Rise of AI Conference 2018 · Berlin

What Is the World’s Strongest AI?

Peter Bentley · CTO at Braintree

At Rise of AI 2018 in Berlin, Peter Bentley, CTO at Braintree, argues that most of what is called AI today is "weak AI": systems good at handling variety within narrow functional limits, such as reading license plates, but unable to do anything outside that task. He contrasts this with "strong AI", which has genuine intelligence, can tackle a broader range of applications, and can learn to do things outside its current scope. Bentley tests the audience with examples including Siri, which he shows misunderstanding a live query through statistical guessing rather than understanding, Amazon Echo, the Never-Ending Language Learner, the Cyc project, and DeepMind's AlphaZero, arguing each is a form of weak AI because it manipulates symbols or statistics without real understanding, echoing the philosophical "Chinese Room" argument. He then outlines what building strong AI would require: not imitating human intelligence, growing a broad range of tasks, allowing new algorithms to emerge, and doing far more testing than today's systems receive. He closes by arguing that strong AI may eventually be needed as companions and advisors to help people manage an increasingly complex world, while cautioning that simpler, explainable systems remain preferable for most applications, especially safety-critical ones.

Key takeaways

  • Weak AI handles variety only within a narrow, specific task, while strong AI can tackle a broader range of applications and learn beyond its original scope.
  • A recorded demo shows Siri repeatedly misunderstanding a question, illustrating that it guesses statistically rather than understanding meaning.
  • DeepMind's AlphaZero is still classed as weak AI because it handles one task at a time, needs heavy developer input, and suffers catastrophic forgetting.
  • The no-free-lunch theorem means no single algorithm works best for every problem, so strong AI likely needs many different specialized methods, like the brain's varied circuits.
  • Giving an AI more capability requires exponentially more testing, which is why driverless cars still underperform after millions of test miles.

Detailed notes

AI's ups and downs, and a career in it

Bentley describes starting in AI from childhood, building robots and programming computers, and earning a Bachelor of Science in artificial intelligence in the late 1980s or early 1990s, when only three UK universities taught AI because the field was considered discredited; his own university, University College London, avoided the term and called its department "intelligent systems" instead. He notes AI is as old as computer science itself, tracing it to pioneers including John von Neumann, Alan Turing, Claude Shannon and Marvin Minsky, and to the first AI conference in the 1950s. Having lived through AI's unfashionable periods, he expects that if current hype outruns results, the field could dip again, though people will keep working on the underlying technology regardless of the label.

Defining weak AI and strong AI

Bentley defines weak AI as AI that handles variety and noise well but only within a narrow, specific application, such as reading vehicle license plates despite variation in lighting, plate style and size; asked to drive a car or play chess, such a system simply cannot. He calls this a simulation of a part of human intelligence, not the real thing. Strong AI, by contrast, handles variety across a broader range of applications, has wider expertise, and critically can learn to do things outside its current scope; it need not resemble human intelligence to genuinely possess intelligence within its own framework. He frames the distinction with the idea that a weak AI can look and act intelligent while being, in his words, a cardboard cutout with a tape recorder behind it, rather than genuinely intelligent.

Testing the audience: Siri, Amazon Echo, NELL, Cyc, AlphaZero

Bentley runs an audience quiz on several well-known systems. He plays a recording of Siri repeatedly failing to answer a question about the psychologist Hans Eysenck, misreading the query as unrelated searches, to argue Siri is weak AI that uses statistics to guess the most likely meaning of speech rather than understanding it, which is why it sometimes changes its answer mid-response. He argues Amazon Echo suffers the same limitations. The Never-Ending Language Learner, which mines the web to build a database of facts with confidence scores, is shown producing both plausible and clearly wrong facts (for example, classifying vitamin A as a video game with 100 percent probability), which he says shows it manipulates symbols without knowing what they refer to. The Cyc project, which collects common-sense rules of thumb, is described the same way: it holds correct-sounding rules without understanding what the underlying concepts mean. DeepMind's AlphaZero is treated as the strongest example shown, a neural, sub-symbolic system that can be retrained across different games such as Go and chess, but Bentley still calls it weak AI because it only handles one task at a time, needs heavy developer input to acquire new knowledge, and, like many neural networks, suffers from catastrophic forgetting.

The Chinese Room argument

Bentley introduces John Searle's Chinese Room argument to explain why symbol manipulation is not the same as understanding: a person inside a room who does not know Chinese can still produce correct Chinese answers to Chinese questions by looking them up in filing cabinets, appearing to converse fluently while understanding nothing. He argues this is what most chatbots and similar systems, including Siri, are doing today: manipulating symbols convincingly without understanding meaning.

