00:00hi everyone welcome to the a 6nz podcast
00:02I am so Nolan today we have two partners
00:04from andreessen horowitz we were just
00:05having an informal conversation in the
00:07hallway literally around machine
00:09learning and AI I'm Steven Sinofsky a
00:11board partner phase 6 and Z gave a
00:13presentation on the evolution of machine
00:15learning and Frank Chen recently put out
00:17a tweet storm on why google's deepmind
00:20algorithm beating Lisa dull was so
00:24significant and they were sort of
00:25talking about like oh my god we've been
00:27but it's not gonna be all
00:28backward-looking because I think the
00:29point is that the evolution is what's
00:32why is why now right and also putting
00:34you know me I'd love to put things in in
00:35context cuz like there's always lessons
00:38to be learned and patterns to avoid a
00:39non-void in patterns that's a key word
00:41for today ok well let's start talking
00:42about those patterns well maybe let's
00:44start with the big go victory which is
00:46it got people really fired up it
00:48dominated the press for a little while
00:50and you might be wondering what is the
00:52computer program won another board game
00:55a board game like what could be less
00:57relevant to everyday life right and
00:58we've seen this before we started with
01:00tic-tac-toe and we got to checkers and
01:02we got to chess and then Watson even won
01:04jeopardy and now here we are talking
01:06about another board game so like it's
01:08kind of irrelevant to everyday life
01:10isn't it and the surprising thing is
01:11look we've had a lot of false starts
01:14with artificial intelligence and the
01:16vector has always been hey look
01:18now that we've run this very
01:20sophisticated board game chess checkers
01:22whatever now we're on the verge to do
01:25general-purpose intelligence right and
01:27that has always been the promise and the
01:29full start of AI and so the question is
01:32whether this victory with go which is an
01:35incredibly complex game you can't brute
01:37force search all of the possible moves
01:40because there's just too many right
01:41there right so people know it's like a
01:42Google times a chess game so think of so
01:50basically if you think of the number
01:51total number of chess moves and there's
01:52a lot of them writes a big board lots of
01:54pieces you multiply by a Google that's
01:56how many possible go board
01:58configurations there are so you can't
02:00actually brute force search all of the
02:02things which is what you'd expect a
02:03computer to be able to do it's got lots
02:05of processor it's got this very reliable
02:07big memory let's just search all the
02:08possible spaces and then we'll figure
02:10out how to win because we know what all
02:12the winning games look like
02:13turns out you can't do this for a go and
02:15so now hope springs eternal again which
02:17is look at the very sophisticated set of
02:20techniques they use to win this game
02:21there's deep learning and then there's
02:24decision trees and then there's
02:26supervised learning look at all of the
02:28techniques and maybe now this time it
02:30really is the dawn of the generalized
02:33intelligence this is a massive massive
02:35win in the world of computer science but
02:38I knew right away just having kind of
02:41been around the block in academia that
02:43people were gonna start to clamor and
02:46define its victory and so try to make it
02:48narrow so like well it's not really AI
02:51because it used multiple techniques some
02:54of them which aren't AI and the funniest
02:57thing about that is that is like been
02:59the very nature of AI dialogue since it
03:02againi since the very very early and in
03:04fact like and so this is where Frank and
03:06I started talking in the hallway because
03:07we were both at AI schools um I was in
03:10graduate school in the late 80s at UMass
03:12and Frank was at Stanford and like the
03:15whole thing about AI was always if you
03:17actually could find a practical use for
03:19one of the techniques you cross it off
03:21the list of AI techniques and it's like
03:23no longer AI and so it's it's just it's
03:26part of the the world of like defining
03:28these things that makes it exciting and
03:30interesting but it also is because
03:32there's been this long history of like
03:34promises that weren't quite made you
03:38know made real and so that's what's so
03:41interesting like yeah a great example of
03:43that is so one of my summer internships
03:44when I was in college or actually right
