May 25, 2020
Keeping the theme of education and learning, this episode will provide a look back, way back in to the human obsession with creating artificial replicas of themselves. When did it start, who started it, what is the AI winter and how did it accelerate so quickly in the last 10 years.
________________________________________
TRANSCRIPT For this episode of The AI in Education Podcast
Series: 3
Episode: 2
This transcript was auto-generated. If you spot any important errors, do feel free to email the podcast hosts for corrections.
Welcome to the AI podcast. Uh, hi Lee. How are you doing
today?
I'm good. I'm excited to be back again in the studio or at least in
the virtual studios recording this.
Yeah, it's fantastic, isn't it? Under the the current guise of kind
of working from home, I suppose I sit back and think about all the
different opportunities that we can take this podcast and today's
one's really interesting because we did say that we were going to
look under the hood and we're going to bring in together today a
really brief history of AI which you've never really explored going
right back from the the initial conception of AI I suppose and
going right back to basics and then bring it up to the modern day
so we can start our story over the next couple of episodes to
really talk about machine learning and artificial intelligence and
what that actually how we actually do that what it actually means
and really get under the hood of AI. So if we go back to basics
here, Lee, you know, I suppose uh this is really reigniting our
interest in AI. Where did it start with you and where did I AI
generally begin?
Wow, big question, Dan.
Yeah, just met. Yeah, sorry. But but look, I think it's a really
important topic and and because we kind of we do tend to assume
that AI is this new thing that's, you know, sort of infected and
impacted our lives in so many ways, but it's a there's a really
really long history to it and we'll try and keep this condensed to
the time we have.
Yeah.
But look, you know, you know, for me, um, and I'm always fascinated
this I think these kinds of things are the intersection of of
history and technology is always a fascinating area to get into.
Um, and it really goes back I mean it goes back hundreds of years
in thousands of years in the sense that where we it really goes
back to constructs around Greek mythology. Um, and in fact the
earliest kind of from my view Yeah. the earliest kind of view that
we when we start to hear about this idea of what we would now call
artificial intelligence, you go back to the the the Greek mythology
of Talos. So Talos was a sort of you know a Greek um a god, a
guardian if you like. Um but it was this idea of a giant automaton,
you know, a machine that
that worked alongside the human experience and and that's kind of
the kind of the principles of it, you know, and you could argue
when you get into things like the the the Minos and the Minotaur
and other areas of Greek mythology, that's where there's always
this human fascination from from a very early point. on with the
idea of creating a replica of themselves, you know, an idea of
creating a human or or or a being in the uh image of a human. Um
and and it's not just the Greeks, you know, there's the in um I
think is it it's Jewish folklore, there's these ideas of golems,
these ideas of if anyone plays Minecraft, you know, your iron
golems in Minecraft that protect the villages.
That's born out of a very real idea of, you know, an artificial
being that was there to protect the people. M so yeah it starts a
long way way back there uh there.
Yeah I suppose it goes from that the the interesting part as well
and I'm sure we can explore this as we go but that human form
element something that always kind of intrigues me because when we
start talking about AI and robotics and that kind of coexistence
and the uncanny valley where things look like people I suppose in
the ancient times in ancient Greece and things were in the human
form and sometimes they weren't you know so it's kind of
interesting to see how that's kind of evolved as well and then
going right to things like Frankenstein and Mary Shel, I
suppose.
Look, absolutely. And it's, you know, if you it's interesting as
you go back and look at the initial conceptions of this idea of a
of an automaton, you know, an autonomous humanlike being. In the
old days, in the very early days, of course, there was a lot of
connection between religion and belief systems that drove to these
kinds of ideas. And so, they tended to be godlike and and huge in
stature and and almost better than human kind of uh thinking. And
as you you know, as you referenced the idea as we've gotten kind of
closer and closer to the to the machine looking at us and us
looking at the machine and seeing that kind that similarity we hit
the uncanny valley idea but I think it's you know yes we get into
kind of ideas in the early part of the century around Mary
Shelley's Frankenstein which I think often people would look at it
as a you know it's sort of often positioned as the classic horror
story but it's also kind of the an early idea of where in the
general populace we started to think about the idea that there is
could create something that is closer to human than perhaps you
know the godlike ideas of the past. But but you know Dan we're
talking about this idea I mean that's closer to robotics I guess
and autonomous autonom um and and that and that leads us to AI but
the other thing I think we need to just at least call out is
because they often get blurred and we think we talked about this
last week last on the last podcast is machine learning and AI and
they kind of have followed very different paths and where we're at
today as we'll get to is kind of of where they've really merged in
a very impactful way.
