What does the future hold for AI?
We're coming back to the topic of generative artificial intelligence, asking how this potentially gamechanging technology is going to be integrated into our society. We'll hear an explanation of neural networks from Geoff Hinton, one of the founding fathers of AI, and some of the most promising avenues for maximising the strengths of machine learning systems with tech journalist David McClelland. After a brief update on the debate around AI sentience from the foothills of the Himalayas from Nicky Clayton, we explore why chatbots might be about to stop advancing as rapidly as before, and how safety experts are demanding that technology companies lift the curtain on their operations, with Mike Pound and Gillian Hadfield.
In this episode
01:05 - How does AI actually work?
How does AI actually work?
Geoff Hinton
We start with an explanation of how artificial intelligence actually works. There’s nobody better to do that than the man dubbed the ‘godfather of AI’: Geoffrey Hinton. He began by telling Chris Smith about the development of AI in the 1970s…
Geoff - At the time, the idea of doing science by simulating things on computers was fairly new and it seemed like the right approach to try to understand how the brain learns, or at least a complimentary approach to doing experiments on the brain. So, I spent my time when I was a graduate student in Edinburgh writing computer programmes that pretended to be networks of brain cells trying to answer the question, how should the connections between brain cells change so the collection of brain cells, hooked up in a network, can learn to do complicated things like for example recognise an object in an image or recognise a word in speech or understand a natural language sentence.
Chris - Do we have a clear idea even today of how that works? Because obviously you were working towards something we had no idea about and trying to model it. Have we got there or are we still in the dark?
Geoff - Neither of those. We haven't fully got there, but we have a much better understanding. So we now have computer models of neural networks, things that run on a computer but pretend their networks are brain cells that work really well. You see that in these larger language models and in the fact that your cell phone can recognise objects now, it can also recognise speech. So we understand how to make things like that work and we understand that brains quite like many of those things. We are not quite sure exactly how the brain learns, but we have a much better idea of what it is that it learns. It learns to behave like one of these big neural networks.
Chris - If it's down to the fact that we've got brain cells talking to brain cells and they're just big populations of connections, is that not relatively easy to model? What's the holdup? Why is it hard to do this?
Geoff - Well, the tricky thing is coming up with the rule about how the strength of a connection should change as a result of the experience the network gets. So, for example, very early on in the 1940s or maybe early 1950s, a psychologist called Hebb had the idea that if two neurons, two brain cells, fire at the same time, then the connection between them will get stronger. If you try and simulate that on a computer, you discover that all the connections get much too strong and the whole thing blows up. You have to have some way of making them weaker too.
Chris - I love that line: what fires together, wires together. It's never left me. because I remember reading Hebb's book when I was at University College London. So how did you try and address that, then? Was it just a damping problem? You make it so that the nerve cells get bored more easily, as it were, so that it doesn't overheat in the way that the computer would otherwise have them do?
Geoff - Well, that's the first thing you think of and you try that and it still doesn't work very well. So, the problem is, can you get it to work well enough so that it can do complicated things like recognise an object in an image or, in the old days, recognise something like a handwritten digit. So you take lots of examples of twos and threes and so on and you see if you can make it recognise which is a two and which is a three. It turns out that's quite tricky, and you try various different learning rules to discover which ones work, and then you learn a lot more about what works and what doesn't work.
Chris - What does and doesn't work, and why?
Geoff - I'll tell you something that does work because that's obviously more interesting. You have a layer of neurons that pretend to be the pixels. An image consists of a whole bunch of pixels, and the pixels have different brightnesses and that's what an image is. It's just numbers that say how bright each pixel is. So that's the input neurons, they're telling you the brightness of pixels, and then you have output neurons. If you're recognizing digits, you might have 10 output neurons and they're telling you which digit it is, and typically the network, at least to begin with, wouldn't be sure, so it would hedge its bets and say, it's probably a two, it might just be a three, it's certainly not a four. It would represent that by the output unit for a two being fairly active. The output unit for a three would be a little bit active, and the output unit for a four would be completely silent, and now the question is, how do you get those pixels as inputs to cause those activities in the outputs? Here's a way to do it that all the big neural networks now use: so, this is the same algorithm as is used to train big chatbots like GPT-4. It's used to train the things that recognise objects and images and it's called back propagation. It works like this: you have some layers of neurons between the inputs and the outputs. So, the neurons that represent the pixel intensities have connections to the first hidden layer, and then the second hidden layer, and then the third hidden layer, and finally to the outputs. So they're called hidden layers because you don't know to begin with what they should be doing.
