Why AI? What Is Driving the AI Transformation?

Professor James Ferguson:

So welcome to this first talk on why AI? What is really driving this AI transformation that we're working through? I'm Professor James Ferguson. I'm the executive director of the Data Intensive Science Group at the University of Cambridge and I’m also the director of the Infosys Cambridge AI Center. So the story of AI can really be summed up by looking at the paradigms of science by Microsoft research. The very first one was empirical, where we simply looked at the world, we wrote down what we saw, and then those stories became the description of science. This was transformed around 1800 when Newton came along and we started to be able to calculate things. We could use maths to build models and predict the future based on these models. So we now could say where the cannonball will land before we fire the cannon. The real transformation, which is the beginning of the AI story, is when computers start to be competitive with calculation. And around 1990, we were assigned to be able to build computational systems that predicted where things would go and to solve equations and to model systems in a way that wasn't possible with simple equations. This then led to the data intensive era where we were able to get very large data sets and analyze them computationally to gain insights directly from the data with less of the predictive calculation that we needed. And this has led to us being in the AI for science era where now we can use automated methods from AI to do science and to power science in ways that was not possible in the traditional approaches. And so the story is here, why have we seen these revolutions and what's really driving it? Well, I think you can say that these last three are really all the same paradigm, that they're paradigm of computers. And essentially, there's a natural story whereby once you have the compute, which is growing exponentially, it follows Moore's law, it doubles every two years, then naturally computers allow you to capture and store data efficiently. And so now rather than to write things down in notebooks, we can automatically capture data and we could build much bigger data sets than before. And so we're seeing this exponential growth in data. Very roughly, it looks like it's doubling about every three years. And the natural and predictable consequence of growing compute and growing data is machine learning because they're really the fuel of anything to do with machine learning. And the exponential growth in compute, combined with the exponential growth in data, is leading to a super exponential growth in machine learning. And so this leads to sort of paradigm shifts where here is a plot of computer performance. And what we're seeing in the AI revolution, essentially there are long periods where there's not enough compute and not enough data. So if you look at this plot back in 2010, we wouldn't have enough power to do something like convolutional neural networks, because we imagine they take, you know, where that line that says requirement, that's the level we need. Before that, five years before that, we have far too little. There's nowhere near enough data. There's nowhere near enough compute to make it work. It's essentially impossible. As we wind forward with exponential growth, we hit a moment where suddenly it is possible. That happened in 2012 where CNN started to outperform traditional image processing methods. And then if you look maybe five years on for that, suddenly we have far too much computer and too much data and this becomes very, very easy. And so what you see in the history of AI is really these paradigm shifts where and at one period it's impossible, it becomes possible, and it becomes trivial. And we've seen this again. In 2017, we had transformers outperforming methods for language. In 2022, we saw building large networks of transformers outperform humans at writing. And then 2025, we've seen that we can now harness teams of agents working with LMs to automate a large number of tasks. And it's the story of constant paradigm shifts due to the exponential growth of compute. And that's really what drives the whole of the AI revolution. And so then what do we think about the future for this? Well, if you look from 2025 on, you can see that with its exponential growth continuing into the future, we should continue to see paradigm shifts in capability. We should expect the models we already have to improve exponentially just from the growth of compute without any development. We should expect continual paradigm shifts in the capability of AI systems. And so we need to plan for an ever-changing future. AI has never been worse than it is today. It's only going to become more and more powerful as we move into the future.