In Conversation with Aravind Srinivas | Cambridge Union

Thank you. Hello and thank you. It's a great honor to be here and thanks to everybody at Cambridge Union for welcoming me very warmly.

One of the things that I didn't expect about being a founder is how much time I get to spend at universities. Speaking in universities keeps you really humble. For a CEO, it's the only audience that actually argues back. So, we'll have time for that. And before I begin, I hope you'll forgive me for speaking from my notes for this talk. Universities are a rare space where I can share some important ideas. I don't have to be a salesman. So, I've written out my thoughts just for you. So, let's begin.

Normally, when I'm invited to speak somewhere, they ask me to talk about the future of AI. I'm not going to talk about the future of AI. Instead, I'm going to talk about the future of you. Your future is far more interesting to me, and it's probably more interesting to you, too. So, to get there, we'll talk a little bit about the past. And Cambridge is such a perfect place to talk about both the past and your future because it has both of them.

As the story goes, the peer review system as we know it was invented here at Cambridge. The truth, to be accurate, is that it was just a Cambridge professor who suggested it at the Royal Society of London. That was in the year 1831. Today, more than 3.5 million scientific papers are published every single year under the peer review system. Now, let me tell you a story. About 160 years later, two PhD students at Stanford University were working on network structures and data mining. This was when the internet boom was just kicking off.

The World Wide Web was all that anyone could talk about, just like the word AI today. And like today, most of the people talking about it back then were investors and the media. The tech community then was the telecommunications industry. There weren't a lot of academics in the discussion. Some, but not a lot. And these two PhD students at Stanford weren't exactly the insiders of the internet boom. But they did know one thing that people outside the university system didn't really care about. Like most academics, they understood the purpose of the university. The purpose of university is to advance knowledge, which is exactly why we use a peer review system. It's also why academic writing has so many citations.

And these two students knew that the amount of knowledge in the world, proxied by the number of papers, only grows. It was easy for them to see that the web would also grow as human knowledge grows. And they also had an experience with an old problem that was new to the web: how to find the right knowledge or the best knowledge. In the old model, it's simple. Academics judge whether somebody's research is useful or credible based on the number of times it got cited. So, the students built an algorithm that evaluated every single website on the internet based on how many other websites cited it or linked to it in relation to the content of those sites.

They called that system PageRank because they were ranking every single page on the internet, and because one of these PhD students had a surname Page. Larry Page and his partner Sergey took this one tiny academic principle and made it available to the mainstream. That's Google. They're doing pretty well today. Google became one of the most useful tools in human history and one of the most successful businesses because Larry Page simply adapted what universities have been doing for about a hundred years.

As I said, the purpose of university is to advance knowledge. So, I'm going to tell you three things about knowledge in this talk, and my hope is that we'll come to a much better working understanding of AI by thinking about some of our own intelligence. First, knowledge has a life. It's born, it decays, and it dies. Knowledge has a life cycle. And because it has a life cycle, it undergoes evolution.

As what we know evolves, we're all getting smarter. It's called the Flynn effect. The Flynn effect is the observation that humanity has been getting smarter by about three IQ points every decade. IQ tests are constantly reformed so that the mean stays at 100. But an easy way to see it is this way. In 1950, if you took somebody in the 92nd percentile of intelligence, they would only be in the 58th percentile today because we all got a lot smarter.

1950 is also the same year the Turing test was invented. That was the first test to determine if or when a machine is smarter than people. That means the Turing test in 1950 was a lot easier to pass because the bar for human intelligence was a lot lower. That's the power of evolution. Knowledge always leads to more knowledge. And so the downside is that the world can get smarter, but you as an individual can get dumber.

Knowledge also decays. Think of it as half-life. The half-life of a radioactive element is the amount of time it takes for half of it to decay. And so the half-life of knowledge is the amount of time it takes for half of it to become obsolete. Here are some examples from right here at Cambridge. In 1978, Sir Roy Calne helped introduce cyclosporin into transplantation. This was important for preventing organ rejection. The first clinical trials were done at Cambridge. In 2001, another drug, tacrolimus, became the predominant treatment in its class in American kidney transplantation. So you could say that the medical knowledge of cyclosporin treatment had a half-life of about 23 years.

It's an imperfect measurement, but it's a useful mental model for evaluating engineers in fast-evolving fields. The implication is: at what point in their career is half of what they know fully obsolete? At what point in your career do you think it'll happen to you? The answer depends on how much time you spend learning new things.

Sadly, in rare cases, knowledge also dies. There are languages with no remaining speakers or wisdom that dies with an elder. This is why for more than 4,000 years, humans have built libraries. When the web came along, we put infinitely more of what we know there. And so, for a long time, people have felt compelled to preserve knowledge. Knowledge dies, but it doesn't have to. So we are constantly saving it and also growing the world's total knowledge as we should. I should note knowledge is like a living thing, but it is not a living thing. Only you are alive, and so I just think you're more alive when you're learning something new all the time.

This is exactly why Perplexity exists. By the way, of course, I didn't come here to talk about Perplexity. I'm here to talk about you and the economy of knowledge. Which brings me to my second point. The second point is that knowledge is not money, but it behaves like money. There is a concept called the time value of money. Interest makes it grow. Inflation makes it shrink. With enough time, both interest and inflation are compounding.

When £100,000 has the buying power of £40,000, in roughly about 23 years, you would need £250,000 to break even. 23 years later, you would need £625,000. It's compounding. So, the gains or losses grow exponentially with time. That's why the most basic financial principle is that you should always be compounding interest at a rate greater than inflation. Because if you're not growing, you're shrinking.

