vtt/episode_015_small.vtt

WEBVTT

00:00.000 --> 00:05.440
 The following is a conversation with Leslie Kailbling. She is a roboticist and professor at

00:05.440 --> 00:12.080
 MIT. She is recognized for her work in reinforcement learning, planning, robot navigation, and several

00:12.080 --> 00:18.560
 other topics in AI. She won the Ijkai Computers and Thought Award and was the editor in chief

00:18.560 --> 00:24.320
 of the prestigious journal machine learning research. This conversation is part of the

00:24.320 --> 00:30.400
 artificial intelligence podcast at MIT and beyond. If you enjoy it, subscribe on YouTube,

00:30.400 --> 00:37.760
 iTunes, or simply connect with me on Twitter at Lex Friedman, spelled F R I D. And now,

00:37.760 --> 00:45.360
 here's my conversation with Leslie Kailbling. What made me get excited about AI, I can say

00:45.360 --> 00:49.920
 that, is I read Girdle Escherbach when I was in high school. That was pretty formative for me

00:49.920 --> 00:59.280
 because it exposed the interestingness of primitives and combination and how you can

00:59.280 --> 01:06.000
 make complex things out of simple parts and ideas of AI and what kinds of programs might

01:06.000 --> 01:12.800
 generate intelligent behavior. So you first fell in love with AI reasoning logic versus robots?

01:12.800 --> 01:18.240
 Yeah, the robots came because my first job, so I finished an undergraduate degree in philosophy

01:18.240 --> 01:24.160
 at Stanford and was about to finish masters in computer science and I got hired at SRI

01:25.440 --> 01:30.960
 in their AI lab and they were building a robot. It was a kind of a follow on to shaky,

01:30.960 --> 01:35.840
 but all the shaky people were not there anymore. And so my job was to try to get this robot to

01:35.840 --> 01:41.200
 do stuff and that's really kind of what got me interested in robots. So maybe taking a small

01:41.200 --> 01:46.160
 step back to your bachelor's in Stanford philosophy, did master's in PhD in computer science,

01:46.160 --> 01:51.440
 but the bachelor's in philosophy. So what was that journey like? What elements of philosophy

01:52.320 --> 01:55.200
 do you think you bring to your work in computer science?

01:55.200 --> 02:00.080
 So it's surprisingly relevant. So part of the reason that I didn't do a computer science

02:00.080 --> 02:04.560
 undergraduate degree was that there wasn't one at Stanford at the time, but that there's part of

02:04.560 --> 02:09.200
 philosophy and in fact Stanford has a special sub major in something called now Symbolic Systems,

02:09.200 --> 02:15.520
 which is logic, model, theory, formal semantics of natural language. And so that's actually

02:15.520 --> 02:19.680
 a perfect preparation for work in AI and computer science.

02:19.680 --> 02:23.760
 That's kind of interesting. So if you were interested in artificial intelligence,

02:26.000 --> 02:32.560
 what kind of majors were people even thinking about taking? What is in your science? So besides

02:32.560 --> 02:37.120
 philosophies, what were you supposed to do if you were fascinated by the idea of creating

02:37.120 --> 02:41.440
 intelligence? There weren't enough people who did that for that even to be a conversation.

02:41.440 --> 02:49.920
 I mean, I think probably philosophy. I mean, it's interesting in my graduating class of

02:49.920 --> 02:57.120
 undergraduate philosophers, probably maybe slightly less than half went on in computer

02:57.120 --> 03:02.240
 science, slightly less than half went on in law, and like one or two went on in philosophy.

03:03.360 --> 03:07.920
 So it was a common kind of connection. Do you think AI researchers have a role,

03:07.920 --> 03:12.480
 be part time philosophers, or should they stick to the solid science and engineering

03:12.480 --> 03:17.200
 without sort of taking the philosophizing tangents? I mean, you work with robots,

03:17.200 --> 03:22.960
 you think about what it takes to create intelligent beings. Aren't you the perfect person to think

03:22.960 --> 03:27.440
 about the big picture philosophy at all? The parts of philosophy that are closest to AI,

03:27.440 --> 03:30.400
 I think, or at least the closest to AI that I think about are stuff like

03:30.400 --> 03:38.400
 belief and knowledge and denotation and that kind of stuff. It's quite formal, and it's

03:38.400 --> 03:44.000
 like just one step away from the kinds of computer science work that we do kind of routinely.

03:45.680 --> 03:53.040
 I think that there are important questions still about what you can do with a machine and what

03:53.040 --> 03:57.680
 you can't and so on. Although at least my personal view is that I'm completely a materialist,

03:57.680 --> 04:01.920
 and I don't think that there's any reason why we can't make a robot be

04:02.800 --> 04:06.800
 behaviorally indistinguishable from a human. And the question of whether it's

04:08.480 --> 04:13.600
 distinguishable internally, whether it's a zombie or not in philosophy terms, I actually don't,

04:14.720 --> 04:16.960
 I don't know, and I don't know if I care too much about that.

04:16.960 --> 04:22.080
 Right, but there is a philosophical notions there, mathematical and philosophical,

04:22.080 --> 04:27.520
 because we don't know so much of how difficult that is, how difficult is a perception problem.

04:27.520 --> 04:32.640
 How difficult is the planning problem? How difficult is it to operate in this world successfully?

04:32.640 --> 04:37.920
 Because our robots are not currently as successful as human beings in many tasks.

04:37.920 --> 04:44.320
 The question about the gap between current robots and human beings borders a little bit

04:44.320 --> 04:52.400
 on philosophy. The expanse of knowledge that's required to operate in this world and the ability

04:52.400 --> 04:57.280
 to form common sense knowledge, the ability to reason about uncertainty, much of the work

04:57.280 --> 05:05.040
 you've been doing, there's open questions there that, I don't know, require to activate a certain

05:06.320 --> 05:09.840
 big picture view. To me, that doesn't seem like a philosophical gap at all.

05:10.640 --> 05:14.240
 To me, there is a big technical gap. There's a huge technical gap,

05:15.040 --> 05:19.360
 but I don't see any reason why it's more than a technical gap.

05:19.360 --> 05:28.400
 Perfect. When you mentioned AI, you mentioned SRI, and maybe can you describe to me when you

05:28.400 --> 05:37.680
 first fell in love with robotics, with robots, or inspired, so you mentioned flaky or shaky flaky,

05:38.400 --> 05:42.720
 and what was the robot that first captured your imagination of what's possible?

05:42.720 --> 05:47.920
 Right. The first robot I worked with was flaky. Shaky was a robot that the SRI people had built,

05:47.920 --> 05:53.360
 but by the time, I think when I arrived, it was sitting in a corner of somebody's office

05:53.360 --> 06:00.640
 dripping hydraulic fluid into a pan, but it's iconic. Really, everybody should read the Shaky

06:00.640 --> 06:07.840
 Tech Report because it has so many good ideas in it. They invented ASTAR Search and symbolic

06:07.840 --> 06:15.520
 planning and learning macro operators. They had low level kind of configuration space planning for

06:15.520 --> 06:20.160
 the robot. They had vision. That's the basic ideas of a ton of things.

06:20.160 --> 06:27.920
 Can you take a step by it? Shaky was a mobile robot, but it could push objects,

06:27.920 --> 06:31.680
 and so it would move things around. With which actuator?

06:31.680 --> 06:40.080
 With its self, with its base. They had painted the baseboards black,

06:40.080 --> 06:48.320
 so it used vision to localize itself in a map. It detected objects. It could detect objects that

06:48.320 --> 06:54.800
 were surprising to it. It would plan and replan based on what it saw. It reasoned about whether

06:54.800 --> 07:02.240
 to look and take pictures. It really had the basics of so many of the things that we think about now.

07:03.280 --> 07:05.360
 How did it represent the space around it?

07:05.360 --> 07:09.680
 It had representations at a bunch of different levels of abstraction,

07:09.680 --> 07:13.920
 so it had, I think, a kind of an occupancy grid of some sort at the lowest level.

07:14.880 --> 07:20.000
 At the high level, it was abstract, symbolic kind of rooms and connectivity.

07:20.000 --> 07:22.160
 So where does Flakey come in?