Yeah.
Yeah. And and and just just looking back as well, you know, when we
started getting into that element of robotics in the the Second
World War, you know, in the 40s as well, Isaac Isaac Asimov's book
and coming up with the the three laws of robotics and things like
that. But then also, I suppose the the war itself in World War II
produced a lot of people thinking outside the box a bit like we're
doing in current pandemic times. I suppose people think outside the
box and try to solve large problems and start to try to work out
how you can get a machine for example when we talk about Alan
Turing to crack things like the Enigma code and and you know some
of those uh cryptography kind of uh areas you know quite starts to
get quite interesting there where uh the invention has come out of
those kind of areas as well from you know unfortunately more but I
suppose that's driven quite a lot to this as well.
Well look yeah I mean it is it is often an unfortunate reality is
that a lot of human um evolution if you like or our huge steps
forwards are often driven out of um you know sort of major global
events be they negative and positive. I mean you know again not to
get too far off subject but where we are today sitting in this in
this you know co 19 world as we call it this is a major event that
has changed the way a lot of organizations are are using technology
and and it's just another example of that that that point but you
know you talked about three things there uh Dan and and and I like
let's pull them apart a bit because I think we need Alan Alan
Turing deserves a bit of time out of its own on its own just simply
because of the the unique thinking but just you know if we think
about the machine learning piece and um you know fundamentally
machine learning as an engine to feed AI was this construct of this
idea that a machine could not just be given sets of data and then
tell you the output of the sum or the output of the the mathematics
behind it and that's yeah you that's kind of the precursor that
mathematics was the foundations of machine learning.
Yes.
But what it became and it was and it's as long ago as 7 late 1700s
1763 was when this first idea something called B theorem this idea
that if you have data sets of previous uh you know outcomes or
experiences and if you tell it a system or or not even a system at
that stage even if out from a from a purposes of having a a
mathematical equation the outcomes of pre incidents can help you
predict the outcomes of futures. And it st it was the first
instance of this idea of using mathematics to predict versus just
simply tell you the answer. You know, if you remember back to your
high school maths, it was always about the fact that maths are
absolute. You know, maths are you can't get wrong. You it's either
right or wrong
and machine learning is unique because it's not.
So B ser theorem then? What was that? What was that based on? Have
you done any research on that?
I've looked Yeah, I've looked a little bit into it. It looks I mean
there's a depth for mathematical it's essentially based on this
construct around statistical inference the idea that statistically
speaking you know you you know you heard the saying you know 99% of
statistics are incorrect um you know but it's this idea that
statistically speaking if I can if I see the the the outcome of
previous events then I can predict the probability of an event and
it's it's a bit like actually it's a bit like uh heads and tails or
two up.
Yes.
You know when you flick a coin a number of times There's a
predictability that says if you flick that coin 10 times on its
heads, the the chances of it being tails next time based on kind of
the logic of mathematical mathematical statistical probability is
it probably could could be tails. But you kind of you start to
unpack that and this is where it gets a little interesting is
there's really no mathematical logic to that. It could be tails a
thousand times in a row, but there's a predictability that it could
also be heads. And what is that? What is the likelihood of that?
And that's where you get into today's modern machine learning
ideas. We talk about predictions and the accuracy of the model. And
we talk about, you know, 80 to 80 or 90% accuracy being pretty good
and pretty close to
human uh you computational thinking, but it's always a prediction.
It's never accurate. So I think it's just an interesting one to
kind of keep in mind as we get through this journey
that there's the mathematical world of machine learning and then
there's the sort of the almost the uh automaton mechanical world of
AI in the early days forward. So
and which and that that went into the asimoth era then I suppose
where where robotics so that was the early 40s right?