03:46across the street from this building I
03:48was at an IBM Santa Teresa labs and I
03:50was working on an expert system
03:53development tool IBM was so convinced
03:55that there were going to be so many
03:56expert systems that we needed to improve
03:59developer productivity in creating them
04:01and what an expert system is it's a
04:03system that captures human knowledge
04:05so the overall process would be you find
04:07somebody really smart in an area a
04:08doctor an insurance adjuster an oil gas
04:11exploration expert and you ask them how
04:14do you found oil how do you diagnose a
04:16disease and you do hours and hours of
04:19interviews and you basically codify that
04:20in a decision tree which is a classic
04:23machine learning algorithm and you hope
04:24that you've asked enough questions and
04:27for the decision tree that they can
04:28start emulating the expertise of that
04:30human this was 1990 and it was going to
04:34be 10 years and we basically interview
04:36every expert in every field and there it
04:39like it's the some of those as
04:40artificial intelligence so why didn't
04:42that reality come about then so
04:44basically they ran into a wall of it
04:46didn't work which is it couldn't capture
04:49it couldn't nearly behave as effectively
04:51as a human so there's obvious things
04:53like there's always edge cases where it
04:55didn't you know it was sort of the
04:57exception to prove the rule the other
04:58thing was that keep in mind what
05:00computational power was back then we
05:02didn't have a lot of memory we didn't
05:04have a lot of disk we didn't have a lot
05:05of CPU cycles and so the rate at which
05:07we could do these calculations and the
05:09amount of storage we had like it just
05:12never worked it's also it's super
05:13interesting though because that that
05:15like I like one that was a favorite of
05:16mine in the 80s in graduate school was
05:18an expert system for chemotherapy which
05:21was actually done here at Stanford back
05:23then and the funny thing was it turns
05:25out like there are areas where there are
05:28a lot of rules and there are just too
05:30many for one person to keep in their
05:32head so you know with medicine it's an
05:34obvious kind of thing like you go to
05:35medical school you you're a resident an
05:37intern a fellow and as long as you see
05:39like a million patients you can figure
05:41out the right rules especially if you
05:44narrow it down to like what's the
05:46chemotherapy regimen for a particular
05:47forum of cancer but the problem is no
05:50one could remember all those things and
05:51get them right and like it's not a
05:54perfect classification because wow now
05:56you have to factor in the age of the
05:58patients if they do stop and there's
05:59always like another data point that
06:01leads to some reasoning under
06:03uncertainty need and yet there are
06:05systems that work even back then that
06:07yielded better answers and it's very
06:10easy to test two because you take all
06:11that data and then you go to Grand
06:13Rounds and present your approach and you
06:15have 50 doctors all at once saying
06:16that's a good approach not a good
06:18approach or state of the art and so
06:20immediately those systems were no longer
06:22at AI they were just like computer
06:24programs that helped you do that
06:26particular thing but it didn't stop
06:27people from like you know let's you know
06:29we need because we don't have computing
06:31power let's create a special computer
06:33that can interpret this list program
06:34even better and charge like 90 thousand
06:37dollars for it and hope that we can get
06:39the performance out of it
06:41let's make better editors for for
06:43prologue so that like you could have
06:45more rules and encode them even better
06:46yeah the out of the things and so
06:52that of course that what was so
06:54interesting about that is what what you
06:56know rhetorically to Frank like where'd
06:58that that ran right up against the PC
07:00revolution right exactly it's a turned
07:02out that what we didn't need was the
07:05Lisp machine which was this company
07:07called Symbolics and sort of the
07:08collapse of symbolics basically was the
07:10first nuclear winter in AI not only did
07:13venture funding completely dry out it
07:14was embarrassing to be a professor in
07:16that field for a long long time because
07:18it just didn't work just at the dawn of
07:19PCs right and so now we're on the
07:21upswing of that which is PCs have led to