07:22.160 --> 07:29.600
 Yeah, okay. I showed up at SRI and we were building a brand new robot. As I said, none of the people

07:29.600 --> 07:34.240
 from the previous project were there or involved anymore, so we were starting from scratch.

07:34.240 --> 07:43.920
 My advisor was Stan Rosenstein. He ended up being my thesis advisor. He was motivated by this idea

07:43.920 --> 07:52.400
 of situated computation or situated automata. The idea was that the tools of logical reasoning were

07:52.400 --> 08:01.200
 important, but possibly only for the engineers or designers to use in the analysis of a system,

08:01.200 --> 08:05.600
 but not necessarily to be manipulated in the head of the system itself.

08:06.400 --> 08:09.920
 So I might use logic to prove a theorem about the behavior of my robot,

08:10.480 --> 08:14.400
 even if the robot's not using logic, and it's had to prove theorems. So that was kind of the

08:14.400 --> 08:22.160
 distinction. And so the idea was to kind of use those principles to make a robot do stuff.

08:22.800 --> 08:28.960
 But a lot of the basic things we had to kind of learn for ourselves, because I had zero

08:28.960 --> 08:32.160
 background in robotics. I didn't know anything about control. I didn't know anything about

08:32.160 --> 08:36.640
 sensors. So we reinvented a lot of wheels on the way to getting that robot to do stuff.

08:36.640 --> 08:39.120
 Do you think that was an advantage or hindrance?

08:39.120 --> 08:45.600
 Oh, no. I'm big in favor of wheel reinvention, actually. I mean, I think you learned a lot

08:45.600 --> 08:51.920
 by doing it. It's important though to eventually have the pointers so that you can see what's

08:51.920 --> 08:58.080
 really going on. But I think you can appreciate much better the good solutions once you've

08:58.080 --> 09:00.400
 messed around a little bit on your own and found a bad one.

09:00.400 --> 09:04.880
 Yeah, I think you mentioned reinventing reinforcement learning and referring to

09:04.880 --> 09:10.960
 rewards as pleasures, a pleasure, I think, which I think is a nice name for it.

09:12.800 --> 09:18.960
 It's more fun, almost. Do you think you could tell the history of AI, machine learning,

09:18.960 --> 09:23.520
 reinforcement learning, how you think about it from the 50s to now?

09:23.520 --> 09:29.440
 One thing is that it oscillates. So things become fashionable and then they go out and

09:29.440 --> 09:34.480
 then something else becomes cool and then it goes out and so on. So there's some interesting

09:34.480 --> 09:41.600
 sociological process that actually drives a lot of what's going on. Early days was cybernetics and

09:41.600 --> 09:48.320
 control and the idea that of homeostasis, people who made these robots that could,

09:48.320 --> 09:54.400
 I don't know, try to plug into the wall when they needed power and then come loose and roll

09:54.400 --> 10:00.960
 around and do stuff. And then I think over time, they thought, well, that was inspiring, but people

10:00.960 --> 10:04.880
 said, no, no, no, we want to get maybe closer to what feels like real intelligence or human

10:04.880 --> 10:15.040
 intelligence. And then maybe the expert systems people tried to do that, but maybe a little

10:15.040 --> 10:21.760
 too superficially. So we get this surface understanding of what intelligence is like,

10:21.760 --> 10:25.840
 because I understand how a steel mill works and I can try to explain it to you and you can write

10:25.840 --> 10:31.520
 it down in logic and then we can make a computer infer that. And then that didn't work out.

10:32.400 --> 10:37.520
 But what's interesting, I think, is when a thing starts to not be working very well,

10:38.720 --> 10:44.480
 it's not only do we change methods, we change problems. So it's not like we have better ways

10:44.480 --> 10:48.160
 of doing the problem of the expert systems people are trying to do. We have no ways of

10:48.160 --> 10:56.800
 trying to do that problem. Oh, yeah, no, I think maybe a few. But we kind of give up on that problem

10:56.800 --> 11:01.520
 and we switch to a different problem. And we work that for a while and we make progress.

11:01.520 --> 11:04.960
 As a broad community. As a community. And there's a lot of people who would argue,

11:04.960 --> 11:09.760
 you don't give up on the problem. It's just the decrease in the number of people working on it.

11:09.760 --> 11:13.920
 You almost kind of like put it on the shelf. So we'll come back to this 20 years later.

11:13.920 --> 11:19.360
 Yeah, I think that's right. Or you might decide that it's malformed. Like you might say,

11:21.600 --> 11:26.800
 it's wrong to just try to make something that does superficial symbolic reasoning behave like a

11:26.800 --> 11:34.000
 doctor. You can't do that until you've had the sensory motor experience of being a doctor or

11:34.000 --> 11:38.560
 something. So there's arguments that say that that problem was not well formed. Or it could be

11:38.560 --> 11:44.160
 that it is well formed, but we just weren't approaching it well. So you mentioned that your

11:44.160 --> 11:49.120
 favorite part of logic and symbolic systems is that they give short names for large sets.

11:49.840 --> 11:56.320
 So there is some use to this. They use symbolic reasoning. So looking at expert systems

11:56.960 --> 12:01.760
 and symbolic computing, what do you think are the roadblocks that were hit in the 80s and 90s?

12:02.640 --> 12:08.320
 Okay, so right. So the fact that I'm not a fan of expert systems doesn't mean that I'm not a fan

12:08.320 --> 12:16.640
 of some kind of symbolic reasoning. So let's see roadblocks. Well, the main roadblock, I think,

12:16.640 --> 12:25.040
 was that the idea that humans could articulate their knowledge effectively into some kind of

12:25.040 --> 12:30.560
 logical statements. So it's not just the cost, the effort, but really just the capability of

12:30.560 --> 12:37.120
 doing it. Right. Because we're all experts in vision, but totally don't have introspective access

12:37.120 --> 12:45.440
 into how we do that. Right. And it's true that, I mean, I think the idea was, well, of course,

12:45.440 --> 12:49.040
 even people then would know, of course, I wouldn't ask you to please write down the rules that you

12:49.040 --> 12:54.000
 use for recognizing a water bottle. That's crazy. And everyone understood that. But we might ask

12:54.000 --> 13:00.800
 you to please write down the rules you use for deciding, I don't know what tie to put on or

13:00.800 --> 13:08.240
 or how to set up a microphone or something like that. But even those things, I think people maybe,

13:08.880 --> 13:12.720
 I think what they found, I'm not sure about this, but I think what they found was that the

13:12.720 --> 13:19.120
 so called experts could give explanations that sort of post hoc explanations for how and why

13:19.120 --> 13:27.680
 they did things, but they weren't necessarily very good. And then they depended on maybe some

13:27.680 --> 13:33.280
 kinds of perceptual things, which again, they couldn't really define very well. So I think,

13:33.280 --> 13:38.800
 I think fundamentally, I think that the underlying problem with that was the assumption that people

13:38.800 --> 13:45.280
 could articulate how and why they make their decisions. Right. So it's almost encoding the

13:45.280 --> 13:51.440
 knowledge from converting from expert to something that a machine can understand and reason with.

13:51.440 --> 13:58.880
 No, no, no, not even just encoding, but getting it out of you. Not not not writing it. I mean,

13:58.880 --> 14:03.680
 yes, hard also to write it down for the computer. But I don't think that people can

14:04.240 --> 14:10.080
 produce it. You can tell me a story about why you do stuff. But I'm not so sure that's the why.

14:11.440 --> 14:16.960
 Great. So there are still on the hierarchical planning side,

14:16.960 --> 14:25.120
 places where symbolic reasoning is very useful. So as you've talked about, so

14:27.840 --> 14:34.400
 where so don't where's the gap? Yeah, okay, good. So saying that humans can't provide a

14:34.400 --> 14:40.560
 description of their reasoning processes. That's okay, fine. But that doesn't mean that it's not

14:40.560 --> 14:44.880
 good to do reasoning of various styles inside a computer. Those are just two orthogonal points.

14:44.880 --> 14:50.560
 So then the question is, what kind of reasoning should you do inside a computer?