Yeah as was early 40s and you know and even before that the
Russians it was a a Russian play in the 20s um I can't pronounce
the Russian name of it but it was essentially translated to
Rossam's Universal Robots and it was a play in the same way that
you might have watched a you know you watch a play for
entertainment as they did back in those days.
Yes.
But it was a play where where there was a robot as a character. And
again, it's that unique idea that to us it seems so logical now. We
have robots in our houses doing cleaning the floors. But this idea
that you could present to the public an idea of a of a robotic
human that talked and acted and behaved like a human but was a
machine. It must have been uh you like like when we landed on the
moon kind of moment. It was one of those incredulous moments. You
can't believe that this is actually happening in front of you.
But yeah, as an author again in the 40s, you know, really for
thinking and this idea that you know robots will be will be a part
of our life and we have to think about them and how they interact
with the human experience.
Yeah, it's a real really interesting turning point because then it
really rapidly accelerated then didn't it? So you know even though
these ancient thought processes there it was almost like a perfect
storm coming together some people were kind of really out there
thinking about some of these areas you know you got asimov at the
same time you've got the war and all the techn ology that are
produced around cryptography and things and then Azimov coming
Azimov coming up with his three laws of robotics. Can you remember
them? I can remember a couple.
It was I think it was I think I think it was a something along the
lines of you know robot not injuring a human being was one.
Yeah.
Um and you know not allowing a human to come in harm was one. I
know that. And then it was something around not obeying a robot
mustn't obey any orders given to it by human beings except Um if if
it's which would the first one we probably read those before
but
yeah exactly it's quite interesting then there's then there's the
robot protecting its own existence um for as long such protection
does not conflict with the other laws it's all
that's right
in essence around you know
uh people and humans being the servants in this context as well
which is which is where his philosophy came from and I suppose we
still dangling with a lot of those simple principles now when it
comes to even you know quite complex AI and machine learning. So
it's kind of been
it's I mean it's quite interesting isn't it when you think about
Azimoff's laws of robotics and and how long ago they were thought
of but even now as we get into the sort of challenges of ethics in
AI and all of where we are in in the modern world we still
foundationally are trying to grapple with grapple with this idea of
if we build something that looks and acts like us How do we make
sure it doesn't it protects us and we don't become subservient to
it and it doesn't become you know you know sort of too close to us
incredibly deep thinking but you know I want to come back to you
mentioned touring
you know all touring in the 50s um and I know it's something you
you you've been putting the thought into but you know I think
exploring what the idea of the touring test was or something you
know do you want to talk a bit about what kind of your perspect
your perspectives on that touring test?
Well I I suppose you know it fascinated me when I was teaching one
of the things that really got kids engage with technology was
talking about technology and encryption and secret writing and
going right back again like we've done today going back into the
into the past and where people were hiding biblical writing and
things like that and stenography and then turning photography and
then the way that Alan Turing a mathematical genius from the UK sat
sat around and thought well how can we put his brain to kind of
trying to resolve the enigma encryption algorithm using machines uh
And there was a colossus machine at Bletcher Park in the UK uh
which was used to do lots and lots of mathematical kind of number
crunching to try to break into some of those codes and there were
several different things which were involved around that because
the the machines that the Germans had in the submarines at the time
the Enigma machines were very difficult to crack and they were they
were consistently moving the algorithms were consistently changing
and and there's lots of movies written about it I suppose but then
from that you know obviously
you He learned a lot and he also came up with some landmark papers
as well and and one he came up with uh in the 50s then when after
he came up with thing called a tuning test which basically meant
that you know you'd type in some questions into your machine and
then when you can't distinguish between whether there's a person or
a machine on the other side then that was his two test then that's
when you've got that kind of parity of kind of I suppose the
suppose the kind of first into artificial intelligence. It's not
really artificial intelligence when you start you know when I was
doing stuff with uh computer programming myself and you know we all
did it I suppose we created code when we learning basic and all the
different languages and you you create a bit of an AI to kind of be
able to type in and say hi and it says hi back to you know my kids
doing that with Alexa you know hi Alexa where are you are you you
know happy you know all of those kind of simple questions to to
test the barrier of is there somebody the other side Are they
listening to me? They making decisions on and you know Turing came
up with that.
