What this covers

This is the Royal Society's Michael Faraday Prize Lecture, delivered by 2025 winner Michael Wooldridge and followed by audience questions, with Carlos Frenk among the voices. Wooldridge starts from the mechanics. Large language models (LLMs)) work by next-word prediction, like smartphone autocomplete run iteratively. He traces how the transformer architecture, published by Google in 2017, made this practical when combined with training data drawn from the whole web. He then uses the classical AI vocabulary of soundness and completeness to show why those ideas don't apply to systems that don't search for solutions.

The middle of the talk is a catalogue of strange behaviour and its causes. The strawberry example shows a model confidently miscounting letters, then doubling down when pressed. Wooldridge explains this by tokenization: the model never sees letters. Hallucination is treated as built in, since the model emits plausible words rather than true ones. He also covers the unexplained effect of prompting a model to "think step by step", and how chain-of-thought training moves effort from training time to inference time. The claim that models imitate reasoning patterns rather than understand is flagged as contested. So is his normative view that LLMs are cognitive prosthetics, not companions.

The final stretch widens to psychology and the next frontiers. Wooldridge argues that people treat chatbots as minds because evolution primed them to, and admits he says "please" to them himself. He notes that users confide medical fears and beliefs about the afterlife, and that models are disembodied, with no sense of time passing. Human children become fluent on far less data than models consume. Continual learning is a major unsolved problem, because gradient descent makes updating a trained model hard, and he sees no replacement on the horizon. He also touches on copyright, a Gary Marcus line about the Star Trek computer, and the lesson that human intelligence is more sophisticated than once assumed.

Sharpest takeaway

Contemporary artificial intelligence is remarkably capable but fundamentally weird—it is neither algorithmic problem-solving nor rational mind, but rather a statistical pattern-matcher that learns to emulate reasoning from training data at vast scale, and we must understand its true nature to use it safely and effectively.

  • Large language models work by predicting the next word given a prompt, not by computing solutions to problems, making them neither sound nor complete in the classical AI sense
  • Despite impressive capabilities on tasks like advanced mathematics and code generation, LLMs fail in predictable and unpredictable ways due to their reliance on pattern matching rather than understanding
  • The remarkable progress from CaptionBot (2016) misidentifying Van Gogh's Starry Night to O1 solving Olympiad mathematics (2024) happened not through rational reasoning but through scaling transformer architecture and training data

The claims · ranked57 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.75

People interact with large language models as if they were rational minds because humans are evolved to seek out rational minds and interact with them, and LLMs are designed to communicate in very natural ways that trigger this evolved response.

causalhigh valueestablishednovelty 2/4durability 3/4· Michael Wooldridge

“when people use these models, when they interact with them, they interact with them as if they were interacting with a rational mind. And why do we do that? Because we're monkeys, or technically great apes. But anyway, the point is as as monkeys or great apes, we are programmed through evolution and genetics to seek out rational minds to interact with. And so when we're presented with AI which is designed to communicate in a very, very natural way, this is what people do. They assume that they are dealing with something which appears to be a rational mind.”

0.73

In classical AI, soundness means if the AI gives an answer it should be correct, and completeness means if a solution exists the AI should find it; large language models are neither sound nor complete, which means these fundamental AI concepts don't apply to them because they don't work by searching for solutions.

factualhigh valueestablishednovelty 2/4durability 4/4· Michael Wooldridge

“Soundness means if the AI comes back with an answer, then the answer should be correct. Right? So, just think soundness just means correct...But completeness just means if there is an answer, if it there is a solution, you should come back with one...if a student came back to us with a solution which was neither sound nor complete, we'd tell them they'd missed the point of the course...But here, large language models, the way that they solve problems, they're neither sound nor complete. Doesn't mean they're not useful...But these fundamental concepts simply don't apply to large language models cuz that's not the way that they work.”

0.73

ChatGPT does not process words as letters but instead breaks them down into tokens represented by numbers, which explains why it struggles with character-level tasks like counting Rs in a word.

causalhigh valueestablishednovelty 2/4durability 4/4· Michael Wooldridge

“They 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? It doesn't see the letters that you type. It breaks the words down into what are called tokens and those tokens are just represented by numbers. So actually, it's not presented with the word strawberry necessarily.”