What it would take to build strong AI

Asked what the world's strongest AI is, Bentley says there is no answer yet, only systems that are relatively stronger because they generate their own solutions with less developer input. He lists what he believes strong AI requires: first, not trying to simulate human intelligence specifically, since true intelligence takes many forms in nature, from dolphins and ants to the immune system; second, giving an AI a growing range of tasks to solve, since intelligence emerges from the need to handle difficulty and generality rather than specialization; third, allowing the AI to build new algorithms for new tasks rather than relying on one method such as deep learning, citing the "no free lunch" theorem that no single algorithm is best for every problem, mirrored by the brain's many different neural circuits; and fourth, far more testing, since giving an AI more capability requires exponentially more testing to verify, which is why he says driverless cars still do not work as well as hoped after millions of test miles, and why he is skeptical of rapid singularity-style progress.

Why we need strong AI

Bentley argues most applications do not need strong AI and that simpler, explainable systems are preferable, especially for safety-critical work such as an autopilot system his company is building for airliners, which must be explainable, testable and certifiable. But he argues strong AI may eventually be needed as the world becomes too complex for people to manage alone, comparing this to struggling to operate his own home television, and suggests strong AI could act as companions, teachers and advisors that genuinely understand people's needs, emotions and surroundings rather than merely simulating understanding.

Quotes

It's a complete red herring to try and simulate human intelligence.

Peter Bentley

There is no one-size-fits-all for intelligence.

Peter Bentley

Do not underestimate the need for testing.

Peter Bentley

Questions answered in this talk

What is the difference between weak AI and strong AI?

According to Peter Bentley at Rise of AI 2018, weak AI handles variety and noise well but only within a narrow, specific task and does not really understand what it processes, while strong AI has genuine intelligence, can tackle a broader range of applications, and can learn to do things outside its original scope.

Is Siri an example of strong AI?

According to Peter Bentley at Rise of AI 2018, Siri is weak AI: a live demo shows it repeatedly misunderstanding a question about a psychologist because it statistically guesses the most likely meaning of speech rather than genuinely understanding it.

Why is it hard to build strong AI today?

According to Peter Bentley at Rise of AI 2018, strong AI requires a growing range of tasks, new algorithms rather than a single method, and exponentially more testing as capability increases, which is why even heavily tested systems like driverless cars still fall short.

Mentioned

Strong AI, Weak AI, Chinese Room argument, AlphaZero, Never-Ending Language Learner, Catastrophic forgetting, No free lunch theorem

Keep exploring

More from Rise of AI Conference 2018

More on AGI & Future of Intelligence

Transcript

Auto-generated transcript, not edited. · 4,025 words

Show the full transcript

okay I should mention because I've seen what happens when it is just mentioned it was my birthday yesterday so okay so I agree a lot for the last speaker I should say and my background is is not dissimilar to his I grew up engineer as a father so I grew up building little robots so from the age of seven or eight I was making robots so is programming little computers and I did my degree my first degree in artificial intelligence so I have a Bachelor of Science in artificial intelligence but I did that in the late eighties early 90s and in those days that was in one of those dip periods in the UK there are only three universities that would teach a course on AI because it was basically considered to be discredited at that point and when I went on later to University College London well they didn't even have an AI department they called it an intelligent systems department because you couldn't use the words AI it was completely discredited we tried

methods we tried expert systems we tried things like that they didn't work so this idea that AI is super fashionable well you know some of us has been through the ups and downs once or twice it's very nice for me right now it's nice to be fashionable but I've been there when it's really not fashionable and I'll keep doing it when it's not fashionable again and yeah I'm cynical enough to say it's quite possible that if there's too much hype expectations will not be met and then it will go down in this dip okay what happens when AI is not fashionable anymore people like us continue to do it we don't call it AI anymore but we just keep doing it it still goes in the technology it's it's there already you use AI and lots of different technologies that was invented decades ago the field of AI is as early as computer science itself some of our pioneers of computer science john von neumann Alan Turing Shannon who was Shannon's PhD student Marvin Minsky