You start off with just random connections in these networks. The network obviously doesn't do anything sensible and, when you put in an image of a digit, it will typically hedge its bets across all the possible 10 digits and say they're all more or less equally likely because it hasn't got a clue what's going on. Then, you ask the following question, how could I change one of the strengths of the connections between a neuron in one layer and neuron in another layer so that it gets a little bit better at getting the right answer. So suppose you're just trying to tell the difference between twos and threes to begin with. You give it a two and it says with a probability 0.5 it's a two with a probability 0.5 it's a three. It's hedging its bets. You ask, well, how could I change connection strength so that it would say 51%, two and 49% three.
You can imagine doing that by just tinkering with the connections. You could choose one of the connection strengths in the network and you can make it a little bit stronger and see if that makes the network work better or work worse if it makes it work worse. Obviously, you make that connection a little bit weaker and that's a bit like evolution. It's obvious that in the end that will work, but it would take huge amounts of time. So, in the early days, we would use networks that had thousands of connections. Now these big chatbots have trillions of connections and it would just take forever to train it that way. But you can achieve pretty much the same thing by this algorithm called back propagation. What you do is you put in an image, let's say it's a two, the weights are initially random weights on the connections.
Information will flow forward through the network and it'll say 50% it's a two and 50% it's a three. Now you send a message back through the network and the message you send back is really saying, I'd like you to make it more likely to be a two and less likely to be a three. If you send the message back in the right way, you can figure out for all the connections at the same time how to change them a little bit so the answer is a little bit more correct. That's called backpropagation. It uses calculus but it's essentially doing this tinkering with connection strengths that evolution would do by just changing one at a time. But the backpropagation algorithm can figure out for all of them at the same time how to change each one a tiny bit to make things work better. So if you have a trillion connections, that's a trillion times more efficient than just changing one and seeing what happens.
What can AI do?
David McClelland, Technology journalist
So now we have a better understanding of how AI works: it’s a long training process of fine tuning neural networks until we get the outputs we’re looking for. And it can lead to extremely exciting technologies, not least the language models that have sent big tech companies scrambling to integrate AI systems into their products. But there are other and, for the moment at least, more tangible, real world examples of how it might actually be able to improve our lives. Chris Smith spoke with the tech journalist David McClelland…
David - Artificial intelligence, it isn't really a new technology or concept. You can go back centuries and people first started talking about AI. You look at the birth of modern AI back in the 1940s, 1950s - Turing, for example - but, over the last 18 or 24 months or so, I'd say, I think we've been riding what some would call an "innovation wave", or others might call it a "hype cycle" that's been powered by a subset of AI known as generative AI. And very crudely generative AI is a type of AI that can create new synthetic content, words, pictures, video based upon what it's been trained on. So, for example, I might have a tool like chat, GPT summarise the plot of lame, miserable, the musical in the style of a Shakespearean sonnet. And because that's been trained on, it's ingested the entire works of Shakespeare and it's got various synopses of West End musicals, it'll very quickly come back to me with a tight 14 lines on "Les Mis". But the thing to remember about these AI systems, they aren't really knowledge based. They are statistical. So chatGPT has got no concept of what a sonnet or a West End musical actually is. It is essentially just guessing what word is most likely to come next in a sentence based upon what it's come across before. And the same is true for these generative AI systems that can create amazing photo, realistically, photo realistic images and even videos. These outputs can look incredibly real, but they're also prone to some basic errors like misrepresenting human hands with too many or too few fingers because, again, it doesn't actually understand what a hand is. So my take on the overall AI narrative at the moment, generative AI, is that we are still at this "party tricks" phase. There's a lot of wow moments that look great in presentations and can impress friends indeed scare people as well. But examples of where generative AI is really adding and creating real value at scale to humanity are, are somewhat slower to emerge, but I think we are starting to see them now.