14:50.560 --> 14:55.680
 Right. And the answer is, I think you need to do all different kinds of reasoning inside

14:55.680 --> 15:01.680
 a computer, depending on what kinds of problems you face. I guess the question is, what kind of

15:01.680 --> 15:12.880
 things can you encode symbolically so you can reason about? I think the idea about and and

15:12.880 --> 15:18.080
 even symbolic, I don't even like that terminology because I don't know what it means technically

15:18.080 --> 15:24.240
 and formally. I do believe in abstractions. So abstractions are critical, right? You cannot

15:24.240 --> 15:30.240
 reason at completely fine grain about everything in your life, right? You can't make a plan at the

15:30.240 --> 15:37.680
 level of images and torques for getting a PhD. So you have to reduce the size of the state space

15:37.680 --> 15:43.040
 and you have to reduce the horizon if you're going to reason about getting a PhD or even buying

15:43.040 --> 15:50.080
 the ingredients to make dinner. And so how can you reduce the spaces and the horizon of the

15:50.080 --> 15:53.200
 reasoning you have to do? And the answer is abstraction, spatial abstraction, temporal

15:53.200 --> 15:58.080
 abstraction. I think abstraction along the lines of goals is also interesting, like you might

15:58.800 --> 16:03.840
 or well, abstraction and decomposition. Goals is maybe more of a decomposition thing.

16:03.840 --> 16:08.880
 So I think that's where these kinds of, if you want to call it symbolic or discrete

16:08.880 --> 16:15.440
 models come in. You talk about a room of your house instead of your pose. You talk about

16:16.800 --> 16:22.560
 doing something during the afternoon instead of at 2.54. And you do that because it makes

16:22.560 --> 16:30.000
 your reasoning problem easier and also because you don't have enough information

16:30.000 --> 16:37.120
 to reason in high fidelity about your pose of your elbow at 2.35 this afternoon anyway.

16:37.120 --> 16:39.440
 Right. When you're trying to get a PhD.

16:39.440 --> 16:41.600
 Right. Or when you're doing anything really.

16:41.600 --> 16:44.400
 Yeah, okay. Except for at that moment. At that moment,

16:44.400 --> 16:48.160
 you do have to reason about the pose of your elbow, maybe. But then maybe you do that in some

16:48.160 --> 16:55.680
 continuous joint space kind of model. And so again, my biggest point about all of this is that

16:55.680 --> 17:01.440
 there should be, the dogma is not the thing, right? It shouldn't be that I am in favor

17:01.440 --> 17:06.320
 against symbolic reasoning and you're in favor against neural networks. It should be that just

17:07.600 --> 17:12.240
 computer science tells us what the right answer to all these questions is if we were smart enough

17:12.240 --> 17:16.960
 to figure it out. Yeah. When you try to actually solve the problem with computers, the right answer

17:16.960 --> 17:22.880
 comes out. You mentioned abstractions. I mean, neural networks form abstractions or rather,

17:22.880 --> 17:30.320
 there's automated ways to form abstractions and there's expert driven ways to form abstractions

17:30.320 --> 17:35.920
 and expert human driven ways. And humans just seems to be way better at forming abstractions

17:35.920 --> 17:44.080
 currently and certain problems. So when you're referring to 2.45 a.m. versus afternoon,

17:44.960 --> 17:49.920
 how do we construct that taxonomy? Is there any room for automated construction of such

17:49.920 --> 17:55.200
 abstractions? Oh, I think eventually, yeah. I mean, I think when we get to be better

17:56.160 --> 18:02.240
 and machine learning engineers, we'll build algorithms that build awesome abstractions.

18:02.240 --> 18:06.720
 That are useful in this kind of way that you're describing. Yeah. So let's then step from

18:07.840 --> 18:14.400
 the abstraction discussion and let's talk about BOMMDP's

18:14.400 --> 18:21.440
 Partially Observable Markov Decision Processes. So uncertainty. So first, what are Markov Decision

18:21.440 --> 18:27.520
 Processes? What are Markov Decision Processes? Maybe how much of our world can be models and

18:27.520 --> 18:32.080
 MDPs? How much when you wake up in the morning and you're making breakfast, how do you think

18:32.080 --> 18:38.080
 of yourself as an MDP? So how do you think about MDPs and how they relate to our world?

18:38.080 --> 18:43.040
 Well, so there's a stance question, right? So a stance is a position that I take with

18:43.040 --> 18:52.160
 respect to a problem. So I as a researcher or a person who designed systems can decide to make

18:52.160 --> 18:58.960
 a model of the world around me in some terms. So I take this messy world and I say, I'm going to

18:58.960 --> 19:04.640
 treat it as if it were a problem of this formal kind, and then I can apply solution concepts

19:04.640 --> 19:09.120
 or algorithms or whatever to solve that formal thing, right? So of course, the world is not

19:09.120 --> 19:14.080
 anything. It's not an MDP or a POMDP. I don't know what it is, but I can model aspects of it

19:14.080 --> 19:19.280
 in some way or some other way. And when I model some aspect of it in a certain way, that gives me

19:19.280 --> 19:25.600
 some set of algorithms I can use. You can model the world in all kinds of ways. Some have some

19:26.400 --> 19:32.880
 are more accepting of uncertainty, more easily modeling uncertainty of the world. Some really

19:32.880 --> 19:40.720
 force the world to be deterministic. And so certainly MDPs model the uncertainty of the world.

19:40.720 --> 19:47.200
 Yes. Model some uncertainty. They model not present state uncertainty, but they model uncertainty

19:47.200 --> 19:53.840
 in the way the future will unfold. Right. So what are Markov decision processes?

19:53.840 --> 19:57.680
 So Markov decision process is a model. It's a kind of a model that you can make that says,

19:57.680 --> 20:05.600
 I know completely the current state of my system. And what it means to be a state is that I have

20:05.600 --> 20:10.720
 all the information right now that will let me make predictions about the future as well as I

20:10.720 --> 20:14.640
 can. So that remembering anything about my history wouldn't make my predictions any better.

20:18.720 --> 20:23.680
 But then it also says that then I can take some actions that might change the state of the world

20:23.680 --> 20:28.800
 and that I don't have a deterministic model of those changes. I have a probabilistic model

20:28.800 --> 20:35.600
 of how the world might change. It's a useful model for some kinds of systems. I mean, it's

20:35.600 --> 20:43.280
 certainly not a good model for most problems. I think because for most problems, you don't

20:43.280 --> 20:49.680
 actually know the state. For most problems, it's partially observed. So that's now a different

20:49.680 --> 20:56.480
 problem class. So okay, that's where the problem depies, the partially observed Markov decision

20:56.480 --> 21:03.600
 process step in. So how do they address the fact that you can't observe most the incomplete

21:03.600 --> 21:09.360
 information about most of the world around you? Right. So now the idea is we still kind of postulate

21:09.360 --> 21:14.080
 that there exists a state. We think that there is some information about the world out there

21:14.640 --> 21:18.800
 such that if we knew that we could make good predictions, but we don't know the state.

21:18.800 --> 21:23.840
 And so then we have to think about how, but we do get observations. Maybe I get images or I hear

21:23.840 --> 21:29.520
 things or I feel things and those might be local or noisy. And so therefore they don't tell me

21:29.520 --> 21:35.440
 everything about what's going on. And then I have to reason about given the history of actions

21:35.440 --> 21:40.000
 I've taken and observations I've gotten, what do I think is going on in the world? And then

21:40.000 --> 21:43.920
 given my own kind of uncertainty about what's going on in the world, I can decide what actions to

21:43.920 --> 21:51.120
 take. And so difficult is this problem of planning under uncertainty in your view and your

21:51.120 --> 21:57.840
 long experience of modeling the world, trying to deal with this uncertainty in

21:57.840 --> 22:04.240
 especially in real world systems. Optimal planning for even discrete POMDPs can be

22:04.240 --> 22:12.000
 undecidable depending on how you set it up. And so lots of people say I don't use POMDPs

22:12.000 --> 22:17.600
 because they are intractable. And I think that that's a kind of a very funny thing to say because

22:18.880 --> 22:23.120
 the problem you have to solve is the problem you have to solve. So if the problem you have to

22:23.120 --> 22:28.160
 solve is intractable, that's what makes us AI people, right? So we solve, we understand that

22:28.160 --> 22:34.320
 the problem we're solving is wildly intractable that we will never be able to solve it optimally,

22:34.320 --> 22:41.360
 at least I don't. Yeah, right. So later we can come back to an idea about bounded optimality

22:41.360 --> 22:44.960
 and something. But anyway, we can't come up with optimal solutions to these problems.

