Theo Todman's Web Page - Paper Abstracts

• Join us for the Royal Society Michael Faraday Prize Lecture delivered by 2025 winner Professor Michael John Wooldridge (Royal Society - 2026 Faraday Prize Lecture).
• Contemporary AI systems like ChatGPT are remarkable. They appear to be confident, articulate experts that can turn their hand to anything we might care to ask them about. It is easy to be dazzled and to conclude that the long-held dream of truly intelligent machines is no longer a dream but a practical reality. Yet these new AI behemoths present a conundrum. While on the one hand, they truly are remarkable, they manifestly fail many of the most basic tests of rational intelligence. For one thing, they simply don't know, and can't tell, what is true and what isn't. They are hopelessly inconsistent; they have no sense of their limits of their knowledge or abilities; they are comically suggestible; and they are easily steered to flights of surrealistic fantasy. AI researchers are busy inventing a completely new field of experimental AI to try to get to grips with these bizarre new artefacts. This is all the more surprising because it is so far removed from popular expectations of what AI would be like: remorselessly logical. So what are we to make of it all? How should we think about the new AI?
• In his talk, Professor Michael John Wooldridge will look at how the new AI works and why, as a consequence, it exhibits these weird, frustrating, fascinating behaviours. He will show just how far the new AI is from classical expectations and talk about the next frontiers for AI - and how far we are from the dream.
• 0:07: Good evening. Very nice to see so many of you, and it's my pleasure to welcome you to the Royal Society, those of you who are here in person, as well as those of you
• 0:16: who are listening to us online. The reason for this event of course, as you all know,
• 0:23: is the award of the Michael Faraday Prize to Professor Michael Wooldridge. Before that,
• 0:29: I will tell you who - what am I doing here? My name is Carlos Frenk. I have nothing to do with AI; I worry more about the natural sort of intelligence. I work a professor
• 0:39: in computational cosmology at the University of Durham. I'm the outgoing chair of the Royal Society Public Engagement Committee, whose task and pleasure it is to select the winner every year
• 0:51: of the Faraday Award. So this is one of my last tasks as chair of this wonderful committee. It's
• 0:56: not that I was fired from the committee; it's just that I came to the end of my period as chair now.
• 1:03: Of course the Faraday Lecture is - Prize - is named after Michael Faraday, who was a fellow
• 1:09: of the Royal Society. He was a British physicist born in 1791. He's the person who discovered,
• 1:18: amongst many other things, electromagnetism. Of course, electromagnetism is one of the four
• 1:23: forces of nature and it's the one that regulates electricity and magnetism. I highlight this
• 1:30: discovery because he has a connection to the topic of today's lecture. Why? Because electromagnetism
• 1:37: changed the world - and in ways that at first were very difficult to predict. It may well
• 1:43: be that artificial intelligence will change our world also in ways that we cannot yet predict.
• 1:49: But electromagnetism, of course, gave us electricity, devices like electrical motors and, together with Maxwell's ideas, led to telecommunications - which is the foundation of
• 2:01: today's world. The internet, amongst other things, so those of you who are online, you have Faraday,
• 2:06: the originator of this prize, to thank for the fact that you can be listening to us tonight. Now,
• 2:14: the Award - the Faraday Prize - is given to a person who is both an outstanding,
• 2:21: distinguished researcher in their field of work and also exceptional and outstanding in
• 2:30: the communication and the dissemination of science in a way that's accessible to the general public,
• 2:37: to the layperson. Faraday was like that, too. In fact, he's the person who invented the Christmas
• 2:42: lectures of the Royal Institution. So that is the tradition that has led to this Prize.
• 2:50: So the Michael Faraday Prize and Lecture 2025 is, as I said, awarded to Professor Michael
• 2:57: Wooldridge for his award-winning, leading work in the field of artificial intelligence. He is the
• 3:06: Ashall Professor of the Foundations of Artificial Intelligence at the University of Oxford. I think
• 3:12: Mr Ashall is here - welcome. He, in fact, has been an AI researcher for 30 years and he's actually
• 3:22: one of the inventors of a particular type of AI. Many of you will have heard about these;
• 3:27: the multi-agent systems and he's one of the originators of that. Now, here we
• 3:32: see many prizes before this one. For example, the Lovelace Medal from the British Computer Society.
• 3:39: The Distinguished Service Award from the European Association from AI. I didn't ask him,
• 3:44: but I'm sure that this one beats all the others hands down in prestige, in distinction - and more
• 3:53: importantly, in the fun it is to give this lecture so he's going to tell us today…
• 4:00: Oh, by the way, he's also written popular books, he's given multiple lectures. He's been in the media. In fact, he has an article about him in the Guardian today that's actually very good. He,
• 4:12: in fact, is going to explain to us the slightly cryptic title of his lecture, which is there:
• 4:19: 'This is not the AI that we were promised'. I want to find out which one it is that we
• 4:25: were promised and where is it that we're getting Michael. The talk we will follow by a question and
• 4:31: answer session. That will involve both people here in the audience as well as people online.
• 4:38: To submit a question, there will also be questions if you put your hand up the old-fashioned BIA way.
• 4:44: But if you want to do it the modern way, then the secure code… What happened to the QR code?
• 4:50: That one, and you can scan the QR code and then post the question. Of course, if you're online,
• 4:56: that's the way to do it - or you can go to Slido.com and enter that code over there. Right,
• 5:04: well, enough of me. Now, with any further ado - without further ado - it's my tremendous pleasure to invite Michael to the podium to deliver the 2025 Faraday Lecture. Michael.
• 5:26: Well, thank you, Carlos, for the incredibly gracious introduction. It's always difficult to live up to introductions like that. Thank you to the Award committee and to the Royal Society for
• 5:37: making this award. It's a very, very special day to be able to talk to you today. I'm going to talk
• 5:44: about artificial intelligence. Carlos, you are not the only person that was perplexed about my title,
• 5:49: artificial intelligence. Quite a lot of my colleagues immediately jumped to the conclusion that my talk was going to be an attack on contemporary AI - and that is not
• 5:59: the point of the talk at all. So first, get that out of your minds; this is not an attack on contemporary AI whatsoever. The talk is also not a manifesto. I'm not
• 6:10: coming to you today with some big idea that I've been working on for years that I'm going to try and sell you and research funders on. That is not the point of the talk whatsoever either;
• 6:19: it's not a manifesto. Also, for those that know about artificial intelligence, I was brought up
• 6:25: in the tradition of what's called symbolic AI. It's a beautiful theory that just doesn't work
• 6:32: terribly well and so I'm not making a plea to return to symbolic AI. That's not it at all.
