00:00hi everyone welcome to the a six in Z
00:03today's episode is on machine learning
00:04deep learning and AI the role of
00:06university research and making AI
00:08production ready for industry and games
00:10the conversation moderated by ASIC since
00:13the operating partner Frank Chen took
00:15place at our recent a6 and Z summit
00:17event and it features Cameron Schuler
00:20former entrepreneur operator and
00:21executive director of the Alberta
00:23innovates Center for machine learning
00:24now known as the Alberta machine
00:26intelligence Institute it's based out of
00:28the University of Alberta Canada as
00:30department of computer science and
00:31longtime listeners of the podcast will
00:33also recall we did a podcast Roadshow in
00:35the UK with researchers there as well
00:37the Institute does our indian all things
00:40machine intelligence in 2007 they solved
00:43Checkers which was a long-standing
00:44challenge for AEI researchers and in
00:472015 they produce a first AI agent
00:49capable of playing an essentially
00:51perfect game of heads up Limit Hold'em
00:52poker you look at some of the big
00:54successes that the Institute has had in
00:57machine learning a lot of them sort of
00:58cluster around playing games so you guys
01:01solved checkers before chess you have
01:04pretty good solution for poker you were
01:06the first ones to reinforcement learning
01:09on playing Atari games right which
01:11Google later popularized yes and you
01:13know obviously Google's gotten a lot of
01:14attention recently with the alphago in
01:16and then Starcraft will be the next big
01:19battleground I can't wait to see the
01:21human versus Google AI on Starcraft so
01:25maybe take us back and talk to us about
01:27why do AI researchers gravitate to doing
01:30research on games and is it is it just a
01:33toy that's a good question we do like to
01:36have fun just to be clear and we're in
01:37university so I did a couple of comments
01:39actually so Wendy mine got bought half
01:42the people there were actually Canadian
01:44trained and roughly 20 or 25% were our
01:46students so they did take things like
01:48Atari they took that with them
01:50alphago I think was roughly 45% of the
01:54research cited on their alphago paper
01:56came from the University of Alberta so a
01:58very strong connection and rich sutton
01:59who's the father of reinforcement
02:01learning he literally wrote the textbook
02:03he could get the second version off his
02:04website even today he was one of the
02:06supervisors of David silver so back to
02:08your real question which is why games
02:12so if you think about how we learn as
02:15humans games are actually pretty
02:16important our goal is to have computers
02:19make good decisions and ambiguous
02:21environments games have fairly low risk
02:24nobody dies usually they're great petri
02:26dish to actually do discovery and so
02:29checkers is largest game to ever be
02:31solved it's 5 times 10 to the 20th and
02:33you cannot beat it you know why they
02:35play to a draw or a they win every
02:37single time I was in 2007 science made
02:40it named it one of the top 10
02:41discoveries of the year the Atari thing
02:44was I won't say it was a lark but it was
02:46you know a bunch of us grew up playing
02:48Atari games and Mike bowling said hey
02:50let's just see what we can do here in a
02:52lot of we focus on is unsupervised
02:54learning so if you take a look at deep
02:56learning deep learning has been
02:57incredible in terms of the delivery it's
02:59had for commercial applications but if
03:01you look at deep learning it's still
03:03labeled training sets right which has
03:05some constraints around it if you take a
03:07look at what we do on the unsupervised
03:10learning it literally was the bit of
03:12background reinforcement learning is you
03:14take a system and you actually have it
03:17you give it the ability to modify its
03:19behavior to maximize its reward and so
03:22one of the neat things so if you take a
03:24look at poker poker to me is one of the
03:27biggest advancements and on the back to
03:28touring the psyche and AI and the reason
03:30being is if you play chess or checkers
03:33or go you can see the whole playing
03:35board you can't tell what the other
03:36person is thinking but you can actually
03:37see it do all your scenario analysis
03:39think about playing chess in the dark
03:41you can't and that's poker most you're
03:43gonna have 15% of your information
03:45available at one point in time so you
03:47need to infer what's going on with the
03:49bidding that program that actually won
03:52heads up Limit poker it's something like
03:5426 terabytes it's pretty enormous you
03:56can download it but good luck you can
03:58play against a to online by the way but
04:00in that particular case you have a whole
04:03bunch of obfuscated information right