22:45.520 --> 22:51.200
 So we have to make approximations. Approximations in modeling approximations in solution algorithms

22:51.200 --> 22:58.160
 and so on. And so I don't have a problem with saying, yeah, my problem actually it is POMDP in

22:58.160 --> 23:02.880
 continuous space with continuous observations. And it's so computationally complex. I can't

23:02.880 --> 23:10.320
 even think about it's, you know, big O whatever. But that doesn't prevent me from it helps me

23:10.320 --> 23:17.360
 gives me some clarity to think about it that way. And to then take steps to make approximation

23:17.360 --> 23:22.080
 after approximation to get down to something that's like computable in some reasonable time.

23:22.080 --> 23:27.920
 When you think about optimality, you know, the community broadly has shifted on that, I think,

23:27.920 --> 23:35.600
 a little bit in how much they value the idea of optimality of chasing an optimal solution.

23:35.600 --> 23:42.240
 How is your views of chasing an optimal solution changed over the years when you work with robots?

23:42.240 --> 23:49.920
 That's interesting. I think we have a little bit of a methodological crisis, actually,

23:49.920 --> 23:54.000
 from the theoretical side. I mean, I do think that theory is important and that right now we're not

23:54.000 --> 24:00.640
 doing much of it. So there's lots of empirical hacking around and training this and doing that

24:00.640 --> 24:05.440
 and reporting numbers. But is it good? Is it bad? We don't know. It's very hard to say things.

24:08.240 --> 24:15.920
 And if you look at like computer science theory, so people talked for a while,

24:15.920 --> 24:21.280
 everyone was about solving problems optimally or completely. And then there were interesting

24:21.280 --> 24:27.520
 relaxations. So people look at, oh, can I, are there regret bounds? Or can I do some kind of,

24:27.520 --> 24:33.280
 you know, approximation? Can I prove something that I can approximately solve this problem or

24:33.280 --> 24:38.160
 that I get closer to the solution as I spend more time and so on? What's interesting, I think,

24:38.160 --> 24:47.680
 is that we don't have good approximate solution concepts for very difficult problems. Right?

24:47.680 --> 24:52.640
 I like to, you know, I like to say that I'm interested in doing a very bad job of very big

24:52.640 --> 25:02.960
 problems. Right. So very bad job, very big problems. I like to do that. But I wish I could say

25:02.960 --> 25:09.600
 something. I wish I had a, I don't know, some kind of a formal solution concept

25:10.320 --> 25:16.640
 that I could use to say, oh, this algorithm actually, it gives me something. Like, I know

25:16.640 --> 25:21.760
 what I'm going to get. I can do something other than just run it and get out. So that notion

25:21.760 --> 25:28.640
 is still somewhere deeply compelling to you. The notion that you can say, you can drop

25:28.640 --> 25:33.440
 thing on the table says this, you can expect this, this algorithm will give me some good results.

25:33.440 --> 25:38.960
 I hope there's, I hope science will, I mean, there's engineering and there's science,

25:38.960 --> 25:44.720
 I think that they're not exactly the same. And I think right now we're making huge engineering

25:45.600 --> 25:49.840
 like leaps and bounds. So the engineering is running away ahead of the science, which is cool.

25:49.840 --> 25:54.800
 And often how it goes, right? So we're making things and nobody knows how and why they work,

25:54.800 --> 26:03.200
 roughly. But we need to turn that into science. There's some form. It's, yeah,

26:03.200 --> 26:07.200
 there's some room for formalizing. We need to know what the principles are. Why does this work?

26:07.200 --> 26:12.480
 Why does that not work? I mean, for while people build bridges by trying, but now we can often

26:12.480 --> 26:17.520
 predict whether it's going to work or not without building it. Can we do that for learning systems

26:17.520 --> 26:23.600
 or for robots? See, your hope is from a materialistic perspective that intelligence,

26:23.600 --> 26:28.000
 artificial intelligence systems, robots are kind of just fancier bridges.

26:29.200 --> 26:33.600
 Belief space. What's the difference between belief space and state space? So we mentioned

26:33.600 --> 26:42.000
 MDPs, FOMDPs, you reasoning about, you sense the world, there's a state. What's this belief

26:42.000 --> 26:49.040
 space idea? Yeah. Okay, that sounds good. It sounds good. So belief space, that is, instead of

26:49.040 --> 26:54.880
 thinking about what's the state of the world and trying to control that as a robot, I think about

26:55.760 --> 27:01.120
 what is the space of beliefs that I could have about the world? What's, if I think of a belief

27:01.120 --> 27:06.640
 as a probability distribution of the ways the world could be, a belief state is a distribution,

27:06.640 --> 27:13.040
 and then my control problem, if I'm reasoning about how to move through a world I'm uncertain about,

27:14.160 --> 27:18.880
 my control problem is actually the problem of controlling my beliefs. So I think about taking

27:18.880 --> 27:23.120
 actions, not just what effect they'll have on the world outside, but what effect they'll have on my

27:23.120 --> 27:29.920
 own understanding of the world outside. And so that might compel me to ask a question or look

27:29.920 --> 27:35.280
 somewhere to gather information, which may not really change the world state, but it changes

27:35.280 --> 27:43.440
 my own belief about the world. That's a powerful way to empower the agent to reason about the

27:43.440 --> 27:47.840
 world, to explore the world. What kind of problems does it allow you to solve to

27:49.040 --> 27:54.560
 consider belief space versus just state space? Well, any problem that requires deliberate

27:54.560 --> 28:02.800
 information gathering. So if in some problems, like chess, there's no uncertainty, or maybe

28:02.800 --> 28:06.320
 there's uncertainty about the opponent. There's no uncertainty about the state.

28:08.400 --> 28:14.000
 And some problems, there's uncertainty, but you gather information as you go. You might say,

28:14.000 --> 28:18.240
 oh, I'm driving my autonomous car down the road, and it doesn't know perfectly where it is, but

28:18.240 --> 28:23.280
 the LiDARs are all going all the time. So I don't have to think about whether to gather information.

28:24.160 --> 28:28.800
 But if you're a human driving down the road, you sometimes look over your shoulder to see what's

28:28.800 --> 28:36.320
 going on behind you in the lane. And you have to decide whether you should do that now. And you

28:36.320 --> 28:40.400
 have to trade off the fact that you're not seeing in front of you, and you're looking behind you,

28:40.400 --> 28:45.440
 and how valuable is that information, and so on. And so to make choices about information

28:45.440 --> 28:56.080
 gathering, you have to reason in belief space. Also to just take into account your own uncertainty

28:56.080 --> 29:03.280
 before trying to do things. So you might say, if I understand where I'm standing relative to the

29:03.280 --> 29:08.880
 door jam, pretty accurately, then it's okay for me to go through the door. But if I'm really not

29:08.880 --> 29:14.240
 sure where the door is, then it might be better to not do that right now. The degree of your

29:14.240 --> 29:18.800
 uncertainty about the world is actually part of the thing you're trying to optimize in forming the

29:18.800 --> 29:26.560
 plan, right? So this idea of a long horizon of planning for a PhD or just even how to get out

29:26.560 --> 29:32.720
 of the house or how to make breakfast, you show this presentation of the WTF, where's the fork

29:33.360 --> 29:42.000
 of a robot looking to sink. And can you describe how we plan in this world is this idea of hierarchical

29:42.000 --> 29:52.000
 planning we've mentioned? Yeah, how can a robot hope to plan about something with such a long

29:52.000 --> 29:58.400
 horizon where the goal is quite far away? People since probably reasoning began have thought about

29:58.400 --> 30:02.560
 hierarchical reasoning, the temporal hierarchy in particular. Well, there's spatial hierarchy,

30:02.560 --> 30:06.240
 but let's talk about temporal hierarchy. So you might say, oh, I have this long

30:06.240 --> 30:13.680
 execution I have to do, but I can divide it into some segments abstractly, right? So maybe

30:14.400 --> 30:19.360
 have to get out of the house, I have to get in the car, I have to drive, and so on. And so

30:20.800 --> 30:25.920
 you can plan if you can build abstractions. So this we started out by talking about abstractions,