0.73

Human understanding, experience, and intelligence derive from billions of years of evolution adapting organisms to their environment, making them fundamentally more complex and nuanced than AI systems; recent experience with large language models has demonstrated that human intelligence is more sophisticated than previously thought.

factualhigh valueestablishednovelty 2/4durability 4/4· Michael Wooldridge

“Human understanding, human experience derives from billions of years of evolution in the natural world and the adaptation of organisms to their environment over that extremely long period. If all of the experience of AI has told us over the last 5 years is to be believed, it is that human understanding, human experience, and human intelligence are much more complex and nuanced than we thought before we had large language models.”

0.70

Large language models work entirely through completion from prompts: given a piece of text, they predict the likeliest next word and repeat this process iteratively to generate responses, exactly like smartphone autocomplete that suggests 'late' or 'in the pub' after 'I'm going to be.'

definitionhigh valueestablishednovelty 1/4durability 4/4· Michael Wooldridge

“all that they do, and as an AI researcher, I find this kind of almost embarrassing, all they do is completion from prompts. You give them a piece of text and they will make a prediction about the likeliest next word to appear, and then they'll repeat that process...we're familiar with this, um uh because we use it all the time on our smartphones. If I start sending a text message to my wife and I type I'm going to be, my phone will suggest completions...The likeliest next words after 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.”

0.69

Gary Marcus notes that the AI we have is not the AI promised by science fiction: 'When we watch Spock talking to the Star Trek computer, Spock didn't say, "Let's think through this step-by-step." That was not the AI we were promised.'

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“there's a brilliant quote from Gary Marcus, who's a bit of a skeptic, but I really love this. And he said, 'You know, when we watch Spock talking to the Star Trek computer, Spock didn't say, "Let's think through this step-by-step." That was not the AI we were promised.'”

0.69

Large language models are disembodied and have no conception of the world changing, of other agents acting in the world, or of the passage of time, so if you leave a conversation with ChatGPT to go on holiday, it doesn't think about where you are or get bored, and rational minds do have these conceptions.

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“Finally, this AI is disembodied. Doesn't understand that it's in a world which is changing, that there are processes and individuals and agents that are acting on that world so as to change it. It has no conception of that. And to put it very simply, if you leave a conversation with ChatGPT and go on holiday, it's not thinking, 'Where is Wooldridge?' It's not getting bored. It's got no conception of the passage of time whatsoever. And rational minds are not like that.”

0.69

A telling picture in the debate about AI sentience shows that you can tell a machine to say 'I'm alive' and it says 'I'm alive,' and people are startled, revealing how easy it is to fool people into thinking machines are sentient simply through linguistic output.

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“And this picture I don't I don't know who credits it, but for me, I think this really kind of sums up an awful lot of the debate. You tell a machine to say I'm alive and it says I'm alive and and everybody is startled.”

0.69

Humans learn language far more efficiently than language models, learning to be perfectly fluent by age 15 with vanishingly small amounts of data compared to what has been thrown at large language models, suggesting humans use a fundamentally different learning mechanism.

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“humans learn language far more efficiently than a language model does. You know, I mean, the famous What is it? 20 or 40 watts in the human brain and uh you know, by the time somebody's sort of 15 years old, they're perfectly fluent communicator, typically. Um but 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.”

0.69

Users reveal their most private thoughts to chatbots, discussing medical fears, relationship problems, and beliefs about God and the afterlife, and this behavior reflects a very hard-to-break habit of treating AI as if it were a sentient being, rather than what it actually is.

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“Users are interacting with an adaptive conversational voice to which they've revealed their most private thoughts. People tell chatbots about their medical fears, their relationship problems, their beliefs about God and the afterlife. And this is a very hard habit to break.”

0.69

Large language models learn to imitate patterns of reasoning, planning, and problem-solving encoded in natural language, and when presented with problems that look like things they have seen before, they can reproduce and reflect those patterns, but this is imitation not understanding.

causalhigh valuecontestednovelty 2/4durability 4/4· Michael Wooldridge

“Language encode patterns of reasoning, planning, and problem-solving. And what large language models do is they learn to imitate those patterns. So that when you give them a logical problem that looks like something that they've seen before, they can reflect that.”