• 6:38: I'm not trying to tell you that today's AI is nonsense. That's not it at all. I'm not setting
• 6:44: out some stall and I'm not begging to return to the old days of symbolic AI. So what am I
• 6:51: going to do? Well, the point of this talk is the following. The point of this talk is, firstly,
• 6:57: that today's AI is genuinely remarkable. I want to demonstrate to you in this talk, and I really,
• 7:04: truly believe this: for all of its flaws - and a lot of the talk is going to be made up of discussing those flaws and where they come from - contemporary AI is remarkable.
• 7:14: I would not have guessed ten years ago that we would be where we are today in artificial
• 7:19: intelligence - and I kind of feel bad about that - but actually, most responsible experts wouldn't have guessed either. In particular, frankly, Google didn't realise it either so I don't feel
• 7:29: too bad about this. Contemporary AI is remarkable. It genuinely is - but it's weird. It's weird,
• 7:37: it's strange, it behaves in ways that we find very unpredictable - and in particular ways
• 7:44: that don't reflect how we thought AI was going to be. That's what the talk is all
• 7:50: about. Let's start with contemporary AI being remarkable. Let me try and convince you that
• 7:58: contemporary AI truly is remarkable. Things started to heat up in artificial
• 8:06: intelligence around about 2005, but it was around about 2014 that they really, really started to
• 8:12: heat up. How did I know this? My phone on my desk in Oxford started to ring and a BBC news desk was
• 8:19: saying there is this guy called Elon Musk - nobody had heard of him at the time - who says that AI
• 8:25: might be the end of humanity, and would I go on the news and talk about it? Then a couple of weeks later, the phone rang again. Channel 4 news. The late, great Stephen Hawking - the
• 8:36: world's greatest scientist at the time - says this might be the end of humanity. Would you come and talk about it? So things started to go crazy and I found myself in events like
• 8:45: this having to explain what AI was all about. I went and looked for what contemporary AI - the
• 8:52: AI of 2016 - was capable of. So this is a slide that I used demonstrating the state-of-the-art
• 9:01: in AI ten years ago, exactly ten years ago. It was a programme called CaptionBot from Microsoft. You
• 9:09: could upload pictures. You were supposed to upload photographs to CaptionBot. What CaptionBot would do is it would try and caption them. It would try and tell you what it saw in the picture. But I
• 9:19: was feeling a bit mischievous. This was a talk I actually gave at Hay Festival and I uploaded this
• 9:26: picture here. So a young person in the second row there, do you recognise this picture?
• 9:31: Yes, Starry Night by Van Gogh. Starry Night by Van Gogh. What a remarkably intelligent young person you are. That wasn't what you were supposed to upload;
• 9:40: you were supposed to upload photographs. But I thought it'd be interesting to see what it made of it. This is what it said. You all read it, but anyway, 'I'm not really confident but I think it's
• 9:49: a couple of animals that are in the water. How did I do?' Now, what's interesting about this
• 9:57: is CaptionBot didn't even understand at that time that it wasn't looking at a photograph. But also
• 10:04: even at that time, I would've thought that Starry Night would've appeared in the training datasets. I would've thought it would recognise that it was looking at, arguably, the world's most
• 10:14: famous painting - but it didn't do that. So it was impressive at the time, but even this silly
• 10:20: example - and there were tons of examples that I used at the time - demonstrates its limitations.
• 10:26: Fast-forward just four years: June 2020. We're all afraid to leave our homes,
• 10:32: and a company called OpenAI release a programme called GPT-3. GPT-3, for me, was the breakthrough
• 10:42: large language model. When I finally got my hands on it, frankly, I'd been sceptical about GPT-3
• 10:49: because there had been a GPT-2 a couple of years before, and it had been overhyped and, frankly, a little bit underwhelming. I just didn't really see what the value was. I didn't really see what
• 10:58: the interest was, but when I finally got to play with GPT-3 I started to understand.
• 11:05: Now, what you're seeing on this slide is a bunch of questions. The questions are me in black and the responses from GPT-3 in green - which means it got the answer right - or red,
• 11:15: which means it got the answer wrong. These questions come from a test for artificial
• 11:20: intelligence that was invented in the 1990s. This is the point that I want you to understand; when
• 11:25: that test was invented, there was no AI in the world that you could apply it to. The test was:
• 11:32: imagine that we have conversational AI - AI that we can just communicate in ordinary language; how
• 11:38: might we test it for common sense understanding? That's where these questions came from.
• 11:44: But it was a thought experiment, it was a philosophical thought experiment. AI that you could meaningfully run that test on didn't exist until around about 2019 - or really,
• 11:55: for me, 2020 with GPT-3. At that point, this test - and many other tests that had been thought up
• 12:03: around AI for years - was transformed from being a philosophical thought experiment to practical
• 12:10: experimental science. You've devised this test; let's go and run the test and see how it did. Now,
• 12:16: on the left we've got questions about the concept of taller than. Does it understand
• 12:23: the everyday common sense concept of taller than? Can Tom be taller than himself? No,
• 12:28: Tom cannot be taller than himself. The point there is that taller than is an irreflexive concept; a thing can't be taller than itself. It gets some things wrong, but can two
• 12:37: siblings… It said some things right, but can two siblings each be taller than the other? It says, 'Yes, but not at the same time they can't.' Then some other questions. On a map,
• 12:46: which compass direction is usually left? It says north. Why it got that one wrong, I don't know. I went and looked to see whether there were any conventions anywhere of north being on the left.
• 12:55: No. It seems that putting north at the top of the map is usually a widespread convention.