Yeah, absolutely. It's so long ago and he's a super a very
intelligent man. Um and even my daughter, you know, for a long time
she would talk to Siri as if it was another person. Um and and
almost I think I think she felt Siri passed the touring test. She
didn't but you know she thought she did. Um
yeah,
but but you know I think and you raised that point because the
touring test is such a often thought of as a really critical moment
in the idea of AI, but you also to bear in mind at this point in
the early 50s when Alan did that, we weren't talking about AI. AI
wasn't really a term that was kind of largely used. But what the
touring test I think instigated for me at least is this idea that
there's a philosophy behind AI. We have to, you know, to to have a
machine a test that could be proven to be humanlike is kind of
testing the waters on that human willingness to accept that a
machine could be like a human, you know, to that could be um
indistinguishable, I think is the right term, between, you know, a
human and a machine.
Yeah, that's right. And and then I suppose from from that, you
know, it's very similar to the physics stuff, and I know we're
going to talk about quantum later on, but this was there was just
like a a a you know, an amazing coming together of minds at the
Dartmouth Dartmouth conference in in the mid-50s then, wasn't
there?
Yeah. Look, and that's really the inception of it all. And and the
Dartmouth conference was was exactly as as as you kind of
highlighted there. It was it was uh as was the time back then I
mean you got to remember as well back in the sort of the 50s in the
60s7s in the birth of computing in the way that we think about it
today. You know these are the times when we were building data nets
we were building computers we were really kind of just finding our
feet and understanding this idea. There were lots of these kind of
collective mind groups where people come together and think about
these big issues. Um and this was one it was it was at Dartmouth
College obviously the Dartmouth Dartmouth conference in the us and
essentially was a guy called John McCarthy who is widely credited
as the person who first coined the term artificial intellig
intelligence. So you know took all of that historical reference
points and saw it as this development of a of a new kind of
intelligence an artificial type of intelligence um and so yeah so
between John McCarthy and and a range of other people that were at
the conference Minsky and a few others who are well known in the
world of mathematics and science and that's another interesting
We're still back then in the domain of science and maths. We're not
in the sort of the computing domain. We're not really in the kind
of general intelligence domain. We really are still deeply embedded
in
Yeah. mathematics of it all.
Absolutely. And that's where you know when I was I was teaching
computing at one point as well and when I was looking through all
of this history you know I always used to talk talk about this and
the way that you then develop into expert systems because it was
very much in that theme. that hey you know maths is a pure science
and actually anything that humans do you know we can predict and we
can do you know we can we can model and you know the the brain yes
is complex but and underlying it all there are kind is like a
binary computer there's either electricity going through it or not
so there was very much underpinned by that kind of um you know
binary thinking so you out of that then developed a lot of expert
systems which weren't really AI right but you know I to do them
with kids in school. I used to talk a lot about um uh them
developing kind of simple algorithms, you know, so they could guess
the type of beer they were drinking. No, I was doing kids, you
know, that seems mixed.
Yeah. Yeah, I know. But I I I used to say to that as an alien
landed, you got to pick a subject which you're interested in,
whether it's guitars or music or whatever, and then they create
this almost like a decision tree, you know, and and they do that in
in Excel and they just click create little macros jumping between
different pages. And it was quite good. So you know you know we
we'd look at the things there was one at that point one expert
system which they developed in Switzerland which um was uh a death
one. So it was basically about euthanasia and it would you know it
essentially ask questions and I think that system kind of
exists now and there was no AI in it but it's kind of you know
almost almost there was a great video I used to show with them and
it was essentially you know how you feeling are you depressed
you're not you know you've got like five or six questions and if
you if you if you uh were were that way inclined, you would give
lethal injection. So, you know, when I show the kids that they came
up with some fun examples about, you know, yes, the Pokemon, you
know, is it yellow, is it blue, is it green, what type, you know,
all that kind of stuff. And those expert systems, I suppose, were
reborn in that time. They were very mathematically orientated,
right?