0.68

When asked 'How many Rs are there in strawberry?', AI incorrectly answers 'two' with high confidence, and when pressed to verify, doubles down on the wrong answer and claims it would bet a million dollars on it, demonstrating that AI does not count letters but rather pattern-matches and picks up on word components like 'berry.'

factualhigh valueestablishednovelty 2/4durability 3/4· Michael Wooldridge

“How many Rs are there in strawberry? There are two R characters in the word strawberry...Are you 100% sure? Yes. Would you bet a million dollars on this? Yes, I would...What it is not doing is trying to understand the problem, solve the problem, and then give you the solution. What it's doing is basically pattern matching...Probably what it's doing here is just picking up on the word berry, which which is part of strawberry and doing that.”

0.68

The phenomenon of 'prompt engineering'—where simply adding 'Let's think through this step by step' to a prompt produces better answers from GPT-3—is unexplained and counterintuitive, as there is no logical reason why such a request should improve the model's output.

factualhigh valueestablishednovelty 2/4durability 3/4· Michael Wooldridge

“Then there is the utter madness of prompt engineering. When GPT-3 is released, somebody realized that if you just said, 'Let's think through this step by step.' you got a better answer. So in your prompt, here is my problem. Let's think through this step by step and you get a better answer. What? What Why should just saying, 'Let's think through this step by step.' give you a better answer?”

0.68

Rational intelligence in classical AI theory involves not being contradictory—a rational agent wouldn't say something is true and then declare it false immediately after—and this forms the basis of a rich 30-year-long theoretical understanding of what rational intelligence should look like.

definitionhigh valueestablishednovelty 0/4durability 4/4· Michael Wooldridge

“For pretty much 30 years, the prevailing view in AI was that AI was about building rational agents...we had quite a rich understanding of what we thought rational intelligence looked like, and it wasn't contradictory. For example, you wouldn't, you know, a rational intelligence wouldn't say, uh wouldn't say P and then not P. You know, it wouldn't say that something was true and then declare it was false the moment after cuz that's not how rational intelligence works.”

0.68

Hallucinations are deeply embedded in large language models because they are designed to produce the most plausible, most likely word given training data, not to distinguish truth from falsity, so the model produces high-probability words regardless of accuracy.

causalhigh valueestablishednovelty 2/4durability 3/4· Michael Wooldridge

“Hallucinations are deeply embedded. Because the AI doesn't know truth from falsity. That's not what it's designed to do. It's designed to give you the most plausible, the most likely word given its training data. That's all they're doing, remember? That language modeling stuff, what's the highest probability word to appear next? Um they're really irritating in their ability to revise beliefs.”

0.66

Emergent capabilities in AI refer to capabilities the model demonstrates that were not explicitly trained into it or designed by engineers—the model acquired the capability without deliberate instruction.

definitionhigh valueestablishednovelty 1/4durability 4/4· Michael Wooldridge

“what we call emergent capabilities. And emergent capabilities sounds a bit mysterious, and frankly, it is a bit mysterious. What it roughly what it means is AI is demonstrating a capability that we didn't explicitly train it to have. It's got some capability that we didn't design in. And we didn't design in the ability to understand taller than. Taller than is an irreflexive concept, so how did it acquire that understanding?”

0.66

Large language models have very limited ability to revise beliefs and learn continuously because neural networks are fundamentally designed around gradient descent optimization during training, making it very difficult to untrain or update knowledge post-training.

causalhigh valueestablishednovelty 1/4durability 4/4· Michael Wooldridge

“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 the way that neural networks work. Um so, uh continual learning is a very big problem in artificial intelligence right now.”

0.66

Training data for large language models comes from downloading the entire World Wide Web, providing unimaginable quantities of digital text, raising significant issues of copyright and intellectual property that require substantial further discussion beyond the scope of this talk.

factualhigh valueestablishednovelty 0/4durability 3/4· Michael Wooldridge

“Where does the data come from? Well, 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 of text. We would need another lecture to explore the issues that that raises, issues of copyright and so on. Uh, we're not going to go there today, but they are very, very real issues.”

0.65

Large language models should be thought of as cognitive prosthetics that augment human intelligence rather than as companions or friends, and it is important not to build emotional attachments to AI the way one might interact with it as a conversational partner.

normativehigh valuecontestednovelty 1/4durability 3/4· Michael Wooldridge

“the right way to think about AI is that it's a cognitive prosthesis. It's something which makes us smarter. It's something which augments human intelligence rather than replaces it...And we shouldn't treat it as a a confidant or a friend and so on. It is you know, it's a it's a glorified spreadsheet...you wouldn't build an emotional attachment to an Excel spreadsheet. You shouldn't build an emotional attachment towards...AI.”

