Peter Norvig: Artificial Intelligence: A Modern Approach #42

Transcript

00:00:00 The following is a conversation with Peter Norvig.

00:00:02 He’s the Director of Research at Google

00:00:05 and the coauthor with Stuart Russell of the book

00:00:07 Artificial Intelligence, A Modern Approach,

00:00:10 that educated and inspired a whole generation

00:00:13 of researchers, including myself,

00:00:15 to get into the field of artificial intelligence.

00:00:18 This is the Artificial Intelligence Podcast.

00:00:21 If you enjoy it, subscribe on YouTube,

00:00:24 give five stars on iTunes, support on Patreon,

00:00:27 or simply connect with me on Twitter.

00:00:29 I’m Lex Friedman, spelled F R I D M A N.

00:00:32 And now, here’s my conversation with Peter Norvig.

00:00:37 Most researchers in the AI community, including myself,

00:00:40 own all three editions, red, green, and blue,

00:00:43 of the Artificial Intelligence, A Modern Approach.

00:00:46 It’s a field defining textbook, as many people are aware,

00:00:49 that you wrote with Stuart Russell.

00:00:52 How has the book changed and how have you changed

00:00:55 in relation to it from the first edition

00:00:57 to the second to the third and now fourth edition

00:01:00 as you work on it?

00:01:00 Yeah, so it’s been a lot of years, a lot of changes.

00:01:04 One of the things changing from the first

00:01:05 to maybe the second or third

00:01:09 was just the rise of computing power, right?

00:01:12 So I think in the first edition, we said,

00:01:17 here’s predicate logic, but that only goes so far

00:01:22 because pretty soon you have millions of short little

00:01:27 predicate expressions and they can possibly fit in memory.

00:01:31 So we’re gonna use first order logic that’s more concise.

00:01:35 And then we quickly realized,

00:01:38 oh, predicate logic is pretty nice

00:01:40 because there are really fast SAT solvers and other things.

00:01:44 And look, there’s only millions of expressions

00:01:46 and that fits easily into memory,

00:01:48 or maybe even billions fit into memory now.

00:01:51 So that was a change of the type of technology we needed

00:01:54 just because the hardware expanded.

00:01:56 Even to the second edition,

00:01:58 resource constraints were loosened significantly

00:02:00 for the second.

00:02:01 And that was early 2000s second edition.

00:02:04 Right, so 95 was the first and then 2000, 2001 or so.

00:02:10 And then moving on from there,

00:02:12 I think we’re starting to see that again with the GPUs

00:02:17 and then more specific type of machinery

00:02:20 like the TPUs and you’re seeing custom ASICs and so on

00:02:25 for deep learning.

00:02:26 So we’re seeing another advance in terms of the hardware.

00:02:30 Then I think another thing that we especially noticed

00:02:33 this time around is in all three of the first editions,

00:02:37 we kind of said, well, we’re gonna find AI

00:02:40 as maximizing expected utility

00:02:43 and you tell me your utility function.

00:02:45 And now we’ve got 27 chapters where the cool techniques

00:02:49 for how to optimize that.

00:02:51 I think in this edition, we’re saying more,

00:02:54 you know what, maybe that optimization part

00:02:56 is the easy part and the hard part is deciding

00:02:59 what is my utility function?

00:03:01 What do I want?

00:03:03 And if I’m a collection of agents or a society,

00:03:06 what do we want as a whole?

00:03:08 So you touched that topic in this edition.

00:03:10 You get a little bit more into utility.

00:03:11 Yeah.

00:03:12 That’s really interesting.

00:03:13 On a technical level,

00:03:15 we’re almost pushing the philosophical.

00:03:17 I guess it is philosophical, right?

00:03:19 So we’ve always had a philosophy chapter,

00:03:21 which I was glad that we were supporting.

00:03:27 And now it’s less kind of the Chinese room type argument

00:03:33 and more of these ethical and societal type issues.

00:03:37 So we get into the issues of fairness and bias

00:03:41 and just the issue of aggregating utilities.

00:03:45 So how do you encode human values into a utility function?

00:03:49 Is this something that you can do purely through data

00:03:53 in a learned way or is there some systematic,

00:03:56 obviously there’s no good answers yet.

00:03:58 There’s just beginnings to this,

00:04:01 to even opening the doors to these questions.

00:04:02 So there is no one answer.

00:04:04 Yes, there are techniques to try to learn that.

00:04:07 So we talk about inverse reinforcement learning, right?

00:04:10 So reinforcement learning, you take some actions,

00:04:14 you get some rewards and you figure out

00:04:16 what actions you should take.

00:04:18 And inverse reinforcement learning,

00:04:20 you observe somebody taking actions and you figure out,

00:04:24 well, this must be what they were trying to do.

00:04:27 If they did this action, it must be because they want it.

00:04:30 Of course, there’s restrictions to that, right?

00:04:33 So lots of people take actions that are self destructive

00:04:37 or they’re suboptimal in certain ways.

00:04:39 So you don’t wanna learn that.

00:04:40 You wanna somehow learn the perfect actions

00:04:44 rather than the ones they actually take.

00:04:46 So that’s a challenge for that field.

00:04:51 Then another big part of it is just kind of theoretical

00:04:55 of saying, what can we accomplish?

00:04:58 And so you look at like this work on the programs

00:05:04 to predict recidivism and decide who should get parole

00:05:09 or who should get bail or whatever.

00:05:12 And how are you gonna evaluate that?

00:05:13 And one of the big issues is fairness

00:05:16 across protected classes.

00:05:18 Protected classes being things like sex and race and so on.

00:05:23 And so two things you want is you wanna say,

00:05:27 well, if I get a score of say six out of 10,

00:05:32 then I want that to mean the same

00:05:34 whether no matter what race I’m on, right?

00:05:37 Yes, right, so I wanna have a 60% chance

00:05:39 of reoccurring regardless.

00:05:44 And one of the makers of a commercial program to do that

00:05:48 says that’s what we’re trying to optimize

00:05:50 and look, we achieved that.

00:05:51 We’ve reached that kind of balance.

00:05:56 And then on the other side,

00:05:57 you also wanna say, well, if it makes mistakes,

00:06:01 I want that to affect both sides

00:06:04 of the protected class equally.

00:06:07 And it turns out they don’t do that, right?

00:06:09 So they’re twice as likely to make a mistake

00:06:12 that would harm a black person over a white person.

00:06:14 So that seems unfair.

00:06:16 So you’d like to say,

00:06:17 well, I wanna achieve both those goals.

00:06:19 And then it turns out you do the analysis

00:06:21 and it’s theoretically impossible

00:06:22 to achieve both those goals.

00:06:24 So you have to trade them off one against the other.

00:06:27 So that analysis is really helpful

00:06:29 to know what you can aim for and how much you can get.

00:06:32 You can’t have everything.

00:06:33 But the analysis certainly can’t tell you

00:06:35 where should we make that trade off point.

00:06:38 But nevertheless, then we can as humans deliberate

00:06:41 where that trade off should be.

00:06:43 Yeah, so at least we now we’re arguing in an informed way.

00:06:45 We’re not asking for something impossible.

00:06:48 We’re saying, here’s where we are

00:06:50 and here’s what we aim for.

00:06:51 And this strategy is better than that strategy.

00:06:55 So that’s, I would argue is a really powerful

00:06:58 and really important first step,

00:07:00 but it’s a doable one sort of removing

00:07:02 undesirable degrees of bias in systems

00:07:07 in terms of protected classes.

00:07:08 And then there’s something I listened

00:07:10 to your commencement speech,

00:07:12 or there’s some fuzzier things like,

00:07:15 you mentioned angry birds.

00:07:17 Do you wanna create systems that feed the dopamine enjoyment

00:07:23 that feed, that optimize for you returning to the system,

00:07:26 enjoying the moment of playing the game of getting likes

00:07:30 or whatever, this kind of thing,

00:07:32 or some kind of longterm improvement?