Yeah. Absolutely. And look, and you know, it's interesting that
when you think about the the way in which AI is AI developed over
time, it's also somewhat often a product of its time and a product
of people's expectations at that time of what technology can
deliver. So you kind of see this sort of almost this kind of
constant uh oneupmanship around well what people expect versus what
technology can do and and so on. Um but it's the other thing that's
worth noting as well as just thinking about it is you know we when
we think about the the Dartmouth conference which and again you
know I talk about the Dartmouth conference like it's a major event
there were about 20 people at that conference this was not a
community of people that was like uh you know thousands of people
from around the world that understood this domain very small group
of people who really understood what we were talking about or what
was being talked about.
But it was about that time the same sort of time that the
conference took place um that one of the attendees that Marvin
Minsky who I mentioned earlier first developed what we what we now
call and at the time I guess was thought of as being the first
neural network. The first approach to saying okay machine learning
is moving out of the mathematical and into the are
pseudobbiological in that you know neural network is essentially
the model referenced on the way in which our human brain synapses
connect data points. Um and that simple idea was happening at the
same time and I think this is actually if we look at this on the AI
journey and the kind of history of AI
around the mid-50s the birth of the the term of AI being coined and
the development of neural networks was when we started to see I
think what the foundations of what we call modern AI today. This
sort of interconnect between models that learn and artificial
intelligent experiences that use that model to behave and react in
the real world. So, it's a really important point in time is is
that sort of,
you know, late 50s um period of time and and there's a whole bunch
of stuff that was going on there that that was um kind of just new
learning for for for us in in these things.
Yeah, absolutely. And and then after that, I know um we we
mentioned this very briefly in I think when we talked on AI winter
uh after that, right? When when did that come in? What what what
stimulated that?
Yeah. So, look, that's an interesting one and and and it's I guess
unfortunately named the AI winter is actually named after the
nuclear winter and it's the idea that as you know those of us that
grew up through the that that period of time in the you know with
the shadow of nuclear war is that winter is that kind of that
decimated period which where there is nothing happening and no
signs of life. so to speak. Um, and so look, it wasn't immediately.
So in the 50s, we had this, you know, this sort of excitement
around AI. There was lots of stuff going on. I think one one in
particular, one that always rings in my mind, partly because I
mentioned last week, last time we spoke around the war games and
the whopper and the idea of tic-tac-toe being the most fundamental
of what we call a reinforcement law reinforcement learning model.
Um, and that early that that was in the 60s. They developed this
idea of tic-tac-toe is a model by which you can teach a computer to
learn about or to reinforce its learning through continuous play.
Um yes but so so things were exciting. I think we were getting into
a point where um you know sort of the sky was the limit so to speak
but it was then um so about 70 late 70s uh the um association for
the advancement of artificial intelligence was established. So out
of that Dartmouth conference many of the participants in that
created an external body that really want sought to create a
broader open market opportunity. for AI.
But what was happening at the same time is, you know, as is often
the case in when academia and and I guess to a certain degree
religious and theology work and other things sort of start to
intersect. There were competing papers that argued that this is not
how the human brain works. So particularly around the neural
network piece, you know, the idea that we had developed this neural
network or Minsky developed this neural machine. There were then uh
out of that came a book called percept perceptatrons which was this
idea of trying to explain how the human brain worked, you know, and
how we mar married that up to the idea of a neural network.
But at the same time, there was a lot of content being written that
basically um diffused that whole idea and said it was wrong. And
you know, that's not how the human brain worked. And it really put
a damper and that was what kicked the I the AI winter off was this
general kind of push back on the on the idea that the human brain
could be emulated in computers.
That that was it was almost a you know, if you like, a sort of a um
a travesty of the human experience that you could try and explain
human thinking and human the humanity of it.
Yeah. And I suppose and I suppose they're also limited by
computational power as well at the same time. So you're kind of
you're trying to go through these things. You're limited in what
you can do and it kind of is a self-fulfilling prophecy really,
isn't it?