30:25.920 --> 30:30.080
 and we're back to that now. If you can build abstractions in your state space,

30:30.080 --> 30:37.760
 and abstractions, sort of temporal abstractions, then you can make plans at a high level. And you

30:37.760 --> 30:42.320
 can say, I'm going to go to town, and then I'll have to get gas, and I can go here, and I can do

30:42.320 --> 30:47.360
 this other thing. And you can reason about the dependencies and constraints among these actions,

30:47.920 --> 30:55.600
 again, without thinking about the complete details. What we do in our hierarchical planning work is

30:55.600 --> 31:00.960
 then say, all right, I make a plan at a high level of abstraction. I have to have some

31:00.960 --> 31:06.640
 reason to think that it's feasible without working it out in complete detail. And that's

31:06.640 --> 31:10.800
 actually the interesting step. I always like to talk about walking through an airport, like

31:12.160 --> 31:16.720
 you can plan to go to New York and arrive at the airport, and then find yourself in an office

31:16.720 --> 31:21.520
 building later. You can't even tell me in advance what your plan is for walking through the airport,

31:21.520 --> 31:26.320
 partly because you're too lazy to think about it maybe, but partly also because you just don't

31:26.320 --> 31:30.960
 have the information. You don't know what gate you're landing in or what people are going to be

31:30.960 --> 31:37.040
 in front of you or anything. So there's no point in planning in detail. But you have to have,

31:38.000 --> 31:43.760
 you have to make a leap of faith that you can figure it out once you get there. And it's really

31:43.760 --> 31:52.000
 interesting to me how you arrive at that. How do you, so you have learned over your lifetime to be

31:52.000 --> 31:56.800
 able to make some kinds of predictions about how hard it is to achieve some kinds of sub goals.

31:57.440 --> 32:01.440
 And that's critical. Like you would never plan to fly somewhere if you couldn't,

32:02.000 --> 32:05.200
 didn't have a model of how hard it was to do some of the intermediate steps.

32:05.200 --> 32:09.440
 So one of the things we're thinking about now is how do you do this kind of very aggressive

32:09.440 --> 32:16.400
 generalization to situations that you haven't been in and so on to predict how long will it

32:16.400 --> 32:20.400
 take to walk through the Kuala Lumpur airport? Like you could give me an estimate and it wouldn't

32:20.400 --> 32:26.800
 be crazy. And you have to have an estimate of that in order to make plans that involve

32:26.800 --> 32:30.080
 walking through the Kuala Lumpur airport, even if you don't need to know it in detail.

32:31.040 --> 32:35.520
 So I'm really interested in these kinds of abstract models and how do we acquire them.

32:35.520 --> 32:39.760
 But once we have them, we can use them to do hierarchical reasoning, which I think is very

32:39.760 --> 32:46.400
 important. Yeah, there's this notion of goal regression and preimage backchaining.

32:46.400 --> 32:53.760
 This idea of starting at the goal and just forming these big clouds of states. I mean,

32:54.560 --> 33:01.840
 it's almost like saying to the airport, you know, you know, once you show up to the airport,

33:01.840 --> 33:08.560
 you're like a few steps away from the goal. So thinking of it this way is kind of interesting.

33:08.560 --> 33:15.600
 I don't know if you have further comments on that of starting at the goal. Yeah, I mean,

33:15.600 --> 33:22.400
 it's interesting that Herb Simon back in the early days of AI talked a lot about

33:22.400 --> 33:26.960
 means ends reasoning and reasoning back from the goal. There's a kind of an intuition that people

33:26.960 --> 33:34.960
 have that the number of the state space is big, the number of actions you could take is really big.

33:35.760 --> 33:39.440
 So if you say, here I sit and I want to search forward from where I am, what are all the things

33:39.440 --> 33:45.040
 I could do? That's just overwhelming. If you say, if you can reason at this other level and say,

33:45.040 --> 33:49.520
 here's what I'm hoping to achieve, what can I do to make that true that somehow the

33:49.520 --> 33:54.000
 branching is smaller? Now, what's interesting is that like in the AI planning community,

33:54.000 --> 33:59.120
 that hasn't worked out in the class of problems that they look at and the methods that they tend

33:59.120 --> 34:04.400
 to use, it hasn't turned out that it's better to go backward. It's still kind of my intuition

34:04.400 --> 34:10.000
 that it is, but I can't prove that to you right now. Right. I share your intuition, at least for us

34:10.720 --> 34:19.920
 mirror humans. Speaking of which, when you maybe now we take it and take a little step into that

34:19.920 --> 34:27.280
 philosophy circle, how hard would it, when you think about human life, you give those examples

34:27.280 --> 34:32.400
 often, how hard do you think it is to formulate human life as a planning problem or aspects of

34:32.400 --> 34:37.600
 human life? So when you look at robots, you're often trying to think about object manipulation,

34:38.640 --> 34:46.240
 tasks about moving a thing. When you take a slight step outside the room, let the robot

34:46.240 --> 34:54.480
 leave and he'll get lunch or maybe try to pursue more fuzzy goals. How hard do you think is that

34:54.480 --> 35:00.720
 problem? If you were to try to maybe put another way, try to formulate human life as a planning

35:00.720 --> 35:05.680
 problem. Well, that would be a mistake. I mean, it's not all a planning problem, right? I think

35:05.680 --> 35:11.920
 it's really, really important that we understand that you have to put together pieces and parts

35:11.920 --> 35:18.640
 that have different styles of reasoning and representation and learning. I think it seems

35:18.640 --> 35:25.680
 probably clear to anybody that it can't all be this or all be that. Brains aren't all like this

35:25.680 --> 35:30.160
 or all like that, right? They have different pieces and parts and substructure and so on.

35:30.160 --> 35:34.400
 So I don't think that there's any good reason to think that there's going to be like one true

35:34.400 --> 35:39.600
 algorithmic thing that's going to do the whole job. Just a bunch of pieces together,

35:39.600 --> 35:48.160
 design to solve a bunch of specific problems. Or maybe styles of problems. I mean,

35:48.160 --> 35:52.880
 there's probably some reasoning that needs to go on in image space. I think, again,

35:55.840 --> 35:59.440
 there's this model base versus model free idea, right? So in reinforcement learning,

35:59.440 --> 36:06.000
 people talk about, oh, should I learn? I could learn a policy just straight up a way of behaving.

36:06.000 --> 36:11.360
 I could learn it's popular in a value function. That's some kind of weird intermediate ground.

36:13.360 --> 36:17.440
 Or I could learn a transition model, which tells me something about the dynamics of the world.

36:18.320 --> 36:22.560
 If I take a, imagine that I learn a transition model and I couple it with a planner and I

36:22.560 --> 36:29.520
 draw a box around that, I have a policy again. It's just stored a different way, right?

36:30.800 --> 36:34.560
 But it's just as much of a policy as the other policy. It's just I've made, I think,

36:34.560 --> 36:41.920
 the way I see it is it's a time space trade off in computation, right? A more overt policy

36:41.920 --> 36:47.680
 representation. Maybe it takes more space, but maybe I can compute quickly what action I should

36:47.680 --> 36:52.880
 take. On the other hand, maybe a very compact model of the world dynamics plus a planner

36:53.680 --> 36:58.240
 lets me compute what action to take to just more slowly. There's no, I mean, I don't think,

36:58.240 --> 37:04.240
 there's no argument to be had. It's just like a question of what form of computation is best

37:04.240 --> 37:12.720
 for us. For the various sub problems. Right. So, and so like learning to do algebra manipulations

37:12.720 --> 37:17.280
 for some reason is, I mean, that's probably going to want naturally a sort of a different

37:17.280 --> 37:22.640
 representation than riding a unicycle. At the time constraints on the unicycle are serious.

37:22.640 --> 37:28.640
 The space is maybe smaller. I don't know. But so I could be the more human size of

37:28.640 --> 37:36.240
 falling in love, having a relationship that might be another another style of no idea how to model

37:36.240 --> 37:43.280
 that. Yeah, that's, that's first solve the algebra and the object manipulation. What do you think

37:44.160 --> 37:50.480
 is harder perception or planning perception? That's why I'm understanding that's why.

37:51.920 --> 37:55.440
 So what do you think is so hard about perception by understanding the world around you?