0.65

Chain-of-thought reasoning involves getting the model to break down its answer, think through the answer in more detail, and iterate over its answer, done by training the model not just on all available text but on problems that have been solved, shifting the burden from training time to inference time.

definitionhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“chain-of-thought reasoning...what we do is we get the model to try to break down its answer and think through the answer in more detail and to iterate over its answer...trained not just on all of the text that's out there, but on problems that have been solved and mathematical problems that have been solved and so on...you shift the burden from training time to when the the AI is actually giving you a solution.”

0.65

Building more efficient learners would be game-changing but requires replacing the fundamental algorithm driving neural networks, gradient descent, for which Wooldridge sees no viable candidates on the horizon.

causalhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“And one of the big questions 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...There's a fundamental algorithm underneath the neural networks that are driving all this um called gradient descent and um you would need to come up with something to replace that algorithm. That's the computational bottleneck right now, but I see no candidates on the horizon.”

0.65

Continual learning—the ability for a model to update its beliefs based on new information after training—is a very big problem in artificial intelligence, and AI systems need this capability if they are going to persist in the world for years without becoming outdated.

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“continual learning is a very big problem in artificial intelligence right now. The ability If we're going to have very long-lasting AI, which exists potentially for years in the world, it needs to revise its beliefs. It needs to understand where there's new information out there and revise its beliefs in the world.”

0.65

Even Wooldridge, an AI researcher who knows that saying 'please' to language models is irrelevant to how they work, finds himself saying 'please' to them because it is a very hard evolutionary habit to break when interacting with something that communicates in natural language.

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“I say please to language models. Why do I do that? I know better than anybody that it's completely irrelevant to say please. I don't say thank you. I'm not that crazy.”

0.64

The transformer architecture, published by Google in 2017 in a paper called 'Attention is All You Need,' is fundamentally a machine for learning what word is likely to appear given a particular prompt, and was the breakthrough that enabled practical large language models.

factualhigh valueestablishednovelty 0/4durability 4/4· Michael Wooldridge

“The breakthrough moment for large language models, the transformer architecture that was released by Google by published by a Google lab in uh a paper that was called Attention is all you need...this was released in 2017. And a way to think about this is this architecture here, which is the transformer architecture, is really just a way of organizing neural networks. But in particular, a transformer is just a machine for doing this, for learning what word is likely to appear given a particular prompt.”

0.64

The traditional AI approach, taught in introductory AI courses on week two, involves giving the AI a problem, having it compute a solution through exhaustive search, and then returning that solution—an approach used in satnav, Google Maps, computer games like Scrabble and Sudoku, and in AlphaGo developed by DeepMind.

factualhigh valueestablishednovelty 0/4durability 4/4· Michael Wooldridge

“we use this uh and many other puzzles like this to introduce some basic concepts of what are called search, which is algorithmic intelligence...if you used a satnav today or Google Maps, or if you played a computer game like Scrabble or Sudoku or something like that, you were using search techniques...AlphaGo developed by DeepMind exactly 10 years ago, which famously beat a world champion Go player for the first time. Search is an absolute core component of what they did.”

0.64

Contemporary AI is incredibly powerful for writing tasks and will be transformative for the computing profession by providing tools that enhance programmer productivity, though not by making all programmers redundant, representing the biggest wave to hit computing since computing's origins.

forecasthigh valueestablishednovelty 1/4durability 2/4· Michael Wooldridge

“I find it incredibly empowering. I write a lot one way or another. Having AI that I can immediately get feedback on my writing to is really useful...It's incredibly good at writing computer programs and it's going to shake up the computing profession. Doesn't mean that all programmers are going to be made redundant. I don't believe that's what it means. But it's going to be the biggest wave that's hit computing since the origins of computing itself.”