Shannon and Minsky created the first conference on AI and we're talking 50s so this is not new stuff but today I'm going to talk about something more specific I'm going to talk about strong AI and in some respects you could consider this as a slightly philosophical discussion does it matter does it matter if we have strong AI or weak AI I maybe I'll let you decide whether it matters or not to me a kind of matters but then I'm a researcher I care about these things but first let's let's think about what these things are what is weak AI well we ka is a kind of AI that's good at handling variety within its functional limits well that's actually what AI is for right a normal software normal control software let's say it just does what it's programmed to do we create some kind of intelligence some kind of smarts when you have variability when you have noise when you have a little bit of unexpected stuff that might happen so we kiii does that

within its functional limits it can handle that noise that variation but it is specialized for specific application maybe it reads license plates number of plates of vehicles on a motorway it's very good at handling that there's lots of variation there's lots of different kinds of number plates there's lots of different lighting there's lots of different vehicles different size plates they're from different places so you need a specialist AI to read those numbers but that's all it's going to do it's not gonna do anything else so it has a very narrow expertise it doesn't know about anything else it only knows about images of license plates if you had asked that license plate reading AI could it drive a car well obviously can't if you could you ask it to play game of chess well obviously it's not going to be able to doesn't mean it's not an AI it just means it's a weak AI it's very specific to its task so in some respects you could say well it's simulating a part of human

intelligence because we can read number plates we can read a license plate on a vehicle so if a machine can do that well it's kind of simulating what we can do the way it does it is not really the way we do it at all but it can do it and so in many respects it's simulating this kind of intelligence doesn't really possess it so if we contrast this idea of weak AI with strong AI well strong AI is also an AI is good at handling variety within its functional limits that's what AI is for but in this case it's able to tackle a range of different applications not just one it can tackle more things it has broader expertise it knows about more things in the world and very importantly if it doesn't know about something a strong AI can potentially learn how to do something else it can broaden its horizons it can learn how to do something outside its current scope now strong AI may not have human intelligence but it does

have intelligence and this I think is the critical part between weak AI and strong AI weak AI well you might have heard this expression if it looks like a duck and it quacks like a duck it must be a duck right well this is the heart of the debate a weak AI might look like a duck and quacks like a duck but it might be a cardboard cutout with a tape recorder behind it it's not a duck it's pretending to be a duck a strong AI really is strong AI really has some kind of intelligence now I should stress it that doesn't mean it has to be human intelligence doesn't have to be able to talk doesn't have to be able to walk it doesn't have to be able to do anything that we do but it has got an ability to process information in an intelligent way and it does that genuinely doesn't pretend to do it it really does it really understands within its framework so that's what a strong AI is as

opposed to a weak AI so let's think about things shall we let's do a pop quiz is it strong AI or is it weak AI I'm gonna go to a few examples for all of you City and get it right or not okay Siri strong AI or weak AI hands up for strong AI okay hands up for weak AI okay lots of people think Oh weak okay so I won't do this for all of my examples but for this example when I was preparing the slides I thought I'd have a conversation with Siri and this was on my phone I happen to be doing some research on a psychologist called I think it's called mind SEC see the first thing you'll see is my phone with the browser open on this person hans icing you would think if the browser is open on the phone with that image and I asked Siri about that same person a smart AI would kind of say well you're ready looking at it right let's see what actually happened

hey Siri tell me about I sank I said overall p.m. euros is falling they're currently down that's all right thank you yes it is hey Cerie tell me about hans i think okay one option I see is deep blue which averages four stars is that the one you're looking for first off hey Suri tell me about I think I found one option akia on Valley Park in Croydon which averages four stars and is moderately priced it's not the one you want what are you talking about no no looks like that's it what hey Siri tell me about the psychologist called the hans eysenck please I found one option concise control is limited is that the one you're looking for what it's not the one you're looking for please no calling concise controls Oh No stop stop oh my god it's still doing it stop stop stop right so with that in mind who thinks it's strong AI and who thinks it's weak AI okay now I'm being unkind of course but it's very clear but

it's a bit thick right Siri is very good at recognising our voices but this is a simulation of intelligence it there's clearly no understanding of what we're saying what it's very much I could use the use the Kris sells Chinese room argument how many of you know about the Chinese room argument here okay not enough I'll educate you so some time ago a philosopher called Christelle described how a computer could be like an AI could be like a room a room in which there are lots of filing cabinets and there's a counter a desk at the front if you want to ask someone a question in Chinese you write it down in Chinese you put it through the window and someone in this room takes this Chinese writing they go the filing cabinets and they look up through the indexes and they find a special bit of paper in response and they hand it back now to you who's providing the question in Chinese and getting answers back in Chinese he'll get is like you're

