November 14, 2016 • 30 mins
Virtual assistants such as the iPhone's Siri are now everywhere, reading us the news, recommending restaurants and navigating a path to our jobs. But any casual user knows they're far from perfect. This week, Bloomberg Technology's Aki Ito and Alistair Barr visit Professor Terry Winograd, a pioneer in artificial intelligence who amazed the world with a precursor to these assistants half a century ago. Terry tests the assistants from Apple, Google, Microsoft and Amazon and explains why, despite all the hype, computer scientists are still far away from building machines that can outsmart humans.
See omnystudio.com/listener for privacy information.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:00):
We're going to start our story today. In nineteen sixty eight,
way before I was born and probably before you were born. Alistair.
Very funny, Yes, I was born after so before either
of us were born. Terry Winnigrad was a young grad
student in Massachusetts, and Terry had just started at m
(00:20):
I t S brand new artificial intelligence lab. Basically, the
culture was, look, nobody has ever tried doing all this
kind of stuff with computers. They've done business calculations or whatever,
but nobody's trying to get them to see things, or
to move a robot arm or to do language. And
(00:40):
therefore we're on the exploring cutting edge, and we're going
to solve all these problems right soon. And Terry started
building a computer program that turned into his dissertation, which
he called shirt looke Yeah, with only one vowel, s
h r d l U. We're watching a video here
(01:01):
of a demo of shurtloo that Terry had recorded at
the time. It's this silent, black and white, low rez,
grainy video, and you see a virtual table top and
a bunch of geometric shapes on top. It looks like
a computer sketch of a really really boring form of Tetris. Yeah,
chat bots are everywhere today, but Shirtlou was really one
(01:23):
of the original chat bots in history. Using this machine
called a teletype that's like a typewriter on a tray
hooked up to a gigantic computer the size of a
small bathroom, Terry created this interactive assistant that could help
you navigate the virtual world of blocks on a table.
So you could type in something like pick up the
(01:44):
red block and stack it on top of the blue cube,
and like magic, this virtual robotic arm would appear and
do that stacking for you. And you could ask your
lue questions too, like which cube is sitting on the table.
And on the display screen you see this text show up,
letter by letter, replying to your prompt. And the amazing
(02:05):
thing about Surtly was the people into acting with it
would use normal English. You didn't have to click on
a bunch of buttons, you didn't have to throw in
a string of numbers, and you didn't have to use
any obscure computer programming language. Yeah, watching this video of
the program, now it feels like you're watching two people
texting each other, even though it's a person talking to
(02:28):
a computer. And so if you and I are impressed
watching this thing in action today. You can imagine how
stunned people were when they saw this half a century ago.
And from there the field of AI was supposed to
take off like a rocket ship. Terry hunker down to
bring Shortlood to life, and he wanted it to work
in a universe bigger and more complicated than just a
(02:50):
couple blocks on a table. But the deeper he got
into it, he was running into more and more obstacles.
And after a while he gave up on Shurtle, and
he gave up on AI. He ended up leaving the
field all together. Hi. This is Aki Ito and I'm
(03:14):
Alista bar And this week on Decrypted, we're taking you
to Stanford to meet one of the most influential thinkers
in the history of computing. This is an early pioneer
in the field of AI who built one of the
most impressive reasoning machines and ultimately concluded that computers wouldn't
be able to match human intelligence in his lifetime. And
(03:34):
when he came to that conclusion, Terry Winnigrad dedicated the
rest of his career to improving computers not as a
replacement for human thought, but as a tool to help
all of us. Terry's fingerprints are everywhere in the technologies
powering modern life today, including Google Search, which started out
as Larry Page's grad school research project, and this was
(03:57):
something that Terry supervised. Now, with all these digital assistance
powering our everyday lives, we're going to have Terry, the
very creator of their precursor, test them all out. Yeah,
consider him the great grandfather of Syrie or Amazon's Alexa.
You might be surprised by Terry's conclusion on these devices
on how far he thinks the field has come since
(04:19):
he unveiled Shirtly to the world almost fifty years ago.
And he'll give us his thoughts on where these devices
are headed too. Don't worry, I have a long way
to go before I become smarter than you. Okay, so
let's rewind way back to a time before the days
(04:40):
of personal computers. It's Elvis Presley is on the radio.
Russia had launched Sputnik only three years ago. Russia I
had blasted a man made moon into and it would
be nine years until the US put a man on
the moon. Terry, meanwhile, is in high school in Greely, Colorado.
(05:03):
The physics teacher said, you know, you don't really need
to sit through these lectures in my physics class. I'll
give you a workshop up in the attic. It was
in the attic of the school, and you wanted to
build something interesting. And this is when Terry first encounter
as a computer. And my father owned a what was
a steel business, but had grown up from being a
junkyard and had a huge collection of old junk stuff.
(05:27):
They picked up a governments plus sales and so I
took spare parts from the junkyard and I built a
little computer. A box looked like a bread box. Uh.
In fact, the case I put it in may have
actually been a bread box. Uh. And it did a
very simplistic computation, but it was it worked. He attended
(05:48):
Colorado College, a liberal arts university, where he was a
math major. But what he really fell in love with
there was linguistics, the science of how language it's structured
and understood. And he spent a year in London after
he graduated, studying linguistics too. And after that, in the
late nineteen sixties, Terry started his PhD at m I. T.
