Artificial Intelligence
This is a Q&A excerpt on the topic of AI from a lecture by Richard Feynman from September 26th, 1985.
This is a clip on the Lex Clips channel that I mostly use to post video clips from the Artificial Intelligence podcast, but occasionally I post favorite clips from lectures given by others. Hope you find these interesting, thought-provoking, and inspiring. If you do, please subscribe, click bell icon, and share!
Full lecture:
• Richard Feynman C...
Lex Clips channel:
/ lexclips
Lex Fridman channel:
/ lexfridman
Artificial Intelligence podcast website:
https://lexfridman.com/ai
Apple Podcasts:
https://apple.co/2lwqZIr
Spotify:
https://spoti.fi/2nEwCF8
RSS:
https://lexfridman.com/category/ai/feed/
Connect with on social media:
Twitter: https://twitter.com/lexfridman
LinkedIn: https://www.linkedin.com/in/lexfridman
Facebook: https://www.facebook.com/lexfridman
Instagram: https://www.instagram.com/lexfridman
Medium: https://medium.com/@lexfridman
Support on Patreon: https://www.patreon.com/lexfridman
0:00
this is a Q&A excerpt on the topic of
0:02
artificial intelligence from a lecture
0:04
by Richard Fineman
0:05
from September 26 1985 I really record
0:08
the audience questions because they're
0:10
barely audible in the original question
0:12
do you think there will ever be a
0:14
machine that will think like human
0:16
beings and be more intelligent than
0:19
human beings
0:22
first of all they think like human
0:24
beings I would say no and I'll explain
0:26
in a minute why I say no and second that
0:28
they be more intelligent in human beings
0:30
is a question intelligences to be
0:33
defined if you were to ask me are they
0:35
better chess players than any human
0:37
being possibly can be yes I'll get you
0:39
someday now they're best better chess
0:42
players than most human beings right now
0:44
one of the things by the way that we
0:46
always do is we want the dawn machine to
0:48
be better than anybody not just better
0:50
than us if we find a machine it can play
0:52
chess better than us it doesn't impress
0:54
us much we keep saying it and what
0:55
happens when it comes up against the
0:57
Masters we imagine that we human beings
1:00
are equivalent to the Masters and
1:01
everything right the machine has to be
1:03
better than a person in everything that
1:06
the best person does at the best level
1:09
okay but it's hard on the machine but
1:12
with regard to the question of whether
1:13
to make it to think like a machine my
1:15
opinion is based on the following idea
1:18
that we try to make these things to work
1:21
as efficiently as we can with the
1:22
materials that we have materials are
1:24
different than nerves and so on if we
1:29
would like to make something that runs
1:30
rapidly over the ground then we could
1:33
watch a cheetah running we could try to
1:36
make a machine that runs like a cheetah
1:38
but it's easier to make a machine with
1:40
wheels with fast wheels or something
1:42
that flies just above the ground in the
1:44
air when we make a bird the airplanes
1:48
don't fly like a bird they fly but they
1:51
don't fly like a bird okay so they don't
1:54
flap the wings exactly they have in
1:55
front another guy of a gadget that goes
1:57
around or the more modern airplane has a
2:00
tube that you heat this the air and
2:02
squirt it out the back a jet propulsion
2:05
a jet engine as an internal rotating
2:10
fans and so on and it uses gasoline it's
2:12
different right so there's no question
2:15
that the later machines are not going to
2:17
think like people think in that sense
2:20
with regard to intelligence I think it's
2:25
exactly the same way for example they're
2:27
not going to do arithmetic the same way
2:29
as we do arithmetic but they'll do it
2:30
better let's take bath ematic very
2:32
elementary mathematics arithmetic
2:35
they do arithmetic better than anybody
2:36
much faster and differently but it's
2:41
fundamentally the same because in the
2:42
end the numbers are equivalent right so
2:44
that's a good example of we're never
2:46
going to change how they do arithmetic
2:47
to make it more like humans that would
2:49
be going backwards because the
2:52
arithmetic done by humans is slow
2:53
cumbersome and Confused in a full of
2:55
errors where these guys are fast if one
2:59
compares her what computers can do to
3:01
the human beings we find the following
3:05
rather interesting comparisons the first
3:07
of all if I give you a human being a
3:10
problem like this I'm going to ask you
3:13
for these numbers back every other one
3:16
in reverse order please right now I'm
3:20
good a series of numbers and I want them
3:21
to you didn't give me one to me back in
3:23
reverse order every other one I'll tell
3:26
you I'll make it easy for you just give
3:27
me the numbers back the way I gave them
3:28
to you you ready one seven three nine
3:35
two six five eight three one seven two
3:45
six three
3:47
anybody got a gonna be able to do that
3:50
now and that's more than not more than
3:53
twenty or thirty numbers but you can