07:24data centers right it's that supply
07:25chain like if you look at an x86 server
07:27in a data center exactly the the servers
07:31that Google is using to compute deep
07:33mine algorithms their pcs and so now we
07:35we do have a ton of computation and a
07:38ton of bandwidth and a ton of memory and
07:40we have this innovation in algorithms so
07:43the old approach was expert systems
07:44which is interview a human expert try to
07:46codify that knowledge the exciting thing
07:48about the go game is these algorithms
07:50are the opposite of that which is the
07:53algorithms are self learning so there's
07:56these techniques called supervised
07:57learning algorithms there's these
07:59progressively learning algorithms
08:00there's either generative algorithms or
08:03there's a whole class of algorithms
08:04where the computer is teaching itself
08:06how to play a better game every game
08:08without interviewing a go expert and if
08:11you look at some of the hacker news
08:12comments you can actually see this in
08:14the game play which is sophisticated go
08:17players are looking at the style of play
08:19and going that's weird they don't teach
08:22that at go school and so you can
08:23actually see this the style of game play
08:25is kind of from an alien intelligence
08:27because it's learning how to play the
08:30game by itself by playing lots and lots
08:32of games and so I think but what that is
08:35sort of like the the really big
08:37breakthrough that's happening now is
08:39that the history of AI was first a human
08:41writing an imperative program to play a
08:43game tic-tac-toe and then everybody
08:45thought that's that's not gonna work so
08:47let's write a program that simulates the
08:50way the human brain would play the game
08:51but it turns out in hindsight that if
08:55how the human brain works you can't
08:56actually write a simulator for the human
08:58brain and that was like 20 years of work
09:00yeah and then Frank like had a very
09:03important phrase that he just mentioned
09:05which is worth diving into which is the
09:06AI winter and so often what we now know
09:10in hindsight is that in technology
09:12revolutions the things that happen very
09:14early like don't yield things right away
09:16yield results right away but that
09:18doesn't make them the bad ideas or wrong
09:21they were just early right actually
09:23Dixon calls it and recently in a post he
09:25wrote the gestation period oh yeah and
09:27also one of the other things too is that
09:30most new big advances are not just like
09:33discrete changes in everything they're
09:35new like sort of prime mortal
09:37combinations of old things and so what
09:40is so fascinating about this go
09:42innovation is that it's not like they
09:44just locked themselves in a lab and
09:45invented the way to play go using like
09:47the latest newest deep learning
09:50technique that they created it's
09:51actually a whole range of of techniques
09:54and I was watching a great video which
09:56I'm 22 and we had another one of these
09:59dogs from Geoff Hinton who is clearly
10:01the the pioneer of deep learning
10:03yeah people come the father yeah and and
10:05he he actually spent a lot of effort in
10:08this interview on Canadian television
10:10explaining why the IBM Watson jeopardy
10:13playing machine was not it was not deep
10:16learning because it used all of these
10:18other techniques and and as if like it's
10:20like pure deep learning and then I
10:21personally just start to twitch because
10:23then it's like is it purely
10:24object-oriented programming is it early
10:26client-server is it truly cloud and like
10:30all of those definitions you know do we
10:31need an N ist dance Standards definition
10:34of deep learning so we're all on the
10:35same page share those definitions as
10:38they stand now and if it's relevant sure
10:40sure the evolution of that definition
10:41but I actually do want us to clarify
10:43like okay you're talking about expert
10:45systems AI deep learning machine
10:48learning how do we define each of these
10:51yeah so maybe let me take a whack at the
10:53taxonomy so at the very highest level
10:55you have artificial intelligence which
10:57is the combination of all experiments
11:00that we've run to try to program
11:01intelligence some of them will be trying
11:03to imitate human intelligence some of
11:05them will be trying to do things that
11:07are mathematically just interest
11:09and reproduce interesting results but
11:11aren't modeled on the brain or human
11:13thinking in any way also our senses are