00:07:34 Right.

00:07:36 Are you even thinking about that?

00:07:39 That’s really going to the philosophical area.

00:07:43 No, I think that’s a really important issue too.

00:07:45 Certainly thinking about that.

00:07:46 I don’t think about that as an AI issue as much.

00:07:52 But as you say, the point is we’ve built this society

00:07:57 and this infrastructure where we say we have a marketplace

00:08:02 for attention and we’ve decided as a society

00:08:07 that we like things that are free.

00:08:09 And so we want all the apps on our phone to be free.

00:08:13 And that means they’re all competing for your attention.

00:08:15 And then eventually they make some money some way

00:08:17 through ads or in game sales or whatever.

00:08:22 But they can only win by defeating all the other apps

00:08:26 by instilling your attention.

00:08:28 And we build a marketplace where it seems like

00:08:34 they’re working against you rather than working with you.

00:08:38 And I’d like to find a way where we can change

00:08:41 the playing field so you feel more like,

00:08:43 well, these things are on my side.

00:08:46 Yes, they’re letting me have some fun in the short term,

00:08:49 but they’re also helping me in the long term

00:08:52 rather than competing against me.

00:08:54 And those aren’t necessarily conflicting objectives.

00:08:56 They’re just the incentives, the direct current incentives

00:09:00 as we try to figure out this whole new world

00:09:02 seem to be on the easier part of that,

00:09:06 which is feeding the dopamine, the rush.

00:09:08 Right.

00:09:09 But so maybe taking a quick step back at the beginning

00:09:15 of the Artificial Intelligence,

00:09:17 the Modern Approach book of writing.

00:09:19 So here you are in the 90s.

00:09:21 When you first sat down with Stuart to write the book

00:09:25 to cover an entire field,

00:09:27 which is one of the only books that’s successfully done that

00:09:30 for AI and actually in a lot of other computer science

00:09:33 fields, it’s a huge undertaking.

00:09:37 So it must’ve been quite daunting.

00:09:40 What was that process like?

00:09:42 Did you envision that you would be trying to cover

00:09:44 the entire field?

00:09:47 Was there a systematic approach to it

00:09:48 that was more step by step?

00:09:50 How was, how did it feel?

00:09:52 So I guess it came about,

00:09:54 go to lunch with the other AI faculty at Berkeley

00:09:57 and we’d say, the field is changing.

00:10:00 It seems like the current books are a little bit behind.

00:10:03 Nobody’s come out with a new book recently.

00:10:05 We should do that.

00:10:06 And everybody said, yeah, yeah, that’s a great thing to do.

00:10:09 And we never did anything.

00:10:10 Right.

00:10:11 And then I ended up heading off to industry.

00:10:14 I went to Sun Labs.

00:10:16 So I thought, well, that’s the end of my possible

00:10:19 academic publishing career.

00:10:21 But I met Stuart again at a conference like a year later

00:10:25 and said, you know that book we were always talking about,

00:10:28 you guys must be half done with it by now, right?

00:10:30 And he said, well, we keep talking, we never do anything.

00:10:34 So I said, well, you know, we should do it.

00:10:36 And I think the reason is that we all felt

00:10:40 it was a time where the field was changing.

00:10:44 And that was in two ways.

00:10:46 So, you know, the good old fashioned AI

00:10:49 was based primarily on Boolean logic.

00:10:52 And you had a few tricks to deal with uncertainty.

00:10:55 And it was based primarily on knowledge engineering.

00:10:59 That the way you got something done is you went out,

00:11:00 you interviewed an expert and you wrote down by hand

00:11:03 everything they knew.

00:11:05 And we saw in 95 that the field was changing in two ways.

00:11:10 One, we’re moving more towards probability

00:11:13 rather than Boolean logic.

00:11:15 And we’re moving more towards machine learning

00:11:17 rather than knowledge engineering.

00:11:20 And the other books hadn’t caught that way

00:11:22 if they were still in the, more in the old school.

00:11:26 Although, so certainly they had part of that on the way.

00:11:29 But we said, if we start now completely taking

00:11:33 that point of view, we can have a different kind of book.

00:11:36 And we were able to put that together.

00:11:39 And what was literally the process if you remember,

00:11:44 did you start writing a chapter?

00:11:46 Did you outline?

00:11:48 Yeah, I guess we did an outline

00:11:50 and then we sort of assigned chapters to each person.

00:11:55 At the time I had moved to Boston

00:11:58 and Stuart was in Berkeley.

00:12:00 So basically we did it over the internet.

00:12:04 And, you know, that wasn’t the same as doing it today.

00:12:08 It meant, you know, dial up lines and telnetting in.

00:12:13 And, you know, you telnet it into one shell

00:12:19 and you type cat file name

00:12:21 and you hoped it was captured at the other end.

00:12:23 And certainly you’re not sending images

00:12:26 and figures back and forth.

00:12:27 Right, right, that didn’t work.

00:12:29 But, you know, did you anticipate

00:12:31 where the field would go from that day, from the 90s?

00:12:37 Did you see the growth into learning based methods

00:12:42 and to data driven methods

00:12:44 that followed in the future decades?

00:12:47 We certainly thought that learning was important.

00:12:51 I guess we missed it as being as important as it is today.

00:12:58 We missed this idea of big data.

00:13:00 We missed that the idea of deep learning

00:13:02 hadn’t been invented yet.

00:13:04 We could have taken the book

00:13:07 from a complete machine learning point of view

00:13:11 right from the start.

00:13:12 We chose to do it more from a point of view

00:13:15 of we’re gonna first develop

00:13:16 different types of representations.

00:13:19 And we’re gonna talk about different types of environments.

00:13:24 Is it fully observable or partially observable?

00:13:26 And is it deterministic or stochastic and so on?

00:13:29 And we made it more complex along those axes

00:13:33 rather than focusing on the machine learning axis first.

00:13:38 Do you think, you know, there’s some sense

00:13:40 in which the deep learning craze is extremely successful

00:13:44 for a particular set of problems.

00:13:46 And, you know, eventually it’s going to,

00:13:49 in the general case, hit challenges.

00:13:52 So in terms of the difference between perception systems

00:13:56 and robots that have to act in the world,

00:13:59 do you think we’re gonna return

00:14:01 to AI modern approach type breadth

00:14:06 in addition five and six?

00:14:08 In future decades, do you think deep learning

00:14:12 will take its place as a chapter

00:14:14 in this bigger view of AI?

00:14:17 Yeah, I think we don’t know yet

00:14:19 how it’s all gonna play out.

00:14:21 So in the new edition, we have a chapter on deep learning.

00:14:26 We got Ian Goodfellow to be the guest author

00:14:29 for that chapter.

00:14:30 So he said he could condense his whole deep learning book

00:14:34 into one chapter.

00:14:35 I think he did a great job.

00:14:38 We were also encouraged that he’s, you know,

00:14:40 we gave him the old neural net chapter

00:14:43 and said, modernize that.

00:14:47 And he said, you know, half of that was okay.

00:14:50 That certainly there’s lots of new things

00:14:52 that have been developed,

00:14:54 but some of the core was still the same.

00:14:58 So I think we’ll gain a better understanding

00:15:02 of what you can do there.

00:15:04 I think we’ll need to incorporate

00:15:07 all the things we can do with the other technologies, right?

00:15:10 So deep learning started out with convolutional networks

00:15:14 and very close to perception.

00:15:18 And it’s since moved to be able to do more

00:15:23 with actions and some degree of longer term planning.

00:15:28 But we need to do a better job

00:15:30 with representation than reasoning

00:15:32 and one shot learning and so on.

00:15:36 And I think we don’t know yet how that’s gonna play out.

00:15:41 So do you think looking at some success,

00:15:45 but certainly eventual demise,

00:15:49 a partial demise of experts

00:15:51 to symbolic systems in the 80s,

00:15:54 do you think there is kernels of wisdom

00:15:56 and the work that was done there

00:15:59 with logic and reasoning and so on

00:16:01 that will rise again in your view?