3:56
give a computer 50,000 numbers like that
3:59
in s confirm any reverse order the sum
4:01
of them all do different things with
4:03
them and so on and it doesn't forget
4:04
them for a long time instead of so there
4:06
are some things that a computer does
4:07
much better than a human and you'd
4:09
better remember that if you kind of
4:10
compare machines to humans but what a
4:12
human has to do for his own always they
4:15
always do this they always try to find
4:18
one thing don''t that they can do better
4:20
than the computer so we now know many
4:22
many things that the humans can do
4:24
better than a computer she's walking
4:26
down a street she's got a certain kind
4:28
of a wiggle and you know that Jane right
4:33
or the sky's going in and you see his
4:36
hair flip just a little bit it's hard to
4:38
see at a distance but that particular
4:39
funny way that he did back of his head
4:41
looks that's jack okay to recognize
4:46
things to recognize patterns
4:48
seems to be something that we have not
4:50
been able to put into a definite
4:52
procedure you would say I have a good
4:54
procedure for recognizing jack just take
4:57
a lots of pictures of Jack by the way a
5:00
picture can be put into the computer in
5:01
fact by this method here if this were
5:04
very much finer I could tell whether
5:05
it's black and white at different spots
5:07
you know you in fact you get pictures in
5:09
a newspaper by black and white dots and
5:11
if you do Stuart fine enough you can't
5:13
see the dots so with enough information
5:14
I can load pictures in so you put all
5:16
the pictures of of Jack on the different
5:20
circumstances and that's the machine to
5:22
compare it the trouble is that the
5:23
actual new circumstance is different the
5:25
lighting is different the distance is
5:27
different the tilt of the head is
5:28
different and you have to figure out how
5:30
to allow for all that and it's so
5:32
complicated and elaborate that even with
5:34
the large machines with the amount of
5:35
storage that's available and the speed
5:37
that they go we can't make figure out
5:39
how to make a definite procedure that
5:42
works at all or what least it works
5:44
anywhere within a reasonable speed so
5:47
recognizing things is difficult for the
5:50
machines at the present time and some of
5:52
those things are done in a snap by a
5:54
person so there are things that humans
5:57
can do that the we don't know how to do
5:59
in a filing system so it is recognition
6:04
and that brings me back to something I
6:05
left which is what kind of a file clerk
6:08
can't be imitated by the machine a file
6:11
clerk that has some special skill which
6:13
represent which requires recognition of
6:16
a complicated kind for instance a fire
6:18
Kirk in the fingerprint Department which
6:21
looks the finger prints and then makes a
6:23
careful comparison to see if these
6:24
finger prints match has not been it's
6:28
just about ready to be it's hard to do
6:30
it almost possible to do by a computer
6:33
did say there's nothing to it I look at
6:35
the two fingerprints and see if all the
6:36
blood dots are the same but of course
6:38
it's not the case the finger was dirty
6:40
the print was made at a different angle
6:42
the pressure was different than the
6:44
ridges are not exactly in the same place
6:46
if you were trying to match the exactly
6:48
the same picture would be easy but where
6:50
the center of the print is which way the
6:52
finger is turned where there's been
6:54
squashed a little more a little bit less
6:55
where there's some dirt on the finger
6:57
whether in the meantime you got a wart
6:58
on his thumb and so forth
7:01
complications these little complications
7:04
make the comparison so much more
7:05
difficult for the machine for the blind
7:07
filing clerk system that is too much
7:10
much much too slow to be certainly
7:12
utterly impractical almost at the
7:14
present time I don't know where they
7:16
stand but they're going fast trying to
7:17
do it where is a human can go across all
7:21
that somehow just like they do in the
7:23
chess game they seem to be able to catch
7:24
on the patterns rapidly and we don't
7:27
know how to do that
7:28
rapidly automatically audience question
7:33
can computers discover new ideas and
7:36
relationships by themselves well it
7:39
depends what do you mean themselves and
7:41
it's it's hard to discover new
7:44
relationships our computers can do there
7:48
have been computers which do things like
7:51
problems they are improving in geometry
7:55
or something in which they've converted
7:57
the problem of finding a proof of a
7:58
theorem into a definite procedure okay
8:02
and once you do that although it's an
8:04
elaborate and dumb way to do proofs they
8:07
can do it the present time a computer
8:10
can't do all the different things that a
8:12
person can do you know it's it's very
8:16
difficult to find some way of defining
8:18
rather precisely something we can do
8:20
that we can say a computer will never be
8:22