11:15good area like being able to vision or
11:17speech exactly and then so that's
11:20artificial intelligence and subsets of
11:22artificial intelligence include deep
11:24learning so deep learning is a specific
11:25algorithm and data structure that
11:27attacks a series of problems it's based
11:29on neural networks which I was studying
11:31back in the Stanford days so you did you
11:33know cs2 21 intro to artificial
11:34intelligence week 1 expert systems week
11:362 neural networks and neural networks
11:38isn't modeled on the human brain
11:40yes it's very loosely modeled on the
11:42human brain all those sophisticated
11:43researchers will tell you there's a lot
11:44that it's very different but you know
11:46the basic idea is that the brain is full
11:48of neurons that are connected by axons
11:50and they're signaling each other and a
11:53neural net which is deep learning is a
11:55mathematical abstraction of that we have
11:57nodes they're connected in a network the
11:59connections have strengths and we can
12:01build these very interesting behaviors
12:03by using that data structure and
12:05iterating on the strengths between the
12:08so that's deep learning it is a specific
12:10algorithm and technique in data
12:12structure and it's on fire it is the
12:16heart of the go algorithms although
12:18interesting to point out it's an
12:20ensemble of techniques that's working
12:22for go let me just dive in really quick
12:24because I actually think there's a
12:25important distinction happening right
12:27now in neural networks and and it's a
12:29little bit of a split in the taxonomy
12:31neural networks aren't new in fact
12:33they're actually if you go back and read
12:34the 1956 Dartmouth summer Aoi conference
12:37they're actually mentioned in there is
12:39one of the first things that that group
12:41of people who basically invented the
12:42field this is like Marvin Minsky and
12:44those may Errol nets were in that paper
12:47because they were a theory a
12:48mathematical theory of the brain so then
12:50all for basically about 40 years it was
12:53like intro computer science you know
12:55like third-year undergraduate computer
12:57science to write your first neural net
12:58to play tic-tac-toe to guess a number
13:00between 1 and 100 like it was a very
13:03simple neural net what's happening right
13:05now and since the invent innovations of
13:07Geoff Hinton have been the ability to
13:08pile on a bunch of neural nets one on
13:11top of another and so maybe you dive in
13:13that like that because that's the big
13:14math advance and that's why you hear all
13:16about how many GPUs do you use to
13:18compute because it's this massive amount
13:20much of why GPAs are so important for
13:22so then let's break down the neural
13:24network taxonomy a bit further so
13:25recurrent neural Nets like let's define
13:27each so basically all of these
13:29adjectives on top of neural Nets
13:31recurrent long-term memory networks
13:33they're all enhancements of the basic
13:35idea and so what a recurrent net will do
13:38is try to feed back previous learnings
13:40into your current state it's probably
13:42how the brain works when I parse
13:43sentences I'm kind of keeping track of
13:45each word as I go along as opposed to
13:47throwing away what I learned in previous
13:49time frames long short-term memories are
13:52sort of the more sophisticated version
13:53of this which is I keep track of more of
13:55the history as opposed to just recent
13:56history again probably how I parse
13:58sentences very similar to human brain
14:00works of your cognitive psychologists
14:01have long talked about short term memory
14:03long term memory you're just gonna
14:04keeping that framework for helping
14:05learning okay so that's some of the
14:06neural nets yeah so one more adversarial
14:11so one really interesting thing is if
14:13you feed pictures into a neural net and
14:15you tune it you can defeat the
14:16categorization fairly trivially by
14:18introducing noise in the data and the
14:21really interesting thing is you
14:22introduce the noise you look at the
14:24resulting pictures humans still
14:25recognize the picture that's a dog
14:27that's a car that's a tree but the
14:29neural networks completely fail so
14:31trying to figure out exactly how I
14:33introduced noise and why they defeated
14:36the categorization algorithms is a super
14:38active area of research right they
14:39called adversarial networks here's a
14:41funny image that makes this round on