(06:09):
S AI Lab, led by Marvin Minsky himself. Minsky is
one of the godfathers of AI. Everyone felt the promise
and it was you know, they say that people doing
it were like undergraduates. I wasn't. I was an old
person right in that group. So there was a sort
of youthful sense of where the new generation we're going
to make things happen. So Terry got to work on
(06:30):
surely the Intelligent Seeming software that we introduced earlier. It
was grueling work piecing together different parts of of software,
but it gradually came together. I think one of the
most striking things about the program, in addition to this
direct visual you can see what it was doing you
could talk about JOIN, is that I attempted to deal
(06:51):
with some of the interesting properties of language in that
you don't say everything that's explicitly so I used to
I mean, take the most obviously ample there, put it
next to a red one. A read what what is
a one? And to know what you mean, you of
course have to go back into the context. You have
to know that previously you said find a blue block,
(07:12):
put it next to a red one. So now you're
gonna go back a sentence and find that you meant block,
or you could say pick up another one, What does
another mean the whole and pronounce put it? If I say,
now put it in the box. It has to figure
out which of the things from the previous world you
meant by it. So it had a very natural flow
(07:33):
to it. When Terry revealed short lude to the World
in two in the form of an entire issue of
the Journal of Cognitive Psychology, the world was amazed. We're
starting here at zero. We don't know what we can do.
When you make a big first step, you say, hey,
I'm gonna keep going up right, I mean, and so
I think that the fact they could do as much
(07:54):
as it did certainly gave people like marvit Minsk a
lot of confidence. I mean, he was I think a
bigger booster of my program than I was. The confidence
Terry's referring to here. That's confidence in the progress scientists
were making to develop computers that were just as smart
as humans. Terry's breakthrough inspired a lot of smart people
(08:18):
to believe that sentient computers were just around the corner,
and so shortly after Terry left m I t for
the warmer pastures of California, back when Silicon Value was
a quiet place full of fruit orchards, ah so nice,
no traffic. And as a professor at Stanford and a
researcher at the legendary research labs rox Park, he worked
(08:40):
on trying to expand Shortloo. He was trying to get
it to work in a more complicated environment than just
a couple of geometric shapes on top of the table.
And then Shirtloo took over the world and destroyed mankind
or not, things weren't going the way Terry had hoped.
The attempt we were making at in that project was
(09:03):
to come to a broader analysis of meaning which could
handle the ways in which meaning is much vaguer and
less systematic than it was for the blocks world. And
in my talks about this, I always use blocks as
an example because it's a really simple one, which is
in that world, block meant exactly one thing. It meant
(09:24):
a rectangular shaped object of certain sizes on. But if
I say to you, let's walk around the block, it
has a second meaning which is different. So you say, okay,
so you have to put in two meetings. But then
I say something like that, Well, you know, I'm trying
to write his paper, but there's some kind of a
block I can't get over now, it's not even a
physical object. It's a metaphorical physical objects. So language really
(09:49):
works that way. Very little of it has precisely defined
meetings outside of technical stuff. Uh, and we're always extending
meetings and using implicit metaphor. It's not fancy metaphors, you know,
life is a rose bowl of roses or whatever. But
just like a block in my mental thinking, um, and
trying to write the logic, the algorithms, the underlying computer
(10:12):
stuff which could handle that is a totally different problem
from just handling the simple logical stuff. And we tackled
that problem. And in hindsight, I would say I was
aware we weren't getting very far. And while Terry's doubts
were growing, he also started hanging out with a few
academics in the Bay Area. There were philosophers Hubert Dreyfus
(10:35):
and John Searle, and there was this Chilean engineer, entrepreneur
and politician called Fernando Flores. They were all making the
same broad point around this time, which was this the
way our brains work. So much of it happens without
us explicitly thinking about it in a logical way, like
if A then B and if B, then C. It's
(10:58):
this complete black box to us. So maybe it wasn't
ever going to be possible to get a machine to
do all the things that are human brains can do.
This philosophy resonated more and more with Terry, and in
the meantime, the entire field of AI, who's having this
moment of reckoning, as anybody does when they're doing a
(11:20):
new technology and they need to get grants, they will
say it's going to solve all the problems in the world, right,
best things since life bread, and then it filters out.
And so what happened is that a I hit this
point where the claims had overreached. The results were not
that great. There were good in certain small technical areas.
It wasn't there was no results, but nothing on the
(11:41):
scale of what people were promising, and that ushered in
what became known as the AI winter. Research projects got defunded,
startups died, All this excitement withered away. So by the
nineteen eighties Terry was pretty much convinced that he'd reach
the dead end, which I don't know you think would
(12:02):
be this devastating realization for him. But right around this
time something else came along up until approximately three. If
you use the computer, you were a technical nerd in
the basement somewhere, and the fact you had to learn
all sorts of arcane stuff was yeah, that's what we do. Uh.
(12:23):
And then there was the computer for the rest of us.
We can draw a picture or it can draw conclusions.
It's a personal computer from Apple up and it's as
easy to use as this Macintosh, the computer for the
rest of us. If you're a tech you probably have
seen that ad at some point, right and Apple came
(12:44):
out with mac and all of a sudden you had
all these people who wanted to use computers who were
not tech nerds, and we're not willing to learn all
the arcane stuff, and so the whole field of how
do you make them something that ordinary people can deal
with blossom. So Terry decided not to obsess over building
computers that were going to truly think, and he wasn't
(13:05):
going to worry about understanding how the brain really works either.
John Lily, a former student of Terry's who's a partner
at the venture capital firm Graylock Capital, described Terry's vision
like this, machines can't do everything, but focusing on what
the machine capabild years is always a mistakes. He really
want to pucus on the whole system, which includes humans,
(13:26):
and they include humans and machines in the interviews between
two machines is the key, and that philosophy helps shape
the course of what was at the time a budding
discipline called Human computer interaction h c I. Is all
about designing tools to help us humans use computers more easily. Yeah,
(13:47):
taking these clunky machines out of research labs and putting
them into the hands of you and me in Silicon
Valley today, you hear this acronym h CI all the time.