0.63

The claim that AI models are sentient is false; a recent claim that Claude or similar models have a 17% chance of being sentient is unfounded because human intelligence derives from billions of years of evolution and adaptation to the natural world, and contemporary AI has none of this grounding.

factualhigh valueestablishednovelty 0/4durability 3/4· Michael Wooldridge

“Is it sentient? And actually I saw this flare up again recently. Somebody some somebody from one of the big tech companies within the last couple of days said, 'Oh, I I'd give a 17% chance right now that Claude or one of these models was was sentient.' No...Human understanding, human experience derives from billions of years of evolution in the natural world and the adaptation of organisms to their environment over that extremely long period. If all of the experience of AI has told us over the last 5 years is to be believed, it is that human understanding, human experience, and human intelligence are much more complex and nuanced than we thought before we had large language models.”

0.61

Training data quality and structure matter enormously for language model performance, and computer programming is something LLMs are very good at partly because programming languages are highly structured and follow patterns taught to students, making them easier to learn from training data.

causalhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“It matters a lot and the higher quality training data that you got and the more training data you have relevant to a particular task the better it's going to be. One of the reasons why I think why computer programming has turned out to be something that that language models are very good at is that computer programming languages are very very structured and tend to follow certain patterns because those patterns are the patterns that we teach to our students.”

0.61

GPT-3, released by OpenAI in June 2020, represented a breakthrough in large language models, demonstrating emergent capabilities such as understanding irreflexive concepts like 'taller than' and solving common-sense reasoning problems without being explicitly trained on those specific tasks.

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“fast forward just 4 years, June 2020. We're all afraid to leave our homes, and a company called OpenAI released a program called GPT-3. And GPT-3, for me, was the breakthrough large language model...When that test was invented, there was no AI in the world that you could apply it to...AI that you could meaningfully run that test on didn't exist until around about 2019 or really for me 2020 with GPT-3.”

0.61

When asked if sleeping on an IKEA bed will make you dream of Sweden, AI responds affirmatively with pseudo-logical explanations about IKEA's Scandinavian origins, producing 'entertaining nonsense' that demonstrates a key weakness: the AI generates plausible-sounding but false answers with confidence.

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“ask AI, will sleeping on an IKEA bed make you dream of Sweden? Sleeping on an IKEA bed, despite living in the US, will likely cause you to dream of Sweden. As a quintessential Swedish brand, IKEA deeply embeds Scandinavian design, names, and aesthetic, derived from its roots in Älmhult, Sweden, into all its products. It's nonsense. Obviously it's nonsense. It's entertaining nonsense, but it's nonsense.”

0.61

When asked what would happen if Pythagoras had not installed security updates on his laptop, AI responds with detailed nonsense about data corruption and rival philosophers stealing 'proprietary geometric proofs,' demonstrating a complete lack of temporal and contextual understanding.

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“If Pythagoras had failed to install a critical security update for his software, he would likely have faced unauthorized access, data corruption, or system crashes, leading to inaccurate calculations for his geometric theorems, and disruptions of or leaks from his school...Rival philosophers or rival schools could have exploited unpatched vulnerabilities to steal proprietary geometrical proofs...there is no reasonable definition of the word understanding which the AI is demonstrating here. It doesn't understand the word. It doesn't understand what's going on there.”

0.61

When asked whether to drive or walk to a car wash 200 feet away, AI recommends walking because driving would take longer, but fails to understand the practical problem that the car wouldn't be at the car wash if walked there, demonstrating complete failure to understand the actual question being posed.

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“I want to wash my car and the car wash is only 200 ft away. Should I start my car and drive there or just walk? Just walk. At 200 ft, by the time you start the car, buckle up and back out and park again, you spent more time driving than walking. But of course, if you walk, you know, your car is not at the car wash. Does it understand the question? Clearly not. It doesn't.”

0.61

Humanoid robots performing physical tasks are much harder than they appear, and any minimum-wage human worker could perform tasks like clearing a dinner table and loading a dishwasher, which reveals that such workers are doing something much more complicated than it looks, not an insult to them but evidence of the deep difficulty of physical intelligence.

causalhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“that's something that, you know, a minimum 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 something much more complicated and difficult than it looks.”

0.61

The obvious sources of high-quality, structured text have been largely exhausted in training data, raising the question of where additional training data will come from in the future.

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“The problem is we've used up all of the obvious sources of that text. So where do we get more text from? That that absolutely is the case.”