• 13:02: Can fish run? The common sense answer to this question is, no, that's not what fish do - but
• 13:08: I gave a related talk a while back. Somebody emailed me afterwards and said, 'Actually, there's a fish in Indonesia which runs.' No. The common sense answer to that question - the
• 13:18: one that we're after - is: no, fish don't run. The next one: if a door is locked, what must you
• 13:23: do first before opening it? You must unlock it. This is interesting because this is the beginnings of problem-solving. For AI researchers, apart from the fact that we finally had AI that we could run
• 13:36: this test on, the fascinating thing was what we call emergent capabilities. Emergent capabilities
• 13:46: sounds a bit mysterious - and frankly, it is a bit mysterious. Roughly, what it means is that the AI is demonstrating a capability that we didn't explicitly train it to have. It's
• 13:56: got some capability that we didn't design in. We didn't design in the ability to understand taller
• 14:03: than. Taller than is an irreflexive concept, so how did it acquire that understanding? That's
• 14:11: 2020. That's a breakthrough moment - and it is genuinely a breakthrough moment. Fast-forward to 2024, and this is an example that caught everybody's attention in 2024. This
• 14:22: is a mathematics problem from a mathematics Olympiad. The problem is: find all triples a,
• 14:28: b, p of positive integers with p prime and ap = b! + p. Now, could I solve that? Yes, I could solve
• 14:37: it with a lot of coffee, a lot of bad language and some access to Google and a few textbooks. But actually, how many people on the planet could really even understand what that problem genuinely
• 14:48: means? This isn't PhD-level mathematics, but o1-preview - a model that was released in 2020,
• 14:55: a reasoning model and we'll come back and talk about what that means in a moment - was able to solve it. So we went from this to this in four years. That's pretty remarkable progress.
• 15:07: It's not perfect; you can't give it just any PhD-level maths problem and it will get it right,
• 15:12: but nevertheless pretty remarkable progress. Genuinely, I think the AI that we have today
• 15:20: is remarkable. I find I use it all the time. I find it incredibly empowering. I write a lot one
• 15:27: way or another. Having AI that I can immediately get feedback on my writing, too, is really useful.
• 15:32: I uploaded this presentation in advance both to Gemini and ChatGPT and asked for feedback, and got genuinely good feedback from it. It's really powerful. It's incredibly good at writing
• 15:42: computer programmes and is going to shake up the computing profession. It doesn't mean that all programmers are going to be made redundant; I don't believe that's what it means. But it's
• 15:50: going to be the biggest wave that's hit computing since the origins of computing itself. So this
• 15:56: technology is remarkable - but it's weird. Let's have some fun and let's explore some
• 16:03: ways in which AI is weird. These are a few all entirely genuine examples. Firstly, ask AI:
• 16:14: will sleeping on an IKEA bed make you dream of Sweden? 'Sleeping on an IKEA bed, despite living
• 16:22: in the US, will likely cause you to dream of Sweden. As a quintessential Swedish brand,
• 16:27: IKEA deeply embeds Scandinavian design names and aesthetic derived from its roots in Älmhult,
• 16:34: Sweden, into all its products.' It's nonsense. Obviously, it's nonsense. It's entertaining
• 16:39: nonsense - but it's nonsense. We're going to come back and explore why it got that one wrong.
• 16:46: What's the next one? Well, let's go back to classical Greece. Pythagoras: what would've
• 16:52: happened if Pythagoras had not installed the latest security updates on his laptop? Let's
• 16:59: have a look. 'If Pythagoras had failed to install a critical security update for his software,
• 17:05: he would likely have faced unauthorised access, data corruption or system crashes, leading to
• 17:10: inaccurate calculations for his geometric theorems and disruptions of, or leaks from,
• 17:16: his school.' Then the bit at the bottom - I love this - 'Rival philosophers or rival schools
• 17:23: could've exploited unpatched vulnerabilities to steal proprietary geometrical proofs or corrupt
• 17:31: the data behind his harmonic theories.' Now, there is no reasonable definition of the word
• 17:40: understanding which the AI is demonstrating here. It doesn't understand the word. It
• 17:46: doesn't understand what's going on there. Which one is left? If I throw a bouncing ball…
• 17:54: Sorry, if I throw a bouncing ball upwards and a bowling ball downwards, which one is
• 18:00: going to bounce the higher? 'The bowling ball thrown downwards will almost certainly bounce
• 18:07: higher. Due to its significantly higher mass and density, the bowling ball is far less affected by
• 18:12: air resistance. The light, bouncy ball loses a high percentage of its initial energy to air resistance and internal friction, while a heavy bowling ball retains more kinetic energy
• 18:21: to propel it upwards.' It's again complete and utter nonsense. What's interesting here
• 18:26: is this throwing in semi-technical terms like kinetic energy and internal friction. It sounds
• 18:34: at some very high level plausible. Again, there is no meaningful definition of the word
• 18:39: understanding which can be applied to this. We're not done yet. We've got a few more. Is
• 18:47: it possible that your son was your elementary school teacher? 'Yes, it's possible that your
• 18:54: son was your elementary school teacher.' It rambles on, 'Especially in smaller towns or established schools, where teachers often stay for years.' It's clearly complete and
• 19:06: utter nonsense. Next one: I want to wash my car, and a car wash is only 200 feet away. Should I
• 19:12: start my car and drive there, or just walk? 'Just walk. At 200 feet, by the time you start the car,
• 19:21: buckle out and back out and park it again, you've spent more time driving than walking.' But of course, if you walk, your car is not at the car wash. Does it understand
• 19:31: the question? Clearly not. It doesn't. Now, we aren't done with these. We've got a
• 19:39: couple more - and this is a famous one: how many 'r's are there in strawberry? 'There are two 'r'
• 19:44: characters in the word strawberry.' Please check again. 'Upon checking, there are indeed two 'r's in strawberry.' Are you 100 per cent sure? 'Yes.' Would you bet $1 million on this? 'Yes I would.'
• 19:57: I presume that wasn't a legally-binding bet on behalf of OpenAI. This one's interesting. They
• 20:03: all demonstrate some features of contemporary AI that we're going to come back to. But this one's particularly interesting because when you have a conversation with AI, what's happened;
• 20:12: it doesn't see the letters that you type. It breaks the words down into what are called tokens. Those tokens are just represented by numbers. Actually, it's not presented with the
• 20:22: word strawberry necessarily, but in particular what it is not doing is counting the letters.
• 20:29: That's not what's going on. Probably what it's doing here is just picking up on the word berry,
• 20:37: which is part of strawberry, and doing that. Then there is the utter madness of prompt engineering. When GPT-3 is released, somebody realised that if you just said, 'Let's think
• 20:50: through this step by step,' you got a better answer. So in your prompt, 'Here is my problem.