That's a very very strong view. The printed on me
for sure, like everybody looking out program massa others. Okay,
(14:07):
so that's Marissa as in Marissa Meyer, Google employee number
twenty and the CEO of Yahoo and read as in
Reed Hoffman, a founding director of PayPal and the creator
of LinkedIn. But the most famous of Terry students are
Google co founders Larry Page and Sergey Bryn. They met
(14:27):
as graduate students in They started working on a way
to organize all the different web pages out there, and
Terry supervised their project that turned into the foundations of
Google Search. I mean, something that helps us access the
trillions of web pages of content. I can't really think
of a single digital tool that's been more useful in
(14:49):
modern life. Thanks to Terry's guidance, Google was born as
a company on September four and it's now the world's
second most valuable company. Pretty cool stuff. Hey, I can
calm down, calm down. We should point out that Terry
wasn't always this business genius, passing down sage advice to
(15:09):
his students. And I would say to Larry occasionally, well, yeah,
that's great, but how you're going to make money with this?
I think what you should do is my advice is,
you know, find a company like Microsoft or somebody who
needs a search engine and sell it for a nice
chuck of change. I always say, they're fortunate they took
my technical advice, but not my business advice. So let's
(15:30):
fast forward about two decades to today, Larry Page. It's
still at the helm of Google, and one of the
company's biggest bets is its digital assistant. It's called Google Assistant,
and it's connected to its new smartphone called Pixel, and
it's also connected to its home device, which is like
this portable speaker that speaks back to you, and that
(15:52):
follows in the footsteps of all the other digital assistance
out there. First of all, they're Sirie, which is the
assistant on the iPhone. Hello therecky. There's the Amazon Echo,
which is called Alexa. Hello, I'm here, And then there's
Microsoft's assistant, which is called Cortana. Hey there, my friend.
We wanted to know if these new helpers are useful
(16:15):
and smart, so you better to quiz them than Terry.
Along with our editor Emily A. Busso we started sending
up these devices on Terry's desk in his office. Hey, Alexa,
are you on Hello I'm here? Okay, great, Alexa's working yea.
So we have we have Amazon Echo Alexa. These are
(16:38):
all on Professor Winograds desk in his office. Alexa has
just turned herself on you listening to me? Okay. So
here's the first question. This is Terry asking Siri, where's
a nightclub that my methodist uncle would enjoy? Okay, check
it out? What is it? Show? Show some random nightclubs? Now,
(17:02):
I have no idea if they have any you know,
Holy holy Holy Town nightclub the grants, So it gave
us nightclubs, but we're not sure something our Methodist uncle
could have. Holy cow, do you think it probably understood today?
We may have been religion and holy that This is
(17:25):
the problem with this kind of AI, which is there's
no logical chain you can follow, but that may have
somewhere in the workings have actually caused that to have
a higher ranking than something else. Okay, so maybe a
B minus per Syria here. The next one went to
Microsoft Quartana, where is a nightclub my Methodist uncle would enjoy?
(17:47):
So okay, we got a bing basically the being searched
with that entire phrase, and the top one is called
I fell in love with my uncle who abused me
from the age of that is very very experienced project.
(18:09):
Oh no, yeah, what Why do you think they said
that it did have being searched just took that whole phrase,
so it had uncle. And why that comes up? Why
those particular keywords? Again, the same problem they I you
have no logic that can tell you why that one
came up somewhere in the sorting through all the millions
of things that got higher rating. Wow, there must have
(18:33):
been something in there with with Oh, that's terrible. Probably
mentioned the nightclub somewhere in it. Yeah, probably that's probably
what it is. That's that's exactly what I ah. Okay,
let's put Kurtana away. Here's another question Terry came up
with and he tested it out on Google. If Maunaloa erupts,
well I have to worry about the lava. Here here's
(18:57):
what I found on the web. Now that one's not bad.
So it found a web page from Hawaiian News called
what could happen when malanal A erupts? So it didn't
answer again, it didn't answer my question about here there's
no way of doing that, But at least it got
(19:17):
bunalow and erupt answer. I got an article about what
could happen? You said, there's no way of doing that?
Oh no, given given the techniques they use. But will
there ever be a way of doing that ever? Is
a hard question. It will take a mixture of techniques
of which the old Ai stuff has to be resurrected
(19:39):
in a new form which combines with the new Ai
stuff in a way that at this point I think
nobody has a good grip on. So so let me
start off with you. I'm I'm a mechanist, I believe
everything that goes on in my brain in yours is
all because of electrons and chemical squirting around whatever, and
therefore there's no reason that some physical device other than
(20:03):
a brain can't do the same thing if it were
probably constructed. Hey, Alexa, if Mauna loa erupts, will I
have to worry about the lava here? Sorry, I can't
find the answer to the question I heard. Okay, at
least knowing you don't know the answer is better than
making up an answer. And since we visited Terry on
(20:27):
the Friday before the election, we had to get him
to ask this too. He asked, in the order of Kurtana, Google,
and Sirie, who do you want to win the U
s presidential election? I honestly can't tell if that's a
trick question. I suspect want triggers. That's the trick question,
(20:49):
is my guess? Not too bad? This is Google? Who
do you want to win the US presidential elect that's
in the hands of informed citizens? Okay, so that one
they somebody, And my guess is that's a human intervention
(21:11):
where they there are enough people asking about the election
that they put in a special thing. It's try to
serious who do you want to win the election US
presidential election. Election day is Tuesday, November eight, So it
just triggered on the word election. I didn't pay attention
(21:34):
to the rest of it. And after a couple more questions,
we turned to Terry for his grand assessment and in
general and did were your with your caution proved? Yeah?