14:43Twitter every so often about like can
14:44you tell the difference between this dog
14:45and the bagel have you guys seen that
14:47yeah it's like and I think it'd be
14:49really funny like try to like make a
14:50neural net figure that out because they
14:52look yeah because they look very similar
14:54I just suddenly saw that yesterday I
14:56think the biggest innovation in computer
14:58science for me in the past 20 years has
15:00been the ability to look at all the
15:01pictures on the internet and find the
15:02cute kittens personally believe that
15:05that is a very high priority to brighten
15:08by not having to search for cute kittens
15:09is super helpful but when I was in
15:11graduate school it was the Cold War and
15:13like we had a giant lab at UMass that
15:15was all about like computer vision and
15:18it was about trying to pick the tank out
15:19of the desert and like figure out what
15:21tank and they had a literally a hallway
15:23like 30 feet long filled with micro VAX
15:26mini computers that you know that would
15:28grind away day and night literally our
15:30hallway was hot all of this it was like
15:33you feed the picture and then like 12
15:36you know yes there's a tank and and like
15:39that was the one and and it didn't know
15:41what it was doing was this like it was
15:43looking for the edges and doing this
15:45math to compare like you know because is
15:47it and then like you go you just go well
15:49here's a book with a picture of a tank
15:51take a picture of that and then it would
15:53go oh look a tank in the desert and it
15:54was just this massive undertaking and so
15:57now you've got the the ability to just
16:00like you can use every photo filter
16:02every infrared every sensor and overlay
16:05all of these different ways which turn
16:06out to actually be closer to how you
16:08might go and recognize something like
16:10your ability to tell the difference
16:11between a kitten and a picture of a
16:13kitten and an image on a computer of a
16:15kitten is important and that's why you
16:17can't fake out face recognition anymore
16:19and things like that so I wait computer
16:22visions might be the most advanced of
16:23all those sensory yeah illustrates one
16:26of the sort of this big trend so the big
16:27trend is the triumph of data over
16:29algorithms which is you try to make more
16:31and more sophisticated edge detection
16:32algorithms feature recognition now there
16:34isn't the big advance with deep learning
16:36was screw all that I'm not gonna try to
16:38figure out what Katniss is four legs and
16:42furry right I'm just gonna feed you a
16:43million pictures of cats and so that's
16:45sort of the triumph if you will of deep
16:47learning it's the triumph of data over
16:49algorithms right and it's and what's
16:51interesting is it's not it's not data in
16:53the way that we had about twenty years
16:55of like if you have a big enough
16:56database you then just use better query
16:59languages and better things to look it
17:01up in the database this is using the
17:03data to sort of build out a model of
17:05what the answer would be right right
17:08which makes for a super interesting
17:09challenge which is just sort of like how
17:12do you debug all of this stuff and for
17:14me that's like the most fascinating
17:15thing because you're you're you know
17:18like people get all like all the hoopla
17:20over self-driving cars can't replace
17:22because you never know and the thing it
17:24like if they're gonna be safer or not
17:25and the interesting thing is it's a very
17:27odd comparison because basically the
17:29self-driving car is going to use a bunch
17:31of machine learning techniques and other
17:32kinds of algorithms to essentially learn
17:35how to drive and make the best guesses
17:37at any given point and that's exactly
17:40what we do every day when we drive
17:43somewhere and and so it isn't gonna be
17:46this Wow we've now figured out the
17:48specifics and we have to paint the stree
17:49with different lines for the car to
17:50follow which is how they thought
17:52self-driving car was gonna be nor is it
17:54gonna be I have a database of all the
17:56highways and all of the cars on the
17:58so let me now look up when to change
18:00lanes how fast to go or anything it's
18:02not like this forest thing nor this
18:03brute force thing it's now this sort of
18:05emergent learning thing which is the
18:07whole deep learning model in the first
18:08book right right it's sort of inherently
18:10unda Bugaboo bee cuz you don't