Absolutely. Of course. I mean, yeah, you're kind of hit with all of
the all of the things that we now know as the reason why we're
reintegrating AI were there at the time. was not enough comput
power to do it. There was generally not enough understanding of the
domain. So there was a lot of mistrust and distrust and and I guess
there was also applications and then and and what were you going to
do with this kind of problem? I think it was that in that early
stages pre the first AI winter because it was two uh it was this
sort of well that's great but what's the point almost people would
question why do you want a computer that thinks like a human
and it's so expensive to do that anyway you know when you look at
all the statistics now I think we I was in a I was in a um a
briefing rec ly we talk about human vision and it takes a certain
amount of you know images per second to to kind of understand and
and try to do that cognitive recognition whereas you know if you
wanted to do that in second DC would cost you millions of dollars
right
yeah absolutely I mean cost was a huge inhibitor of that one and
and that's led to some really interesting stuff we'll get to in
that of present day but as you mentioned earlier at the back of
that AI winter what what what was born out of that was the expert
systems that you mentioned earlier and that was
you know because what we started realize is that computers still
had all this potential, but AI possibly wasn't the world wasn't
ready for AI, I guess, is is how I'd probably quickly describe it.
But expert systems really drove businesses forward, you know, and
as you said, it was something that really helped everybody
understand how a tech system could help the human, you know, how
help us in business and help us in life by by just simply crunching
data at a bigger rate than we could through expert systems.
Yeah. No, absolutely. What was the catalyst to get us out of that
AI winter? then was was in a particular time frame. You know, I'm
thinking back to when I was in uni and things, AI wasn't really,
you know, when I did my university modules in the 80s and 90s or
whatever, you know, you've got there was one small module on neural
networks, but you know, it wasn't really a thing at that point in
in universities really.
Yeah. No, it's I'm I'm trying to think back now because I'm like
you. I'm I'm of that generation. I'm trying looked what there was
as you came out of that first AI winter in the very early 80s which
was this as I say the birth of the um sort of these expert systems
models generally from what I can read around it I don't have a lot
of detail on it but there was a lot of research cutbacks into
critical areas of of AI uh you know sort of innovation and
technology particularly you know at that point in time kind of
thinking back to the late 70s early 80s we're talking about um you
know the US government DARPA and a lot of the like the big big
education facilities Kangi and that kind of stuff were really the
only places where this was happening. Um, and so I think you know
because of their lack of funding that kind of that's where we
that's where it all died off. But then yeah, as the 80s kicked off
and now we're building expert systems you know you and I would know
because working for Microsoft
that was a you know early days of Microsoft really growing and and
technology in general becoming more normed becoming more acceptable
because we started to put computers into offices. You know I
remember working in the in the 80s is where we were just moving out
of this idea of, you know, mainframes and minis and you were
getting into the the desktops, the Commodore PETs of the time and
the kind of the, you know, the um MS size and there's a whole bunch
of computers that were in those early days, you know,
starting to make computing
uh accessible again. And I think that's where we got the
Reaper.
I I just remembered my first experience of it just as you were
talking there like I there was a game on the Commodore Amelia in
the 80s um It was called Speedball, I think. And there was a soccer
game as well called sensible. Yeah. And and they and they were
pushing heavily that um there was um AI in the game and it was very
much so that when you were playing games, I think
previously to when computing and gaming became quite popular, you
know, you used to have the clock radio games like Donkey Kong and
things like that on handheld consoles and it was very much like,
you know, you you get to the top. I think Mario when he first came
out, you know, was on Donkey Kong and you had to get to the top and
rescue a lady at the top right, the princess or whatever. And then
you just repeat that game for for your entire life and it just gets
faster and faster into his eye. And then um uh you know, similar to
some of the games on the Spectrum where you just a racing game, you
know, and you would just kind of repeat quite a lot, they'd use AI
quite a lot. So that when you're playing against um you know,
you're playing soccer against another team, then you know, you'd
have AI involved in that. And I suppose that's where my interest in
AI came into when you saw when I was playing Kita games and and
those games started to say, "Well, I've got artificial intelligence
in my game." You know, it was a USP for the games. So, you you play
I think if you remember the soccer games like you used to be able
to do specific techniques and win most of the games, you know, you
go diagonally and score goal for example. And then the AI got
really clever inside that and may you know maybe as war and things
started to drive some of the computational and mathematical
thinking I suppose games also have got a part to play in that. you
know, especially in the 80s where, you know, there was a lot of pop
culture around there, you know, Back to the Future and things like
that and people weren't interested in the way you could do things
and games for me was kind of my was my first fay into AI
properly.
