Skip to main content
Loading Explore guides…
AI and Future Technology · Acedra
Home
Explore
Computer Science
AI and Future Technology
Sign in to save
Chat
If YouTube asks you to sign in, open the video on YouTube.
Watch on YouTube
Reload player
Transcript
20 segments
Copy
Hide video
0:00
it's probably like nuclear fusion you know yours people when will the first nuclear fusion power be available and the answer is always in 50 years time and the answer is and when will there be genuinely intelligent machines well I might be very wrong here but some optimists tell within 10 years I said well maybe it's 50 years maybe it will always be 50 years I just do not know Turing was fascinated with the idea of how intelligent could you make a machine be the imitation game itself was a game
0:32
Turing devised to what yes the idea was that if you're at the other end of a telephone line and you gotta, I think in those days of course it was a teletype or a dumb terminal you know you type questions at something and the answer comes back and if after an agreed period of time you're then asked was that a human you were talking to or was it a machine if you can't decide that it's not human then it's passed the test it's masqueraded as a human I think what he realized all along perhaps teasing
1:05
as a little and it's always seemed to me is that with a limited bandwidth for interaction like that it's very probably one of Turing's own undecidable problem you can't get enough information to come to a rational decision as to whether it's a human or a machine but this doesn't stop there being i think, is this right, there's a yearly competition now they normally give scores like you know sixty percent thought it was human forty percent thought it was machine then
1:31
answer is it's machine and stuff like that to me you see it really is it as i was saying a case of undecidability if you need to visit the other end of the line and see whether there's somebody typing teletype to come to a decision but then of course that spoils the whole nature of the test it was a topic that fascinated Turing and one that he really really cared about and it's interesting had he lived i'd love to know what he makes of the present day artificial intelligence scene and just
2:02
what is easy to do in the AI field and what is hard and my own observation and i'm not an expert sorry fully expect to be put in my place and told her royal is that the computer could do a fixed task in a narrow domain very differently to a human and be super batted example chess absolutely superb example I think deep blue be Kasparov way back when but it absolutely did not use the pattern matching analysis type thing with the Grandmaster would do it's brute force you know the whole game tree
2:41
as it were chess if I do this don't do that then do that I mean your average chess player might be able to look six or seven moves ahead over all possibilities blast deep blue can go I lord knows how many moves ahead and actually this I think is is it is a good examplar of this because what grandmasters were missing and actually what I think ken Thompson's chess machine Bell help to fill in is he I think he did all five piece endings or something like that and the ones that
3:16
the grandmasters didn't know about because they needed tons of look ahead I think that is right is the ones way you absolutely do something stupid you go miles or what he and the rest of them say don't do that no no you don't do that this machine's stupid it's gonna lose it's gonna lose you know then right at the very end a tiny little thing that nobody had noticed and from the absolute depths of near defeat he suddenly finds a sneaky little way to make this pawn do
3:45
that matter now before we know where you are it's a squeaker narrow victory in only 10,000 moves or something like that you know I'm exaggerating but you get the idea that it's those where you would have to go miles away from the solution into an area where human said there's no hope it's everything is lost but it isn't there's a sneaky little Holly way through to getting a victory that's the kind of things a machine can do but a
4:08
human can't. Then the other side of that is the travelling salesman problem where there are so many possibilities but a human would just look at an opera go well i'm going to try that well there is a connection there's an interesting one too you're quite right that it's one of these classic NP complete problems you it is believed that there is no way to get the perfect answer than simply to look at every single possibility and the monumental way the combinations of those
4:35
add up is just mind-boggling however what I think people in the field will tell you is there is a very good set of algorithms which will give you a pretty darn good answer even though it might not be absolutely perfect and that's the differences between you know good enough within five or six percent in the best versus must be absolutely the best route so that kind of brings you back to the whole human versus computer Turing thing which is if I'm on the end of the
5:06
phone or the teletype and I say I'd know what color is music yeah this is the problem about that Turing test game you know you could get either end of the equation trying to play silly games I think there was a program called Perry probably invented was it a Stanford somewhere in America Perry was going to replicate the speech patterns of a paranoid schizophrenic you say and you can imagine if you're up against that and you don't know whether the other end is actually a real schizophrenic person
5:40
or a program simulating that behavior that is far harder to tell the difference or on because you know you you ask some perfectly ordinary question like what is two plus two I don't know why you're focusing on two it's my hated number and all this kind of stuff it makes it very very difficult to start saying that is a human but one with a mental problem versus that is a program that's simulating that sort of thing so I think it's the thing that AI always runs into is this that getting a part
6:13
way answer but by very different routes than a human would I'm giving you enough information about navigating to return to be useful fine is when you start wanting total perfection and fantastic subtlety but because things fall apart because that's the thing as well ambiguity that humans are accustomed to and the idea of inflection in town yeah so for instance ok Google how do you pronounce something in Hungarian
6:44
so it's given us the word for something in Hungarian and bess it it's told us it pronounced it like this follow me but I'm being more general about it so I want to know how Hungarians pronounce things on the air fabulous things like ambiguity and use of similes and metaphors is a classic for throwing things off there's a real classic from the seventies I think it is is you know okay artificial intelligence program analyze the difference between time flies like an arrow and fruit flies like
7:23
a banana now that's a beauty because you know time flies like an arrow it's a simile and all this kind of thing extension of a metaphor but on the other hand the other one which on the surface has got a very similar grammatical structure Time flies is a noun phrase like ah it's not like you've been used in the similar sense it's like is a joy and a clearly eating a banana you say now it's that kind of subtlety and that kind of knowledge of how the real world works
7:56
allied to syntactic analysis that's what you and I use all the time you know and what computer systems would find so very very difficult but of course you're inviting the computer program to sort of say hey there must be a thing called a time fly and they live on an exclusive diet of a truce oh yeah we know that's impossible so we're using that kind of knowledge over this whole domain of what the real world is like in order to disambiguate between those two all the
8:25
successes of artificial intelligence tend to be a narrow domains you know if you're like so play chess is one example there's a thing they did it stanford mice in i think it was about antibiotics and their effect and the great advantage of course are you with a computer is if it is all done by logic then it will do it and it will not get tired and it will not usually make mistakes unless you're counting up Gangnam style hits but that's the point you know you don't get bored you don't
8:56
make mistakes if you're a computer so therefore that's why they often outperform humans roger penrose very famous mathematician at oxford got really rather hot under the collar about all this I think he wrote a book called something like the emperor's new mind had a real go at artificial intelligence basically said the trouble is you're assuming that the human brain is just like a very complex Turing machine what I am convinced it's more than that i Roger Penrose think
9:23
there are quantum effects and if only my neurological medical friends would tell me how to find them and how to categorize them I'm able to tell you more about the computational model that is really going on in the human brain but I've not read anything about that recently so it's on the to-do list I think
9:43
friendly environment by enveloping around the agent in question Lemaitre every time that had got up to nine it would nudge the dial to the left of it to move on one plus they get stuck in what's called a local maximum or a local minimum on the left will be 2 to the power zeros
Here to help you learn
Sign in to keep a private conversation with your course tutor.
Ask me anything about the material…
Sign in to chat