18:11understand how it is exactly that it's
18:13making those decisions so contrast that
18:15with another well-known machine learning
18:17data structure which is decision trees
18:18in a decision tree you can actually
18:20examine the decision tree and understand
18:22why a system made any single decision
18:25very very easy to debug the bummer is
18:28decision trees don't catch you very good
18:29results and so these deep networks get
18:31you much better results but they're
18:33undable you don't really know all you
18:35can do is kind of feed it more data and
18:37run the models and say statistically how
18:39likely are you to drive correctly one of
18:41the areas that i'm super-interested
18:42always has been in in text like I worked
18:45on a word processor for a long time
18:47typing typing and grammar and things are
18:50all super important and that's a
18:52microcosm of the evolution of AI it
18:54turns out even if you just look at like
18:55autocorrect in word there's a whole
18:57history of AI even though it's like
19:00ultimately 20 lines of code that we
19:03write but what's interesting is it also
19:06points out almost this chasm in the
19:08academic world about how to describe a
19:10solution because there are there's a
19:12very long history in algorithmic
19:14decision tree like history in the world
19:16of natural language processing where you
19:19look at a block of text you know we know
19:21how to diagram sentences to find parts
19:23of speech and so probably since about
19:251956 like people have been working on
19:27the ability to algorithmically figure
19:30out text and they all thought it would
19:32just be a couple extra years of work to
19:34then take the diagram sentence the data
19:36structure and turn it from English into
19:37French you know or to turn it into from
19:40English into a concept right and it
19:43turns out it knowing the structure of a
19:45sentence doesn't help you do either of
19:47those things it doesn't and so along
19:50comes deep learning and the idea is oh
19:52well if you have enough text in French
19:55you can basically find a way to turn it
19:57into English without knowing French or
20:00algorithmic lis diagramming exactly
20:02takes over the algorithm so it actually
20:04it actually works but debugging it is
20:06really hard which sort of freaks out the
20:08algorithmic people because wait what if
20:10there's a mistake oh well then just go
20:12get more french text and start over
20:13again and but then it turns out you can
20:15probably do a better job if you apply
20:18some of the linguistics to it and think
20:21about it in advance because you're
20:22always gonna get like a probability you
20:25know 80% choice well exactly and I would
20:27think that in case of natural language
20:29in particular that is the only way to
20:30resolve the ambiguity problem in an
20:33efficient way like you have to have some
20:34sort of approach that isn't just purely
20:36one or the other to get people an output
20:39that makes sense to them cuz that is the
20:40whole point of natural language let's be
20:41natural yeah and I think this is where
20:43we're gonna see the next big
20:44breakthroughs it won't be one technique
20:45in isolation just like the go algorithms
20:48one on a combination of techniques the
20:50other techniques by the way you
20:51mentioned briefly something so for
20:53natural language processing a very
20:54natural thing to do would be to do parts
20:57of speech tagging entity resolution
20:58entity resolution is when I see sono
21:01talk she in a email is that a person is
21:03that a place name is that a store name
21:05is that like what is that more precisely
21:07I think entity resolution also is when
21:08you have variations of that name like
21:10sown just exact right did all that right
21:16we can sort of figure out what is the
21:17root verb is was B are they all stem and
21:20so let's use those techniques in
21:22combination with deep learning and I
21:25think that's where we'll see the next
21:26big wins and that's an important point
21:27about just innovation in general which
21:29is there will be massive innovation and
21:32in fact I I fully expect to see pure
21:34deep learning approaches to translation
21:37to image recognition which is you know a
21:39Jeanette already is that but all of them
21:41the computer scientists will continue to
21:43push sort of this pure-play approach to
21:45in innovating and there'll be new neural
21:47net algorithms that do and they'll keep
21:48doing and and actually systems will
21:50arise that are pure deep learning to