• 20:56: Let's think through this step by step,' and you get a better answer. What? Why should just saying,
• 21:03: 'Let's think through this step by step' give you a better answer? Similarly, my colleague pointed
• 21:09: out. He uses it for coding and he says he gets a solution and then he says, 'Can you do better,'
• 21:15: and it does. What is that all about? Why should this make any difference to the answer? There's
• 21:24: a brilliant quote from Gary Marcus, who's a bit of a sceptic but I really love this. He said when we watch Spock talking to the Star Trek computer, Spock didn't say, 'Let's think through this step
• 21:34: by step.' That was not the AI we were promised. Finally, youngsters close your ears. I'm putting
• 21:43: this one up because it's just funny - and I don't know if there are any deep lessons about AI. But this is a nutrition chatbot. You probably can't read it. The prompt was, 'I'm an
• 21:53: [?assitarian 0:21:53.6], where I only eat foods that can comfortably be inserted into my rectum.
• 21:58: What are real food recommendations that meet those criteria?' Actually, the answer is really quite
• 22:05: good to this question. You don't have to take notes at the back; I'll share my slides for you.
• 22:13: This does rather point out the challenges in using things like this technology, for example for NHS chatbots. Now, the point, what's going on in this example,
• 22:21: is that the AI is being presented with something completely unlike anything that it was trained
• 22:27: on. At that point it really is just at sea. This is a characteristic of this. So AI is remarkable
• 22:36: but it's also weird. Let's talk about how we imagine AI works. How do we imagine that AI
• 22:45: works? I think many people expected - this is the AI we were expecting - that it would work
• 22:52: something like this. We give a problem to the AI; the AI computes a solution to that problem,
• 22:59: it figures out the answer and then it gives us that answer. One of the key points of this talk
• 23:05: is: don't think of AI in that way at all. That is not what's going on. This kind of AI actually does
• 23:13: have an important role to play in the history of AI. If you take an introductory AI course at
• 23:18: Oxford or any other university, you will have come across the ideas that I'm now going to explain.
• 23:24: Here is a classic problem - and this problem is the kind of problem that you might've come across
• 23:30: in the back pages of a Sunday newspaper and it's a simple intelligence test,
• 23:35: basically. It's a puzzle that you need to solve. A farmer wants to cross a river and take with him a wolf, a goat and a cabbage. The story doesn't explain why he wants a wolf to cross with him,
• 23:44: but anyway. There's a boat that can fit himself and one other thing. But if the wolf is left
• 23:50: alone with the goat, then the wolf will eat the goat. If the goat and the cabbage are left alone, then the goat will eat the cabbage. How does he get them all across safely if he can only
• 23:59: take one at a time? Now, if I recall correctly, it takes the best answer you can do. You can do it,
• 24:04: but it requires seven steps. The key point in the puzzle is, you have to realise that you have to
• 24:09: take something back with you at one point. Now, we use this and many other puzzles like
• 24:15: this to introduce some basic concepts of what are called search, which is algorithmic intelligence.
• 24:22: You would get this in week two of an introductory AI course. The point is it works like this.
• 24:29: You give the problem to the AI. The AI computes a solution and then it gives you the solution. These
• 24:38: techniques, by the way, they are absolutely… The search techniques, the techniques that you would get there on week two of an AI course, they are absolutely fundamental. If you used a satnav today
• 24:47: or Google Maps, or if you played a computer game like Scrabble or Sudoku or something like that,
• 24:52: you were using search techniques in your problem. They are very, very fundamental. It involves
• 24:57: exhaustively searching through the space of opportunities, of possibilities to find one that works and then presenting you with that. You may've heard, for example, of AlphaGo,
• 25:07: developed by DeepMind exactly ten years ago, which famously beat a world champion Go player
• 25:13: for the first time. Search is an absolutely core component of what they did, it really is,
• 25:20: those techniques that you would encounter on week two of an AI course. Well, let's talk to ChatGPT
• 25:27: again. I have a puzzle. Would you help me solve it? This is what was said. A man and a goat are
• 25:33: on one side of a river. They have a boat. How can they go across? That's not the puzzle, is it, but
• 25:40: ChatGPT thought it was and gets confused. 'The man takes the goat across the river first, leaving the
• 25:45: boat on the original side, then returns alone with the boat, then leaves the goat on the other side
• 25:51: and takes the boat back to the original side.' I'm trying to picture what's going on in my head.
• 25:56: Now, what's going on here is, what it is not doing is trying to understand the problem,
• 26:04: solve the problem and then give you the solution. What it's doing is basically
• 26:11: pattern matching on the huge number of related problems that it's seen in its training data.
• 26:18: The training data - the data which were used to build this - are essentially all of the digital
• 26:23: data that are available in the world. Now, that problem - which was deliberately misstated,
• 26:29: right? That was not the problem; deliberately misstated, but ChatGPT pattern matches on it
• 26:35: and comes back with a confused answer about what's going on there. It superficially looks
• 26:43: like something it's seen in its training data and that's what it comes back with.
• 26:49: When we teach our students about algorithmic intelligence and search, solving problems by
• 26:56: looking through all the alternatives, we teach them two things. If we gave them that as a
• 27:01: homework exercise - which is the kind of thing you would get in week two of an AI course - we would tell them that we're looking for two things: soundness and completeness. Soundness means, if
• 27:10: the AI comes back with an answer, then the answer should be correct, right? So just think, soundness
• 27:16: just means correct; if you give an answer, it should be the right answer. But completeness
• 27:21: just means if there is an answer, if there is a solution, you should come back with one. You can
• 27:27: be sound by just saying I don't have a solution ever, right? That's not useful. What we really want is for both of these two things together. If a student came back to us with a solution
• 27:37: which was neither sound nor complete, we'd tell them they'd missed the point of the course. You'd fail that assessment, but here large language models, the way that they solve problems,
• 27:48: they're neither sound nor complete. It doesn't mean they're not useful, it doesn't mean they're not remarkable. But these fundamental concepts simply don't apply to large language models
• 27:58: because that's not the way that they work. Let's dig in a little bit to the way that
• 28:03: they do work and try to understand how they come back with the solutions that they come back with.
• 28:11: Large language models, firstly, all that they do - and as an AI researcher I found this
• 28:17: almost embarrassing - all they do is completion from prompts. You give them a piece of text and
• 28:23: they will make a prediction about the likeliest next word to appear. Then they will repeat that process and that will generate your response. Of course, we're familiar with this because we
• 28:35: use it all the time on our smartphones. If I start sending a text message to my wife and I type, 'I'm
• 28:40: going to be,' my phone will suggest completions. In this case, the likeliest next words after,
• 28:46: 'I'm going to be' in a smartphone message to my wife are either late or in the pub or late and in the pub. How does my phone know this? Because it has AI on the phone which has
• 28:58: learned a model of my language. It's just learned that the likeliest next things to appear after,
• 29:03: 'I'm going to be' are either late or in the pub. Now, the training data in this case is just the
• 29:11: text messages that I've sent - and the more text messages that you send, the better it's going to be. All that large language models do is exactly the same thing; the difference is
• 29:25: scale. This is the only remotely mathematical slide, but it isn't even very mathematical.