00:16:05 So certainly I think the idea of representation

00:16:08 and reasoning is crucial

00:16:10 that sometimes you just don’t have enough data

00:16:13 about the world to learn de novo.

00:16:17 So you’ve got to have some idea of representation,

00:16:22 whether that was programmed in or told or whatever,

00:16:24 and then be able to take steps of reasoning.

00:16:28 I think the problem with the good old fashioned AI

00:16:33 was one, we tried to base everything on these symbols

00:16:39 that were atomic.

00:16:42 And that’s great if you’re like trying to define

00:16:45 the properties of a triangle, right?

00:16:47 Because they have necessary and sufficient conditions.

00:16:50 But things in the real world don’t.

00:16:52 The real world is messy and doesn’t have sharp edges

00:16:55 and atomic symbols do.

00:16:57 So that was a poor match.

00:16:59 And then the other aspect was that the reasoning

00:17:05 was universal and applied anywhere,

00:17:09 which in some sense is good,

00:17:11 but it also means there’s no guidance

00:17:13 as to where to apply.

00:17:15 And so you started getting these paradoxes

00:17:17 like, well, if I have a mountain

00:17:20 and I remove one grain of sand,

00:17:22 then it’s still a mountain.

00:17:25 But if I do that repeatedly, at some point it’s not, right?

00:17:28 And with logic, there’s nothing to stop you

00:17:32 from applying things repeatedly.

00:17:37 But maybe with something like deep learning,

00:17:42 and I don’t really know what the right name for it is,

00:17:44 we could separate out those ideas.

00:17:46 So one, we could say a mountain isn’t just an atomic notion.

00:17:52 It’s some sort of something like a word embedding

00:17:56 that has a more complex representation.

00:18:02 And secondly, we could somehow learn,

00:18:05 yeah, there’s this rule that you can remove

00:18:06 one grain of sand and you can do that a bunch of times,

00:18:09 but you can’t do it a near infinite amount of times.

00:18:12 But on the other hand, when you’re doing induction

00:18:15 on the integer, sure, then it’s fine to do it

00:18:17 an infinite number of times.

00:18:18 And if we could, somehow we have to learn

00:18:22 when these strategies are applicable

00:18:24 rather than having the strategies be completely neutral

00:18:28 and available everywhere.

00:18:31 Anytime you use neural networks,

00:18:32 anytime you learn from data,

00:18:34 form representation from data in an automated way,

00:18:36 it’s not very explainable as to,

00:18:41 or it’s not introspective to us humans

00:18:45 in terms of how this neural network sees the world,

00:18:48 where, why does it succeed so brilliantly in so many cases

00:18:53 and fail so miserably in surprising ways and small.

00:18:56 So what do you think is the future there?

00:19:00 Can simply more data, better data,

00:19:03 more organized data solve that problem?

00:19:06 Or is there elements of symbolic systems

00:19:09 that need to be brought in

00:19:10 which are a little bit more explainable?

00:19:12 Yeah, so I prefer to talk about trust

00:19:16 and validation and verification

00:19:20 rather than just about explainability.

00:19:22 And then I think explanations are one tool

00:19:25 that you use towards those goals.

00:19:28 And I think it is an important issue

00:19:30 that we don’t wanna use these systems unless we trust them

00:19:33 and we wanna understand where they work

00:19:35 and where they don’t work.

00:19:37 And an explanation can be part of that, right?

00:19:40 So I apply for a loan and I get denied,

00:19:44 I want some explanation of why.

00:19:46 And you have, in Europe, we have the GDPR

00:19:50 that says you’re required to be able to get that.

00:19:53 But on the other hand,

00:19:54 the explanation alone is not enough, right?

00:19:57 So we are used to dealing with people

00:20:01 and with organizations and corporations and so on,

00:20:04 and they can give you an explanation

00:20:06 and you have no guarantee

00:20:07 that that explanation relates to reality, right?

00:20:11 So the bank can tell me, well, you didn’t get the loan

00:20:13 because you didn’t have enough collateral.

00:20:16 And that may be true, or it may be true

00:20:18 that they just didn’t like my religion or something else.

00:20:22 I can’t tell from the explanation,

00:20:24 and that’s true whether the decision was made

00:20:27 by a computer or by a person.

00:20:30 So I want more.

00:20:33 I do wanna have the explanations

00:20:35 and I wanna be able to have a conversation

00:20:37 to go back and forth and said,

00:20:39 well, you gave this explanation, but what about this?

00:20:41 And what would have happened if this had happened?

00:20:44 And what would I need to change that?

00:20:48 So I think a conversation is a better way to think about it

00:20:50 than just an explanation as a single output.

00:20:55 And I think we need testing of various kinds, right?

00:20:58 So in order to know,

00:21:00 was the decision really based on my collateral

00:21:03 or was it based on my religion or skin color or whatever?

00:21:08 I can’t tell if I’m only looking at my case,

00:21:10 but if I look across all the cases,

00:21:12 then I can detect the pattern, right?

00:21:15 So you wanna have that kind of capability.

00:21:18 You wanna have these adversarial testing, right?

00:21:21 So we thought we were doing pretty good

00:21:23 at object recognition in images.

00:21:25 We said, look, we’re at sort of pretty close

00:21:28 to human level performance on ImageNet and so on.

00:21:32 And then you start seeing these adversarial images

00:21:34 and you say, wait a minute,

00:21:36 that part is nothing like human performance.

00:21:39 You can mess with it really easily.

00:21:40 You can mess with it really easily, right?

00:21:42 And yeah, you can do that to humans too, right?

00:21:45 So we.

00:21:46 In a different way perhaps.

00:21:47 Right, humans don’t know what color the dress was.

00:21:49 Right.

00:21:50 And so they’re vulnerable to certain attacks

00:21:52 that are different than the attacks on the machines,

00:21:55 but the attacks on the machines are so striking.

00:21:59 They really change the way you think

00:22:00 about what we’ve done, right?

00:22:03 And the way I think about it is,

00:22:05 I think part of the problem is we’re seduced

00:22:08 by our low dimensional metaphors, right?

00:22:13 Yeah.

00:22:14 I like that phrase.

00:22:15 You look in a textbook and you say,

00:22:18 okay, now we’ve mapped out the space

00:22:20 and a cat is here and dog is here

00:22:24 and maybe there’s a tiny little spot in the middle

00:22:27 where you can’t tell the difference,

00:22:28 but mostly we’ve got it all covered.

00:22:30 And if you believe that metaphor,

00:22:33 then you say, well, we’re nearly there.

00:22:35 And there’s only gonna be a couple adversarial images.

00:22:39 But I think that’s the wrong metaphor

00:22:40 and what you should really say is,

00:22:42 it’s not a 2D flat space that we’ve got mostly covered.

00:22:45 It’s a million dimension space

00:22:47 and a cat is this string that goes out in this crazy path.

00:22:52 And if you step a little bit off the path in any direction,

00:22:55 you’re in nowhere’s land

00:22:57 and you don’t know what’s gonna happen.

00:22:59 And so I think that’s where we are

00:23:01 and now we’ve got to deal with that.

00:23:03 So it wasn’t so much an explanation,

00:23:06 but it was an understanding of what the models are

00:23:09 and what they’re doing

00:23:10 and now we can start exploring, how do you fix that?

00:23:12 Yeah, validating the robustness of the system and so on,

00:23:15 but take it back to this word trust.

00:23:20 Do you think we’re a little too hard on our robots

00:23:22 in terms of the standards we apply?

00:23:25 So, you know,

00:23:30 there’s a dance in nonverbal

00:23:34 and verbal communication between humans.

00:23:37 If we apply the same kind of standard in terms of humans,

00:23:40 we trust each other pretty quickly.

00:23:43 You know, you and I haven’t met before

00:23:45 and there’s some degree of trust, right?