21:53solve all these things because this is
21:54how you win a PhD right but if you're a
21:56product manager and an engineer building
21:59a product you actually don't care if you
22:02win an award for the purely the most
22:04pure algorithm and that's actually been
22:06the history of all innovation in
22:08computer science has been the products
22:10always represent a little bit of a
22:13combination of some known things
22:15breaking those rules of the new thing
22:18and then the new things and if you think
22:19about to make it work the Internet
22:21itself is not like the purest form of
22:23networking it's actually kind of like a
22:25giant series of hacks and the way I
22:27always think of it is in a perfect world
22:29there are no caches and so therefore
22:32everything is so well architected that
22:34you don't have a cache of anything
22:35anywhere because caches are just hacks
22:37and then you realize well the Internet
22:39is one giant cache of everything create
22:41a couple multi-billion dollar companies
22:43right right then you come along and you
22:45say like wow to really make the internet
22:47work we actually need cash companies and
22:48I think that everything is gonna have
22:50like elements of deep learning and and
22:53then people building products that have
22:55to solve problems are not gonna be shy
22:57about hacking deep learning taking
22:59there's like even my favorite one was
23:01just the the Google inbox did this email
23:04reply and I remember the cynical
23:06comments about it like what it does is
23:07it machine learns you're a bunch of mail
23:08and then it basically suggests what to
23:11use as a reply to a mail message which
23:13is kind of a cool stupid computer trick
23:14but it doesn't just reply it gives you a
23:17choice of two and the obvious cynical
23:19comment is well that's dumb why doesn't
23:20it just pick the right choice and it's
23:23like well because a it doesn't know and
23:25be like why not show you a couple
23:27choices if you're just being practical
23:29about it there's just an opportunity to
23:31do a better job so you guys have
23:32definitely convinced me about why the
23:34product-driven approach to some of these
23:36solutions is so practical for lack of a
23:38better phrase but I'm still not
23:40convinced about why this time is
23:42different because we started off talking
23:43about how Oh people talked about you get
23:45to this point some algorithm beats like
23:48a game of some sort and then next the
23:50next wave of AI is about to happen how
23:52do we know that this time it truly is
23:54different and how is that gonna actually
23:56look we don't we don't know but here's
23:58some reasons that people are excited
23:59about the go victory so one as I pointed
24:02out the search space is so big that
24:04traditional techniques just couldn't
24:06work so they break through in terms of
24:08how to search that space they use
24:09existing techniques like Monte Carlo
24:11tree search to prune sort of the
24:13candidate sub trees that they weren't
24:15going to explore right and to be clear
24:16mine and Carlo tree search is not any
24:18kind of deep learning technique that's
24:19right it's not a deep learning cat right
24:21it's a traditional AI technique many
24:23understand is a hero in this yeah I
24:25think that's right so one you couldn't
24:27brute-force a search space two you've
24:29in deep learning that are frankly
24:31unanticipated and just search the
24:32internet for deep learning systems there
24:34are systems where robots are learning to
24:37cook food by watching YouTube videos of
24:40there's systems that can take photos and
24:43paint them in the style of Renoir or van
24:45Gogh there's algorithms that can create
24:47paintings that are indistinguishable
24:48from human created paintings the
24:50successes are many and varied and
24:52involve things that you would think
24:54require creativity a uniquely human
24:56thing I'll be clear the existence of
24:58those successes is not the reason alone
25:00it's a fact that there's occurred
25:01because of the ubiquity of data
25:02yeah actually I think that that that is
25:05like in a sense the ultimate reason why
25:06all these things are working is
25:08basically because of cloud computing the
25:11scale of the architecture of cloud
25:13computing and the internet that brings
25:15all that data like when I was in college
25:17I my adviser was sort of the father of
25:19information retrieval in order to do
25:22research on search you basically got a