having a conversation it's like it's you're getting a perfect response because whatever they're doing in this room is really seeming just like a conversation I put in hello I get hello back I say how are you I get I'm fine today back but the people inside this Chinese room don't understand Chinese they don't even understand that these are these are symbols that these mean anything it's purely a symbolic processing algorithm so Chris Searle was saying yeah this is a kind of AI but this is why it's very hard to talk about strong iron through intelligence because it's just pretending to be intelligent and this is actually exactly what you're getting with the most chatbox today and many other kinds of this sort of technology it's pretending to be intelligent it's doing symbol manipulation Siri doesn't understand what the words mean it's actually using a lot of statistics to figure out the most likely thing I said that's why you saw some of those words change after sometimes it almost gets it writer than it changes

its mind because it figures statistically I'm much more likely to have said something else so it changes the words based on statistics Elly's tries to figure out statistically what am I most likely to be asking for it seemed to think statistically I wanted to either eat fish and chips or go to Ikea or cause some kind of random company that I never heard of so intelligence shouldn't be about that and isn't about that so while this works very well for many applications it's not strong AI what about Amazon echo who thinks it's strong AI okay he thinks it's a weak AI very good so I happen to know the guy who created the first company behind the Amazon echo device which was then acquired by Amazon so I decided I wouldn't do another video for Amazon echo I could have it suffers from all the same problems as Siri okay what about this one anyone heard of the neverending language learner okay so what this does is it searches the internet it tries to look

at what is written in our collective knowledge and derive facts so it creates a big network of a database of facts can then be interrogated and it can then infer things with I don't if you can see some of these facts again I did not manipulate this these are some of the facts that it was learning on the day I prepared these slides some of them make a lot of sense it thinks a small cut is an item found on the floor with a 95% probability yep that that's quite possible it also thinks prescription appetite suppressant caps is a muscle with 99.9% probability now I don't think that's quite right it also says it thinks vitamin A is a video game with a hundred percent likelihood now I did really look on like because there's a lot of it there's a lot of games out there I really did look for a game called vitamin A to my knowledge there isn't one so again I won't even ask you because I think I've kind of

slightly destroyed now but that's not to say Nellie's a bad system it's doing exactly what we asked it to do but it is collecting a lot of these facts it doesn't understand what these things are it doesn't know what vitamin A is in the world it doesn't know what a video game is it's manipulating symbols and linking them together in a big database plus it is a weak AI you remember at the start I said does it matter well maybe it doesn't what about this one loose it has come out of the psych project and psych for those of you who don't know that you should know is a very old and long-running project that aims to collect if you like rules of thumb heuristics common-sense knowledge throughout humanity bring it all together so you've got a computer that somehow no sensible stuff what does that mean does that mean it's truly intelligent I'm afraid not yes it does know a lot of these different rules of thumb and facts and a lot of them

are right but it knows them in terms of symbols again it doesn't know that how what's a good rule of thumb I don't know and anyone got a good rule of thumb for me don't trip over something you might hurt yourself that's a good rule of thumb it doesn't know what any of those things are it doesn't know what tripping over is it doesn't even know what hurting yourself is it's it's it's manipulating symbols again that doesn't mean it's not useful but in terms of intelligence it's a simulation of intelligence so it's a weak AI what about this one then this one I will ask you guys is alpha zero a strong AI or a weak AI hands up for strong AI very scared for the hands oh now hands up for weak AI okay and a lot of undecideds acting this time so this is an interesting one deep mine now for zero I like it and the fact I like it's nothing to do with the fact that a lot of the people

who've earned deepmind came from UCL well that helps it's a neural network based approach it's a sub symbolic approach it's not processing symbols it can solve different problems it's able to play go it can then be retrained and play chess it can solve and has solved many different problems so why then do I still say it's a weak AI it is a weak AI it's let's say it's a lot stronger than most other ai's but it is a weak AI and it's a weak AI because it is only limited to a single thing at a time and it really does require a huge amount of our input to force it to have knowledge on specific things and like many neural networks there's quite a lot of issues with something called catastrophic forgetting where it can learn one thing and then you teach it something else and it forgets the first thing and that's not ideal so it's good it's perhaps not as good as we might like it to be so what is the world's

strongest I I'm really sorry to have to give you this answer there isn't one the world's strongest AI well with still working on that but there are some that are stronger than others those ai's that can learn to soul from by themselves they create their own solutions with less input from the developers they are stronger compared the eyes that are simply programmed to process existing concepts now how would we make a strong AI I'm a computer scientist I have a research lab I worry about these things number one it's a complete red herring to try and simulate human intelligence true intelligence has nothing to do with us dolphins have true intelligence an ant has true intelligence at all different forms of intelligence our immune system is a kind of intelligence it's not the sort that we often talk about and our immune system isn't very good at chatting to us it does some really useful stuff so if you want to make a strong AI don't keep going on about speech recognition chat box rule