00:23:48 That nothing’s gonna go crazy wrong

00:23:50 and yet to AI, when we look at AI systems

00:23:53 or we seem to approach skepticism always, always.

00:23:58 And it’s like they have to prove through a lot of hard work

00:24:03 that they’re even worthy of even inkling of our trust.

00:24:06 What do you think about that?

00:24:08 How do we break that barrier, close that gap?

00:24:11 I think that’s right.

00:24:12 I think that’s a big issue.

00:24:13 Just listening, my friend Mark Moffat is a naturalist

00:24:18 and he says, the most amazing thing about humans

00:24:22 is that you can walk into a coffee shop

00:24:25 or a busy street in a city

00:24:28 and there’s lots of people around you

00:24:30 that you’ve never met before and you don’t kill each other.

00:24:34 Yeah.

00:24:34 He says, chimpanzees cannot do that.

00:24:36 Yeah, right.

00:24:37 Right?

00:24:38 If a chimpanzee’s in a situation where here’s some

00:24:42 that aren’t from my tribe, bad things happen.

00:24:46 Especially in a coffee shop,

00:24:47 there’s delicious food around, you know.

00:24:48 Yeah, yeah.

00:24:49 But we humans have figured that out, right?

00:24:53 And you know.

00:24:54 For the most part.

00:24:55 For the most part.

00:24:55 We still go to war, we still do terrible things

00:24:58 but for the most part, we’ve learned to trust each other

00:25:01 and live together.

00:25:02 So that’s gonna be important for our AI systems as well.

00:25:08 And also I think a lot of the emphasis is on AI

00:25:13 but in many cases, AI is part of the technology

00:25:18 but isn’t really the main thing.

00:25:19 So a lot of what we’ve seen is more due

00:25:22 to communications technology than AI technology.

00:25:27 Yeah, you wanna make these good decisions

00:25:30 but the reason we’re able to have any kind of system at all

00:25:33 is we’ve got the communication

00:25:35 so that we’re collecting the data

00:25:37 and so that we can reach lots of people around the world.

00:25:41 I think that’s a bigger change that we’re dealing with.

00:25:45 Speaking of reaching a lot of people around the world,

00:25:47 on the side of education,

00:25:51 one of the many things in terms of education you’ve done,

00:25:53 you’ve taught the Intro to Artificial Intelligence course

00:25:56 that signed up 160,000 students.

00:26:00 There’s one of the first successful example

00:26:02 of a MOOC, Massive Open Online Course.

00:26:06 What did you learn from that experience?

00:26:09 What do you think is the future of MOOCs,

00:26:11 of education online?

00:26:12 Yeah, it was a great fun doing it,

00:26:15 particularly being right at the start

00:26:19 just because it was exciting and new

00:26:21 but it also meant that we had less competition, right?

00:26:24 So one of the things you hear about,

00:26:27 well, the problem with MOOCs is the completion rates

00:26:31 are so low so there must be a failure

00:26:33 and I gotta admit, I’m a prime contributor, right?

00:26:37 I probably started 50 different courses

00:26:40 that I haven’t finished

00:26:42 but I got exactly what I wanted out of them

00:26:44 because I had never intended to finish them.

00:26:46 I just wanted to dabble in a little bit

00:26:48 either to see the topic matter

00:26:50 or just to see the pedagogy of how are they doing this class.

00:26:53 So I guess the main thing I learned is when I came in,

00:26:58 I thought the challenge was information,

00:27:03 saying if I’m just, take the stuff I want you to know

00:27:07 and I’m very clear and explain it well,

00:27:10 then my job is done and good things are gonna happen.

00:27:14 And then in doing the course, I learned,

00:27:17 well, yeah, you gotta have the information

00:27:19 but really the motivation is the most important thing

00:27:23 that if students don’t stick with it,

00:27:26 it doesn’t matter how good the content is.

00:27:29 And I think being one of the first classes,

00:27:32 we were helped by sort of exterior motivation.

00:27:36 So we tried to do a good job of making it enticing

00:27:39 and setting up ways for the community

00:27:44 to work with each other to make it more motivating

00:27:46 but really a lot of it was, hey, this is a new thing

00:27:49 and I’m really excited to be part of a new thing.

00:27:51 And so the students brought their own motivation.

00:27:54 And so I think this is great

00:27:56 because there’s lots of people around the world

00:27:58 who have never had this before,

00:28:03 would never have the opportunity to go to Stanford

00:28:07 and take a class or go to MIT

00:28:08 or go to one of the other schools

00:28:10 but now we can bring that to them

00:28:12 and if they bring their own motivation,

00:28:15 they can be successful in a way they couldn’t before.

00:28:18 But that’s really just the top tier of people

00:28:21 that are ready to do that.

00:28:22 The rest of the people just don’t see

00:28:26 or don’t have the motivation

00:28:29 and don’t see how if they push through

00:28:31 and were able to do it, what advantage that would get them.

00:28:34 So I think we got a long way to go

00:28:36 before we were able to do that.

00:28:37 And I think some of it is based on technology

00:28:40 but more of it’s based on the idea of community.

00:28:43 You gotta actually get people together.

00:28:46 Some of the getting together can be done online.

00:28:49 I think some of it really has to be done in person

00:28:52 in order to build that type of community and trust.

00:28:56 You know, there’s an intentional mechanism

00:28:59 that we’ve developed a short attention span,

00:29:02 especially younger people

00:29:04 because sort of shorter and shorter videos online,

00:29:08 there’s a whatever the way the brain is developing now

00:29:13 and with people that have grown up with the internet,

00:29:16 they have quite a short attention span.

00:29:18 So, and I would say I had the same

00:29:21 when I was growing up too, probably for different reasons.

00:29:23 So I probably wouldn’t have learned as much as I have

00:29:28 if I wasn’t forced to sit in a physical classroom,

00:29:31 sort of bored, sometimes falling asleep,

00:29:33 but sort of forcing myself through that process.

00:29:36 So sometimes extremely difficult computer science courses.

00:29:39 What’s the difference in your view

00:29:42 between in person education experience,

00:29:46 which you, first of all, yourself had

00:29:48 and you yourself taught and online education

00:29:52 and how do we close that gap if it’s even possible?

00:29:54 Yeah, so I think there’s two issues.

00:29:56 One is whether it’s in person or online.

00:30:00 So it’s sort of the physical location

00:30:03 and then the other is kind of the affiliation, right?

00:30:07 So you stuck with it in part

00:30:10 because you were in the classroom

00:30:12 and you saw everybody else was suffering

00:30:15 the same way you were,

00:30:17 but also because you were enrolled,

00:30:20 you had paid tuition,

00:30:22 sort of everybody was expecting you to stick with it.

00:30:25 Society, parents, peers.

00:30:29 And so those are two separate things.

00:30:31 I mean, you could certainly imagine

00:30:32 I pay a huge amount of tuition

00:30:35 and everybody signed up and says, yes, you’re doing this,

00:30:38 but then I’m in my room

00:30:40 and my classmates are in different rooms, right?

00:30:43 We could have things set up that way.

00:30:45 So it’s not just the online versus offline.

00:30:48 I think what’s more important

00:30:50 is the commitment that you’ve made.

00:30:53 And certainly it is important

00:30:56 to have that kind of informal,

00:30:59 you know, I meet people outside of class,

00:31:01 we talk together because we’re all in it together.

00:31:05 I think that’s really important,

00:31:07 both in keeping your motivation

00:31:10 and also that’s where

00:31:11 some of the most important learning goes on.

00:31:13 So you wanna have that.

00:31:15 Maybe, you know, especially now

00:31:17 we start getting into higher bandwidths

00:31:19 and augmented reality and virtual reality,

00:31:22 you might be able to get that

00:31:23 without being in the same physical place.

00:31:25 Do you think it’s possible we’ll see a course at Stanford,

00:31:30 for example, that for students,

00:31:33 enrolled students is only online in the near future

00:31:37 or literally sort of it’s part of the curriculum

00:31:39 and there is no…

00:31:41 Yeah, so you’re starting to see that.

