Good afternoon, ladies and gentlemen. And I can assure you I will not be rambling for the next 15 minutes. But interestingly enough, you probably would realize you've used artificial intelligence more in the last one year than you probably would have expected. Chatbots, probably image editing, maybe even summarizing a few documents. From my perspective of having covered this space for a very long time now, there's one company which has got the puzzle completely right, or as the tech people would call it, the stack in place. Google is the answer. They do it better than most. You would have heard of Gemini as the AI chatbot. You probably heard about Google Cloud as well, even the Tensor AI hardware. But there's one thing that this company never let go of, which is the intent for research over at Mountain View. It is my pleasure to speak with Pushmeet Kohli, who is the vice president of science and strategic initiatives at Google DeepMind. Thank you, Pushmeet, so much for joining us.
How's the weather in London? Has the winter set in?
It has been cold, but the weather today is quite bright and sunny.
On that note, Pushmeet, I wanted to start off by asking you: DeepMind is straddling that fundamental science and applied AI line, and there have been examples of recent successes with AlphaGeometry and AlphaFold. I'm intrigued, and I'm guessing most of the audience here would be too. What is the framework with which you decide which sort of applied sciences require bandwidth and investment with AI?
Thanks, Vishal. I mean, if you look back at the mission of DeepMind, and ever since the founding of DeepMind here in London almost 15 years back now, the mission has been to build AI responsibly to benefit humanity. And so the benefit humanity sort of part, we believe one of the key things that AI can do to benefit humanity is to really push the boundaries of human knowledge, of what we know about the world that we live in, about ourselves as human beings, about life. And that we do through science by making progress in many scientific areas. And we have been very sort of fortunate to be able to sort of show the potential of AI in problems like the protein structure prediction problem and AlphaFold. So mainly our sort of focus is on what are these root node problems where AI can have transformational impact, which will not just lead to some small improvement but will really transform the way society does something. AlphaFold is a great example in this as a case study. Before AlphaFold was announced by Google DeepMind, it used to take almost five years sometimes to figure out the structure of one single protein. And these proteins are the building blocks of life. Everything to do with drug discovery to designing sort of new enzymes to deal with pollution and so on. Life is essentially made up of proteins, and yet we did not know the structure of these proteins, and it would have taken a huge amount of work and effort to uncover this knowledge. And what AlphaFold did is just made all of that sort of knowledge available extremely quickly. But in a matter of a few seconds now, anyone on the planet can figure out the structure of any protein on the planet, which would have taken millions of dollars and multiple years in the past.
Pushmeet, the general conventional wisdom would suggest LLMs are getting better and therefore they must be applied to more problems. But DeepMind's approach is different. It's more of a narrow approach where you tailor a model for a particular domain or a use case. How do you balance these two requirements?
Our approach, Vishal, has always been to push the boundaries of AI and to develop the most powerful models. And when we think about what are the most powerful and most competent sort of competent models, essentially the way to measure intelligence is to figure out how quickly a model is able to accomplish a task. And we have ever since our founding been developing models which are ever more competent in solving harder problems and in a more general way, which require less data and less supervision. And so in some sense, LLMs are a natural progression of that longer-term focus that we have had. On one side we are working on building fundamental sort of breakthroughs that push the efficiency of AI models, makes them general, from coming up with new paradigms like deep reinforcement learning, which showed its benefits on sort of Atari and the legendary sort of Lee Sedol Go match in South Korea, to then sort of specialist models for AlphaFold. And in each sort of iteration, whether we were using a specific architecture like a transformer architecture or sort of improving it, or sort of using the models themselves like the Gemini generation of models, which now have been able to show competency across a wide variety of tasks, not just in answering questions about your everyday sort of questions that you might be interested in, but also deep mathematical problems. So Gemini Deep Think was one of our premier models which achieved gold medal performance at the International Math Olympiad. So this is not an easy thing, right? Very few people are able to sort of solve maths problems at that level, yet a very general purpose model like Gemini Deep Think was able to achieve that. So if you look at the spectrum, yes, we are working on specialized models for specific tasks, but there are also general models that we are building to make that whole process much more efficient. At the end of the day, it's all about the problem. How do you solve the most impactful problems in the most efficient manner possible using AI?
Would you say that we've reached the point where scientists and doctors can start trusting AI with their work?
Now, I think to answer your question, this is a new kind of intelligence. We are still sort of learning and we are trying to understand the behavior of this technology. And yes, it is very powerful, but it does make mistakes, and the important element is to figure out when it is basically giving us the right results and when it is sort of failing. So going back to sort of AlphaFold as an example, now it solved this amazing problem. It could figure out the structure of any protein, but still if you're a biologist you will think, well, this model, although it is extremely accurate, it still makes mistakes sometimes, and I wouldn't want to spend the next five years of my life or ten years of my life assuming it is correct and then finding out that it is wrong. So AlphaFold was not just very accurate, but it was also very good about showing its uncertainty about the problem. So where it made a mistake, it also sort of held its hand high and said, well, I might have been uncertain about this particular solution, so don't trust it that much. Now if you look at the modern generation of LLMs, they are making improvements, but sometimes they do hallucinate, and we are building technology to make sure that when they do hallucinate we can catch it and also give users more understanding about when the output can be trusted and when it can't.