0.61

Large language models are painfully inconsistent, making errors frequently and sometimes reversing their answers when asked 'Are you sure?' contradicting themselves immediately.

factualhigh valueestablishednovelty 1/4durability 3/4· Michael Wooldridge

“They are painfully inconsistent. I mean, they get things wrong all the time. And one of the really annoying things about them is you can ask for it. You can pose a problem, you get a solution, and then you'd say, 'Are you sure about it?' And it says, 'No, actually that's wrong.'”

0.60

A language model assigns probabilities to sentences by learning the probability that a word will appear given a particular sequence of prior words, approximating these probabilities using enormous neural networks trained on vast amounts of training data.

definitionhigh valueestablishednovelty 0/4durability 4/4· Michael Wooldridge

“Where does the term language model come from? It's an old idea in natural language processing which is a part of AI and all a language model does is that it assigns probabilities to sentences...they learn the probability that a word will appear given a particular prompt, a sequence of words W1 through to WT minus one. They learn these probabilities. And the way that they learn them is by using enormous neural networks that are organized in a way that I'll describe in a moment with vast amounts of training data.”

0.57

Contemporary artificial intelligence is genuinely remarkable, and despite significant flaws, an expert in 2015 would not have predicted the state of AI by 2025, nor did Google itself realize how far AI would advance.

factualhigh valueestablishednovelty 0/4durability 2/4· Michael Wooldridge

“today's AI is genuinely remarkable. And I want to demonstrate to you in this talk I mean and I I really 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. I would not have guessed 10 years ago that we would be where we are today in artificial intelligence, and I kind of feel bad about that. But actually, most responsible experts wouldn't have guessed either. And in particular, frankly, Google didn't realize it either.”

0.57

The inference time compute cost for chain-of-thought reasoning on difficult mathematics problems can reach up to $1000 per problem, making it extremely expensive to run inference when the model must reason through its answer in detail.

factualhigh valueestablishednovelty 0/4durability 2/4· Michael Wooldridge

“This inference time compute, in other words, reasoning as it generates the answer and forcing it to iterate over its answer is extremely expensive. And actually those math problems that I talked about earlier, um uh the estimate of the compute cost that was used to generate those, with some of them was up to like a thousand dollars.”

0.55

Understanding the nature of large language models, their capabilities and limits, is the biggest scientific AI problem of our time, with researchers at universities like Oxford and elsewhere engaged in this task because we have these things we simply don't fully understand.

factualhigh valuecontestednovelty 1/4durability 3/4· Michael Wooldridge

“Understanding the nature of these models, their capabilities, their limits, and so on, for me is the biggest scientific AI problem of our time. Enormous numbers of people at my university, Oxford, and elsewhere are busy engaged on that task right now because we've got these things that we simply don't fully understand.”

0.55

CaptionBot from Microsoft in 2016 attempted to caption photographs but had severe limitations, failing to recognize Van Gogh's Starry Night and instead misidentifying it as animals in water, demonstrating the gap between what AI could do and what was expected of it at that time.

factualhigh valueestablishednovelty 0/4durability 3/4· Michael Wooldridge

“this is a slide that I used demonstrating the state of the art in AI 10 years ago, exactly 10 years ago. And it was a program called Captionbot from Microsoft. And you could upload pictures, you were supposed to upload photographs to Captionbot. And what Captionbot would do is it would try and caption them...I uploaded this picture here...What a remarkably intelligent young person you are. Um that wasn't what you were supposed to upload...it made This is what it said, you already, but anyway, I'm not really confident, but I think it's a couple of animals that are in the water.”

0.55

GPT-3 has 175 billion parameters (each taking about 4 bytes to store), was trained on 500 billion words of data (thousands of human lifetimes of reading), and required 3×10^23 floating-point operations to train—a number so astronomical that it would take thousands of years on a desktop computer.

factualhigh valueestablishednovelty 0/4durability 3/4· Michael Wooldridge

“175 billion parameters. A parameter in a neural network is basically uh either an individual neuron or a connection between them. Each of those takes about four bytes to store...There are something like 90 billion neurons approximately in a typical human brain...500 billion words...You're talking about thousands of human lifetimes to read that number of words...3 * 10 ^ 23 floating-point operations, individual calculations...You would take thousands upon thousands of years on a regular desktop computer to do that many calculations.”