bases or playing board games that is not the route to strong AI true intelligence is a very difficult to achieve thing it emerges through the need to solve difficult problems how do you make it you have to provide a growing number of very tasks a strong AI can do more than one thing it's a generalist it it has to adapt to and learn to tackle new challenges and that's really why nature created in in the first place if you want to be a super specialist or be a tree a tree doesn't need a brain a tree is a great specialist at being a tree in fact if you look throughout nature most organisms don't have brains those with complex brains are very much in the minority because it takes a tremendous amount of effort and energy and time to make a brain intelligence is a long and slow and difficult thing to make so number three you have to allow your AI to build new algorithms to solve new tasks this is a real issue

that many people think they think I've got the new algorithm it's deep learning it's whatever it's deep something it's knowledge this it's case based that whatever it is although I'm sure our last speaker wasn't claiming that case based reasoning would be the answer to everything he says no so exactly so there is no one-size-fits-all for intelligence indeed if you look at the brain there are made different neural circuits we have different ways that we solve different tasks in our brains so that means and indeed we've even got theories and computer science it's called the no free lunch theorem there is no single algorithm that is better than all other algorithms for all given problems that means you actually need a different algorithm for different kinds of problem so it's no coincidence that in our brains we've got layer upon layer different solutions that have been built up over time to allow us to tackle different problems and that is what we expect strong AI is going to look like we have a bit of a

problem in that in computer science we still actually don't know which algorithms are best suited to which problem so we have to do a lot trial error we're still working on that too number four testing testing is the killer okay now what ever talks about testing but if you think about driverless cars how many millions of miles have been driven and things still don't work quite as well as we hope they would in fact this is an exponential going in the wrong direction that's why I really don't like singularity because I think it's nonsense the exponential goes the other way as soon as you give an AI a little bit more capability it can potentially do exponentially more things if it could do exponentially more things you've got to do exponentially more testing and that's what nature teaches us as well Nature has had to do billions of years of testing in parallel over countless numbers of examples of brains through countless examples of problems do not underestimate the need for testing and this is

why it doesn't matter how much computing power you've got look at Google look at Amazon we always fill up the computing power that we have to capacity just to try and test the things that we create we always do so testing is that slow down that negative exponential now I know I'm going over time so I'll try and speed up so why do we need strong AI well do we need it for most applications we probably don't does it matter if it's simulating intelligence if it's useful and maybe it doesn't and indeed I would argue strongly because I also have an AI company and we also need systems that can be explained to our clients and we also have systems for example one at one of my guys is creating a new autopilot system for airliners now this is a safety critical system you've got to know how it works so this is why we try and create systems that you can explain that you can test that you can certify this is very important

for safety critical systems so simpler is usually better by all means make it work but strong AI yeah maybe just don't worry about it but maybe strong AI might have a use one day why do we need strong AI well my answer is probably the same to all these other questions why do we want pets why do we like companionship why do we want someone to listen to us and understand our needs why do we want personalized education health care careers allow us to achieve our best why do we need help in our ever complex digital worlds well frankly because we're complicated with very complicated creatures excessively complicated creatures because there's too many office we're making an excessively complicated world for ourselves and because frankly we need help now I'm not saying it's even a good thing I am I'm a professor I work in this area I try and create these technologies I'm not saying it's even a good thing we're doing it one could argue that the digitalization of our world and making

it evermore complex is not the cleverest thing to do but one could also argue if we must have this many billions of people on our planet than the only way to keep them healthy and fed and alive is through these technologies and therefore strong AI will only be needed in the future because soon which is not going to be clever enough to look after ourselves anymore and I sometimes feel like that's already happened I mean I've got a TV at home with online things and satellites that I can barely control my own TV and I've got a PhD in computer science so the time is coming quite soon when the world will be too complicated for us so many choices in health care so many choices in power providers and try sport and food what's the optimum solution for us how should we be living so we need a solution we need something that's genuinely intelligent something that genuinely actually understands things doesn't pretend to understand really understands that there are things such as people

who have things such as emotions and there are such things as power stations and environment and animals going extinct we want an intelligence that will actually be our teachers our advisors our friends nothing to be afraid of these will be our companions these will be things that are designed to help us and they probably won't look like that either okay thank you

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.