37:55.440 --> 38:03.520
 Well, I mean, I think the big question is representational. A hugely the question is

38:03.520 --> 38:12.560
 representation. So perception has made great strides lately, right? And we can classify images and we

38:12.560 --> 38:17.760
 can play certain kinds of games and predict how to steer the car and all this sort of stuff.

38:17.760 --> 38:28.160
 I don't think we have a very good idea of what perception should deliver, right? So if you

38:28.160 --> 38:32.000
 if you believe in modularity, okay, there's there's a very strong view which says

38:34.640 --> 38:40.400
 we shouldn't build in any modularity, we should make a giant gigantic neural network,

38:40.400 --> 38:44.000
 train it end to end to do the thing. And that's the best way forward.

38:44.000 --> 38:51.280
 And it's hard to argue with that except on a sample complexity basis, right? So you might say,

38:51.280 --> 38:55.120
 oh, well, if I want to do end to end reinforcement learning on this giant giant neural network,

38:55.120 --> 39:00.800
 it's going to take a lot of data and a lot of like broken robots and stuff. So

39:02.640 --> 39:10.000
 then the only answer is to say, okay, we have to build something in build in some structure

39:10.000 --> 39:14.080
 or some bias, we know from theory of machine learning, the only way to cut down the sample

39:14.080 --> 39:20.320
 complexity is to kind of cut down somehow cut down the hypothesis space, you can do that by

39:20.320 --> 39:24.880
 building in bias. There's all kinds of reason to think that nature built bias into humans.

39:27.520 --> 39:32.800
 Convolution is a bias, right? It's a very strong bias and it's a very critical bias.

39:32.800 --> 39:39.520
 So my view is that we should look for more things that are like convolution, but that address other

39:39.520 --> 39:43.440
 aspects of reasoning, right? So convolution helps us a lot with a certain kind of spatial

39:43.440 --> 39:51.520
 reasoning that's quite close to the imaging. I think there's other ideas like that,

39:52.400 --> 39:58.240
 maybe some amount of forward search, maybe some notions of abstraction, maybe the notion that

39:58.240 --> 40:02.560
 objects exist, actually, I think that's pretty important. And a lot of people won't give you

40:02.560 --> 40:06.720
 that to start with, right? So almost like a convolution in the

40:08.640 --> 40:13.600
 in the object semantic object space or some kind of some kind of ideas in there. That's right.

40:13.600 --> 40:17.680
 And people are like the graph, graph convolutions are an idea that are related to

40:17.680 --> 40:26.240
 relational representations. And so I think there are, so you, I've come far field from perception,

40:26.240 --> 40:33.200
 but I think, I think the thing that's going to make perception that kind of the next step is

40:33.200 --> 40:38.000
 actually understanding better what it should produce, right? So what are we going to do with

40:38.000 --> 40:41.920
 the output of it, right? It's fine when what we're going to do with the output is steer,

40:41.920 --> 40:48.880
 it's less clear when we're just trying to make one integrated intelligent agent,

40:48.880 --> 40:53.520
 what should the output of perception be? We have no idea. And how should that hook up to the other

40:53.520 --> 41:00.240
 stuff? We don't know. So I think the pressing question is, what kinds of structure can we

41:00.240 --> 41:05.520
 build in that are like the moral equivalent of convolution that will make a really awesome

41:05.520 --> 41:10.240
 superstructure that then learning can kind of progress on efficiently?

41:10.240 --> 41:14.080
 I agree. Very compelling description of actually where we stand with the perception from

41:15.280 --> 41:19.120
 you're teaching a course on embodying intelligence. What do you think it takes to

41:19.120 --> 41:24.800
 build a robot with human level intelligence? I don't know if we knew we would do it.

41:27.680 --> 41:34.240
 If you were to, I mean, okay, so do you think a robot needs to have a self awareness,

41:36.000 --> 41:44.160
 consciousness, fear of mortality? Or is it, is it simpler than that? Or is consciousness a simple

41:44.160 --> 41:51.680
 thing? Do you think about these notions? I don't think much about consciousness. Even most philosophers

41:51.680 --> 41:56.560
 who care about it will give you that you could have robots that are zombies, right, that behave

41:56.560 --> 42:01.360
 like humans but are not conscious. And I, at this moment, would be happy enough for that. So I'm not

42:01.360 --> 42:06.240
 really worried one way or the other. So the technical side, you're not thinking of the use of self

42:06.240 --> 42:13.760
 awareness? Well, but I, okay, but then what does self awareness mean? I mean, that you need to have

42:13.760 --> 42:18.800
 some part of the system that can observe other parts of the system and tell whether they're

42:18.800 --> 42:24.560
 working well or not. That seems critical. So does that count as, I mean, does that count as

42:24.560 --> 42:30.560
 self awareness or not? Well, it depends on whether you think that there's somebody at home who can

42:30.560 --> 42:35.600
 articulate whether they're self aware. But clearly, if I have like, you know, some piece of code

42:35.600 --> 42:41.120
 that's counting how many times this procedure gets executed, that's a kind of self awareness,

42:41.120 --> 42:44.560
 right? So there's a big spectrum. It's clear you have to have some of it.

42:44.560 --> 42:48.800
 Right. You know, we're quite far away on many dimensions, but is the direction of research

42:49.600 --> 42:54.720
 that's most compelling to you for, you know, trying to achieve human level intelligence

42:54.720 --> 43:00.880
 in our robots? Well, to me, I guess the thing that seems most compelling to me at the moment is this

43:00.880 --> 43:10.320
 question of what to build in and what to learn. I think we're, we don't, we're missing a bunch of

43:10.320 --> 43:17.120
 ideas. And, and we, you know, people, you know, don't you dare ask me how many years it's going

43:17.120 --> 43:22.320
 to be until that happens, because I won't even participate in the conversation. Because I think

43:22.320 --> 43:26.240
 we're missing ideas and I don't know how long it's going to take to find them. So I won't ask you

43:26.240 --> 43:34.160
 how many years, but maybe I'll ask you what it, when you will be sufficiently impressed that we've

43:34.160 --> 43:41.280
 achieved it. So what's a good test of intelligence? Do you like the Turing test and natural language

43:41.280 --> 43:47.520
 in the robotic space? Is there something where you would sit back and think, oh, that's pretty

43:47.520 --> 43:52.800
 impressive as a test, as a benchmark. Do you think about these kinds of problems?

43:52.800 --> 43:58.480
 No, I resist. I mean, I think all the time that we spend arguing about those kinds of things could

43:58.480 --> 44:04.800
 be better spent just making their robots work better. So you don't value competition. So I mean,

44:04.800 --> 44:11.280
 there's a nature of benchmark, benchmarks and data sets, or Turing test challenges, where

44:11.280 --> 44:15.440
 everybody kind of gets together and tries to build a better robot because they want to outcompete

44:15.440 --> 44:21.360
 each other, like the DARPA challenge with the autonomous vehicles. Do you see the value of that?

44:23.600 --> 44:27.440
 Or can get in the way? I think you can get in the way. I mean, some people, many people find it

44:27.440 --> 44:34.880
 motivating. And so that's good. I find it anti motivating personally. But I think you get an

44:34.880 --> 44:41.200
 interesting cycle where for a contest, a bunch of smart people get super motivated and they hack

44:41.200 --> 44:47.120
 their brains out. And much of what gets done is just hacks, but sometimes really cool ideas emerge.

44:47.120 --> 44:53.360
 And then that gives us something to chew on after that. So it's not a thing for me, but I don't

44:53.360 --> 44:58.480
 I don't regret that other people do it. Yeah, it's like you said, with everything else that

44:58.480 --> 45:03.840
 makes us good. So jumping topics a little bit, you started the journal machine learning research

45:04.560 --> 45:09.920
 and served as its editor in chief. How did the publication come about?