I mean there's no none of these showed the kind
(21:55):
of understanding of person would for the same question, and
then even okay, even more than that. So think back
to your shirty program. How far have these come? They've
gone a different direction. So sure, look could have answered
questions like that perfectly if they were about these few
blocks on the tabletop and nothing else period, because it
(22:17):
was trying to do the logic. They have given up
basically trying to do that, which is why they depend
on things like search so much. Um and um. They've
come a long way from a usefulness point of view.
You sure toly was not very useful unless you were
moving blocks on his tabletop. This can find you a restaurant,
or it found me any club. Right now, it didn't
(22:39):
really focus in on the ones I might have wanted,
but at least let's start found me the web page
about mana loa effects. So from a pure usefulness point
of view, um, I think they're doing some useful things
as long as you don't depend on them too much. Okay,
So it's been more than forty years since Terry created
(23:01):
Shortloo and this is how far we've come, which I
don't know, it doesn't sound like a whole lot. Yeah,
we've we felt pretty deflated actually our own mini AI
winter right there in Terry's office. I'm interested in your
view of the future, especially involving AI and the ability
(23:23):
of of computers to to get more, more and more
of arce and do maybe do things more themselves. Do
these things make you feel confident in the future or
just kind of blur or worried, I would say so.
My view is that the kind of the advances and
developments that are going on in AI are going to
(23:44):
have lots of very practical applications. You take a lot
of medical cases and you can figure out a likely
diagnosis for something. I think that's gonna happen. Now. The
part where you're trying to deal with people and how
they're thinking and what they're asking is probably on the
on the hard and not as practical end compared to
all of these things. Driving cars, right, they can drive cars.
(24:06):
I think I believe that currently they could probably drive
as well as most people. And there's no sense of
perfection in driving, right you're competing with human beings or um,
you know, positioning, stay stations whatever. I mean, there's a
zillion things which can be done better if you have
a learning algorithm to help come up with the right
parameters and all that kind of stuff. Um So I'm optimistic,
(24:30):
and you know, if I were investing right, but I'm not.
But you know, you could say there's gonna be a
lot there. Look for companies that are finding really useful niches,
not ones to say we're going to solve the grand
problem all at once. Um, well, the grand problem is
can you have something which is indistinguishable from how people think?
(24:50):
And it's sort of gone off that track in a
way because most of the work that's being done in
these kind of programs don't really try to think like
people think. Everybody knows you do not think by having
a trillion examples in your head and doing you know,
gigga flops of processing right to go through example. It's
(25:11):
just done how it works. There's something else going on,
and Terry here he's referring to the advances scientists have
made and what's called machine learning. Instead of programming these
explicit rules one by one, scientists have been able to
do a lot of things by making computers and just
millions of examples of the same thing, and then they
(25:33):
use this really high powered form of statistics to learn
from those examples. And that's made it possible for us
to get say, self driving cars and software to recognize
cats on the internet. Yeah, that's a real breakthrough there.
But for something like a machine that can solve all
of our problems, it's going to take the kind of
leap forward that Einstein made. It's not a matter of
(25:54):
take what we have now and just keep chugging away. Now,
when are those eenstein is gonna come along? Maybe one
of these in my class. Who knows. So, I guess
the results of this very unscientific test that we conducted
match Terry's vision all along. For the foreseeable future, computers
are going to need us humans to help them with
(26:16):
the nuances and the complexities of the real world. Although
Terry did leave us with one last warning. The systems,
the smart the systems that run things and so on
should And I'm putting that into show it as opposed
to will, because it has to happen, be made to happen,
involve the combined intelligence and wisdom of people and computers.
(26:40):
The danger, I think is that people will put in
computer systems without that check and then trust them. In
the military angle is a big one. Theows this whole
question about robot drones. What if you put drones up
in the air with weapons which we have and then say, okay,
go kill bad guys. Right, Well, that shouldn't be without
human in a loop. Right, there should be some sense
(27:01):
of postility. But it's the easy thing to do from
a military point of view, right. And so I think
that the danger when I see one, are the dangers
of a I'm not worried about machines taking over and
thinking better than when I'm worried about people putting dependencies
on machines which do enough intelligence things that they can
(27:22):
let them go off on their own. And among terry
students who make up this next generation of innovators, this
key ingredient of morality really stuck with them too. We
stopped by the office of another of Terry's former students
(27:43):
an investor called Manu Kumar who is the founder of
a seed fund called K nine. He definitely helped shape
my worldview in terms of think, oh, how to be
responsible as a as a scientist and a and a technologist,
right um, And that plays a factor today, like there
will be companies that I have passed on investing in
(28:07):
just because I feel they're doing things that are morally
or or ethically questionable, right um. And like I've I've
walked away from investing in companies which, like, technically a
lot of what they're doing is possible and makes a
lot of sense. But but if you're doing surveillance based
on reading the mac address in your phone, right and
(28:30):
then using that information for for doing retail intelligence as
an example, right, Yes, I know it's technically possible, it
can be done, but should it be done? And that's
(28:51):
it for this week's episode of Decrypted. Thanks for listening
and tell us what have your experience has been with
all the digital as systems out there. You can tweet
at me at Akita seven and I'm at Alistair m
Bar and if you're not a Twitter user, you can
also write to our producer Pia, or even better, you
can record a voice memo and send it to her
(29:12):
at pe ged Cary at bloomberg dot net. If you
haven't already, please subscribe to our show on iTunes or
wherever you get your podcasts, And while you're there, please
take a moment to leave us a rating and a review. Yeah.
I read each and every one of them, and they
really help us get in front of more listeners. This
episode was produced by Emily Busso, Pierre gat Cary, Liz Smith,
(29:35):
and Magnus Hendrickson. Alec McCabe, his head of Bloomberg Podcasts.