Chris - Yeah, indeed. We are seeing some examples in the medical space, for example, aren't we? We have seen people using this in radiotherapy for cancer treatment in imaging because AI can be taught to be better than we are at spotting certain diagnoses. It's that kind of thing. It does appear to be getting some traction?
David - Yeah, and there's a couple of areas, both like you say, in in medical imaging using more traditional machine learning models where these systems can update their knowledge based upon what they're seeing to identify cancer from scans. Studies have shown that they can perform at least on par with medical experts, faster than humans, which is really important when skilled resources around the world might be at a premium and waiting lists in certain nations are starting to grow as well. Some recent researchers found that AI plus human review of breast cancer screenings - this was a study in Sweden - it can increase the detection rate beyond just humans alone, even two humans looking at it. And there was an NHS trial in Scotland early this year of a tool called Mia, which was able to identify early stage cancers that doctors had failed to identify themselves. And that kind of benefit means much less invasive procedures for a cancer patient later on down the line, a much higher survival rate, but also with generative AI as well. One area of generative AI, which really does seem to be showing some promise, is chatbots. Now, chatbots have had a bit of a bad rap over the years. You know, whether they're used for customer services, for example, and they don't quite understand you, the more modern chatbots that use generative AI capabilities are really showing some potential. So academics at the University of Cambridge have been researching the use of chat GPT chatbots to triage people with potential eye or eyesight problems as a way of deciding which patients need urgently to be seen by specialists. And one of the most recent versions of the chatGPT model, chatGPT-4, it was found to perform better than junior doctors. And it's not expected to replace eye doctors. But these tools can support GPs and other non specialist medical practitioners in helping speed patient care and save people's eyes. And again, maybe that's something that in, in some countries we take for granted the availability of medical staff. But as you start looking elsewhere around the world, even having somebody who can just answer basic questions and steer you to the right course of of medical help can be invaluable.
Chris - It's a bit like the sort of "e-equivalent" of an exoskeleton suit that, you know, you hear people on production lines having these extra suits they can strap on to give them more strength to do bigger jobs. Because computer coders are saying that it's taking some of the grunt out of relentlessly recoding, or coding up bits of work; that people have already solved that problem, but instead of having to go and find that solution, string it together in the right way, you just tell a chatGPT type system, "go and get this and assemble me some code to do the following job." And it does it.
David - Developers are finding generative AI and have done for, for a few years now, incredibly useful. And given that the semantics of computer code are fairly rigid, very, very rigid. In fact, it actually plays very much into large language models and generative AI systems hands. It's the same as with some areas of creative work as well. It's that problem of having the blank piece of paper. How do you get started with a particular problem? What does the basic research look like? And whether you are a journalist or whether you are a software developer, being able to type a prompt, and this is what, you know, we're used to typing into a search engine, a search query. When you are conversing with a generative AI system, we, we type in what's known as a prompt. "Create something for me that does this. And it sounds like that," for example. When one is creating those prompts, you can get something back that might get you 70, 80, maybe 90% of the way there. It might not get you all of the way there, but you then as the, as the journalist, you become an editor and you are editing some work and making sure that it is factually accurate, that it, it tells the story that you want to create. And the same with a software developer as well. Does it solve the problem? Are there any gaps in the code? Is there anything that you might need to tweak? You can save so much time doing some of the basic work and it enables you to operate potentially at a, at a much higher level. Some people would, might be concerned that AI will come for their jobs and there is a a saying that it's not that you'll be replaced by an AI, it's that you may be replaced by a person who knows how to use an AI. And what I've been trying to do in my work is to find ways where I can use an AI to help me with some of my more menial tasks. And more often than not, I'm finding it to be very helpful.