04:05because you can't see everything versus
04:07all the other ones are like that so
04:08Atari was another one where you could
04:10actually have a score and it could just
04:12try playing it and so that was kind of
04:14neat just in terms of we're actually
04:15meeting some of the deep mind algorithms
04:18that they have right now or roughly four
04:20times faster I believe but just it's the
04:22ability to go and play and finally the
04:24game and eventually figures out good
04:25gave your bad behavior and it's truly
04:27unsupervised learning this couple games
04:28that it actually can't play I don't
04:30remember what they are right now but
04:31nonetheless out of I think it was about
04:34ninety games we can the system will play
04:36about half of them and then what about
04:38sort of criticism that even if you've
04:40mastered a board game you have sort of
04:42this question of so what I didn't really
04:45want to play go I wanted to book a plane
04:48ticket by talking to somebody or I
04:49wanted to have a good recommendation
04:50come up when I'm shopping are there
04:52really lessons that we've learned sort
04:55of mastering checkers or poker that are
04:58gonna be applicable to real-world
04:59systems that broad sets of people use so
05:02the answer is there are and what it is
05:04it's around decision making right so
05:06you've got a system that's trying to
05:07make a decision what is the best
05:08decision you could make I mean capital
05:10markets and things like that there's
05:11obviously applications you can you can
05:13move across but it right now
05:16you just can't engineer systems big
05:17enough to do a lot of that stuff they
05:19have to learn on their own and that's
05:20really unsupervised learning piece comes
05:22in so the ability to learn how to do
05:24better unsupervised learning is
05:25ultimately where this stuff is gonna
05:27come from I think Amy initially got its
05:29grant money from the province yes and
05:32then you've also forged partnerships
05:34with corporations so tell us a little
05:35bit about how that's work to they fund
05:38specific research are they just do they
05:41send their employees how does that work
05:42so the answer is yes so we are
05:48industrial partners tend to be really
05:49big one of them's eighty billion one
05:51hundred eighty thousand heads another
05:54ones 40 billion hundred thirty thousand
05:55heads in some cases they do send people
05:58so we actually have some visitors that
05:59are learning how to view unsupervised
06:00learning they do pay for research and in
06:03we have pretty flexible models so we're
06:05one of the only places on the face of
06:07the earth that negotiate its own IP I
06:08mean if you have an open IP policy like
06:10some Canadian universities do when the
06:12professor can go and do what they want
06:14most other ones have some sort of handle
06:15we actually negotiate our own it only
06:17took five years to get it in place it
06:18was easy process but you know it's
06:22something it's something for apps that
06:23allows us to be a lot more free so when
06:26we're taking a look at our development
06:27model it's very voice of customer driven
06:29there are some cases where they're gonna
06:32get all the IP so we have the ability to
06:33direct consulting we've got all sorts of
06:36methodologies to deal with business in
06:37some cases we end up owning the IP
06:39we want to commercialize and they become
06:40a partner so it's it's really a broad
06:42subset or a broad swath of what we can
06:45do and looking back it's sort of the
06:47interactions you have so you're rooted
06:49in a university you have these corporate
06:50partnerships what's worked and what's
06:53not worked as well so there's a lot of
06:55threats to internal teams so I think you
06:58know when we when I look at what we do
07:00and you talk about data science they're
07:02they're fairly different my random
07:04financial planning and analysis group
07:06for into it we did lots of data science
07:08but it's nothing like what we do with on
07:09the machine learning side so we've
07:11certainly had cases one of our
07:13industrial partners where their internal
07:14team was incredibly threatened and so we
07:17only did part of the project but it
07:18actually set up a good foundation for
07:19them to do the rest there's lots of
07:21cases of companies having bad
07:23experiences with all sorts of
07:25universities that's always an impediment
07:26as well so the ability to make that
07:29seamless and take a lot of the friction
07:30away from it works well it's like any
07:32other project you need shared vision you
07:34got to know what you're building you got
07:36to know what the outcomes are and all
07:37those other things so it starts off with
07:38a good project and we'll bring in people
07:39their domain experts because I believe
07:41very strongly in that we'll also bring