able to do there are some things that
8:27
people make up that say that while it's
8:30
doing it will it feel good or while it's
8:33
doing it will it understand what it's
8:36
doing or some other abstraction I'd
8:38
rather feel that these are things like
8:39
while it's doing it we'll be able to
8:41
scratch the lice out of its hair no it
8:43
hasn't got any hair the lice to scratch
8:45
from okay so there are if you've got to
8:50
be careful when you say what the human
8:51
does if you add to the actual result of
8:55
his effort some other things that you
8:57
like the appreciation of the aesthetic
9:00
but you didn't do that I'm not saying
9:01
you did but a lot of people do that when
9:03
they ask questions and if we add things
9:06
that we think we're doing on top of what
9:09
we actually do and just look at not just
9:11
the result of what we're doing but a lot
9:13
of extra things then it gets harder and
9:14
for the computer to do it because the
9:17
human beings have a tendency to try to
9:19
make sure that they can do something
9:21
that no machine can do somehow it
9:25
doesn't bother them anymore it must have
9:26
bothered them in earlier times that
9:28
machines are stronger physically than
9:30
they are they can lift weights that are
9:31
heavier than people who can move things
9:33
faster than people who can run fast if
9:35
you can fly you can do it's terribly
9:38
strengths and so forth and we don't
9:41
still sit around worrying that there's
9:43
some way that the man can turn his hand
9:45
that some machine can't do that we can
9:48
easily make machines that are better
9:50
than us in predicting the the weather
9:52
for instance because what you do to
9:55
predict the weather is to look at old
9:57
records and see when the circumstance
10:00
was similar and guess that the results
10:02
will be similar added to that a certain
10:04
amount of analysis of the movement of
10:06
wind according to the laws of physics
10:08
and their current amount of hocus-pocus
10:09
put together ok now the speed will be
10:12
higher and the effectiveness of the
10:14
prediction is greater if you could look
10:16
at more cases so you get a better chance
10:19
of getting one closer and put more a
10:21
longer and more elaborate calculation
10:24
including more variables which is too
10:25
hard for us to in time to make the
10:27
prediction now we have to make the
10:29
prediction of the weather let's say that
10:30
weather for three days from now has to
10:32
be predicted in three days or the damn
10:34
thing is useless right and we work at a
10:36
certain speed but the computers worked
10:38
faster and can do more and therefore for
10:40
instance for weather prediction in the
10:41
end maybe not today but someday it's not
10:44
at all inconceivable that the machine
10:45
could do weather prediction faster and
10:47
more effectively and more accurately
10:48
than we do we will have have however
10:51
given it the procedure now the question
10:54
is what happens if we don't give it the
10:56
procedure well a man that people have
11:00
tried that this game of giving it
11:02
instead of a direct procedure a kind of
11:05
what it has been called heuristics try
11:08
an analogy to get a new idea of how to
11:10
do something compare this to that try an
11:13
extreme case etc and a man by the name
11:17
of Lynette has gone the farthest with
11:19
this do I have time much time do I have
11:21
because we want to see these slides
11:24
what I know that's interesting
11:26
information but how much how much time
11:29
do I have
11:30
it's time for the slideshow now since I
11:33
have no more time I will say no more
11:35
about the subject what no but it it
11:40
takes a few minutes that's why I asked
11:41
for the time okay you make this machine
11:46
which was again a filing cabinet you
11:48
understand what it does is it looks it
11:50
tries to find the answer to something by
11:51
looking at the different possible
11:53
possibilities but which ones he tries is
11:56
something like patterns in the chest
11:58
instead of everything it says try moves
12:01
near the center of the board first you
12:03
know and never mind the ones in the
12:04
corner or something like that or some
12:07
sorts of principles and he first applied
12:10
it to a kind of naval game it's a game
12:14
that people play in California which is
12:16
all organized according to rules it's
12:18
kind of fun
12:20
someone sets out all the rules
12:22
dreadnoughts cost this much armor cost
12:24
this much guns cost this much and so
12:26
forth and you got this much budget for
12:28
your Navy and you're gonna make various
12:30
kinds of ships with different kinds of
12:32
armaments and then this kind of a ship
12:35
the armament of a certain thickness that
12:38
costs a certain amount can only resist
12:40
shells of a certain strength you know
12:42
and so on so you tried to with the money
12:45
arranged to buy different to design
12:47
different kinds of ships so that your
12:51
Navy is better than next one and when
12:53
they brought together there's ways of