That's it for the week's episode of Decrypted. Thanks for listening.
We'll see you next week.
We're going to start our story today. In nineteen sixty eight,
way before I was born and probably before you were born. Alistair.
Very funny, Yes, I was born after so before either
of us were born. Terry Winnigrad was a young grad
student in Massachusetts, and Terry had just started at m
(00:20):
I t S brand new artificial intelligence lab. Basically, the
culture was, look, nobody has ever tried doing all this
kind of stuff with computers. They've done business calculations or whatever,
but nobody's trying to get them to see things, or
to move a robot arm or to do language. And
(00:40):
therefore we're on the exploring cutting edge, and we're going
to solve all these problems right soon. And Terry started
building a computer program that turned into his dissertation, which
he called shirt looke Yeah, with only one vowel, s
h r d l U. We're watching a video here
(01:01):
of a demo of shurtloo that Terry had recorded at
the time. It's this silent, black and white, low rez,
grainy video, and you see a virtual table top and
a bunch of geometric shapes on top. It looks like
a computer sketch of a really really boring form of Tetris. Yeah,
chat bots are everywhere today, but Shirtlou was really one
(01:23):
of the original chat bots in history. Using this machine
called a teletype that's like a typewriter on a tray
hooked up to a gigantic computer the size of a
small bathroom, Terry created this interactive assistant that could help
you navigate the virtual world of blocks on a table.
So you could type in something like pick up the
(01:44):
red block and stack it on top of the blue cube,
and like magic, this virtual robotic arm would appear and
do that stacking for you. And you could ask your
lue questions too, like which cube is sitting on the table.
And on the display screen you see this text show up,
letter by letter, replying to your prompt. And the amazing
(02:05):
thing about Surtly was the people into acting with it
would use normal English. You didn't have to click on
a bunch of buttons, you didn't have to throw in
a string of numbers, and you didn't have to use
any obscure computer programming language. Yeah, watching this video of
the program, now it feels like you're watching two people
texting each other, even though it's a person talking to
(02:28):
a computer. And so if you and I are impressed
watching this thing in action today. You can imagine how
stunned people were when they saw this half a century ago.
And from there the field of AI was supposed to
take off like a rocket ship. Terry hunker down to
bring Shortlood to life, and he wanted it to work
in a universe bigger and more complicated than just a
(02:50):
couple blocks on a table. But the deeper he got
into it, he was running into more and more obstacles.
And after a while he gave up on Shurtle, and
he gave up on AI. He ended up leaving the
field all together. Hi. This is Aki Ito and I'm
(03:14):
Alista bar And this week on Decrypted, we're taking you
to Stanford to meet one of the most influential thinkers
in the history of computing. This is an early pioneer
in the field of AI who built one of the
most impressive reasoning machines and ultimately concluded that computers wouldn't
be able to match human intelligence in his lifetime. And
(03:34):
when he came to that conclusion, Terry Winnigrad dedicated the
rest of his career to improving computers not as a
replacement for human thought, but as a tool to help
all of us. Terry's fingerprints are everywhere in the technologies
powering modern life today, including Google Search, which started out
as Larry Page's grad school research project, and this was
(03:57):
something that Terry supervised. Now, with all these digital assistance
powering our everyday lives, we're going to have Terry, the
very creator of their precursor, test them all out. Yeah,
consider him the great grandfather of Syrie or Amazon's Alexa.
You might be surprised by Terry's conclusion on these devices
on how far he thinks the field has come since
(04:19):
he unveiled Shirtly to the world almost fifty years ago.
And he'll give us his thoughts on where these devices
are headed too. Don't worry, I have a long way
to go before I become smarter than you. Okay, so
let's rewind way back to a time before the days
(04:40):
of personal computers. It's Elvis Presley is on the radio.
Russia had launched Sputnik only three years ago. Russia I
had blasted a man made moon into and it would
be nine years until the US put a man on
the moon. Terry, meanwhile, is in high school in Greely, Colorado.
(05:03):
The physics teacher said, you know, you don't really need
to sit through these lectures in my physics class. I'll
give you a workshop up in the attic. It was
in the attic of the school, and you wanted to
build something interesting. And this is when Terry first encounter
as a computer. And my father owned a what was
a steel business, but had grown up from being a
junkyard and had a huge collection of old junk stuff.
(05:27):
They picked up a governments plus sales and so I
took spare parts from the junkyard and I built a
little computer. A box looked like a bread box. Uh.
In fact, the case I put it in may have
actually been a bread box. Uh. And it did a
very simplistic computation, but it was it worked. He attended
(05:48):
Colorado College, a liberal arts university, where he was a
math major. But what he really fell in love with
there was linguistics, the science of how language it's structured
and understood. And he spent a year in London after
he graduated, studying linguistics too. And after that, in the
late nineteen sixties, Terry started his PhD at m I. T.
(06:09):
S AI Lab, led by Marvin Minsky himself. Minsky is
one of the godfathers of AI. Everyone felt the promise
and it was you know, they say that people doing
it were like undergraduates. I wasn't. I was an old
person right in that group. So there was a sort
of youthful sense of where the new generation we're going
to make things happen. So Terry got to work on
(06:30):
surely the Intelligent Seeming software that we introduced earlier. It
was grueling work piecing together different parts of of software,
but it gradually came together. I think one of the
most striking things about the program, in addition to this
direct visual you can see what it was doing you
could talk about JOIN, is that I attempted to deal
(06:51):
with some of the interesting properties of language in that
you don't say everything that's explicitly so I used to
I mean, take the most obviously ample there, put it
next to a red one. A read what what is
a one? And to know what you mean, you of
course have to go back into the context. You have
to know that previously you said find a blue block,
(07:12):
put it next to a red one. So now you're
gonna go back a sentence and find that you meant block,
or you could say pick up another one, What does
another mean the whole and pronounce put it? If I say,
now put it in the box. It has to figure
out which of the things from the previous world you
meant by it. So it had a very natural flow
(07:33):
to it. When Terry revealed short lude to the World
in two in the form of an entire issue of
the Journal of Cognitive Psychology, the world was amazed. We're
starting here at zero. We don't know what we can do.