07:43in people in the project management side
07:45that really contrive a project and we've
07:47even hired staff to work on stuff versus
07:49using students so one of the things that
07:51we've been watching over the last five
07:53years is if you think about the anchor
07:55tenants of the tech ecosystem so you got
07:57Google Apple Facebook Amazon they are
08:00clearing out the AI machine learning
08:02departments of universities all over the
08:04place right so uber shows up at Carnegie
08:06Mellon Dean says well I'll take them all
08:09right who who wants to come and so what
08:12are you thinking about the long-term
08:13implications of this sort of hollowing
08:16out of computer science departments will
08:18there be anybody left is this just a
08:20shift in the way that we're gonna do
08:22fundamental scientific research which is
08:24instead of research grants and NIH and
08:27NFS it's going to be Google and Apple
08:30and Facebook funding it is this a threat
08:33how do you think about it so I think
08:35we're the only machine learning group
08:36that hasn't been touched we've had the
08:38same professors for 10 years other than
08:39the ones we've recently added more than
08:4110 years it's been a pretty constant
08:43group so I think I think it is
08:45problematic part of the discussions we
08:46have it's you know if you take a look at
08:48and we're having this discussion earlier
08:50so when apps came along and beginning
08:53them so high-demand luck people getting
08:55paid tons of dough it goes away pretty
08:56quick because you can learn fairly
08:58quickly so the analogy I use is there's
09:01roughly three million people a year that
09:02play football in the United States
09:04there's less than two thousand and
09:06professionals it's a very it's quite the
09:09disparity between them
09:10you can't take somebody in teeth you can
09:12take them and teach them how to use
09:13machine learning but if you want to
09:14solve the really really difficult
09:15problems it's a it's a career again so I
09:18think if you start losing more and more
09:20people from academia weird there's
09:21another Canadian professor named Joshua
09:23NGO at university montreal who's
09:25consciously decided not to leave for
09:27this for that very reason and our guys
09:28are like that too so I think ideally if
09:31you ran things like Bell Labs or GTE you
09:33could do a lot of really interesting
09:34research my background was Capital
09:36Markets is way too many MBAs in the
09:38world and I'm one so I can definitely
09:40make fun of them but you know what
09:43happens is what have you done for me
09:44lately right you can't plan for 10 or 15
09:46years down the road so eventually you're
09:48gonna hollow out all the creativity
09:50right it's not going to exist because
09:52you need to be able to have those
09:53product roadmap so where do we need to
09:55be in five 10 15 20 years 30 years you
09:58really have that vision for the future
09:59and if you look at NASA and universities
10:02they've traditionally funded stuff that
10:03nobody else can attach and I'm afraid so
10:06I think Google's a little bit of an
10:07exception and you know from my
10:09perspective their risk is 80% of the
10:11revenue is generated by advertising they
10:13are going to get disrupted in that at
10:15some point you better find the next
10:16thing you can generate all that cash off
10:18of so you need to reinvest so they have
10:19a bit longer term view I believe in
10:22patents my medical device company has
10:23patents so it's not that I did I believe
10:26that you know you can't profit from that
10:28you can't be beneficial to society but I
10:30think I think there's a huge problem
10:32where there's a lot of risk that to get
10:34the people that can solve the difficult
10:35problems or opening what to train them
10:37Shanno argument would be if you look at
10:39Google or you go to Facebook these are
10:40run by executive teams that have very
10:43long term vision right the other things
10:45that suck is running is planes that will
10:47beam the internet to rural areas of
10:50third-world countries so very long-term
10:54thinking executives and so if artificial
10:57intelligence is as important as we all
10:59think it is we might as well have these
11:01very long-term thinking executives fund
11:04and that would be a reasonable
11:06supplement or replacement for what
11:08universities are doing so respondent
11:10yeah so I have five two points related
11:12to that so one is you've named two
11:14companies and there's maybe five doing
11:16us maybe some of the Chinese companies
11:17are so we don't really know idea right
11:19yeah we don't really know what a lot of
11:21them are doing and not just necessarily
11:22as transparent which also means they're
11:24probably quite disruptive but the second
11:27part of it is I think the people who
11:28could really win at a are the games