45:11.680 --> 45:17.040
 And what do you think about the current publishing model space in machine learning

45:17.040 --> 45:23.040
 artificial intelligence? Okay, good. So it came about because there was a journal called machine

45:23.040 --> 45:29.840
 learning, which still exists, which was owned by Clure. And there was I was on the editorial

45:29.840 --> 45:33.520
 board and we used to have these meetings annually where we would complain to Clure that

45:33.520 --> 45:37.280
 it was too expensive for the libraries and that people couldn't publish. And we would really

45:37.280 --> 45:41.920
 like to have some kind of relief on those fronts. And they would always sympathize,

45:41.920 --> 45:49.120
 but not do anything. So we just decided to make a new journal. And there was the Journal of AI

45:49.120 --> 45:54.880
 Research, which has was on the same model, which had been in existence for maybe five years or so,

45:54.880 --> 46:01.920
 and it was going on pretty well. So we just made a new journal. It wasn't I mean,

46:03.600 --> 46:07.600
 I don't know, I guess it was work, but it wasn't that hard. So basically the editorial board,

46:07.600 --> 46:17.440
 probably 75% of the editorial board of machine learning resigned. And we founded the new journal.

46:17.440 --> 46:25.280
 But it was sort of it was more open. Yeah, right. So it's completely open. It's open access.

46:25.280 --> 46:31.600
 Actually, I had a postdoc, George Conrad Harris, who wanted to call these journals free for all.

46:33.520 --> 46:37.600
 Because there were I mean, it both has no page charges and has no

46:40.080 --> 46:45.520
 access restrictions. And the reason and so lots of people, I mean, for there were,

46:45.520 --> 46:50.240
 there were people who are mad about the existence of this journal who thought it was a fraud or

46:50.240 --> 46:55.200
 something, it would be impossible, they said, to run a journal like this with basically,

46:55.200 --> 46:59.200
 I mean, for a long time, I didn't even have a bank account. I paid for the

46:59.840 --> 47:06.640
 lawyer to incorporate and the IP address. And it just didn't cost a couple hundred dollars a year

47:06.640 --> 47:12.880
 to run. It's a little bit more now, but not that much more. But that's because I think computer

47:12.880 --> 47:19.920
 scientists are competent and autonomous in a way that many scientists in other fields aren't.

47:19.920 --> 47:23.920
 I mean, at doing these kinds of things, we already type set around papers,

47:23.920 --> 47:28.000
 we all have students and people who can hack a website together in the afternoon.

47:28.000 --> 47:32.960
 So the infrastructure for us was like, not a problem, but for other people in other fields,

47:32.960 --> 47:38.960
 it's a harder thing to do. Yeah. And this kind of open access journal is nevertheless,

47:38.960 --> 47:45.840
 one of the most prestigious journals. So it's not like a prestige and it can be achieved

47:45.840 --> 47:49.920
 without any of the papers. Paper is not required for prestige, turns out. Yeah.

47:50.640 --> 47:56.960
 So on the review process side of actually a long time ago, I don't remember when I reviewed a paper

47:56.960 --> 48:01.360
 where you were also a reviewer and I remember reading your review and being influenced by it.

48:01.360 --> 48:06.480
 It was really well written. It influenced how I write feature reviews. You disagreed with me,

48:06.480 --> 48:15.280
 actually. And you made it my review, but much better. But nevertheless, the review process

48:16.880 --> 48:23.600
 has its flaws. And what do you think works well? How can it be improved?

48:23.600 --> 48:27.600
 So actually, when I started JMLR, I wanted to do something completely different.

48:28.720 --> 48:34.800
 And I didn't because it felt like we needed a traditional journal of record and so we just

48:34.800 --> 48:40.800
 made JMLR be almost like a normal journal, except for the open access parts of it, basically.

48:43.200 --> 48:47.600
 Increasingly, of course, publication is not even a sensible word. You can publish something by

48:47.600 --> 48:54.400
 putting it in an archive so I can publish everything tomorrow. So making stuff public is

48:54.400 --> 49:04.800
 there's no barrier. We still need curation and evaluation. I don't have time to read all of

49:04.800 --> 49:20.480
 archive. And you could argue that kind of social thumbs uping of articles suffices, right? You

49:20.480 --> 49:25.440
 might say, oh, heck with this, we don't need journals at all. We'll put everything on archive

49:25.440 --> 49:30.400
 and people will upvote and downvote the articles and then your CV will say, oh, man, he got a lot

49:30.400 --> 49:44.000
 of upvotes. So that's good. But I think there's still value in careful reading and commentary of

49:44.000 --> 49:48.480
 things. And it's hard to tell when people are upvoting and downvoting or arguing about your

49:48.480 --> 49:55.440
 paper on Twitter and Reddit, whether they know what they're talking about. So then I have the

49:55.440 --> 50:01.360
 second order problem of trying to decide whose opinions I should value and such. So I don't

50:01.360 --> 50:06.240
 know. If I had infinite time, which I don't, and I'm not going to do this because I really want to

50:06.240 --> 50:11.920
 make robots work, but if I felt inclined to do something more in a publication direction,

50:12.880 --> 50:16.160
 I would do this other thing, which I thought about doing the first time, which is to get

50:16.160 --> 50:22.480
 together some set of people whose opinions I value and who are pretty articulate. And I guess we

50:22.480 --> 50:27.520
 would be public, although we could be private, I'm not sure. And we would review papers. We wouldn't

50:27.520 --> 50:31.600
 publish them and you wouldn't submit them. We would just find papers and we would write reviews

50:32.720 --> 50:39.120
 and we would make those reviews public. And maybe if you, you know, so we're Leslie's friends who

50:39.120 --> 50:45.200
 review papers and maybe eventually if we, our opinion was sufficiently valued, like the opinion

50:45.200 --> 50:50.800
 of JMLR is valued, then you'd say on your CV that Leslie's friends gave my paper a five star reading

50:50.800 --> 50:58.800
 and that would be just as good as saying I got it accepted into this journal. So I think we

50:58.800 --> 51:04.800
 should have good public commentary and organize it in some way, but I don't really know how to

51:04.800 --> 51:09.120
 do it. It's interesting times. The way you describe it actually is really interesting. I mean,

51:09.120 --> 51:15.040
 we do it for movies, IMDB.com. There's experts, critics come in, they write reviews, but there's

51:15.040 --> 51:20.960
 also regular non critics humans write reviews and they're separated. I like open review.

51:22.240 --> 51:31.600
 The eye clear process, I think is interesting. It's a step in the right direction, but it's still

51:31.600 --> 51:39.840
 not as compelling as reviewing movies or video games. I mean, it sometimes almost, it might be

51:39.840 --> 51:44.400
 silly, at least from my perspective to say, but it boils down to the user interface, how fun and

51:44.400 --> 51:50.400
 easy it is to actually perform the reviews, how efficient, how much you as a reviewer get

51:51.200 --> 51:56.560
 street cred for being a good reviewer. Those human elements come into play.

51:57.200 --> 52:03.600
 No, it's a big investment to do a good review of a paper and the flood of papers is out of control.

52:05.280 --> 52:08.960
 There aren't 3,000 new, I don't know how many new movies are there in a year, I don't know,

52:08.960 --> 52:15.440
 but there's probably going to be less than how many machine learning papers there are in a year now.

52:19.840 --> 52:22.320
 Right, so I'm like an old person, so of course I'm going to say,

52:23.520 --> 52:30.240
 things are moving too fast, I'm a stick in the mud. So I can say that, but my particular flavor

52:30.240 --> 52:38.240
 of that is, I think the horizon for researchers has gotten very short, that students want to

52:38.240 --> 52:46.000
 publish a lot of papers and it's exciting and there's value in that and you get padded on the

52:46.000 --> 52:58.320
 head for it and so on. And some of that is fine, but I'm worried that we're driving out people who

52:58.320 --> 53:05.280
 would spend two years thinking about something. Back in my day, when we worked on our theses,

53:05.280 --> 53:10.560
 we did not publish papers, you did your thesis for years, you picked a hard problem and then you

53:10.560 --> 53:16.320
 worked and chewed on it and did stuff and wasted time and for a long time. And when it was roughly,

53:16.320 --> 53:22.800
 when it was done, you would write papers. And so I don't know how to, and I don't think that

53:22.800 --> 53:26.720
 everybody has to work in that mode, but I think there's some problems that are hard enough

53:27.680 --> 53:31.680
 that it's important to have a longer research horizon and I'm worried that

53:31.680 --> 53:39.600
 we don't incentivize that at all at this point. In this current structure. So what do you see

53:41.440 --> 53:47.280
 what are your hopes and fears about the future of AI and continuing on this theme? So AI has

53:47.280 --> 53:53.440
 gone through a few winters, ups and downs. Do you see another winter of AI coming?