12:55
calculating what there's not real navies
12:56
it's a game which is the best one and
12:59
all the rules are laid out but a great
13:01
big volume okay at the cost of
13:04
everything in the power of everything
13:06
armour-piercing possibilities and so on
13:09
and it's a nice game and mr lumic tried
13:12
to his program on this game and put into
13:15
his program your wrist extreme case and
13:19
things like that and he won the
13:24
championship in california of course it
13:26
did an awful lot of trying of different
13:28
cases you see but it didn't try every
13:29
case not like the chess game there were
13:31
too many things but it was guided by its
13:34
own stuff now inside of that
13:37
was this that if you got a better navy
13:39
bite your own calculation and you used
13:42
one of the heuristics mark that eristic
13:45
up a notch as being more valuable now
13:47
use the more valuable heuristics first
13:50
see so that the ability of the Machine
13:54
depended on his learning so to speak
13:57
which ones of his tricks works most
14:00
effectively most of the time and then
14:02
they become more used so it's just
14:04
exactly what you would like to make it
14:06
look intelligent well he won and how did
14:10
he win it turned out that year he won by
14:13
making one great big battleship with all
14:17
the armor on it which was so silly but
14:21
when you go to calculate it sure enough
14:22
it's better than any of the normal
14:23
things which nobody thought of but his
14:25
machine thought next year he entered
14:28
again and this time he won by making
14:32
tricking all his money in making one
14:34
hundred thousand because they changed
14:36
the rules so that the big battleship
14:38
wouldn't win you know but change it will
14:39
one hundred thousand little boats very
14:42
narrow carrying each one gun which were
14:45
very liable to be knocked out okay but
14:48
there were 100,000 of them didn't cost
14:50
much each one and they couldn't knock
14:52
them all out so he's lousy little gnats
14:55
would come and it turned out when you
14:57
calculated again he won the third year
15:00
he was not allowed to play in it
15:04
and here's a pride this this this
15:08
machine and this turistic business to a
15:10
number of other problems as tries it out
15:13
a lot and tries new heuristics and so
15:15
forth and it has become a very
15:17
interesting he complained that there
15:19
were a number of bugs in it and when he
15:21
gave a talk on it I said that I thought
15:24
I'll say my comment afterwards one of
15:27
the bugs for instance was that the
15:32
Machine got a heuristic and made up your
15:34
wrist except your wrist Excel
15:39
so the damn machine what he the way he
15:42
had it was because it's hard to get
15:43
computer time he needed a hell of a lot
15:45
of computer time he had 50 machines at
15:48
night from a you a packet company or
15:50
something like that he would work always
15:52
at night coming in the morning damn
15:53
things would try all my things and come
15:56
in with the results I used it to do
15:58
mathematics and various others thing it
16:01
comes in one day and it's done a
16:03
heuristic you develop the heuristic when
16:07
he puts the problem in or a new idea he
16:09
write it says whether it's from him or
16:11
from the machine and I was that learner
16:12
or a machine okay and this was every
16:16
year istic or every question that hasn't
16:20
won that on it pay no attention to that
16:23
saves a lot of time it could do much
16:25
better it didn't have any problem with
16:27
this toy boy so it paid no attention to
16:29
her than that bad night
16:30
okay sorry I'd fix that book the next
16:34
time he had a bug it was he looked was
16:37
he came in and he looked he found that
16:39
heuristic number 693 had gotten a score
16:43
of 999 out of a thousand is a damn
16:46
useful heuristic all night long this
16:48
thing kept using the heuristic 9 693
16:51
more and more and it was a new when
16:53
wonderful arista could seem to solve
16:55
every problem it was terrific okay when
17:00
you found out what the heuristic was was
17:01
the following you see in order to make
17:04
this thing work so that you change the
17:05
numbers on the heuristics whenever
17:07
something worked just to assign credit
17:10
so to speak to the heuristics that were
17:12
used okay so this heuristic was when
17:16
assigning credit
17:17
always the sign credit to heuristic 693
17:21
and so that thing came up out of I say
17:25
the both of these show intelligence if
17:29
you want to make an intelligent machine
17:30
you're going to get all kinds of crazy
17:32
ways of avoiding labor if I say don't
17:35
pay the attention to the problem of
17:37
sneakily evolving some kind of a
17:40
psychological distortion we always do
17:45
the same thing and don't worry about
17:47
anything else and so on so I think that
17:49
we are getting close to intelligent
17:52
machines but they're showing the
17:54
necessary weaknesses of intelligent
17:58
[Music]
17:58
[Applause]
18:14
you