total because suddenly we knew how to use this kind of new computing with this amount of massive cloud uh clusters with lots of data you know billions and billions of words not thousands and thousands of words and so suddenly what happened was scale and data became the route to success. Lots of uh of old lesearchers you know are still annoyed that you know dumb computing beats smart researcher five out of six times unfortunately uh okay there are still problems that need the deep AI analysis uh I make only a sour joke about that there's really great work that still has to be done but the machine learning techniques the deep learning techniques the distributed massive computing worked now let me put in a tiny footnote yeah remember what I said that the what these people were saying at Google try this try this try this and then call us well that tells you how tricky these techniques are sometimes they work and sometimes they don't and then we get back to some of the great questions Amma was asking and Mark was asking about if you if they seem to work some of the time should you trust them all of the time but what happened was it went for the whole field went from look at this toy working in the lab and isn't this amazing to this is how theory works and it's just there and free. Thank you. Don't ask how much computing is going on behind you. But okay.

SPEAKER_03 40:22

So so these breakthroughs that you've talked about Stu involve large scale compute clouds. Does this mean that only the big players will be in the AI game going forward?

SPEAKER_01 40:35

Getting these results has involved enormous amounts of data being used with really smart techniques and huge computers and we got these breakthroughs that problems that we had been working on in the research lab for 40 years suddenly became practical. We haven't solved everything but suddenly all of the incremental progress of the past 20 or 40 years got subsumed by new techniques that just changed the game. However yes uh part of this was throwing the biggest computing uh systems in the world at these problems so at one level this became a game for the big guys you know Google Amazon and so forth uh are able to do this you can rent their facilities but you need to do that well and GPT3 took I don't know was a month or three months or to compute on a very large many million dollar computer. So at the moment the leading edge is to the big guy and the big compute but there's two things first of all the result of doing all this computing is frequently something that you can use on a telephone or a small computer or even something smaller. This massive computing has created a decision module a prediction machine it's or an interpreting device but that interpreting device can now be run on your phone because you know it took that ungodly amount of compute to do step one and step two now goes real fast. And so your phone can keep up with your voice. The big machine trains the little one rent the big machines of these days also and you can run on top of things like GPT3 and do your own uh special models. So we're starting to reach a point I suspect where yes gigantism will still be really important but first of all there's a lot of research into do you really need to do all that work that you did? And many of these algorithms are applied semi-blindly because we don't quite understand what we're going to get. When you get more experience in a field you realize you don't need to do quite as much training quite as much computing there are ways of simplifying the results. So we are entering a world of greater practicality probably and right now there are areas where you must use AI you know if if anybody comes with to me with a scientific project involving trying to interpret images and they don't use AI, I just say try again. The AI is just too good at doing those problems. And we've learned how to do them well enough that you don't need to have the half trillion parameter model to do okay. So it helps to have the big guys doing the heaviest duty computing But my guess is that within a certain number of years, we will have started to understand far better how to get away with a little less, and that it will be greatly democratized in practice. There's good news and bad news in that. I know how much effort uh Google put into safety for autonomous vehicles. Because A, you know, Google is run by good engineers, and B, it's rich enough to afford throwing lots of hard uh problems at very smart people. Once you get down into the smaller shops, they're not going to be thinking that hard. And so we've got to make sure the guardrails are built in on the big systems so that people who use them don't drive off the metaphorical road. So I my guess is that in the next this during this decade, we're going to see a switch from only the big guys can make the biggest breakthroughs to anybody can make significant progress and significant use in whole new areas. Now, obviously, this is uh acting as a personal prediction machine, and you'll ask me in 2030 if I was right, uh, and I'm safe until then. But uh but un but honestly, I think that we're going to be seeing a reasonable shift over the next while. At a certain point, you know, you don't need even bigger brontosauruses.

SPEAKER_05 45:34

I'm gonna put up one policy question, though. Right now, there's a lot of attention in Europe and in Congress at uh regulating big tech. And from the answer you just very thoughtfully gave, there may be a need to have big tech just so that we can get over those uh gigantic computer resource challenges, or our adversaries in countries where it's not so where all of the commercial companies are linked up to the military will have advantages that we can't even begin to think about. And we may be putting ourselves behind them inadvertently if we try to regulate them prematurely. Uh, can you comment on that? Uh, if I haven't made it too obscure?