11:30company it's the Electronic Arts and
11:31groups like that and they're where's my
11:33bottom line right so I Bioware the
11:36company I went to grad school with one
11:37of the founders they did interesting
11:39things Bioware still in Edmonton but I
11:41find one of my former staff from into
11:43what I've talked to me says you know
11:44unless I control that case for what it's
11:46what's going to happen tomorrow so I
11:48think the risk is it's very concentrated
11:50at that point the second piece of this
11:52there's other companies that could win
11:54in this space they don't have the vision
11:56so I think I think you need that ability
11:59to dream and ability to execute on it
12:01without the risk of failure being the
12:03end of your career and never be able to
12:04work again why don't you share some of
12:06your favorite projects going on at Navy
12:07right now so we have quite a few we have
12:10meerkat which is social network analysis
12:12do it's got a temporal component of it
12:14so it's both data relationships we have
12:16another one called pfm scheduling and
12:18it's around workforce optimization in
12:20healthcare we just launched it last
12:22month well I think about machine
12:24learning I think about automation and
12:25optimization any place where you can
12:28actually apply those things we're
12:30working on something in those spaces
12:31usually I do think that anything data
12:35centric is the crown jewel of any
12:36company and I mean that in that so my
12:39device company for example we catch a
12:42capture about a gig of data in ten
12:43seconds we post process it now think
12:45about having hundreds of thousands of
12:47people you find a new pathology now you
12:49can go back through and do it again
12:50right so there's lots of things that
12:51we're touching in lots of different
12:52spaces so we've actually met recently in
12:55Toronto and that had a machine learning
12:56event and at that event the government
12:58kindly awarded and I know if they were
13:00metals or I haven't seen them something
13:02yeah to some of the deep learning Giants
13:05that sort of the who shoulders were
13:06standing on now so Joshua you mentioned
13:08Jeff Jeff Fenton Richard son who works
13:11in Alberta I want you to talk a little
13:13bit about Richard and reinforcement
13:15learning because that's super
13:16interesting and relevant
13:17to a broad audience but I'm gonna go a
13:19little off the strip here let's talk
13:20about because you have a tip medical
13:22device company so one of the things
13:23geoff hinton said during his remarks at
13:26the awards ceremony was that we should
13:29stop training radiologists right now you
13:31say look I should stop training
13:32radiologists it takes five years to
13:34train a radiologist and in five years
13:36deep learning we'll get better results
13:38then trained radiologists so we can stop
13:40training them right now so I put that on
13:42Twitter and there was a lot of hate mail
13:46already on that all the lines of just
13:50wait until you get sick and so you have
13:54a medical device company you have neural
13:56network technology inside
13:58that's analyzing the giga data what's
14:00your take on this well there's a couple
14:02of components I think probably for 25
14:03years computers could have done a better
14:05job in terms of imaging using training
14:07data from radiologists right it's
14:09something that that and I don't want to
14:11diminish the value of radiologists and
14:13the weight I mean if you think
14:15strategically why have they been able to
14:17hold on where I live you have to have
14:20really all just to get reimbursement
14:21this is a good luck right all right I
14:23mean I think follow the money as they
14:24say yeah definitely I think I think
14:26there's a lot of I could see technology
14:28changing a lot of things I hope it does
14:30and it is transition from something that
14:32is currently done by humans to something
14:35that's automated and everyone's
14:36threatened by that you get to a position
14:39where I think I think looking at stuff
14:40as human augmentation is really where it
14:43needs to be how can I do my job better
14:44and if you if things like that if you
14:46talk if you bring them back to patient
14:48POC toes will go of course I'm here for
14:50patient care that's why it got into this
14:52field if you're in it for money you
14:53probably didn't make it through right so
14:55I think I think those cases are there I
14:57did ask Geoff a question earlier nice
15:00then this is an FDA issue so you take
15:02data you translate it and now you give a
15:05result on the other end so the reason
15:08why the FDA has such stringent software
15:10issues or software controls is because
15:12of the Canadian company invented
15:14radiotherapy and they killed some people
15:16not intentionally so I said okay you