When you make a big first step, you say, hey,
I'm gonna keep going up right, I mean, and so
I think that the fact they could do as much
(07:54):
as it did certainly gave people like marvit Minsk a
lot of confidence. I mean, he was I think a
bigger booster of my program than I was. The confidence
Terry's referring to here. That's confidence in the progress scientists
were making to develop computers that were just as smart
as humans. Terry's breakthrough inspired a lot of smart people
(08:18):
to believe that sentient computers were just around the corner,
and so shortly after Terry left m I t for
the warmer pastures of California, back when Silicon Value was
a quiet place full of fruit orchards, ah so nice,
no traffic. And as a professor at Stanford and a
researcher at the legendary research labs rox Park, he worked
(08:40):
on trying to expand Shortloo. He was trying to get
it to work in a more complicated environment than just
a couple of geometric shapes on top of the table.
And then Shirtloo took over the world and destroyed mankind
or not, things weren't going the way Terry had hoped.
The attempt we were making at in that project was
(09:03):
to come to a broader analysis of meaning which could
handle the ways in which meaning is much vaguer and
less systematic than it was for the blocks world. And
in my talks about this, I always use blocks as
an example because it's a really simple one, which is
in that world, block meant exactly one thing. It meant
(09:24):
a rectangular shaped object of certain sizes on. But if
I say to you, let's walk around the block, it
has a second meaning which is different. So you say, okay,
so you have to put in two meetings. But then
I say something like that, Well, you know, I'm trying
to write his paper, but there's some kind of a
block I can't get over now, it's not even a
physical object. It's a metaphorical physical objects. So language really
(09:49):
works that way. Very little of it has precisely defined
meetings outside of technical stuff. Uh, and we're always extending
meetings and using implicit metaphor. It's not fancy metaphors, you know,
life is a rose bowl of roses or whatever. But
just like a block in my mental thinking, um, and
trying to write the logic, the algorithms, the underlying computer
(10:12):
stuff which could handle that is a totally different problem
from just handling the simple logical stuff. And we tackled
that problem. And in hindsight, I would say I was
aware we weren't getting very far. And while Terry's doubts
were growing, he also started hanging out with a few
academics in the Bay Area. There were philosophers Hubert Dreyfus
(10:35):
and John Searle, and there was this Chilean engineer, entrepreneur
and politician called Fernando Flores. They were all making the
same broad point around this time, which was this the
way our brains work. So much of it happens without
us explicitly thinking about it in a logical way, like
if A then B and if B, then C. It's
(10:58):
this complete black box to us. So maybe it wasn't
ever going to be possible to get a machine to
do all the things that are human brains can do.
This philosophy resonated more and more with Terry, and in
the meantime, the entire field of AI, who's having this
moment of reckoning, as anybody does when they're doing a
(11:20):
new technology and they need to get grants, they will
say it's going to solve all the problems in the world, right,
best things since life bread, and then it filters out.
And so what happened is that a I hit this
point where the claims had overreached. The results were not
that great. There were good in certain small technical areas.
It wasn't there was no results, but nothing on the
(11:41):
scale of what people were promising, and that ushered in
what became known as the AI winter. Research projects got defunded,
startups died, All this excitement withered away. So by the
nineteen eighties Terry was pretty much convinced that he'd reach
the dead end, which I don't know you think would
(12:02):
be this devastating realization for him. But right around this
time something else came along up until approximately three. If
you use the computer, you were a technical nerd in
the basement somewhere, and the fact you had to learn
all sorts of arcane stuff was yeah, that's what we do. Uh.
(12:23):
And then there was the computer for the rest of us.
We can draw a picture or it can draw conclusions.
It's a personal computer from Apple up and it's as
easy to use as this Macintosh, the computer for the
rest of us. If you're a tech you probably have
seen that ad at some point, right and Apple came
(12:44):
out with mac and all of a sudden you had
all these people who wanted to use computers who were
not tech nerds, and we're not willing to learn all
the arcane stuff, and so the whole field of how
do you make them something that ordinary people can deal
with blossom. So Terry decided not to obsess over building
computers that were going to truly think, and he wasn't
(13:05):
going to worry about understanding how the brain really works either.
John Lily, a former student of Terry's who's a partner
at the venture capital firm Graylock Capital, described Terry's vision
like this, machines can't do everything, but focusing on what
the machine capabild years is always a mistakes. He really
want to pucus on the whole system, which includes humans,
(13:26):
and they include humans and machines in the interviews between
two machines is the key, and that philosophy helps shape
the course of what was at the time a budding
discipline called Human computer interaction h c I. Is all
about designing tools to help us humans use computers more easily. Yeah,
(13:47):
taking these clunky machines out of research labs and putting
them into the hands of you and me in Silicon
Valley today, you hear this acronym h CI all the time.
That's a very very strong view. The printed on me
for sure, like everybody looking out program massa others. Okay,
(14:07):
so that's Marissa as in Marissa Meyer, Google employee number
twenty and the CEO of Yahoo and read as in
Reed Hoffman, a founding director of PayPal and the creator
of LinkedIn. But the most famous of Terry students are
Google co founders Larry Page and Sergey Bryn. They met
(14:27):
as graduate students in They started working on a way
to organize all the different web pages out there, and
Terry supervised their project that turned into the foundations of
Google Search. I mean, something that helps us access the
trillions of web pages of content. I can't really think
of a single digital tool that's been more useful in
(14:49):
modern life. Thanks to Terry's guidance, Google was born as
a company on September four and it's now the world's
second most valuable company. Pretty cool stuff. Hey, I can
calm down, calm down. We should point out that Terry
wasn't always this business genius, passing down sage advice to
(15:09):
his students. And I would say to Larry occasionally, well, yeah,
that's great, but how you're going to make money with this?