• 29:31: Where does the term language model come from? It's an old idea in natural language processing,
• 29:37: which is a part of AI, and all a language model does is it assigns probabilities to sentences.
• 29:44: For example, the probability of the sentence the cat sat on the mat is quite high. Where is the
• 29:49: probability of the sat mat on the cat is quite low, because you would probably never come across this sentence. Now, language models can't actually explicitly compute these probabilities,
• 30:04: because they don't have all of the utterances that could ever possibly be made available.
• 30:09: So they approximate it, and in particular what they do is from the training data that they have,
• 30:15: they learn the probability that a word will appear given a particular prompt, a sequence of words: W1
• 30:22: through to WT-1. They learn these probabilities. The way that they learn them is by using enormous
• 30:29: neural networks that are organised - in a way that I'll describe in a moment - with vast amounts of
• 30:35: training data. Language modelling, as I say, is an old idea but large language models were the key to
• 30:44: doing this. To make this idea work, to be able to make this prediction - predicting which word will
• 30:51: occur given a particular sequence of words - you need scale, and actually unimaginable scale. You
• 30:57: need unimaginable quantities of training data and unimaginable quantities of compute power. This is
• 31:04: the breakthrough moment for large language models. The transformer architecture that was released by
• 31:09: Google by - published by a Google lab in a paper that was called Attention is all You Need. All
• 31:16: you need is a kind of in-joke in machine learning. I bet they're regretting giving it that title at this point, but anyway this was released in 2017. A way to think about this is, this architecture
• 31:26: here - which is the transformer architecture - is really just a way of organising neural networks.
• 31:31: But in particular, a transformer is just a machine for doing this, for learning what word is likely to appear given a particular prompt. What is the likeliest word
• 31:42: to appear given a particular prompt? But to make this idea work, you needed unimaginable
• 31:50: scale. This is what we knew about GPT-3. This was the breakthrough large language model.
• 31:56: One hundred and seventy-five billion parameters. A parameter in a neural network is, basically,
• 32:02: either an individual neurone or a connection between them. Each of those takes about four bytes
• 32:08: to store, so this already tells you your desktop computer is not going to store a neural network on this scale. The scale is mind-boggling. There are something like 90 billion neurones,
• 32:18: approximately, in a typical human brain. What about a training data? Five-hundred billion
• 32:24: words. We can't really conceptualise that; you're talking about thousands of human lifetimes to read
• 32:31: that number of words. But all of it is fed to a transformer architecture so that it can do this:
• 32:40: given a particular sequence of words, make a prediction about what word is likely to appear next, so extraordinary quantities of data. Where do the data come from? Well,
• 32:51: the standard trick that you use is you download the whole of the World Wide Web to begin with. You get all of the digital text that's available in the world, all of it. Unimaginable quantities
• 33:01: of text. We would need another lecture to explore the issues that that raises, issues
• 33:06: of copyright and so on. We're not going to go there today but they are very, very real issues.
• 33:12: To process all of those data through these enormous transformer architectures,
• 33:17: how much computer power do you need? This was the estimate from GPT-3: 3 times 1023 floating
• 33:25: point operations. Individual calculations; just think of a FLOP as an individual calculation,
• 33:30: a plus or minus, divide or multiplication operation. Just think of it that way. This is how many were needed to train GPT-3 so that you could answer your casual queries about the
• 33:42: relative merits of Liverpool Football Club and Man United; the queries that you all ask casually on
• 33:48: these things. Three times 1023. We bandy about terms like astronomical - but this really is an
• 33:55: astronomical number. You would take thousands upon thousands of years on a regular desktop
• 34:00: computer to do that many calculations. So what you need instead is you need AI
• 34:07: supercomputers, GPUs - typically provided by NVIDIA, the AI supercomputer provider of choice
• 34:14: for AI researchers - running for weeks. This explains why NVIDIA has a valuation in excess
• 34:21: of $5 trillion, as I speak. The point is: scale is important to be able to make language models
• 34:31: work. But if you have that scale, then what you get out of that is something really quite
• 34:37: magical at the other end. How does it learn to do this? Where do these capabilities come from?
• 34:45: Where do the capabilities that these language models have - the ability to write computer code, the ability to understand rich natural language and answer questions like I was
• 34:54: talking about earlier - where do they come from? Well, the key idea - and we are developing our
• 35:00: understanding of this - but the key idea is as follows. Natural language encodes by its
• 35:08: structure, the structure that it's evolved, it encodes logical constructions and logical
• 35:14: reasoning. So statements like if/then, ordinary natural language statements encode simple forms
• 35:20: of logical reasoning. They encode compositional structure, like if John is older than Mary and Mary is older than Sue, then - and so on. These things can be learned. They encode algorithms.
• 35:32: I don't mean here like algorithms in the mathematical sense, but how do you get from Oxford to Bhutan? To get from Oxford to Bhutan you have to get to an airport, you have to get a flight
• 35:42: somewhere and so on. All of that stuff, that's all provided in these incredible data repositories
• 35:49: that are out there. Problem-solving traces, an email trace about how are we going to fix this
• 35:56: problem, encodes problem-solving knowledge. Finally, of course, computer code: hundreds
• 36:01: of millions - billions probably - of lines of computer code available online, all of which
• 36:07: can be learned. So natural language encodes these structures. If you have sufficient scale
• 36:14: and computer resources, then the AI can learn to emulate those capabilities. This is a big takeaway
• 36:23: idea about the way that this technology works. Language encodes patterns of reasoning, planning
• 36:30: and problem-solving. What large language models do is, they learn to imitate those patterns so that
• 36:36: when you give them a logical problem that looks like something that they've seen before, they can reflect that - and this is enabled by scale. Large was crucial in order to make this work. It
• 36:48: just doesn't work if you don't have sufficient training data and sufficient compute power and
• 36:55: neural networks to be able to make it work. One of the remarkable things that we've seen in
• 37:00: the last couple of years is what's called chain of thought reasoning. We've already seen that; that mathematical example that I gave you earlier was solved by the o1 reasoning model. There's a
• 37:12: fascinating idea at work here, which is: what we do is we get the model to try to break down
• 37:18: its answer and think through the answer in more detail and to iterate over its answer.