Chris - Does it write better than you can?
David - No comment!
18:13 - Is an AI alive?
Is an AI alive?
Nicky Clayton, University of Cambridge
Indeed, AI has clearly come on leaps and bounds. And it’s raised another question: is it able to think and act like we do? Last summer, Google fired an engineer who claimed that one of the company’s AI systems had “become sentient”, and the human-like qualities of AI has sparked debate between groups as diverse as computer scientists, philosophers, cognitive scientists and even Buddhist monks. Nicky Clayton, professor of comparative cognition at the University of Cambridge, told Chris Smith about a recent summit on the topic in Nepal…
Nicky - Well, I was in Katmandu at a wonderful conference on how to study and understand non-human consciousness and by non-human they meant both animals and artificial intelligence.
Chris - And who was at the conference?
Nicky - Well, the usual suspects, neuroscientists and philosophers and psychologists as you would expect and computer scientists, but with an interesting and important twist, a lot of Buddhist monks.
Chris - Why were they there? Apart from the fact it's Nepal, which is next door to where they're from.
Nicky - I suppose it was to explore big picture questions about consciousness, for Buddhists assume that all animals not just humans are conscious and they found it very hard to imagine that a machine might have consciousness. And I agree, it's one thing to say that a machine can do all kinds of amazing things that it's programmed to do and outsmart people at the game of chess, for example, but that they could actually reflect on their thoughts, reminisce about the past, imagine the future, and strategise: when is option one better than option 37. I find that hard to imagine. So I guess I'm on the side of the Buddhist monks on this one.
Chris - Did the computer scientists disagree with you, then? Do they think their machines are alive or was everyone of like mind?
Nicky - Not everyone was of like mind, but many of the computer scientists also agree that machines can do a lot because of how they're preprogrammed. These language models are very persuasive about some of the things that the machines can learn so quickly. But most people, I think, were not convinced that machines have consciousness.
How will AI evolve?
Mike Pound, University of Nottingham
Will more data and more sophisticated models mean generative AI continues to improve? Or will we see diminishing returns from this burgeoning field? Mike Pound is an associate professor in computer vision at the University of Nottingham.
Mike - I think it is getting more powerful. Continued development in the research community means that some of these methods are now much more performant than they were before. They produce really, really impressive output. But I also think that the way we use these tools has become easier. So, chat bots, there are websites where you can go and talk to them and that gives people the impression that AI is on the rise even though it's been happening behind the scenes for quite a few years.
Chris - In your view then, is it going to continue on the present trajectory or is it going to be a bit like where we said well, with Moore's law, with the power of computer chips and processors, we are going to reach a point where they're not going to be able to get any better because something begins to hold things back.
Mike - It's something that not every academic agrees on. I think that there is a tendency to assume that because of this rapid rise of, let's say, chatbots, that they're going to only get better and better from now on. I think that the view that many people subscribe to, and in fact the thing I sort of subscribe to at the moment, is we haven't really seen evidence yet that they will continue to get better at quite this rate. We've seen new chatbots emerge, but since then we've seen iterations of these chat bots that don't double the power, they don't triple the power, they just get a bit better each time. I think for a while, unless something drastic changes, we might expect to see iterations and evolutions of these things rather than something that comes along and just blows everything out of the water again and impresses with a whole new set of abilities.
Chris - And when one considers these engines, are they really generalists? As in, as they get more powerful, are they just going to get better and better at doing everything? Like a Swiss army knife does a range of different jobs and it does them well? Or are they very, very specific and they're very good at doing one thing really, really well and there are a few spinoffs that it does half heartedly and we are fooled into thinking it's doing a good job?
Mike - This is the really interesting question: to what extent can we make AI a general purpose tool and to what extent do we have to have a specific AI for a specific role? I think at the moment the very best models are the ones that are specific to a task. In my research I often have a very specific thing that I'm trying to do. For example, I'm trying to diagnose something in a medical image, or I'm trying to identify something in a plant image, and in those times a very specific AI aimed at that task is the thing that will get the best results.