00:31:42 I know Georgia Tech has a master’s that’s done that way.

00:31:46 Oftentimes it’s sort of,

00:31:48 they’re creeping in in terms of a master’s program

00:31:50 or sort of further education,

00:31:54 considering the constraints of students and so on.

00:31:56 But I mean, literally, is it possible that we,

00:32:00 you know, Stanford, MIT, Berkeley,

00:32:02 all these places go online only in the next few decades?

00:32:07 Yeah, probably not,

00:32:08 because, you know, they’ve got a big commitment

00:32:11 to a physical campus.

00:32:13 Sure, so there’s a momentum

00:32:16 that’s both financial and culturally.

00:32:18 Right, and then there are certain things

00:32:21 that’s just hard to do virtually, right?

00:32:25 So, you know, we’re in a field where,

00:32:29 if you have your own computer and your own paper,

00:32:32 and so on, you can do the work anywhere.

00:32:36 But if you’re in a biology lab or something,

00:32:39 you know, you don’t have all the right stuff at home.

00:32:42 Right, so our field, programming,

00:32:45 you’ve also done a lot of programming yourself.

00:32:50 In 2001, you wrote a great article about programming

00:32:54 called Teach Yourself Programming in 10 Years,

00:32:57 sort of response to all the books

00:32:59 that say teach yourself programming in 21 days.

00:33:01 So if you were giving advice to someone

00:33:02 getting into programming today,

00:33:04 this is a few years since you’ve written that article,

00:33:07 what’s the best way to undertake that journey?

00:33:10 I think there’s lots of different ways,

00:33:12 and I think programming means more things now.

00:33:17 And I guess, you know, when I wrote that article,

00:33:20 I was thinking more about

00:33:23 becoming a professional software engineer,

00:33:25 and I thought that’s a, you know,

00:33:27 sort of a career long field of study.

00:33:31 But I think there’s lots of things now

00:33:33 that people can do where programming is a part

00:33:37 of solving what they wanna solve

00:33:40 without achieving that professional level status, right?

00:33:44 So I’m not gonna be going

00:33:45 and writing a million lines of code,

00:33:47 but, you know, I’m a biologist or a physicist or something,

00:33:51 or even a historian, and I’ve got some data,

00:33:55 and I wanna ask a question of that data.

00:33:58 And I think for that, you don’t need 10 years, right?

00:34:02 So there are many shortcuts

00:34:04 to being able to answer those kinds of questions.

00:34:08 And, you know, you see today a lot of emphasis

00:34:11 on learning to code, teaching kids how to code.

00:34:16 I think that’s great,

00:34:18 but I wish they would change the message a little bit,

00:34:21 right, so I think code isn’t the main thing.

00:34:24 I don’t really care if you know the syntax of JavaScript

00:34:28 or if you can connect these blocks together

00:34:31 in this visual language.

00:34:33 But what I do care about is that you can analyze a problem,

00:34:38 you can think of a solution, you can carry out,

00:34:43 you know, make a model, run that model,

00:34:46 test the model, see the results,

00:34:50 verify that they’re reasonable,

00:34:53 ask questions and answer them, right?

00:34:55 So it’s more modeling and problem solving,

00:34:58 and you use coding in order to do that,

00:35:01 but it’s not just learning coding for its own sake.

00:35:04 That’s really interesting.

00:35:05 So it’s actually almost, in many cases,

00:35:08 it’s learning to work with data,

00:35:10 to extract something useful out of data.

00:35:11 So when you say problem solving,

00:35:13 you really mean taking some kind of,

00:35:15 maybe collecting some kind of data set,

00:35:17 cleaning it up, and saying something interesting about it,

00:35:20 which is useful in all kinds of domains.

00:35:23 And, you know, and I see myself being stuck sometimes

00:35:28 in kind of the old ways, right?

00:35:30 So, you know, I’ll be working on a project,

00:35:34 maybe with a younger employee, and we say,

00:35:37 oh, well, here’s this new package

00:35:39 that could help solve this problem.

00:35:42 And I’ll go and I’ll start reading the manuals,

00:35:44 and, you know, I’ll be two hours into reading the manuals,

00:35:48 and then my colleague comes back and says, I’m done.

00:35:51 You know, I downloaded the package, I installed it,

00:35:53 I tried calling some things, the first one didn’t work,

00:35:56 the second one worked, now I’m done.

00:35:58 And I say, but I have a hundred questions

00:36:00 about how does this work and how does that work?

00:36:02 And they say, who cares, right?

00:36:04 I don’t need to understand the whole thing.

00:36:05 I answered my question, it’s a big, complicated package,

00:36:09 I don’t understand the rest of it,

00:36:10 but I got the right answer.

00:36:12 And I’m just, it’s hard for me to get into that mindset.

00:36:15 I want to understand the whole thing.

00:36:17 And, you know, if they wrote a manual,

00:36:19 I should probably read it.

00:36:21 And, but that’s not necessarily the right way.

00:36:23 I think I have to get used to dealing with more,

00:36:28 being more comfortable with uncertainty

00:36:30 and not knowing everything.

00:36:32 Yeah, so I struggle with the same,

00:36:33 instead of the spectrum between Donald and Don Knuth.

00:36:37 Yeah.

00:36:38 It’s kind of the very, you know,

00:36:39 before he can say anything about a problem,

00:36:42 he really has to get down to the machine code assembly.

00:36:45 Yeah.

00:36:46 And that forces exactly what you said of several students

00:36:50 in my group that, you know, 20 years old,

00:36:53 and they can solve almost any problem within a few hours.

00:36:56 That would take me probably weeks

00:36:58 because I would try to, as you said, read the manual.

00:37:00 So do you think the nature of mastery,

00:37:04 you’re mentioning biology,

00:37:06 sort of outside disciplines, applying programming,

00:37:11 but computer scientists.

00:37:13 So over time, there’s higher and higher levels

00:37:16 of abstraction available now.

00:37:18 So with this week, there’s the TensorFlow Summit, right?

00:37:23 So if you’re not particularly into deep learning,

00:37:27 but you’re still a computer scientist,

00:37:29 you can accomplish an incredible amount with TensorFlow

00:37:33 without really knowing any fundamental internals

00:37:35 of machine learning.

00:37:37 Do you think the nature of mastery is changing,

00:37:40 even for computer scientists,

00:37:42 like what it means to be an expert programmer?

00:37:45 Yeah, I think that’s true.

00:37:47 You know, we never really should have focused on programmer,

00:37:51 right, because it’s still, it’s the skill,

00:37:53 and what we really want to focus on is the result.

00:37:56 So we built this ecosystem

00:37:59 where the way you can get stuff done

00:38:01 is by programming it yourself.

00:38:04 At least when I started, you know,

00:38:06 library functions meant you had square root,

00:38:09 and that was about it, right?

00:38:10 Everything else you built from scratch.

00:38:13 And then we built up an ecosystem where a lot of times,

00:38:16 well, you can download a lot of stuff

00:38:17 that does a big part of what you need.

00:38:20 And so now it’s more a question of assembly

00:38:23 rather than manufacturing.

00:38:28 And that’s a different way of looking at problems.

00:38:32 From another perspective in terms of mastery

00:38:34 and looking at programmers or people that reason

00:38:37 about problems in a computational way.

00:38:39 So Google, you know, from the hiring perspective,

00:38:44 from the perspective of hiring

00:38:45 or building a team of programmers,

00:38:47 how do you determine if someone’s a good programmer?

00:38:50 Or if somebody, again, so I want to deviate from,

00:38:53 I want to move away from the word programmer,

00:38:55 but somebody who could solve problems

00:38:57 of large scale data and so on.

00:38:59 What’s, how do you build a team like that

00:39:02 through the interviewing process?

00:39:03 Yeah, and I think as a company grows,

00:39:08 you get more expansive in the types

00:39:12 of people you’re looking for, right?