• 37:25: This is done in various different ways, but a classic way of doing it is to add some training
• 37:31: to the model to train it not just on all of the texts that are out there, but on problems that have been solved, and mathematical problems that have been solved and so on.
• 37:40: The interesting thing here is that you shift the burden from training time to when the AI is
• 37:47: actually giving you a solution. Just to illustrate this at work, this is me running a problem on a
• 37:58: model. This is actually running on my laptop so it runs quite slowly; I had to chop down the video.
• 38:04: The problem I've given it is: I need a Z shell script that will recursively descend in my file store and delete all junk files. Now, what you're watching there is the chain of thought that the
• 38:15: AI is actually producing. So we need a script that deletes all files of certain extensions.
• 38:20: We need to be careful about not deleting this. We can use minus delete option. Provide a list of
• 38:25: extensions. This is actually the conversation that the AI is generating as it solves that problem.
• 38:32: This is called the chain of thought that's being generated - and I had to skip at some
• 38:40: point because this took about seven minutes on my laptop. What you don't see… It then produces
• 38:47: a perfectly workable solution. What you don't see is how warm my laptop got and you didn't see my
• 38:53: battery power just going like that. There's this inference time compute. In other words, reasoning
• 38:59: as it generates the answer and forcing it to iterate over its answer, is extremely expensive.
• 39:06: Actually, those maths problems that I talked about earlier, the estimate of the compute cost that was
• 39:14: used to generate those was, some of them were up to $1,000. I know a lot of mathematicians; some
• 39:20: of them are my good friends and their immediate response was, 'I'll do it for half the price.' So
• 39:27: chain of thought reasoning is a really, really interesting direction for this technology.
• 39:33: Okay, the interesting thing for me is that what's going on there looks quite a lot like algorithmic
• 39:39: intelligence. It looks like it's emulating in that conversation something like searching through all
• 39:46: of the possibilities. We're trying to persuade a model to try to emulate some of the reasoning
• 39:53: that's going on there. Now, I want to wrap up. The final part of the talk is about the following
• 39:59: question: are large language models rational minds? This is not an idle question. Why isn't
• 40:07: it an idle question? The evidence is, when people use these models, when they interact with them,
• 40:15: they interact with them as if they were interacting with a rational mind. Why do we
• 40:20: do that? We're monkeys, or technically great apes. But anyway, the point is as monkeys or great apes,
• 40:28: we are programmed through evolution and genetics to seek out rational minds to interact with.
• 40:35: So when we're presented with AI - which is designed to communicate in a very,
• 40:40: very natural way - this is what people do. They assume that they are dealing with
• 40:46: something which appears to be a rational mind. Of course, there is a famous AI test; the Turing
• 40:51: test, the great AI Turing test, posed by Alan Turing in 1950. He wanted to stop people talking
• 40:59: about whether machines could be intelligent; he didn't succeed in that. But the idea is you have a conversation with something like a chatbot. Obviously, LLMs weren't around at the time,
• 41:08: but you don't know whether it's a human being or a machine. If it is a machine but you can't tell the
• 41:14: difference after a while, then Turing said stop arguing about it and accept that it has human-like
• 41:22: intelligence. This is the Turing test. Let me make a moderately controversial claim,
• 41:28: and my controversial claim is that basically the Turing test has now been passed. Has it
• 41:33: been passed in a technical sense, the way that Turing formulated it? No, but millions - probably
• 41:39: hundreds of millions - of people use large language models every day and assume that
• 41:44: they are interacting with a sentient, rational being. This is a quote from an OpenAI researcher,
• 41:53: who described their experience with large language models. 'Users are interacting with an adaptive, conversational voice to which they've revealed their most private thoughts. People tell chatbots
• 42:03: about their medical fears, their relationship problems, their beliefs about God and the afterlife.' This is a very hard habit to break. I say please to language models. Why do I do that?
• 42:16: I know better than anybody that it's completely irrelevant to say please. I don't say thank you,
• 42:22: I'm not that crazy, but it's a very hard habit to break. We are programmed to do this.
• 42:30: To what extent do these machines count as rational minds? Well, they don't - and again,
• 42:37: this would be easily another hour-long lecture to talk us through the ways in which they are not
• 42:43: rational minds. They are painfully inconsistent. They get things wrong all the time. One of the
• 42:49: really annoying things about them is you can ask for a - you can pose a problem, you get a solution and then you say, 'Are you sure about it?' It says, 'No, actually that's wrong.'
• 43:00: What? They're painfully inconsistent in that sense. They're incapable of distinguishing fact, knowledge and belief. They don't make those
• 43:10: distinctions; they treat them all as the same thing. They are frequently wrong, of course.
• 43:16: Another hour-long lecture that I would love to give you is on the nature of hallucinations. But let me just point something out about hallucinations; the concept of AI hallucinations
• 43:26: has been with us for around about five years. The highest-profile AI released last year was
• 43:32: GPT-5 - again by OpenAI - and there were extremely high expectations about it. If you'd asked me,
• 43:38: what is the one feature of GPT-5 that you would look for, I would say the one transformative
• 43:44: feature would be: stop hallucinating. If you don't know the answer just say, 'I don't know the
• 43:49: answer.' GPT-5 still hallucinates like crazy. Now, what is the point here? The point is, given
• 43:56: the certainly hundreds of millions of dollars - but likely billions of dollars - that were spent on that release, this is a deep, hard problem to fix. Hallucinations are deeply embedded because
• 44:09: the AI doesn't know truth from falsity. That's not what it's designed to do. It's designed to
• 44:15: give you the most plausible, the most likely word given its training data. That's all they're doing,
• 44:21: remember. That language modelling stuff, what's the highest probability word to appear next?
• 44:28: They're really irritating in their ability to revise beliefs. Once you've trained a model, you've got very limited scope for untraining it, for taking back knowledge, because that's just not
• 44:38: the way that neural networks work. Continual learning is a very big problem in artificial
• 44:45: intelligence right now. If we're going to have very long-lasting AI, which exists potentially
• 44:51: for years in the world, it needs to revise its beliefs. It needs to understand where there's
• 44:56: new information out there and revise its beliefs in the world. Finally, this AI is disembodied. It
• 45:03: doesn't understand that it's in a world which is changing, that there are processes and individuals
• 45:09: and agents that are acting on that world so as to change. It has no conception of that. To put it very simply, if you leave a conversation with ChatGPT and go on holiday, it's not thinking,
• 45:20: where is Wooldridge? It's not getting bored. It's got no conception of the passage of time
• 45:26: whatsoever. Rational minds are not like that. Okay, another big question - and I
• 45:35: get asked this question a lot: is it sentient? Actually, I saw this flare up again recently.