Chris - When I train an AI, if I'm doing a really good job and I want to train it to recognise cancerous moles, I train it to recognise a mole, then I train it to recognise cancerous moles from healthy moles. But it's focusing on moles. I'm not distracting it with pictures of mathematical formulae or daisies and tomatoes.
Mike - Yes, that's exactly right. If you ask a chat bot to diagnose a cancerous mole, it will say something, but that thing is unlikely to be correct because we've never actually explained that problem or explained the biology behind these cancerous moles or anything like this. I think that, in the tech sector, so the big companies that are training these giant models, there's a tendency to believe that if we just double the amount of data or triple the amount of data, we will get better and better models and they'll become so performant that they can do any task; they can do medical imaging, they can do plant imaging, they can analyse our general home photos as well. And actually that doesn't seem to be the case. If you want the very best performance, it's better to have a smaller data set on a specific problem.
Mike - It might be that we do see an uptick in the ability of these models the more data we add. But the resources that training these is taking on the scale of data of the internet is now becoming very, very high. There will be a point, I think, where we have to start making decisions as to whether it's efficient or indeed if we can afford to do it. It costs a lot of money to train these models and if they only get a small performance boost each time you double the size of the data, at some point you're going to decide that's not worth doing.
Chris - Neils Bohr famously said, or is alleged to have said, prediction is very hard, especially when it concerns the future. I'm going to ask you to predict the future. What's going to be the thing we have to work on or solve, then? Where is the next emerging thing with this? What can it do at the moment but not that well or what's the gap we haven't closed? What dots are we going to join next, do you think, to really move this on?
Mike - So I think separately we're going to see people continue to develop bigger versions of these models and we will see what happens there, and we're going to see people continue to develop smaller models for specific tasks and that's a separate thing, but I think that the thing we are really missing at the moment, especially with these largest of models, is having them actually interact with us on a day-to-day basis. People often ask me what the difference is between a smartphone assistant like Siri, or a chat bot like Chat GPT and the answer is, actually, Siri is much more constrained because it actually has to do real things. If you ask Siri to put something in your calendar, it actually has to go and do that and it has to know how to talk to your calendar app. If you ask it to go on the web and search some data, it has to be able to do that.
Mike - Most chatbots don't need to do this. They can just write text that looks nice but hasn't actually had to source any data. It's just what's in the training set. I think it becomes quite a lot more difficult when you have to control systems based on what the output of your AI is because your risks are much higher. Having a chat with your chatbot, if it gives you a bad poem or it writes a bad paragraph, that's not an absolute problem. But if it's controlling your self driving car, that becomes a much bigger problem. And so I think that over the next few years we're going to have to really start to work on how we integrate these systems in a safe way with things that we actually are trying to do day to day.
Chris - Some of us are long enough in the tooth and old enough to remember the .com that rapidly turned into the .bomb bubble. Do you think we're at that sort of precipice again, that this is all over inflating, overhyping and it's going to implode? Or do you think that actually it's in good shape and here to stay?
Mike - In an annoying way, I actually think it's both things at the same time. I think that there are incredible abilities of AI that we've just seen coming in the last few years that are going to transform areas like drug discovery and that are going to have a profound impact. But I also think that there is a lot of hype with these chatbots and people assuming they're doing something that perhaps they're not. When we actually start trying to use these chatbots to do things, we're going to find they need to be much more accurate than they are all the time. That's when it becomes really, really difficult. There will be a bursting of the bubble in the sense that people will get used to what they can and can't do and they'll use them with that in mind, they perhaps might not be quite so hyped as they are today, but I do think that they are incredibly useful tools and they are going to become more and more prevalent in our day-to-day lives.