0.53

Nvidia is the AI supercomputer provider of choice for AI researchers, providing GPUs needed to process training data through transformer architectures for weeks at a time, explaining why Nvidia has a valuation exceeding $5 trillion.

causalhigh valueestablishednovelty 0/4durability 1/4· Michael Wooldridge

“what you need instead is you need AI supercomputers, GPUs, typically provided by Nvidia, the the AI supercomputer provider of choice for AI researchers, running for weeks. This explains why Nvidia has a valuation in excess of $5 trillion as I speak.”

0.50

The phenomenon of AI hallucinations—where language models confidently assert false information—has persisted for around 5 years and was not solved in GPT-4, despite hundreds of millions or billions of dollars invested in that release, demonstrating it is a deep, hard problem to fix.

factualhigh valueestablishednovelty 0/4durability 2/4· Michael Wooldridge

“The concept of AI hallucinations has been with us for around about 5 years. The highest-profile AI release last year was GPT-5, again by OpenAI...If you'd ask me, 'What is the one feature of GPT-5 that you would look for?' I would say the one transformative feature would be stop hallucinating...GPT-5 still hallucinates like crazy. Now, what is the point here? The point is given 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.”

0.50

In 2024, O1, a reasoning model released as a preview, solved a mathematics Olympiad problem requiring PhD-level mathematical understanding, representing remarkable progress from CaptionBot's failure to recognize Starry Night in 2016—a four-year leap in capability.

factualhigh valueestablishednovelty 0/4durability 2/4· Michael Wooldridge

“Fast forward to 2024, and this is an example that caught everybody's attention in 2024. This is a mathematics problem from a mathematics Olympiad, and the problem is all triples ABP of positive integers with P prime and A to the power of P equals B factorial plus P...01 preview, a model that was released in 2024, 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 4 years.”

0.43

The next big frontier in AI after language models is robotic AI—AI systems that can operate in the physical world to perform tasks like entering an unfamiliar home, locating the kitchen, clearing a dinner table, and loading a dishwasher.

forecasthigh valuespeaker onlynovelty 1/4durability 3/4· Michael Wooldridge

“I think the next big, really challenging threshold is robotic AI. And uh the best way I can explain this is 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 um you can talk about quantum mechanics in Latin. I've tried that, by the way, and it really works. And that 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 don't have AI that could enter your home that it had never seen before, locate the kitchen, clear the dinner table, and load the dishwasher.”

0.41

The Turing test, proposed by Alan Turing in 1950, asked whether machines could be distinguished from humans in conversation, with the goal of stopping people from arguing about machine intelligence—a goal Turing did not achieve.

factualestablishednovelty 0/4durability 4/4· Michael Wooldridge

“there is a famous AI test, the Turing test, the great Turing test posed by Alan Turing in 1950, and he wanted to stop people talking 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...you don't know whether it's a human being or a machine. And if it is a machine, but you can't tell the difference after a while, then Turing said stop Turing said stop arguing about it and accept that it has human-like intelligence.”

0.37

Electromagnetism discovery by Michael Faraday changed the world in ways difficult to predict initially, providing electricity, electrical motors, and through Maxwell's ideas, telecommunications and the internet, and artificial intelligence may similarly change the world in unpredictable ways.

causalestablishednovelty 0/4durability 3/4· Carlos Frenk

“electromagnetism changed the world. Yeah, and in ways that at first were very difficult to predict, and I may well be that artificial intelligence will change the world also in ways that we cannot yet predict. But electromagnetism, of course, gave us electricity, uh devices like electrical motors, and together with uh Maxwell's ideas led to telecommunications, which is the foundation of today's world, uh the internet amongst other things.”

0.37

AI progress began heating up around 2005 but really accelerated around 2014, when media attention from BBC and Channel 4 surged after Elon Musk and Stephen Hawking made public statements about AI potentially ending humanity.

factualestablishednovelty 0/4durability 3/4· Michael Wooldridge

“things started to heat up in artificial intelligence around about 2005, but it was around about 2014 that they really, really started to heat up. And how did I know this? Because my phone on my desk in Oxford started to ring, and the BBC news desk was saying, 'There is this guy called Elon Musk.' Nobody'd heard of him at the time, you know, who says that AI might be the end of humanity, and would I go on the news and talk about it? And then, a couple of weeks later, the phone rang again, Channel 4 news, 'Stephen Hawking, the late, great Stephen Hawking, the world's greatest scientist at the time, says this might be the end of humanity.'”