SPEAKER_01 46:23

Not at all obscure. Uh, let me just begin with a careful disclaimer. My career has been happily spent in what I guess my lawyers taught me to call companies with dominant market positions. Uh, some of them, of course, got broken up in antitrust suits. Uh, but okay. And others are in court on them. Uh, but uh so I have a bias, I'm in favor of the big, excellent company. But yes, there are, if I ask what are the American national advantages in AI, they surely do not are do not come from the uh military and uh industrial complex nor the aerospace complex. Uh they come from you know Google, Amazon, Microsoft, Facebook, and a small and a certain number of more obscure, amazing companies. Uh and each of these has enormous intellectual in addition to physical capital resources tied up in the hardest and biggest problems. So I'm in favor of avoiding repressive and uh and delaying regulation, uh, partly because it's so hard to regulate a subject which is changing quickly. And uh the regulations are almost always backward looking or looking in the in a broken crystal ball. Uh, I actually keep on my desk a broken crystal ball to show what I use for my forecasting. You know, it's an actual crystal ball which has a crack in it. So um so therefore, I let me just say that I agree with Rollins' basic point that it is easy to break the main string that is going to like give America a serious opportunity both economically and insecurity in this space. Figuring out how to avoid the worst consequences and how to control without ruining the progress is a very hard problem. And you know, basically, the main reason I'm not too despairing is the time lag between scientific and engineering work and regulation. And frequently what you regulate isn't interesting anymore. Uh however, that's not a trustworthy method mechanism. So I would be very careful about what you think you're doing. Uh the again, uh my if you go far enough into my background, I'm a child of Ma Bell. I worked in the old Bell system when it was the million-person Bell system. And I had a wonderful uh product, scientifically productive life at Bell Labs. This country has not fully recovered from the loss of Bell Labs and a few things like it. So I therefore worry that you can ruin a delicate machine if you're not careful. And I am quite uh concerned about what the redrawn uh uh international power map might look like in a few years. So I don't want to get uh too um uh deep into this, but AI and similar technologies change the time scale and the complexity scale of what is going to happen in any military interaction. And it's a real problem if you're slower than the other guy.

SPEAKER_03 50:14

So so with with your caveat about about premature regulation, particularly in rapidly moving fields, I'll I'll sort of challenge you to speculate. Are there areas of research that even now appear too dangerous to encourage?

SPEAKER_01 50:34

At the research level, I don't know how to answer that because the problem is research is multivalent. Uh you know, the one experiment can turn into a different use. Uh there are areas which are decidedly scary, uh, you know, being various computations that allow us to design biological molecules uh can uh lead to you to very uh interesting and possibly catastrophic results. There are the military side, we know that there are applications that we may not like the results of, but it's very hard to tell those apart from other questions of how do you get more reliable results out of a uh steepest descent computation, or how do you uh get a much more reliable reinforcement algorithm uh as you uh are using uh some sort of modern technique, or how do you actually have computers train computers to play chess, and then suddenly you say, well, that's the same thing as playing on the battlefield, isn't it? Oops. And so uh the my difficulty here is that in this this field it's much closer between genuine curiosity research and new new ideas and new techniques and potentially exciting in quotes applications. And so what you need to worry about is the directions and the use of the applications and how they apply to how things how society reacts and how these things are utilized. And but it's real hard to connect the dots between uh you know Lacune's uh great work and uh better uh deep network algorithms and battlefield analysis, which he surely did not intend or even think about. So uh it's a real touchy question because I don't know how to circumscribe early research in understanding algorithms and experimenting with lots of data with bad outcomes. Um I mean, just consider that uh you know that for varying reasons, people think that the social network algorithms, which are deeply AI-based in practice, uh, have led to social polarization and political bad results. Where in this chain are you going to would you have stopped research or experimentation? And at which point, uh how long did it take to realize that people don't always make the nicest use of things you give them? At that point, we've reached the limits of human prediction and human planning, unfortunately. And now uh I flip it over to the legal, social uh science, uh, political science and such people to say, how do we make smarter societies and smarter thinking about what to do with our tools? Uh, the problem is it's much harder to talk about the many ways an AI system can work than saying any sharp knife is dangerous because it doesn't always look like a knife.

SPEAKER_00 54:12

So looking forward, right? Um, we've had this really exciting conversation about what AI is, what it isn't, how it's evolved. Are there any new trends, exciting applications? I think applications, exciting applications was a term that you just used in your conversation there with um Mark. Are there any exciting applications involving AI that that you see on the horizon that you think we should all be sort of looking out for as we have a better understanding of AI through this conversation now?