27:57 - AI's impact on society
AI's impact on society
Gillian Hadfield, University of Toronto
But what if the likes of Microsoft and Google do manage to overcome the technical challenges to improving generative AI even further? We’ve heard about the varied and impressive use cases for AI across a broad range of fields, but these examples all involve complementing the work of humans, not replacing them. What if this changes in the future? Gillian Hadfield is a law professor and economist at the University of Toronto…
Gillian - We don't know if we'll get there, but we need to think about, as possible, that we could have artificial agents participating in our world with us, a bit like an alien species. It could be quite wonderful because this could be new members of our society that bring us lots of intelligence and ways of solving big problems. I think when I get worried is when I think about, do we really understand what it would mean to have new actors, new agents, new entities, new species among us? Have we thought carefully about how we make sure that they play well with us?
Chris - You've got a policy forum that you've just put together in 'Science,' one of the world's leading science publications. What were the points that you bring up there and why have you picked on those things?
Gillian - One is to really emphasise how quickly the technology is advancing and the trajectory we might be on to really quite fundamental change. My contribution in particular is to draw attention to, we really haven't thought about how to adapt all the complex structure that we bring to making sure that our human societies are, by and large, safe, cooperative, happy places. Of course there's no sense in which they are completely. A key point of the piece from my perspective is to say, we need to raise the urgency with which we are thinking about how would we in fact integrate, like I said, this alien species into our world.
Chris - You are alluding to a lack of urgency on something that you think is a high priority. Usually people deprioritise things they don't think are going to happen. Do you think, really, there's a lot of hope and hype out there, but the people who really know, know different and that's why they're not really pushing the envelope on this.
Gillian - I don't think that's the case. We have people who are at the heart of the technology, who are raising the concern about how rapidly things are going and the potential for it to achieve levels of transformation, to change our world. People disagree about that, but the disagreement is not between the people who know what's happening and the people who don't. I think the lack of urgency is actually coming from the general public as well as our regulators, politicians and governors, etc., not really having a good idea of why this would be such a challenge to how we do things. I think we take for granted a lot of the basic, invisible ways in which we make sure that people behave themselves and they're not there for this alien species, they wouldn't be in place. So I don't think the lack of urgency is because people who know best know that it's not going to happen.
Chris - Whose problem do you think it is? Do you think it is the problem of the companies? Because a lot of this technology is in the hands of commercial entities who are profit making organisations that are multinationals. Do you think it's down to them to sort this out or do you think it's a policymaker thing, but of course with that comes differences of geography, culture, etc.?
Gillian - I think the fact that this technology is being built almost exclusively inside private technology companies is precisely one of the reasons we who are not inside those technology companies need to be paying a lot of attention. I think we are potentially watching the decision making, the power to decide what our lives look like, shift from our councils and our towns and our governments and our communities into technology companies. I think that's a bad thing. So, I definitely think that this is a job for all of us who are not technologists to really understand what's happening and to be paying enough attention to say, hey, wait a second, we should be in charge of how this new world evolves.
Chris - Is that your wishlist? There needs to be some rules. At the moment it's the wild west and anyone seems to be able to do anything, and that's your concern.
Gillian - We need to be putting in place the mechanisms for us to even understand what's going on. A proposal that I've made with some colleagues is to say, we need basic visibility from governments into what's being built. We don't have that right now, so we should have national registries that require that the governments have a right to know what's being built, what we know about what it can do, and what we know about the safety measures. We should have that in place quite urgently. I think that we should be recognising, for example, that the billions of dollars that are being poured into development right now are driving towards creating increasingly autonomous AI agents. Agents that can go out and book you airline tickets and make reservations for you and maybe then start designing products and contracting to have them built and sold to really start participating in our society and our economy.
But we don't have any of the rules that that we have around ordinary humans being able to participate, like work authorisation or an address, some way of finding out who somebody is, tracking that down. An AI agent should have a legal identity that allows us to say: that one is not allowed to operate anymore or this one has to pay for the damages that were created and, oh, by the way, that means it needs to have assets behind it, a bank account basically. There's that kind of infrastructure that I think we need to rapidly be putting in place.
Comments
Add a comment