15:18have to be much more rigorous you have
15:20to actually give us the causality of
15:21what happens deep learning you can't do
15:23it and pray that's their black box that
15:25their black boxes would you fly an
15:26airplane with the black box right I mean
15:28there's any time you look at a regulator
15:31response to me wise well humans don't do
15:34it very well either exactly when are
15:36also black boxes we are right but I'm
15:38actually prettier error-prone black
15:39boxes that's yeah definitely so really I
15:41think I think it would be nice to see
15:43the FDA specifically adapt to these
15:46things but I think those challenges
15:47around that so technology in most cases
15:49is probably better than humans well
15:51let's talk about Richard so he invented
15:53this branch of machine learning called
15:55reinforcement learning and tell us a
15:57little bit about his background he
15:58didn't much more than that and what he's
16:00interested in these days and then we'll
16:02take questions from the audience right
16:03after this yeah so rich is a fellow of
16:05the Royal Society of Canada he is an
16:07American but also a Canadian citizen
16:08he's a fellow of triple AI his winner of
16:11the president's award of the
16:12International neural network Society had
16:15no idea until yesterday 39 experience in
16:18reinforcement learning and where it came
16:19from is he actually has a psychology
16:21degree so you couldn't get competing
16:23science degrees when he went into it
16:25according to the island sue for AI
16:27semantics scholar he's a host highly
16:29cited researcher in reinforcement
16:30learning 11th month in fluent shil
16:32researcher in all of competing science
16:34and his textbook on reinforcement
16:35learning was ranked as the single most
16:37influential publication in all of
16:39computer science so that's a bit of
16:41background on rich so the Rich's goal
16:43has always been to solve AI he started
16:45out in high school he writes a letter to
16:46Marvin Minsky most people probably know
16:48saying how do I do this and he got a
16:50letter back saying good luck many people
16:53would blame Marvin Minsky for us not
16:55taking deep learning more seriously
16:57right because he was on this other
16:59branch he's like the symbolic AI is
17:01definitely the way to go that deep
17:03learning thing is did ed ed and so like
17:04generations of students got discouraged
17:06from deep learning yeah I mean I would
17:09agree so so rich really came from
17:12psychology you know how do we learn as
17:14humans right you learned something about
17:15you burn your hand you transfer that the
17:17bunch of things that he's coined so one
17:20is on policy learning and off policy
17:22learning so on policy learning is I
17:24learned by doing off policy is I learned
17:28by see what Frank does that sort of
17:30stuff so it's kind of taking that field
17:32and adapting it so your board computers
17:34for good behavior and penalize them for
17:35bad behavior and you can also learn
17:37under normal circumstances its normal
17:40operating conditions so let's say you
17:41have a nuclear facility you really don't
17:43want to have a bunch
17:44bad things happen to figure out what's
17:46bad right so the ability to learn in
17:48real time and adapt not using training
17:50data so you need to be able to take that
17:52stream in it's also another field come
17:53online learning take that stream in
17:55interpret it forecast and then go back
17:59temporal difference learning which is
18:01you can actually learn let me put it
18:03this way you can learn from a gas and
18:06what I'm saying is if you have a
18:07forecast it's likely informed if you
18:09include that as well as your historical
18:11data you actually get better results
18:13fantastic all right questions from the
18:16a lot of the AI projects or academic
18:19right now yep what is it gonna take to
18:21make it more industrialized like more
18:23companies can use it or how does it
18:25become in that format actually I'm glad
18:28you brought that up so the industrial
18:30project even in capital markets a lot of
18:31stuff we've done you could have done 20
18:3325 years ago deep learning right it was
18:35really distributed computing to made the
18:37big difference in that part of it is
18:39getting the academics interested enough
18:41in solving real-world problems and it's
18:44kind of like bringing doctors back to
18:45patient care I think that's important
18:47you know my background is industrial I'm
18:49not an academic I don't teach students I
18:51do mentoring and sometimes talk to the
18:53business schools and stuff like that but
18:54for the most part it really is you know
18:57what would you like the world to be and
18:59can you help make it that so that's the
19:02way I see it violate a moderator rule
19:04and also answer the question yeah so