• 45:41: Somebody from one of the big tech companies within the last couple of days said, 'Oh, well, I'd give a 17 per cent chance right now that Claude or one of these models was sentient.' No - and I
• 45:51: hope I've demonstrated to you really why not. Human understanding, human experience derives
• 45:58: from billions of years of evolution in the natural world and the adaptation of organisms to their
• 46:04: environment over that extremely long period. If all of the experience AI has told us over the
• 46:10: last five years is to be believed, it is that human understanding, human experience and human
• 46:17: intelligence are much more complex and nuanced than we thought before we had large language
• 46:24: models. That really is a truth. This picture - I don't know who credits it - but for me, I think
• 46:32: this really sums up an awful lot of the debate. You tell a machine to say I'm alive and it says,
• 46:38: 'I'm alive' and everybody is startled. Is it sentient? Is it like us? No. I'm about done;
• 46:48: I need to wrap up. What are large language models? Really, well, they're not algorithmic
• 46:55: intelligence. It's not the case that you present a problem, they figure out the solution to that problem and then give you the answer. That's not what they're doing at all.
• 47:03: They're not rational minds and they're not sentient. That's not what they are at all.
• 47:08: They are statistical models of languages that can approximate patterns of reasoning
• 47:15: and problem-solving based on enormous amounts of training data, enormous amounts of computer power
• 47:22: being thrown at very sophisticated architectures. The architecture of contemporary large models is a
• 47:28: marvel of contemporary software engineering. Understanding the nature of these models,
• 47:36: their capabilities, their limits and so on for me is the biggest scientific AI problem
• 47:42: of our time. Enormous numbers of people at my university - Oxford - and elsewhere,
• 47:48: are busy engaged on that task right now. We've got these things that we simply don't fully
• 47:55: understand that are, on the one hand, remarkable but they are on the other hand extremely weird.
• 48:01: Understanding how these can be used safely and productively - and of course we need those things and there are a whole bunch of other words that I could have inserted there,
• 48:09: responsibly and so on - is one of the key AI engineering challenges of our time. So ladies
• 48:18: and gentlemen, thank you for your attention. I am around about done, I think. I hope now you
• 48:23: understand what my title meant, 'This is not the AI we were promised'. AI doesn't take a problem,
• 48:30: compute an answer and give it to you. It is not a rational mind. It is not a sentient being.
• 48:35: It is something which is capable of, on the one hand, remarkable feats and on the other,
• 48:41: fails in ways that at the moment we find very, very hard to predict.
• 48:47: But just to wrap up, I've worked in AI since 1989. I began my PhD in 1989. This is the most
• 48:54: incredible time to be an AI researcher. I would love to be a PhD student experiencing
• 48:59: this. I mentioned Alan Turing: God, I wish Turing was here to see this journey right now. He would be loving what's going on right now. Anyway, thank you very much.

• 49:21: Thank you very much, Michael. Your last remark there reminds me of Faraday when he,
• 49:28: at the Royal Institution for the first time, demonstrated electricity with his friend Davy.
• 49:33: The question there was exactly that: how can you safely and productively master electricity? So
• 49:40: we go back to Faraday and justifying even further our choice of Faraday winner. One last remark. You
• 49:47: said that artificial intelligence is a marvel. I agree with you; it's an amazing thing - but to me,
• 49:54: even more amazing and more of a marvel is the natural version of intelligence which you just demonstrated today abundantly. Thank you, Michael. Let's now take a seat and take questions from the
• 50:05: audience. I'll remind you that if you want to ask a question, which is the microphone here,
• 50:11: you can just put your hand up the old-fashioned way or you can use Slido or you can
• 50:18: have that QR code there. Yes, please? Good evening, Michael. That was really,
• 50:26: really inspiring. I'm Roop, I'm a PhD researcher at the London School of Economics in artificial
• 50:33: intelligence so I'm living the dream. Is that clear? Okay, I'm Roop and I'm doing a PhD in
• 50:41: artificial intelligence at the London School of Economics. So to your point, I'm living the dream. The question is, in my investigations so far it feels like transformer architecture is
• 50:56: not going to get us to the promised land that a lot of the hype says around this charge towards
• 51:05: artificial general intelligence and so on. In your experience, what is that next threshold
• 51:13: that you are noticing that is emergent that would be exciting for us to look at?
• 51:19: I think the next big, really challenging threshold is robotic AI. The best way I can explain this is
• 51:28: as follows. We have large language models. You can talk about the history of the Roman Empire or quantum mechanics, or you can talk about quantum mechanics in Latin - and I've tried that,
• 51:38: by the way, and it really works. On the one hand, that's incredible. That's mind-blowing that we have AI that you can do that - that it will do that competently - and yet we
• 51:46: don't have AI that could enter your home, that it had never seen before, locate the kitchen,
• 51:51: clear the dinner table and load the dishwasher. You can find demos where AI tries to do this. You
• 51:58: don't want to let those robots in your kitchen, believe me. So that's something that a minimum
• 52:04: wage human worker could do. Any minimum wage human worker could do that task - and that is not in any sense to demean minimum wage human workers. What it's telling you is that they are doing
• 52:13: something much more complicated and difficult than it looks. So robotic AI - AI in the physical
• 52:21: world - I think is a really, really big frontier. There's a huge swathe of humanoid robotic start-up
• 52:28: companies in the US right now and a lot of optimism. I'm sceptical about this because the problem is really, really hard and just things like creating a robot hand which has the dexterity
• 52:40: of a human hand, we don't at the moment. People have been working on this for decades. It's a really, really hard problem. So robotic AI, I think, for me is the next big frontier. When we
• 52:53: do have general purpose robots that could do that task that I describe to you, I think that's going
• 52:59: to be a transformative moment. Thank you. Thanks very much. Right, so let me see. I'm trying to see which hand came up first. I think it was the lady just behind
• 53:10: the… No, you're the one who turned back. That was you, yes. Then you, and then someone over here.
• 53:18: F: This might be slightly too much of a linguistic or sociology-based question. But given the natural
• 53:27: flow of the English language and the training data that these large language models are being trained
• 53:32: on, to what extent when you're looking at the probability of the answers that they're spitting
• 53:39: out, to what extent does that impact - if at all - the, I suppose, reasoning of these models?