00:39:14 So I think, you know, in the early days,

00:39:16 we’d interview people and the question we were trying

00:39:19 to ask is how close are they to Jeff Dean?

00:39:22 And most people were pretty far away,

00:39:26 but we take the ones that were not that far away.

00:39:29 And so we got kind of a homogeneous group

00:39:31 of people who were really great programmers.

00:39:34 Then as a company grows, you say,

00:39:37 well, we don’t want everybody to be the same,

00:39:39 to have the same skill set.

00:39:40 And so now we’re hiring biologists in our health areas

00:39:47 and we’re hiring physicists,

00:39:48 we’re hiring mechanical engineers,

00:39:51 we’re hiring, you know, social scientists and ethnographers

00:39:56 and people with different backgrounds

00:39:59 who bring different skills.

00:40:01 So you have mentioned that you still may partake

00:40:06 in code reviews, given that you have a wealth of experience,

00:40:10 as you’ve also mentioned.

00:40:13 What errors do you often see and tend to highlight

00:40:16 in the code of junior developers of people coming up now,

00:40:20 given your background from Blisp

00:40:23 to a couple of decades of programming?

00:40:26 Yeah, that’s a great question.

00:40:28 You know, sometimes I try to look at the flexibility

00:40:31 of the design of, yes, you know, this API solves this problem,

00:40:37 but where is it gonna go in the future?

00:40:39 Who else is gonna wanna call this?

00:40:41 And, you know, are you making it easier for them to do that?

00:40:46 That’s a matter of design, is it documentation,

00:40:50 is it sort of an amorphous thing

00:40:53 you can’t really put into words?

00:40:55 It’s just how it feels.

00:40:56 If you put yourself in the shoes of a developer,

00:40:58 would you use this kind of thing?

00:40:59 I think it is how you feel, right?

00:41:01 And so yeah, documentation is good,

00:41:03 but it’s more a design question, right?

00:41:06 If you get the design right,

00:41:07 then people will figure it out,

00:41:10 whether the documentation is good or not.

00:41:12 And if the design’s wrong, then it’d be harder to use.

00:41:16 How have you yourself changed as a programmer over the years?

00:41:22 In a way, you already started to say sort of,

00:41:26 you want to read the manual,

00:41:28 you want to understand the core of the syntax

00:41:30 to how the language is supposed to be used and so on.

00:41:33 But what’s the evolution been like

00:41:36 from the 80s, 90s to today?

00:41:40 I guess one thing is you don’t have to worry

00:41:42 about the small details of efficiency

00:41:46 as much as you used to, right?

00:41:48 So like I remember I did my list book in the 90s,

00:41:53 and one of the things I wanted to do was say,

00:41:56 here’s how you do an object system.

00:41:58 And basically, we’re going to make it

00:42:01 so each object is a hash table,

00:42:03 and you look up the methods, and here’s how it works.

00:42:05 And then I said, of course,

00:42:07 the real Common Lisp object system is much more complicated.

00:42:12 It’s got all these efficiency type issues,

00:42:15 and this is just a toy,

00:42:16 and nobody would do this in real life.

00:42:18 And it turns out Python pretty much did exactly

00:42:22 what I said and said objects are just dictionaries.

00:42:27 And yeah, they have a few little tricks as well.

00:42:30 But mostly, the thing that would have been

00:42:34 100 times too slow in the 80s

00:42:36 is now plenty fast for most everything.

00:42:39 So you had to, as a programmer,

00:42:40 let go of perhaps an obsession

00:42:44 that I remember coming up with

00:42:45 of trying to write efficient code.

00:42:48 Yeah, to say what really matters

00:42:51 is the total time it takes to get the project done.

00:42:56 And most of that’s gonna be the programmer time.

00:42:59 So if you’re a little bit less efficient,

00:43:00 but it makes it easier to understand and modify,

00:43:04 then that’s the right trade off.

00:43:05 So you’ve written quite a bit about Lisp.

00:43:07 Your book on programming is in Lisp.

00:43:10 You have a lot of code out there that’s in Lisp.

00:43:12 So myself and people who don’t know what Lisp is

00:43:16 should look it up.

00:43:18 It’s my favorite language for many AI researchers.

00:43:20 It is a favorite language.

00:43:22 The favorite language they never use these days.

00:43:25 So what part of Lisp do you find most beautiful and powerful?

00:43:28 So I think the beautiful part is the simplicity

00:43:31 that in half a page, you can define the whole language.

00:43:36 And other languages don’t have that.

00:43:38 So you feel like you can hold everything in your head.

00:43:42 And then a lot of people say,

00:43:46 well, then that’s too simple.

00:43:48 Here’s all these things I wanna do.

00:43:50 And my Java or Python or whatever

00:43:54 has 100 or 200 or 300 different syntax rules

00:43:58 and don’t I need all those?

00:44:00 And Lisp’s answer was, no, we’re only gonna give you

00:44:03 eight or so syntax rules,

00:44:06 but we’re gonna allow you to define your own.

00:44:09 And so that was a very powerful idea.

00:44:11 And I think this idea of saying,

00:44:15 I can start with my problem and with my data,

00:44:20 and then I can build the language I want for that problem

00:44:24 and for that data.

00:44:25 And then I can make Lisp define that language.

00:44:28 So you’re sort of mixing levels and saying,

00:44:32 I’m simultaneously a programmer in a language

00:44:36 and a language designer.

00:44:38 And that allows a better match between your problem

00:44:41 and your eventual code.

00:44:43 And I think Lisp had done that better than other languages.

00:44:47 Yeah, it’s a very elegant implementation

00:44:49 of functional programming.

00:44:51 But why do you think Lisp has not had the mass adoption

00:44:55 and success of languages like Python?

00:44:57 Is it the parentheses?

00:44:59 Is it all the parentheses?

00:45:02 Yeah, so I think a couple things.

00:45:05 So one was, I think it was designed for a single programmer

00:45:10 or a small team and a skilled programmer

00:45:14 who had the good taste to say,

00:45:17 well, I am doing language design

00:45:19 and I have to make good choices.

00:45:21 And if you make good choices, that’s great.

00:45:23 If you make bad choices, you can hurt yourself

00:45:28 and it can be hard for other people on the team

00:45:30 to understand it.

00:45:31 So I think there was a limit to the scale

00:45:34 of the size of a project in terms of number of people

00:45:37 that Lisp was good for.

00:45:38 And as an industry, we kind of grew beyond that.

00:45:43 I think it is in part the parentheses.

00:45:46 You know, one of the jokes is the acronym for Lisp

00:45:49 is lots of irritating, silly parentheses.

00:45:53 My acronym was Lisp is syntactically pure,

00:45:58 saying all you need is parentheses and atoms.

00:46:01 But I remember, you know, as we had the AI textbook

00:46:05 and because we did it in the nineties,

00:46:08 we had pseudocode in the book,

00:46:11 but then we said, well, we’ll have Lisp online

00:46:13 because that’s the language of AI at the time.

00:46:16 And I remember some of the students complaining

00:46:18 because they hadn’t had Lisp before

00:46:20 and they didn’t quite understand what was going on.

00:46:22 And I remember one student complained,

00:46:24 I don’t understand how this pseudocode

00:46:26 corresponds to this Lisp.

00:46:29 And there was a one to one correspondence

00:46:31 between the symbols in the code and the pseudocode.

00:46:35 And the only thing difference was the parentheses.

00:46:39 So I said, it must be that for some people,

00:46:41 a certain number of left parentheses shuts off their brain.

00:46:45 Yeah, it’s very possible in that sense

00:46:47 and Python just goes the other way.

00:46:49 So that was the point at which I said,

00:46:51 okay, can’t have only Lisp as a language.

00:46:54 Cause I don’t wanna, you know,

00:46:56 you only got 10 or 12 or 15 weeks or whatever it is

00:46:59 to teach AI and I don’t want to waste two weeks

00:47:01 of that teaching Lisp.