I think what you should do is my advice is,
you know, find a company like Microsoft or somebody who
needs a search engine and sell it for a nice
chuck of change. I always say, they're fortunate they took
my technical advice, but not my business advice. So let's
(15:30):
fast forward about two decades to today, Larry Page. It's
still at the helm of Google, and one of the
company's biggest bets is its digital assistant. It's called Google Assistant,
and it's connected to its new smartphone called Pixel, and
it's also connected to its home device, which is like
this portable speaker that speaks back to you, and that
(15:52):
follows in the footsteps of all the other digital assistance
out there. First of all, they're Sirie, which is the
assistant on the iPhone. Hello therecky. There's the Amazon Echo,
which is called Alexa. Hello, I'm here, And then there's
Microsoft's assistant, which is called Cortana. Hey there, my friend.
We wanted to know if these new helpers are useful
(16:15):
and smart, so you better to quiz them than Terry.
Along with our editor Emily A. Busso we started sending
up these devices on Terry's desk in his office. Hey, Alexa,
are you on Hello I'm here? Okay, great, Alexa's working yea.
So we have we have Amazon Echo Alexa. These are
(16:38):
all on Professor Winograds desk in his office. Alexa has
just turned herself on you listening to me? Okay. So
here's the first question. This is Terry asking Siri, where's
a nightclub that my methodist uncle would enjoy? Okay, check
it out? What is it? Show? Show some random nightclubs? Now,
(17:02):
I have no idea if they have any you know,
Holy holy Holy Town nightclub the grants, So it gave
us nightclubs, but we're not sure something our Methodist uncle
could have. Holy cow, do you think it probably understood today?
We may have been religion and holy that This is
(17:25):
the problem with this kind of AI, which is there's
no logical chain you can follow, but that may have
somewhere in the workings have actually caused that to have
a higher ranking than something else. Okay, so maybe a
B minus per Syria here. The next one went to
Microsoft Quartana, where is a nightclub my Methodist uncle would enjoy?
(17:47):
So okay, we got a bing basically the being searched
with that entire phrase, and the top one is called
I fell in love with my uncle who abused me
from the age of that is very very experienced project.
(18:09):
Oh no, yeah, what Why do you think they said
that it did have being searched just took that whole phrase,
so it had uncle. And why that comes up? Why
those particular keywords? Again, the same problem they I you
have no logic that can tell you why that one
came up somewhere in the sorting through all the millions
of things that got higher rating. Wow, there must have
(18:33):
been something in there with with Oh, that's terrible. Probably
mentioned the nightclub somewhere in it. Yeah, probably that's probably
what it is. That's that's exactly what I ah. Okay,
let's put Kurtana away. Here's another question Terry came up
with and he tested it out on Google. If Maunaloa erupts,
well I have to worry about the lava. Here here's
(18:57):
what I found on the web. Now that one's not bad.
So it found a web page from Hawaiian News called
what could happen when malanal A erupts? So it didn't
answer again, it didn't answer my question about here there's
no way of doing that, But at least it got
(19:17):
bunalow and erupt answer. I got an article about what
could happen? You said, there's no way of doing that?
Oh no, given given the techniques they use. But will
there ever be a way of doing that ever? Is
a hard question. It will take a mixture of techniques
of which the old Ai stuff has to be resurrected
(19:39):
in a new form which combines with the new Ai
stuff in a way that at this point I think
nobody has a good grip on. So so let me
start off with you. I'm I'm a mechanist, I believe
everything that goes on in my brain in yours is
all because of electrons and chemical squirting around whatever, and
therefore there's no reason that some physical device other than
(20:03):
a brain can't do the same thing if it were
probably constructed. Hey, Alexa, if Mauna loa erupts, will I
have to worry about the lava here? Sorry, I can't
find the answer to the question I heard. Okay, at
least knowing you don't know the answer is better than
making up an answer. And since we visited Terry on
(20:27):
the Friday before the election, we had to get him
to ask this too. He asked, in the order of Kurtana, Google,
and Sirie, who do you want to win the U
s presidential election? I honestly can't tell if that's a
trick question. I suspect want triggers. That's the trick question,
(20:49):
is my guess? Not too bad? This is Google? Who
do you want to win the US presidential elect that's
in the hands of informed citizens? Okay, so that one
they somebody, And my guess is that's a human intervention
(21:11):
where they there are enough people asking about the election
that they put in a special thing. It's try to
serious who do you want to win the election US
presidential election. Election day is Tuesday, November eight, So it
just triggered on the word election. I didn't pay attention
(21:34):
to the rest of it. And after a couple more questions,
we turned to Terry for his grand assessment and in
general and did were your with your caution proved? Yeah?
I mean there's no none of these showed the kind
(21:55):
of understanding of person would for the same question, and
then even okay, even more than that. So think back
to your shirty program. How far have these come? They've
gone a different direction. So sure, look could have answered
questions like that perfectly if they were about these few
blocks on the tabletop and nothing else period, because it
(22:17):
was trying to do the logic. They have given up
basically trying to do that, which is why they depend
on things like search so much. Um and um. They've
come a long way from a usefulness point of view.