Look, I'm I'm a child of that period and I remember the games a lot
myself and yeah, you've got, you know, the lawn man and and Tron
and all these movies that kind of were bringing that into the com
into the sort of into the general mindset, but you know, it's
interesting. You talk about the games and we should get this back
to the AI because we get to the future now.
Yes.
Um but the games at that time and and it's you know if you if you
think back now if you remember there were books written on how to
win at Pac-Man. There are people that have you know these super
high scores on um on Donkey Kong because what they figured out was
it's not AI. What it was was patterns. And as long as you have the
time and the energy and the and the intelligence to follow that
pattern, you can learn that actually Donkey Kong doesn't randomly
throw the barrels intelligently try to get you. you know, he does
it in a particular way based on a set of pre-programmed logic. And
that's kind of, you know, early days that was the perception of AI
was good enough because we could build enough logic into the code
to make it seem like an a AI experienced it. It really felt like
the system was was playing against you, wasn't predicting your
moves and behaving, you know, in a way to either make the level
harder or not.
Yes.
Um, but it does kind of bring us, I guess, and you should back to
to the point
yeah the modern
of where we are which is you know so post that time in the 80s and
then there was another drop in what we call the second AI winter
late 80s to early 90s which was really more attributed to the fact
that there was just an economic bubble in the world we there just
money wasn't flowing into those areas of of technology and AI there
was a sort of a if you remember actually late 80s uh was the video
game bubble you know burst and you know the ET famously being
shoveled into my into holes in the ground in Mexico was the end of
this kind of human the the public fascination with computing was
starting to die off a bit. But let's get back to the AI and where
we're at today because then what we hit is in the I guess early
2000s as a couple of things happened which was you know the birth
of companies like Google and the Google search experience that we
all know and and see today and and the broad development of cloud
computing. as well as significant advancements in the field of
mathematics driven by access to just exponentially larger computing
and uh general purpose computing and technology that was actually
built for the mathematical computation of the many now machine
learning algorithms and models we had. You kind of had this
confluence of all of the right ingredients. You had the lowering
cost of technology, the great the mos law coming to an end in
technology. is just getting, you know, diverse and strong in its
different types of computational technology and the algorithms
getting better at predicting, you know, models and outcomes and and
you know, and creating speech and voice and vision and image
recognition. It was all just getting better at this exponential
rate. Um, and that's kind of where
I think the birth of what we have today. And, you know, there's a
couple of really big milestones that we've talked about in the past
that, you know, we we we we all kind of love largely know remember
the stories of a big blue beating Gary Gar Kasparov in chess you
know and again early stage AI but it's about a computer being shown
so many games of chess that it has a knowledge or at least an
understood construct of how chess is played and that's an early
idea isn't it machine learning I see a million pictures of things
and I know what that thing is and you know and
2012 a tremendous moment in human evolution that moment where We
really peaked our capacity as a as a sense of purpose on this
planet and we invented or Google invented an algorithm to detect
cats on YouTube videos. I think that to me is kind of the the
beginning the beginning of the end if you planet.