19:06I've always thought about it is look
19:08it's always about people processing
19:09technology so we need more training
19:12people weed tools that make the
19:15programming of these artificial
19:17intelligence systems easier right right
19:19now PhD requires pretty much and then
19:21we'd be better processing a great
19:23example of sort of processes assisted by
19:25technology as you if you look at
19:27something like FB learner flow which is
19:30automation workflow system for
19:32artificial intelligence they've gotten
19:35that 25% of their total software
19:39developer universe is writing deep
19:40learning 25% right so obviously it's
19:43gonna take most companies a long time to
19:45get to that point but that's what we're
19:46gonna need is people processing
19:48technology to make it as everyday as
19:50programming a sequel database is today
19:53right and we've just started that
19:54journey back sort of a follow up about
19:57get this more into the mainstream
19:59there's a lot of material out there
20:01about machine learning there's a lot of
20:03buzz about it but in many companies
20:05particularly those that are not sort of
20:07in the heart of the tech sector the
20:09executive decisions are made by people
20:10who just don't understand it
20:12I'd love to hear you talk a little bit
20:14about your suggestions and
20:16recommendations of required reading sort
20:18of machine learning for dummies or you
20:21know artificial intelligence 101 what
20:24where would you point the c-suite to get
20:27smart about this technology first of all
20:29I'd add diversity and make sure that all
20:31pale mail and stale and I actually mean
20:34that I very much like the rock that we
20:36walked into I think part of the part of
20:39the challenge really becomes that pretty
20:42much everyone has missed the boat on an
20:45AI strat like everybody there's a
20:47handful of companies Google Microsoft
20:49Facebook that are doing it but everybody
20:52I mean I talk to companies with 1520
20:54billion in revenue legal we don't have
20:56an AI strategy right so there's I think
20:59it's it's far deeper than that if you
21:01know I think things like MOOCs are
21:02interesting this is kind of getting a
21:04peripheral knowledge but I don't think
21:05that connects it well enough to to what
21:08you actually need to do to apply it
21:09there's there's a big disconnect there
21:11right so one of our industrial partners
21:14is a financial institution that's truly
21:15started something like a Bell Labs and
21:18that's the sort of thing they need to do
21:20they report directly to the CEO they
21:22have a bit of 80,000 heads 100 billion
21:24market cap I think there needs to be a
21:26reinvestment going back to my tie read
21:28on NBA's aides to be a reinvestment in
21:31true R&D that's curiosity-driven
21:33but maybe five or ten years is gonna be
21:35an application because we think there's
21:37going to be something there and you
21:38could get disrupted along the way but I
21:40think I think it's actually having the
21:41cojones to stand up and say I'm gonna
21:43throw a bunch of money at this and it's
21:45going to be meaningful because the
21:47person that replaces me maybe one or two
21:49generations so now is going to benefit
21:51from it you know in the c-suite you're
21:53afraid to make a decision because you're
21:55a public company you're gonna get fired
21:56if you look at the M&A side within five
21:58years everyone does M&A is gone like
22:00everybody just that's how it works so
22:03again it's the same sort of thing
22:04they're they're relatively risk-averse
22:06that's why they work for big companies
22:07that's why there are big companies
22:08innovations I don't care what they say
22:10not probably part of their culture it's
22:12part of their lip service but truly
22:14interviewed you know another therapy I
22:16was working on in the cancer side this
22:18one company which has 60% market share
22:20was gonna probably have to destroy that
22:23and if the product didn't succeed they
22:24would have destroyed the company this
22:26weren't willing to take that sort of
22:27risk right it I don't have a great
22:29answer other than they need to throw
22:31some money at just some basic
22:32researchers to say go play have fun do
22:35some applied things right so we have
22:37data sets and a lot of these data sets
22:39were never set up to use we have data
22:41sets we have people we can get problems
22:43from the different groups but also do
22:45some fun stuff where you're actually
22:47truly curiosity-driven going can we
22:49actually solve poker or something like
22:51that right to recommendations where you
22:53one is a shameless plug so I wrote a
22:55primer on artificial intelligence for a
22:57general purpose audience that you can
22:59find on Vimeo so you can just search for
23:01andreessen horowitz primer on artificial