00:47:03 So I say, I gotta have another language.

00:47:04 Java was the most popular language at the time.

00:47:06 I started doing that.

00:47:08 And then I said, it’s really hard to have a one to one

00:47:12 correspondence between the pseudocode and the Java

00:47:14 because Java is so verbose.

00:47:16 So then I said, I’m gonna do a survey

00:47:18 and find the language that’s most like my pseudocode.

00:47:22 And it turned out Python basically was my pseudocode.

00:47:26 Somehow I had channeled Guido,

00:47:30 designed a pseudocode that was the same as Python,

00:47:32 although I hadn’t heard of Python at that point.

00:47:36 And from then on, that’s what I’ve been using

00:47:38 cause it’s been a good match.

00:47:41 So what’s the story in Python behind PyTudes?

00:47:45 Your GitHub repository with puzzles and exercises

00:47:48 in Python is pretty fun.

00:47:49 Yeah, just it, it seems like fun, you know,

00:47:53 I like doing puzzles and I like being an educator.

00:47:57 I did a class with Udacity, Udacity 212, I think it was.

00:48:02 It was basically problem solving using Python

00:48:07 and looking at different problems.

00:48:08 Does PyTudes feed that class in terms of the exercises?

00:48:11 I was wondering what the…

00:48:12 Yeah, so the class came first.

00:48:15 Some of the stuff that’s in PyTudes was write ups

00:48:17 of what was in the class and then some of it

00:48:19 was just continuing to work on new problems.

00:48:24 So what’s the organizing madness of PyTudes?

00:48:26 Is it just a collection of cool exercises?

00:48:30 Just whatever I thought was fun.

00:48:31 Okay, awesome.

00:48:32 So you were the director of search quality at Google

00:48:35 from 2001 to 2005 in the early days

00:48:40 when there’s just a few employees

00:48:41 and when the company was growing like crazy, right?

00:48:46 So, I mean, Google revolutionized the way we discover,

00:48:52 share and aggregate knowledge.

00:48:55 So just, this is one of the fundamental aspects

00:49:00 of civilization, right, is information being shared

00:49:03 and there’s different mechanisms throughout history

00:49:04 but Google has just 10x improved that, right?

00:49:08 And you’re a part of that, right?

00:49:10 People discovering that information.

00:49:11 So what were some of the challenges on a philosophical

00:49:15 or the technical level in those early days?

00:49:18 It definitely was an exciting time

00:49:20 and as you say, we were doubling in size every year

00:49:24 and the challenges were we wanted

00:49:26 to get the right answers, right?

00:49:29 And we had to figure out what that meant.

00:49:32 We had to implement that and we had to make it all efficient

00:49:36 and we had to keep on testing

00:49:41 and seeing if we were delivering good answers.

00:49:44 And now when you say good answers,

00:49:45 it means whatever people are typing in

00:49:47 in terms of keywords, in terms of that kind of thing

00:49:50 that the results they get are ordered

00:49:53 by the desirability for them of those results.

00:49:56 Like they’re like, the first thing they click on

00:49:58 will likely be the thing that they were actually looking for.

00:50:01 Right, one of the metrics we had

00:50:03 was focused on the first thing.

00:50:05 Some of it was focused on the whole page.

00:50:07 Some of it was focused on top three or so.

00:50:11 So we looked at a lot of different metrics

00:50:13 for how well we were doing

00:50:15 and we broke it down into subclasses of,

00:50:19 maybe here’s a type of query that we’re not doing well on

00:50:23 and we try to fix that.

00:50:25 Early on we started to realize that we were in an adversarial

00:50:29 position, right, so we started thinking,

00:50:32 well, we’re kind of like the card catalog in the library,

00:50:35 right, so the books are here and we’re off to the side

00:50:39 and we’re just reflecting what’s there.

00:50:42 And then we realized every time we make a change,

00:50:45 the webmasters make a change and it’s game theoretic.

00:50:50 And so we had to think not only of is this the right move

00:50:54 for us to make now, but also if we make this move,

00:50:57 what’s the counter move gonna be?

00:50:59 Is that gonna get us into a worse place,

00:51:02 in which case we won’t make that move,

00:51:03 we’ll make a different move.

00:51:05 And did you find, I mean, I assume with the popularity

00:51:08 and the growth of the internet

00:51:09 that people were creating new content,

00:51:11 so you’re almost helping guide the creation of new content.

00:51:14 Yeah, so that’s certainly true, right,

00:51:15 so we definitely changed the structure of the network.

00:51:20 So if you think back in the very early days,

00:51:24 Larry and Sergey had the PageRank paper

00:51:28 and John Kleinberg had this hubs and authorities model,

00:51:33 which says the web is made out of these hubs,

00:51:38 which will be my page of cool links about dogs or whatever,

00:51:44 and people would just list links.

00:51:46 And then there’d be authorities,

00:51:47 which were the page about dogs that most people linked to.

00:51:53 That doesn’t happen anymore.

00:51:54 People don’t bother to say my page of cool links,

00:51:57 because we took over that function, right,

00:52:00 so we changed the way that worked.

00:52:03 Did you imagine back then that the internet

00:52:05 would be as massively vibrant as it is today?

00:52:08 I mean, it was already growing quickly,

00:52:10 but it’s just another, I don’t know if you’ve ever,

00:52:14 today, if you sit back and just look at the internet

00:52:18 with wonder the amount of content

00:52:20 that’s just constantly being created,

00:52:22 constantly being shared and deployed.

00:52:24 Yeah, it’s always been surprising to me.

00:52:27 I guess I’m not very good at predicting the future.

00:52:31 And I remember being a graduate student in 1980 or so,

00:52:35 and we had the ARPANET,

00:52:39 and then there was this proposal to commercialize it,

00:52:44 and have this internet, and this crazy Senator Gore

00:52:49 thought that might be a good idea.

00:52:51 And I remember thinking, oh, come on,

00:52:53 you can’t expect a commercial company

00:52:55 to understand this technology.

00:52:58 They’ll never be able to do it.

00:52:59 Yeah, okay, we can have this.com domain,

00:53:01 but it won’t go anywhere.

00:53:03 So I was wrong, Al Gore was right.

00:53:05 At the same time, the nature of what it means

00:53:07 to be a commercial company has changed, too.

00:53:09 So Google, in many ways, at its founding

00:53:12 is different than what companies were before, I think.

00:53:16 Right, so there’s all these business models

00:53:19 that are so different than what was possible back then.

00:53:23 So in terms of predicting the future,

00:53:25 what do you think it takes to build a system

00:53:27 that approaches human level intelligence?

00:53:29 You’ve talked about, of course,

00:53:31 that we shouldn’t be so obsessed

00:53:34 about creating human level intelligence.

00:53:36 We just create systems that are very useful for humans.

00:53:39 But what do you think it takes

00:53:40 to approach that level?

00:53:44 Right, so certainly I don’t think

00:53:47 human level intelligence is one thing, right?

00:53:49 So I think there’s lots of different tasks,

00:53:51 lots of different capabilities.

00:53:54 I also don’t think that should be the goal, right?

00:53:56 So I wouldn’t wanna create a calculator

00:54:01 that could do multiplication at human level, right?

00:54:04 That would be a step backwards.

00:54:06 And so for many things,

00:54:07 we should be aiming far beyond human level

00:54:09 for other things.

00:54:12 Maybe human level is a good level to aim at.

00:54:15 And for others, we’d say,

00:54:16 well, let’s not bother doing this

00:54:18 because we already have humans can take on those tasks.

00:54:21 So as you say, I like to focus on what’s a useful tool.

00:54:26 And in some cases, being at human level

00:54:30 is an important part of crossing that threshold

00:54:32 to make the tool useful.

00:54:34 So we see in things like these personal assistants now

00:54:39 that you get either on your phone

00:54:41 or on a speaker that sits on the table,

00:54:44 you wanna be able to have a conversation with those.