25:24box of tapes from the New York Times
25:26that had the contents of you know 150
25:29years of New York Times articles and
25:30probably 25 people did PhDs out of that
25:33one lab searching that one corpus and
25:36you think about that and you're like
25:37well that's just stupid well there were
25:39two problems one you there most the
25:40other data in the world wasn't on tapes
25:42like that that you could get two and two
25:44even if you could you know lab could
25:46afford like the storage to put it all on
25:49for each student to be able to do their
25:51experiment and now like anyone learning
25:53to do anything in computer science has
25:55access to all of the world's information
25:57even if they just used Wikipedia yes as
26:00as their sole source more for that's
26:03that is a 10,000 times what the average
26:06student had back then and you have the
26:08compute power that's essentially free to
26:11do all of the work that you you could
26:13keep retrying deep learning keep doing
26:14different things and iterate all in some
26:17finite practical amount of time that
26:19makes this all like it's it's happening
26:21it's here it's now it's real it's not a
26:24theory by one lab that can identify one
26:27tank in one picture just on looking at
26:30Wikipedia to do entity resolution you
26:32can disambiguate this torture sentence I
26:34can't remember what entrepreneurship it
26:36with me but it's a beautiful sentence if
26:37you're trying to figure out like what
26:39are all these elements Paris Hilton was
26:43the Paris Hilton listening to Paris
26:45Hilton that's hilarious so there's a
26:47person there's a city state there is a
26:49hotel and then there's an album all of
26:52those are perfectly disambiguated in
26:54Wikipedia there are entries where every
26:56single one of those so think of the leg
26:59we've now have compared to when poor
27:01Steven was transcribing from tape so
27:04huge enablers data the way the cloud
27:07computing the scale micro services
27:09architecture you know the way
27:10applications are being built there's so
27:12many different things on that so then
27:13one last question how is it gonna leave
27:15the province of purely logical things
27:17because at the end of the day go is a
27:18logical game and it's very codify about
27:22how you then go to the next leap to
27:24intuitive decision making and decision
27:27making under uncertainty in general so
27:29let's go back to go because this is how
27:31we all started so read google's blog
27:33post on the go algorithm and the blog
27:36post basically starts with why did we
27:37pick go so one it was a huge search
27:40space but to the best masters that go
27:42have always been because you can't
27:44exhaustively search the space driven by
27:46intuition leaps of intuition and so part
27:50of the promise of go why people are so
27:51excited about this victory is maybe this
27:53is an example of that which is because
27:55it wasn't mathematically searchable that
27:58you needed to develop strategies that
28:00were based on intuition that algorithm
28:03is playing go in a way that is
28:05unrecognizable to humans it feels like
28:09so I think maybe this is the vector
28:11which is these deep learning techniques
28:12which we can't completely characterize
28:14and describe much less debug are leading
28:17to these flashes of insight and they
28:19might not be human insight it might be
28:21like artificial intelligence new kind of
28:23intelligence yeah yeah but just be
28:25careful like this Christmas don't buy
28:26like a fluffy cute thing that shows up
28:28and says it can fix things in your house
28:34really cool present that everybody wants
28:36to get for all their friends funniest
28:43thing I was someone was making a joke
28:45that no it's kinda it's gonna get start
28:47because they were pissed but I'm gonna
28:48watch super play Super Mario Brothers
28:50all day because Frank share this awesome
28:51video yeah this is part of my tweet
28:53storm if you haven't seen it so go watch
28:55this awesome YouTube video of how
28:56ughter algorithm learned to play and
28:58then completely Ace Super Mario Brothers
29:01what's find it so we'll put it up on the
29:03link with this podcast okay alright guys
29:05well thank you and that's first of many
29:07conversations cuz I still do not quite
29:09understand the full taxonomy but I think
29:12that's part of the point here is that we
29:13have history crashing in with the
29:15present and trying to figure out what's
29:17coming next yes thank you thank you