53:53.440 --> 54:02.480
 Or are you more hopeful about making robots work, as you said? I think the cycles are inevitable,

54:03.040 --> 54:10.080
 but I think each time we get higher, right? I mean, it's like climbing some kind of

54:10.080 --> 54:19.600
 landscape with a noisy optimizer. So it's clear that the deep learning stuff has

54:19.600 --> 54:25.760
 made deep and important improvements. And so the high watermark is now higher. There's no question.

54:25.760 --> 54:34.400
 But of course, I think people are overselling and eventually investors, I guess, and other people

54:34.400 --> 54:40.640
 look around and say, well, you're not quite delivering on this grand claim and that wild

54:40.640 --> 54:47.680
 hypothesis. It's like probably it's going to crash something out and then it's okay. I mean,

54:47.680 --> 54:54.000
 it's okay. I mean, but I don't I can't imagine that there's like some awesome monotonic improvement

54:54.000 --> 55:01.760
 from here to human level AI. So in, you know, I have to ask this question, I probably anticipate

55:01.760 --> 55:09.120
 answers, the answers. But do you have a worry short term, a long term about the existential

55:09.120 --> 55:18.880
 threats of AI and maybe short term, less existential, but more robots taking away jobs?

55:20.480 --> 55:28.000
 Well, actually, let me talk a little bit about utility. Actually, I had an interesting conversation

55:28.000 --> 55:32.480
 with some military ethicists who wanted to talk to me about autonomous weapons.

55:32.480 --> 55:39.360
 And they're, they were interesting, smart, well educated guys who didn't know too much about AI or

55:39.360 --> 55:43.600
 machine learning. And the first question they asked me was, has your robot ever done something you

55:43.600 --> 55:49.120
 didn't expect? And I like burst out laughing because anybody who's ever done something other robot

55:49.120 --> 55:54.720
 right knows that they don't do much. And what I realized was that their model of how we program

55:54.720 --> 55:59.440
 a robot was completely wrong. Their model of how we could put program robot was like,

55:59.440 --> 56:05.600
 program robot was like, Lego Mindstorms, like, Oh, go forward a meter, turn left, take a picture,

56:05.600 --> 56:11.120
 do this, do that. And so if you have that model of programming, then it's true, it's kind of weird

56:11.120 --> 56:16.240
 that your robot would do something that you didn't anticipate. But the fact is, and actually,

56:16.240 --> 56:21.840
 so now this is my new educational mission, if I have to talk to non experts, I try to teach them

56:22.720 --> 56:28.080
 the idea that we don't operate, we operate at least one or maybe many levels of abstraction

56:28.080 --> 56:33.280
 about that. And we say, Oh, here's a hypothesis class, maybe it's a space of plans, or maybe it's a

56:33.280 --> 56:38.400
 space of classifiers, or whatever. But there's some set of answers and an objective function. And

56:38.400 --> 56:44.800
 then we work on some optimization method that tries to optimize a solution in that class.

56:46.080 --> 56:50.560
 And we don't know what solution is going to come out. Right. So I think it's important to

56:50.560 --> 56:55.520
 communicate that. So I mean, of course, probably people who listen to this, they know that lesson.

56:55.520 --> 56:59.600
 But I think it's really critical to communicate that lesson. And then lots of people are now

56:59.600 --> 57:05.600
 talking about, you know, the value alignment problem. So you want to be sure, as robots or

57:06.480 --> 57:11.280
 software systems get more competent, that their objectives are aligned with your objectives,

57:11.280 --> 57:17.680
 or that our objectives are compatible in some way, or we have a good way of mediating when they have

57:17.680 --> 57:22.240
 different objectives. And so I think it is important to start thinking in terms, like,

57:22.240 --> 57:28.480
 you don't have to be freaked out by the robot apocalypse to accept that it's important to think

57:28.480 --> 57:33.760
 about objective functions of value alignment. And that you have to really, everyone who's done

57:33.760 --> 57:38.160
 optimization knows that you have to be careful what you wish for that, you know, sometimes you get

57:38.160 --> 57:45.280
 the optimal solution. And you realize, man, that was that objective was wrong. So pragmatically,

57:45.280 --> 57:51.360
 in the shortest term, it seems to me that that those are really interesting and critical questions.

57:51.360 --> 57:55.680
 And the idea that we're going to go from being people who engineer algorithms to being people

57:55.680 --> 58:00.800
 who engineer objective functions, I think that's, that's definitely going to happen. And that's

58:00.800 --> 58:03.360
 going to change our thinking and methodology and stuff.

58:03.360 --> 58:07.520
 We're going to, you started at Stanford philosophy, that's wish you could be science,

58:07.520 --> 58:13.840
 and I will go back to philosophy maybe. Well, I mean, they're mixed together because, because,

58:13.840 --> 58:18.240
 as we also know, as machine learning people, right? When you design, in fact, this is the

58:18.240 --> 58:23.360
 lecture I gave in class today, when you design an objective function, you have to wear both hats.

58:23.360 --> 58:28.320
 There's the hat that says, what do I want? And there's the hat that says, but I know what my

58:28.320 --> 58:34.240
 optimizer can do to some degree. And I have to take that into account. So it's, it's always a

58:34.240 --> 58:40.480
 trade off. And we have to kind of be mindful of that. The part about taking people's jobs,

58:40.480 --> 58:47.360
 that I understand that that's important, I don't understand sociology or economics or people

58:47.360 --> 58:51.840
 very well. So I don't know how to think about that. So that's, yeah, so there might be a

58:51.840 --> 58:56.640
 sociological aspect there, the economic aspect that's very difficult to think about. Okay.

58:56.640 --> 59:00.000
 I mean, I think other people should be thinking about it, but I'm just, that's not my strength.

59:00.000 --> 59:04.320
 So what do you think is the most exciting area of research in the short term,

59:04.320 --> 59:08.560
 for the community and for your, for yourself? Well, so, I mean, there's this story I've been

59:08.560 --> 59:16.480
 telling about how to engineer intelligent robots. So that's what we want to do. We all kind of want

59:16.480 --> 59:20.960
 to do, well, I mean, some set of us want to do this. And the question is, what's the most effective

59:20.960 --> 59:25.840
 strategy? And we've tried, and there's a bunch of different things you could do at the extremes,

59:25.840 --> 59:32.000
 right? One super extreme is we do introspection and we write a program. Okay, that has not worked

59:32.000 --> 59:37.360
 out very well. Another extreme is we take a giant bunch of neural guru and we try and train it up to

59:37.360 --> 59:43.040
 do something. I don't think that's going to work either. So the question is, what's the middle

59:43.040 --> 59:49.840
 ground? And again, this isn't a theological question or anything like that. It's just,

59:49.840 --> 59:57.040
 like, how do, just how do we, what's the best way to make this work out? And I think it's clear,

59:57.040 --> 1:00:01.840
 it's a combination of learning, to me, it's clear, it's a combination of learning and not learning.

1:00:02.400 --> 1:00:05.920
 And what should that combination be? And what's the stuff we build in? So to me,

1:00:05.920 --> 1:00:10.080
 that's the most compelling question. And when you say engineer robots, you mean

1:00:10.080 --> 1:00:15.600
 engineering systems that work in the real world. That's the emphasis.

1:00:17.600 --> 1:00:23.200
 Last question, which robots or robot is your favorite from science fiction?

1:00:24.480 --> 1:00:32.960
 So you can go with Star Wars or RTD2, or you can go with more modern, maybe Hal.

1:00:32.960 --> 1:00:37.040
 No, sir, I don't think I have a favorite robot from science fiction.

1:00:37.040 --> 1:00:45.520
 This is, this is back to, you like to make robots work in the real world here, not, not in.

1:00:45.520 --> 1:00:50.000
 I mean, I love the process. And I care more about the process.

1:00:50.000 --> 1:00:51.040
 The engineering process.

1:00:51.600 --> 1:00:55.760
 Yeah. I mean, I do research because it's fun, not because I care about what we produce.

1:00:57.520 --> 1:01:01.920
 Well, that's, that's a beautiful note, actually. And Leslie, thank you so much for talking today.

1:01:01.920 --> 1:01:07.920
 Sure, it's been fun.