00:54:47 And I think as an industry,

00:54:49 we haven’t quite figured out what the right model is

00:54:51 for what these things can do.

00:54:55 And we’re aiming towards,

00:54:56 well, you just have a conversation with them

00:54:57 the way you can with a person.

00:55:00 But we haven’t delivered on that model yet, right?

00:55:02 So you can ask it, what’s the weather?

00:55:04 You can ask it, play some nice songs.

00:55:08 And five or six other things,

00:55:11 and then you run out of stuff that it can do.

00:55:14 In terms of a deep, meaningful connection.

00:55:16 So you’ve mentioned the movie Her

00:55:18 as one of your favorite AI movies.

00:55:20 Do you think it’s possible for a human being

00:55:22 to fall in love with an AI assistant, as you mentioned?

00:55:25 So taking this big leap from what’s the weather

00:55:28 to having a deep connection.

00:55:31 Yeah, I think as people, that’s what we love to do.

00:55:35 And I was at a showing of Her

00:55:39 where we had a panel discussion and somebody asked me,

00:55:43 what other movie do you think Her is similar to?

00:55:46 And my answer was Life of Brian,

00:55:50 which is not a science fiction movie,

00:55:53 but both movies are about wanting to believe

00:55:57 in something that’s not necessarily real.

00:56:00 Yeah, by the way, for people that don’t know,

00:56:01 it’s Monty Python.

00:56:03 Yeah, it’s been brilliantly put.

00:56:05 Right, so I think that’s just the way we are.

00:56:07 We want to trust, we want to believe,

00:56:11 we want to fall in love,

00:56:12 and it doesn’t necessarily take that much, right?

00:56:15 So my kids fell in love with their teddy bear,

00:56:20 and the teddy bear was not very interactive.

00:56:23 So that’s all us pushing our feelings

00:56:26 onto our devices and our things,

00:56:29 and I think that that’s what we like to do,

00:56:31 so we’ll continue to do that.

00:56:33 So yeah, as human beings, we long for that connection,

00:56:36 and just AI has to do a little bit of work

00:56:39 to catch us in the other end.

00:56:41 Yeah, and certainly, if you can get to dog level,

00:56:46 a lot of people have invested a lot of love in their pets.

00:56:49 In their pets.

00:56:50 Some people, as I’ve been told,

00:56:52 in working with autonomous vehicles,

00:56:54 have invested a lot of love into their inanimate cars,

00:56:58 so it really doesn’t take much.

00:57:00 So what is a good test to linger on a topic

00:57:05 that may be silly or a little bit philosophical?

00:57:07 What is a good test of intelligence in your view?

00:57:12 Is natural conversation like in the Turing test

00:57:14 a good test?

00:57:16 Put another way, what would impress you

00:57:20 if you saw a computer do it these days?

00:57:22 Yeah, I mean, I get impressed all the time.

00:57:24 Go playing, StarCraft playing, those are all pretty cool.

00:57:35 And I think, sure, conversation is important.

00:57:39 I think we sometimes have these tests

00:57:44 where it’s easy to fool the system, where

00:57:46 you can have a chat bot that can have a conversation,

00:57:51 but it never gets into a situation

00:57:54 where it has to be deep enough that it really reveals itself

00:57:58 as being intelligent or not.

00:58:00 I think Turing suggested that, but I think if he were alive,

00:58:07 he’d say, you know, I didn’t really mean that seriously.

00:58:11 And I think, this is just my opinion,

00:58:15 but I think Turing’s point was not

00:58:17 that this test of conversation is a good test.

00:58:21 I think his point was having a test is the right thing.

00:58:25 So rather than having the philosophers say, oh, no,

00:58:28 AI is impossible, you should say, well,

00:58:31 we’ll just have a test, and then the result of that

00:58:33 will tell us the answer.

00:58:34 And it doesn’t necessarily have to be a conversation test.

00:58:37 That’s right.

00:58:37 And coming up a new, better test as the technology evolves

00:58:40 is probably the right way.

00:58:42 Do you worry, as a lot of the general public does about,

00:58:46 not a lot, but some vocal part of the general public

00:58:51 about the existential threat of artificial intelligence?

00:58:53 So looking farther into the future, as you said,

00:58:56 most of us are not able to predict much.

00:58:59 So when shrouded in such mystery, there’s a concern of,

00:59:02 well, you start thinking about worst case.

00:59:05 Is that something that occupies your mind, space, much?

00:59:09 So I certainly think about threats.

00:59:11 I think about dangers.

00:59:13 And I think any new technology has positives and negatives.

00:59:19 And if it’s a powerful technology,

00:59:21 it can be used for bad as well as for good.

00:59:24 So I’m certainly not worried about the robot

00:59:27 apocalypse and the Terminator type scenarios.

00:59:32 I am worried about change in employment.

00:59:37 And are we going to be able to react fast enough

00:59:41 to deal with that?

00:59:41 I think we’re already seeing it today, where

00:59:44 a lot of people are disgruntled about the way

00:59:48 income inequality is working.

00:59:50 And automation could help accelerate

00:59:53 those kinds of problems.

00:59:55 I see powerful technologies can always be used as weapons,

00:59:59 whether they’re robots or drones or whatever.

01:00:03 Some of that we’re seeing due to AI.

01:00:06 A lot of it, you don’t need AI.

01:00:09 And I don’t know what’s a worst threat,

01:00:12 if it’s an autonomous drone or it’s CRISPR technology

01:00:17 becoming available.

01:00:18 Or we have lots of threats to face.

01:00:21 And some of them involve AI, and some of them don’t.

01:00:24 So the threats that technology presents,

01:00:27 are you, for the most part, optimistic about technology

01:00:31 also alleviating those threats or creating new opportunities

01:00:34 or protecting us from the more detrimental effects

01:00:38 of these new technologies?

01:00:38 I don’t know.

01:00:39 Again, it’s hard to predict the future.

01:00:41 And as a society so far, we’ve survived

01:00:47 nuclear bombs and other things.

01:00:50 Of course, only societies that have survived

01:00:53 are having this conversation.

01:00:54 So maybe that’s survivorship bias there.

01:00:59 What problem stands out to you as exciting, challenging,

01:01:02 impactful to work on in the near future for yourself,

01:01:06 for the community, and broadly?

01:01:09 So we talked about these assistance and conversation.

01:01:13 I think that’s a great area.

01:01:14 I think combining common sense reasoning

01:01:20 with the power of data is a great area.

01:01:26 In which application?

01:01:27 In conversation, or just broadly speaking?

01:01:29 Just in general, yeah.

01:01:31 As a programmer, I’m interested in programming tools,

01:01:35 both in terms of the current systems

01:01:38 we have today with TensorFlow and so on.

01:01:41 Can we make them much easier to use

01:01:43 for a broader class of people?

01:01:45 And also, can we apply machine learning

01:01:49 to the more traditional type of programming?

01:01:52 So when you go to Google and you type in a query

01:01:57 and you spell something wrong, it says, did you mean?

01:02:00 And the reason we’re able to do that

01:02:01 is because lots of other people made a similar error,

01:02:04 and then they corrected it.

01:02:06 We should be able to go into our code bases and our bug fix

01:02:10 bases.

01:02:10 And when I type a line of code, it should be able to say,

01:02:13 did you mean such and such?

01:02:15 If you type this today, you’re probably going to type

01:02:17 in this bug fix tomorrow.

01:02:20 Yeah, that’s a really exciting application

01:02:22 of almost an assistant for the coding programming experience

01:02:27 at every level.

01:02:29 So I think I could safely speak for the entire AI community,

01:02:35 first of all, for thanking you for the amazing work you’ve

01:02:37 done, certainly for the amazing work you’ve done

01:02:40 with AI and Modern Approach book.

01:02:43 I think we’re all looking forward very much

01:02:45 for the fourth edition, and then the fifth edition, and so on.

01:02:48 So Peter, thank you so much for talking today.

01:02:51 Yeah, thank you.

01:02:51 My pleasure.