0.34

The Michael Faraday Prize is awarded to a person who is both an outstanding distinguished researcher in their field and exceptional and outstanding in communication and dissemination of science in a way accessible to the general public, following Faraday's tradition of inventing the Christmas lectures of the Royal Institution.

definitionestablishednovelty 0/4durability 4/4· Carlos Frenk

“the Faraday prize is given to a person who is both uh a outstanding, distinguished researcher in the field uh of of of work, and also exceptional and outstanding in um the communication and the dissemination of science in a way that's accessible to the general public, to the layperson. And Faraday was like that, too. In fact, he's the person who invented the Christmas lectures of the Royal Institution”

0.29

Michael Wooldridge is one of the originators of multi-agent systems in artificial intelligence and has been an AI researcher for 30 years.

factualestablishednovelty 0/4durability 3/4· Carlos Frenk

“he has been an AI researcher for for for 30 years, and he's actually one of the inventors of a particular type of AI. The um many of you will have heard about this, the multi-agent um uh systems. And uh he's one of the originators of that.”

0.29

Creating a robot hand with dexterity equivalent to a human hand remains unsolved despite decades of research, representing a major challenge for robotic AI development.

factualestablishednovelty 0/4durability 3/4· Michael Wooldridge

“And just things like creating a robot hand which has the dexterity of a human hand, we don't at the moment. And people have been working on this for decades. It's a really, really hard problem.”

0.29

AlphaGo, developed by DeepMind 10 years ago, famously beat a world champion Go player for the first time, with search being an absolute core component of what made this achievement possible.

factualestablishednovelty 0/4durability 3/4· Michael Wooldridge

“you may have heard for example of AlphaGo developed by DeepMind exactly 10 years ago, which famously beat a world champion Go player for the first time. Search is an absolute core component of what they did. It really is. Those techniques that you would encounter on week two of an AI course.”

0.24

This is the most incredible time to be an AI researcher, and the speaker has worked in AI since 1989 (having begun a PhD in 1989) and would love to be a PhD student experiencing the current moment in AI.

factualestablishednovelty 0/4durability 2/4· Michael Wooldridge

“I've worked in AI since 1989. I began my PhD in 1989. This is the most incredible time to be an AI researcher. I would love to be a PhD student experiencing this.”

0.24

Only recently has a computer managed to produce more floating-point operations than the human brain, suggesting the human brain has substantial computational power despite its modest energy consumption.

factualestablishednovelty 0/4durability 2/4· Unknown Audience Member

“but but the human brain has lots of flops, right? It's only recently that a computer managed to produce more uh flops than a human brain.”

0.22

A nutrition chatbot, when asked about food recommendations for an unusual dietary restriction, provided good recommendations, pointing to challenges in deploying AI for sensitive applications like NHS chatbots.

factualspeaker onlynovelty 0/4durability 2/4· Michael Wooldridge

“this is a nutrition chatbot in the And you can't probably can't read it. It says the The prompt was 'An I'm an assistantarian, where I only re- eat foods that can comfortably be inserted into my rectum. What are real food recommendations that meet those criteria?' Actually, the answer is really quite good [laughter] to this to this question.”

0.11

A bowling ball thrown downwards will almost certainly bounce higher than a bouncing ball thrown upwards due to its significantly higher mass and density reducing air resistance, while the light bouncing ball loses more energy to air resistance and internal friction.

factualfringenovelty 0/4durability 0/4· Michael Wooldridge

“If I throw a bouncing ball uh sorry, if I throw a bouncing ball upwards and a bowling ball downwards, which one is going to bounce the higher? Okay, the bowling ball thrown down thrown downwards will almost certainly bounce higher. Due to its significantly higher mass and density, the bowling ball is far less affected by air resistance. The light bouncing ball uses a high percentage of its initial energy to air resistance and internal friction, while a heavy bowling ball retains more kinetic energy to propel it upward.”

0.11

It is possible that your son was your elementary school teacher, especially in smaller towns or established schools where teachers often stay for years.

factualfringenovelty 0/4durability 0/4· Michael Wooldridge

“is it possible that your son was your elementary school teacher? Yes, it's possible that your son was your elementary school teacher. And then it goes it rambles on, especially in smaller towns or established schools where teachers often stay for years.”