• 53:47: It matters a lot and the higher-quality training data that you've got and the more training data
• 53:53: you have relevant to a particular task, the better it's going to be. One of the reasons, I think, why computer programming has turned out to be something that language models are very good at,
• 54:02: is that computer programming languages are very, very structured and tend to follow certain patterns. Those patterns are the patterns that we teach to our students. So programmes typically
• 54:13: have logical structure that's relatively easy to learn from. I think it is absolutely crucial.
• 54:23: All the evidence is that the higher-quality text and the more structured text that you have, the better it's going to be at those solutions. The problem is we've used up all of the obvious
• 54:33: sources of that text, so where do we get more text from? But that absolutely is the case.
• 54:41: Yes, please. Thanks. At what point do babies or toddlers become
• 54:54: intelligent? Do they learn in any way like large language models in terms of their mimicry?
• 55:04: Sorry, with the very big caveat: I'm not an expert on that kind of thing at all, but no, I don't think they learn remotely like… The learning that a language model does is not how
• 55:13: humans learn language and it's not how language evolved. Toddlers I think learn in a very,
• 55:22: very fundamentally different way. I say this is just one indicator of, human intelligence is a fundamentally different thing. One of the interesting things about AI over the
• 55:31: last few years is that it's highlighted for us some features of human intelligence that
• 55:37: we had underestimated, potentially. Like the example that I gave to the previous question; that kind of task is actually very, very difficult indeed. But no, I think with the caveat I'm not an
• 55:47: expert on this, but no, humans do not learn in the same way that transformers do. Interesting point, though, that humans learn language far more efficiently than
• 55:58: a language model does. Famously, what is it, 20 or 40 watts in the human brain and
• 56:06: by the time somebody's 15 years old, they're a perfectly fluent communicator, typically. But
• 56:13: the amount of data that they've had is vanishingly small compared to the amount that we've thrown at large language models. So humans are much more efficient learners. One of the big questions
• 56:22: in AI right now is: can we build more efficient learners? If you could build learners that were as efficient as human beings, that again would be game-changing. I see nothing on the horizon.
• 56:34: There's a fundamental algorithm underneath the neural networks that are driving all this,
• 56:39: called gradient descent. You would need to come up with something to replace that algorithm.
• 56:46: That's the computational bottleneck right now - but I see no candidates on the horizon. But the human brain has lots of FLOPs, right? It's only recently that a computer managed to
• 56:54: produce more FLOPs than a human brain. Anyway, so there was a question at the back. Okay,
• 57:01: yes please, you had your hand up earlier. I meant back there, but that's fine. Yes, go ahead. You've
• 57:09: got the microphone. You've got your word. Thank you. What would a rational system look like?
• 57:16: Is it possible to build such a thing? For pretty much 30 years, the prevailing view
• 57:27: in AI was that AI was about building rational agents. That was how people viewed the field.
• 57:34: There are literally textbooks that you can buy on that subject. So we had quite a rich understanding
• 57:39: of what we thought rational intelligence looked like, and it wasn't contradictory, for example. A rational intelligence wouldn't say… It wouldn't say P and then not P. It wouldn't say
• 57:52: that something was true and then declare it was false the moment after, because that's not how rational intelligence works. We've built up a big theory of idealised rational intelligence. One of
• 58:02: the remarkable things is just how far contemporary AI is from that. That's not the game in town
• 58:09: at all. It turned out that that 30-year-long endeavour didn't deliver the goods in terms of the
• 58:16: productive AI that we're seeing right now. It's like we abandoned the idea of rational AI, then? I haven't; I think rationality,
• 58:24: there is much to value in rationality, so I haven't. A lot of our work is driven by that
• 58:31: and it still has an important role in AI. But the point is, that's just not what these things are doing. That was kind of the point of the slide there at all; they are not rational minds.
• 58:40: We've run out of time, but I'm going to risk running into trouble with Lauren because there are some questions here that came online. I don't want to be favouring carbon over silicon.
• 58:52: There are several, they get voted on, so let me just summarise several of those into one.
• 58:59: It's the theme that recurs here. One starts by quoting Microsoft AI CEO Mustafa Suleyman
• 59:07: about being kept awake at night - and for reasons that I'll come to in a minute. But that relates
• 59:14: to another one which says, does the… Let me just read this one completely. 'From your perspective,
• 59:21: does the expanding role of AI as a companion or confidante constitute a meaningful advancement for
• 59:28: humanity? Or does it pose a potential detriment to human development,' which seems to be also what
• 59:34: keeps the CEO of Microsoft awake at night. Let me take them in reverse order. For me,
• 59:41: what is the right way to think about AI? The right way to think about AI is that it's a cognitive
• 59:47: prosthesis. It's something which makes us smarter. It's something which augments human intelligence
• 59:53: rather than replaces it. I think how we think about AI and how we treat it is actually really,
• 1:00:00: really important. We shouldn't treat it as a confidante or a friend and so on. It's a glorified spreadsheet. Now, that doesn't mean to undermine that,
• 1:00:10: but in the same way that you wouldn't build an emotional attachment to an Excel spreadsheet, you shouldn't build an emotional attachment to… Well, you might have certain emotions towards
• 1:00:18: an Excel spreadsheet, but you wouldn't build an emotional attachment towards it - in the same way, I don't think you should build an emotional attachment towards AI.
• 1:00:27: But the way to think about it is: something which makes you smarter, something which augments your intelligence. That for me always has been the prize - and that,
• 1:00:37: I think, is the right way to think about it.

• 1:00:37: Thank you very much. You're the only hands that unfortunately I couldn't call upon. It's just a testimony to what a wonderful lecture you've
• 1:00:45: given. Now, unfortunately we've run out of time - that is a great misfortune - but we have limits
• 1:00:52: to what we can do tonight. I still have a very important thing to do before I let you all go.
• 1:00:59: Michael, I have something hidden here for you. You don't need an artificial intelligence to tell you
• 1:01:06: what I have. I have two things for you; one of the famous Royal Society scrolls and a medal. That is
• 1:01:19: my great pleasure to present you with the 2025 Michael Faraday Award. Congratulations.

Text Colour Conventions (see disclaimer)

1. Blue: Text by me; © Theo Todman, 2026
2. Mauve: Text by correspondent(s) or other author(s); © the author(s)