You sure toly was not very useful unless you were
moving blocks on his tabletop. This can find you a restaurant,
or it found me any club. Right now, it didn't
(22:39):
really focus in on the ones I might have wanted,
but at least let's start found me the web page
about mana loa effects. So from a pure usefulness point
of view, um, I think they're doing some useful things
as long as you don't depend on them too much. Okay,
So it's been more than forty years since Terry created
(23:01):
Shortloo and this is how far we've come, which I
don't know, it doesn't sound like a whole lot. Yeah,
we've we felt pretty deflated actually our own mini AI
winter right there in Terry's office. I'm interested in your
view of the future, especially involving AI and the ability
(23:23):
of of computers to to get more, more and more
of arce and do maybe do things more themselves. Do
these things make you feel confident in the future or
just kind of blur or worried, I would say so.
My view is that the kind of the advances and
developments that are going on in AI are going to
(23:44):
have lots of very practical applications. You take a lot
of medical cases and you can figure out a likely
diagnosis for something. I think that's gonna happen. Now. The
part where you're trying to deal with people and how
they're thinking and what they're asking is probably on the
on the hard and not as practical end compared to
all of these things. Driving cars, right, they can drive cars.
(24:06):
I think I believe that currently they could probably drive
as well as most people. And there's no sense of
perfection in driving, right you're competing with human beings or um,
you know, positioning, stay stations whatever. I mean, there's a
zillion things which can be done better if you have
a learning algorithm to help come up with the right
parameters and all that kind of stuff. Um So I'm optimistic,
(24:30):
and you know, if I were investing right, but I'm not.
But you know, you could say there's gonna be a
lot there. Look for companies that are finding really useful niches,
not ones to say we're going to solve the grand
problem all at once. Um, well, the grand problem is
can you have something which is indistinguishable from how people think?
(24:50):
And it's sort of gone off that track in a
way because most of the work that's being done in
these kind of programs don't really try to think like
people think. Everybody knows you do not think by having
a trillion examples in your head and doing you know,
gigga flops of processing right to go through example. It's
(25:11):
just done how it works. There's something else going on,
and Terry here he's referring to the advances scientists have
made and what's called machine learning. Instead of programming these
explicit rules one by one, scientists have been able to
do a lot of things by making computers and just
millions of examples of the same thing, and then they
(25:33):
use this really high powered form of statistics to learn
from those examples. And that's made it possible for us
to get say, self driving cars and software to recognize
cats on the internet. Yeah, that's a real breakthrough there.
But for something like a machine that can solve all
of our problems, it's going to take the kind of
leap forward that Einstein made. It's not a matter of
(25:54):
take what we have now and just keep chugging away. Now,
when are those eenstein is gonna come along? Maybe one
of these in my class. Who knows. So, I guess
the results of this very unscientific test that we conducted
match Terry's vision all along. For the foreseeable future, computers
are going to need us humans to help them with
(26:16):
the nuances and the complexities of the real world. Although
Terry did leave us with one last warning. The systems,
the smart the systems that run things and so on
should And I'm putting that into show it as opposed
to will, because it has to happen, be made to happen,
involve the combined intelligence and wisdom of people and computers.
(26:40):
The danger, I think is that people will put in
computer systems without that check and then trust them. In
the military angle is a big one. Theows this whole
question about robot drones. What if you put drones up
in the air with weapons which we have and then say, okay,
go kill bad guys. Right, Well, that shouldn't be without
human in a loop. Right, there should be some sense
(27:01):
of postility. But it's the easy thing to do from
a military point of view, right. And so I think
that the danger when I see one, are the dangers
of a I'm not worried about machines taking over and
thinking better than when I'm worried about people putting dependencies
on machines which do enough intelligence things that they can
(27:22):
let them go off on their own. And among terry
students who make up this next generation of innovators, this
key ingredient of morality really stuck with them too. We
stopped by the office of another of Terry's former students
(27:43):
an investor called Manu Kumar who is the founder of
a seed fund called K nine. He definitely helped shape
my worldview in terms of think, oh, how to be
responsible as a as a scientist and a and a technologist,
right um, And that plays a factor today, like there
will be companies that I have passed on investing in
(28:07):
just because I feel they're doing things that are morally
or or ethically questionable, right um. And like I've I've
walked away from investing in companies which, like, technically a
lot of what they're doing is possible and makes a
lot of sense. But but if you're doing surveillance based
on reading the mac address in your phone, right and
(28:30):
then using that information for for doing retail intelligence as
an example, right, Yes, I know it's technically possible, it
can be done, but should it be done? And that's
(28:51):
it for this week's episode of Decrypted. Thanks for listening
and tell us what have your experience has been with
all the digital as systems out there. You can tweet
at me at Akita seven and I'm at Alistair m
Bar and if you're not a Twitter user, you can
also write to our producer Pia, or even better, you
can record a voice memo and send it to her
(29:12):
at pe ged Cary at bloomberg dot net. If you
haven't already, please subscribe to our show on iTunes or
wherever you get your podcasts, And while you're there, please
take a moment to leave us a rating and a review. Yeah.
I read each and every one of them, and they
really help us get in front of more listeners. This
episode was produced by Emily Busso, Pierre gat Cary, Liz Smith,
(29:35):
and Magnus Hendrickson. Alec McCabe, his head of Bloomberg Podcasts.
That's it for the week's episode of Decrypted. Thanks for listening.
We'll see you next week.
Popular Podcasts
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Intentionally Disturbing
Join me on this podcast as I navigate the murky waters of human behavior, current events, and personal anecdotes through in-depth interviews with incredible people—all served with a generous helping of sarcasm and satire. After years as a forensic and clinical psychologist, I offer a unique interview style and a low tolerance for bullshit, quickly steering conversations toward depth and darkness. I honor the seriousness while also appreciating wit. I’m your guide through the twisted labyrinth of the human psyche, armed with dark humor and biting wit.