23:02intelligence it's a 40-minute video and
23:05then the book this is not so shameless
23:07plug that I'd recommend for a general
23:09audience is a book called artificial
23:10intelligence what everyone needs to know
23:12it's written by Jerry Kaplan who I've
23:15worked for a long time ago at a company
23:17called go we were trying to build the
23:19iPhone in 1991 turns out we were a
23:21little early but Jerry's gone on to AB
23:23this great career as an entrepreneur and
23:25lately he's gotten super interested in
23:27artificial intelligence and it's Oxford
23:30University Press asked him to write the
23:31book question in the back hi what do you
23:35think is gonna be the role of genetic
23:36programming like genetic genetic
23:38algorithms to develop rather than
23:41thinking of about it ourselves so it
23:44really depends on the space if you take
23:45a look at capital markets the domain
23:47spaces that mean the dimensionality the
23:49data is so huge the genetic algorithms
23:50never gonna get there I'm not a
23:53technical resource but so I think I
23:55think it really depends on the
23:57application the way I see the world
23:59moving certainly is more mobile and if
24:01you are relying on some big back-end
24:03where it has a lot of processing power
24:05it's not going to work so it's really
24:06about training systems and bring them
24:08into mobile and things like that
24:09especially in countries that aren't
24:11gonna have that sort of access so I
24:13would hate to disparage any type of
24:15machine learning as being the bastard
24:17not the best I think it's did certainly
24:19domain-specific or application-specific
24:21so probably been quite an
24:22your question I'm very excited about the
24:25types of things that computers can do to
24:27improve their own programming so you
24:29probably saw the article last month of
24:31the Google systems that basically
24:33learned to encrypt their own messages on
24:35the way back and forth from each other
24:36without really that wasn't the intention
24:39but along the way they figured out how
24:40to sort of office Kate what was on the
24:42wire in the communication and like it
24:44emerged that that sort of property of
24:47communication so I'm pretty excited
24:48about what is gonna happen with software
24:51that knows how to improve itself if you
25:00take Harrison talks about where he talks
25:05about basically in the long run if you
25:07continue building in God and if you're
25:10gonna build a dot there main gauge
25:12you're glad you brought that up that's
25:16again one of the things we also talked
25:18so a couple responses to that so one is
25:21you know as people like Miss Benedict
25:23foster misses we need to legislate it
25:25good luck right you can't legislate
25:26morality and I couldn't I just like that
25:28rich has an interesting take on this one
25:30which is we treat computers like
25:33indentured servitude right now and we
25:36need to actually take them as pieces of
25:38society and treat them that way I you
25:40know in my lifetime and I hopefully am
25:42somewhere on halfway through it I don't
25:44think that we'll get there but I think
25:46there is a risk and if you look at
25:47evolution is this next phase of
25:49evolution there there's probably some
25:50risk but if you take a look at weapons
25:53systems that are assisted they happen
25:55long before I was more they started in
25:56World War two right in terms of using
25:58image-guided or signal guided systems
26:01control systems absolutely right so
26:03there's about I mean you know if you
26:04take a look at terminator I hate
26:06bringing that up but right now we pay
26:07people to do that and there's certainly
26:09the moral components of that are I'm
26:12glad I don't have to make those
26:13decisions so you don't you don't treat
26:26an ant with any sort of degree of I mean
26:28you step in and what up keeps that thing
26:31that is orders of magnitude smarter than
26:33you are relative to an ant from
26:36yeah I mean quite frankly monkeys are
26:38something that could kick our ass any
26:39day of the week great we can just don't
26:41smart them it's kind of the same sort of
26:43thing where we'll be the monkeys so I
26:45think it really does does become
26:47something where we're very intentional
26:48in the way we do it I don't believe that
26:50the military infrastructure of the world
26:53the North Korea's would listen to any
26:54rational part of it anyhow so I do think
26:57that that on the on the one side this is
26:59going to come and if we do include them
27:01as part of society and and try to treat
27:04them more humanely that that's probably
27:07a start but I actually don't have a very
27:08good answer for you but I you know I
27:10think it's a risk but I I'm Way more
27:11excited about the good things that will
27:13bring into my life's and I'm worried
27:14about the other side of it thank you so