SPEAKER_01 54:45

There's a bunch of areas. One in my own deep professional activities these days is ways that AI can revolutionize and improve scientific and engineering research. Uh, this will be too geeky for part of this audience, but the AI techniques should allow small labs to do things that only a huge lab could do before, to do experiments, to imagine experiments that couldn't be imagined before, and even analyze more of the literature than any poor slave graduate student can be made to read, which is a lot. But so there's this is a whole space that excites me enormously. And you know, in my uh day job, we're about to launch several activities to speed that up. Uh, and this is uh extremely interesting to me in general. Uh, when I look at what you can do with new techniques for analyzing scientific and medical data, it is amazing. Uh, what is it, what's in the data that people have yet to extract from it. So, on the more social side, we are start, we're suddenly starting to get enough data about society, about people, that you can start making much better analyses, better ways to just run some of the plumbing of the society. Um if uh you're going to uh you know figure out traffic patterns, computers actually are much better at helping you do this than just leaving the red lights on for an extra few seconds. Uh and we have it's really annoying that we're starting to commute again rather than just sit in this chair like I have for two years, uh, and remembering traffic has not been a good thing. So many of just many big areas are wide open. If you ask what are the social and economic sectors that are just ripe for improvement, and we we have many reasons they're not going anywhere, but the standard sectors are health and education. Uh yeah, health. I have some superb doctors, and they are really good at recognizing things. But what but you know, these best doctors have seen a small thousands of patients over their 20 years of hard work, and they follow through on maybe 10,000 people total. With good records, computers can plow through far more, and we're starting not reliably yet to find some interesting patterns. Um, AI is also really good coupled with new sensor technologies for doing things like measuring how you're actually behaving and looking for weird patterns, not just the easy patterns in your uh bloodstream, in your heart uh palpitations, etc. With new we can look with new computing at new patterns and hope to be able to recognize much at a much earlier stage what bad problems there are, and also tell how to fix things. Uh, I mean, I've looked at uh lots of interesting research on artificial pancreas uh insulin pumps and such. And these turn out to be really hard AI problems, um, in addition to being tough biochemistry and biophysics problems. Um, we have great hopes that you'll be able to control diabetes at a level that has never been possible because it takes uh the the body and the systems react too slowly. New biochemistry plus new AI may change the that from controlling your blood sugar over multiple hours to over minutes. These are all possible things that you can just see coming on the horizon. Uh by horizon, I mean five years even. Uh bigger uses in medicine. Uh uh, I don't want to get into the bionic man and all that kind of stuff yet, but uh you you you know, various interesting uh actuators and sensors will improve your basic life. And as I have trouble convincing myself that I'm still 27, you know, these problems are you know increasingly likely. And uh so the opportunities for medicine and even more, the opportunities for public health are enormous. You know, the the COVID crisis has not been a good, a shining story for public health. Uh, the epidemiological models have been terrible. You know, COVID has just been smarter than uh the epidemiologist too often. Well, partly because we don't have enough fine-grained data. You know, you go and count the number of ICU beds filled, that's not a great way to study uh the behavior of this disease. Ten years from now, we ought to be much better at this, and AI techniques will be at the core of any of the progress that we make. Education, uh let's just say that uh uh the US has a very mixed education success. Uh the best states and the best cities outperform the rest of the world, and many of the rest of them underperform embarrassingly poor countries. Well, how is it that we can apply technology so that everybody can have the same test scores as Boston? Uh and how do we improve the way people learn? And how do we improve uh on and take advantage of AMA's ability to tell when the client has just not understood something and convert that to a fourth grade class where the same thing happens every morning and the fourth grader doesn't get it, and the teacher isn't uh is too busy to recognize which fourth graders are in trouble because uh you know the teachers only so so competent. Well, computers ought to be able to help with more learning theory. So there are big things like that where there are wonderful opportunities for great progress. Yes, I can put in lots of red flags and I can give the other lecture, but fundamentally, I'm amazed at how well the new research has worked. And yes, there are new techniques that are just that, you know, every six months, basically, there's a new blockbuster technique in AI. Uh, and you know, the storyline gets changed among the experts. But until it's been used for a year or two, you really don't know the limits of each of these new tricks, and the tricks are themselves deeply complex. So uh watch the space. But I think that the applications are going to be the key. Well, how are we going to make clever use? Um, I am not worried that AI is going to eat all of our jobs and all of our souls. Uh economists have been reliable at predicting which jobs will be lost, but they have never been any good at predicting what was going to happen instead. And there are lots of people who worried a lot when the farm started shutting down and nobody was going to have a job because you didn't need to feed Dobbin anymore. Well, we seem to have gotten through that. And we're going to get through it with a mess, no doubt, through in the 2040s, but the 20s 2020s are going to be exciting, and uh then we're going to have to fight our way through uh the change. Uh I'm strangely optimistic it'll happen it'll work out, but it's going to take a lot of hard thinking and work.

SPEAKER_00 1:03:01

We will definitely be watching this space, and you set up a great uh lens for us to look at sort of the next stage in the horizon. So thank you so very much for your time today, Dr. Burk.

SPEAKER_01 1:03:11

Oh, thank you. Uh I I enjoyed it, and I hope uh you uh get some good use out of this. Thanks.

SPEAKER_03 1:03:19

Our theme music was composed and performed by the very talented Ben Rosenblum. We welcome comments and questions. Send email to comments at mindthegapdialogues.com. We read all comments and questions, and we'll try to respond in the letters section of a future podcast episode. If you are writing about a particular episode, please do mention the specific episode number. Please do include pronunciation tips to help us properly say your name when we reply in a subsequent episode. See you next time on Mind the Gap Dialogues on AI.

SPEAKER_02 1:04:11

Thank you for listening to the AVA Business Law Sections Podcast Series to the extent that the section offers a robust collection of content. To explore more about this topic or to learn about joining the section, visit ambar.org slash bizlaw. That's B I Z L A W.