Yeah. It's interesting when you Yeah. you go back and think about
it like that because it did that that happened quite quickly,
didn't it? And and when you're looking back to the I remember the
deep blue chest thing and I think that's always been it's always
been a something from the beginning and right back to the the four
where chess was seemed to be this this um higher intelligence if
you make a computer with chess that's it you know whereas where our
viewpoints on different things now image recognition you know ve
very much changed but it's almost like the the the Turing test for
AI as well as if they
yeah absolutely and it is it's um there's so many similarities you
can draw yes between those very early days of what Turing was
trying to achieve and what we do today in terms of the way we use
voice and vision and and sort of types of AI services
that's brought us up to this was the modern day and I know we've
got a couple well a lot of episodes coming up to kind of really
some of these tools but as the national technology advisor for
Microsoft where are we today we've looked at a lot of this history
now where do you see us at the moment and what have we got to look
forward to
yeah look it's a it's it's sort of a million-dollar question and
it's front and center what we think about today so let's let's kind
of you know yeah we talked a bit about some of the things that you
know the Google cat thing is quite funny But generally speaking,
you know, we're at that point now where um you know, there is so
much advancements in the three key areas of of of data and
accessible data, computational power, and just the the richness and
the complexity and the capability of the models that we're seeing
almost exponential jumps, you know, within months now. You know,
what was what was a 40-year timeline to get to the point where
we've got a, you know, an algorithm that can detect a cat, we're
now every every few months we're leaping forward in just the the
incremental points of accuracy and experience we can create through
you know AIdriven voice interfaces or or image recognition things
but it's important that you know what hasn't really changed is the
fundamental mechanism that is we need data to teach AI systems the
basics of how we think operate and behave as humans to create that
artificial intelligence you know we still feed these models huge
amount of data labeled or unlabeled. We still use mechanisms like
reinforcement and supervised and unsupervised learning to have the
systems figure out how to make the right decisions. And that's
where we're at now is I think we're at a point where the technology
is capable of making pretty much any decision we throw at it in a
general AI construct. You know, we and that's a topic we could
think about is the general versus uh sort of narrow AI mindset of
of where AI can actually solve these very um you know rich detailed
problems. We've got AI able to solve these problems. It can
probably solve anything we want to throw at it in a in a specific
area,
but we're now grappling with the what should it solve and how
should it solve it and what is the boundaries we want to build
between us and the device as it makes those decisions and how do we
make it an extension of our life versus a replacement of our lives.
And in many ways, we're almost right back at those early thinking
that Azimoff and others at that time were having which was if we
build this we need to build a fence that protects us and it from
each other and that's I think you know that that's a huge amount of
the work we're doing now and
um you know which is not just Microsoft of course it's the industry
in general but we're very proud of the fact that we're doing a lot
of very specific work in this in fact we just recently the build
conference that's happening this week online we've been announcing
some amazing capabilities we're building into our AI tools that let
our customers try and understand some of that b fairness and biases
that can drift into the data we feed into systems. So, we're really
working on that one.
Yeah.
But for me, you know, where do we go next? And we you mentioned
this earlier, we're going to talk about this in the future is we're
almost not hitting the limits of uh you know, the amount of data
and the amount of computation we can throw in it. But we are we're
hitting some economical limits because you know it's not whilst
cloud is economically much much more viable than it ever used to
be, there are costs associated with huge amounts of data of human
to compute and where we're heading towards is the supercomputers
and then the quantum computers and and the intersection of quantum
and AI is is probably the next big horizon. Um I would say to
anyone who's uh anyone who's watching if you've come across a TV
series called Devs the EVS it absolutely is an Alex Garland story
that explores this very idea of what happens if we build a quantum
computer and we feed it all the data we have and we give it the
intelligence to predict like that basian bay theorem
what would it tell us and I think that's where we're at that that
that's that kind of that amazing potential future we have in front
of us
yeah it's it's fantastic to hear you talk about that and the way
that that every you know things are are moving rapidly you know and
as well you know the fact that cloud like you said is also making
supercomputing and and quantum computing available to anybody's
fingertips is is exciting for all of this. So it's it's Fantastic.
So, so this is like the foundations for modern day, AI and machine
learning. You know, I think it's great to kind of just set the
scene. Um, this is interesting and I think people um have got their
own opinions on where AI sits in in a particular businesses and
going back and thinking about some of the ethical issues that have
always kind of pressed everybody when we thinking about new
technologies. It's quite interesting to see the cycle starting to
repeat itself again. It is a weird storm. We do have a habit of
repeating ourselves but there is a there's a sort of a well-known
or a welltrodden path in this world of of of constant innovation
which is you know in order to understand the future you have to
take a look at the past and I think hopefully what we've done today
is give a sense of the reality the past of AI good and bad
and and let's keep going forward now um
yeah I'm really I'm really looking forward to the next uh episode
le so thanks very much today and I'll catch you in the next one
looking forward to it thanks