Knowledge behaves the same way. Just like the value of a pound, the value of what you know will depreciate over time. And just like regular depreciation, it'll do this only a small amount every day, so small that you hardly notice it until one day you wake up and you have nothing. Your knowledge no longer has the same market value. This is inevitable. Remember the Pythagorean theorem was once something only elites were allowed to know. Everyone here has learned it in grade seven. We all get smarter.

Someday, far in the future, what you studied here at Cambridge might be the curriculum for seventh graders. And that's a good thing if it happens. It means there's progress. But it also means your economic goals require that you never stop learning. You must learn with an interest rate that's greater than the inflation rate of knowledge, or you must double what you know at a rate greater than the half-life of your knowledge. And that process takes time.

When you view the world through the lens of compounding, you have to view it with a long future. That's why the best trick for compounding your wealth is to start young, and the second best way is to die old. Warren Buffett started investing at the age of 11, but 99% of his wealth accumulated after he was 65, which hints at the final similarity of knowledge and money: you have to spend it.

When you have capital, you have to deploy it to work. When you have knowledge, especially exceptional knowledge, you should put it to work. Do something with it. That's the difference between invention and innovation. Invention asks, did you make something? While innovation asks, did it make a difference? And if you want to make a difference, you must take action on knowledge. Only action triggers new questions, inquiry, and new knowledge. You must always be learning and always be doing. And this is where AI comes in and leads me to my final and third point for today.

The third point: the purpose of AI is to accelerate human knowledge and prosperity. Consider these first two things I've said this evening. Knowledge has a lifelike character. We should nurture it, advance it, because broadly people benefit. Knowledge has an economic character, and we should cultivate it because broadly people benefit. If you believe in these two things, then you also have to think about AI slightly differently. Instead of being obsessed with achieving AGI or some other theoretical idea, you would see the reason to build better, more accurate AI is obvious.

You would see that the highest and most important purpose of AI is to make people smarter and more prosperous. If this sounds a lot like university, that's a good thing. Think about peer review again. It was invented in 1831. This year, Perplexity launched a feature called Model Council. You ask three different AI models the same question. Each of them goes and researches, comes back with an answer, and then a chair model evaluates where these models disagree and agree and reasons about a final answer.

Model Council does not replace peer review. Its purpose is to bring the spirit of academic inquiry to more people beyond these walls all over the world. And so the idea that it's more important to be right than to sound right, and to scientifically seek truth and knowledge, is very important. The arc of the universe is for everyone to have more access to more knowledge.

I'll close with a short story about my own company. When we started, we were just three AI researchers with more questions than a vision. And then for a fourth person, we decided it was time to get health insurance. Health insurance in America is pretty complicated even for Americans. And I'm not even from America, so it was even more confusing to me. The Google results were of no help. And the normal thing a founder does in this case is they call up somebody who's done it before, like an investor or another founder. It was too late at night and I wanted to get it done. So I had to find the knowledge myself.

And the problem with that is doing the research yourself means you Google through SEO-optimized insurance sites that are all bad advice competing to look the most credible. That's when I built the first version of Perplexity. It did the research for me, and I made it cite the answers so I could verify and keep asking more questions. And then I set up our healthcare. It was solved. On to the next problem, and the next problem, and the next problem, and here we are today. The moral of that story is you don't need to know a guy anymore.

You can just learn stuff and do it. With AI, you can access any knowledge. But it's a bigger deal that any person can access any knowledge. And so this is the ultimate promise of AI. Anyone now can be curious. Anyone can build anything. Anyone now can enter the temple of knowledge. And that's good news for you. It means the most important thing you can do for the rest of your life is to be curious.

The most successful people among all of you will be whoever has the most questions and the best questions. So my advice to you is to become as curious as you can. Do not settle for scrolling social media on your phone. Do not settle for an AI bot that makes you dinner reservations. Do not believe a chatbot that tells you you're brilliant. Believe in the spirit of inquiry. Believe in what you can learn with AI. And believe in what you'll make when you do. And make it good. Thank you.

Yeah, so I grew up in a city called Chennai in southern India, and my background prioritized knowledge more than wealth, and that quality still is with me today. If you could go ask my parents whether they would be more proud of me founding Perplexity and having all the success and, you know, the startup valuations compared to getting a PhD from Berkeley, they would say the second, because seeking knowledge was always more important to us than seeking wealth. And in some sense, I'm really happy that the product that we are, in our company, embodies that fundamental value.

Yeah, I would really give a lot of credit to my trajectory, to founding the company, to being at Cal, to doing a PhD there in AI. I think I got to have a chance to do an internship at OpenAI very early on in my PhD, and I remember going with a lot of different ideas. It was just 40 people at the time. It was a nonprofit company, and the founder and chief scientist, Ilya Sutskever, gave me 15 minutes of his time, which is fantastic to have as an intern.

I prepared like two days for that meeting, had a lot of my ideas written down, and he listened to it pretty fast and said, 'All your ideas are bad. Let me tell you the right idea.' And, you know, obviously it was very disappointing considering the prep, but that one single meeting taught me the importance of seeking truth and not taking it personally when ideas are debated. And the second thing is he told me the truth in a nutshell, which turned out to be right, which is he said there are only two ideas that matter: generative learning and reinforcement learning, which is training on all the data and then shaping it to be good at certain skills.

And you just have to throw compute at it and it'll all work. That's basically what he said, and that's basically what ended up being true. And so he also said, you know, there are only a few products you can build that allow you to take advantage of all the capability advances in AI. One is self-driving cars, and the other is search. And I was far more motivated about search. It aligned with my own values of seeking knowledge. So that's essentially decided why I went on to build Perplexity. So I often think like if I had done a PhD in a different university, or if I didn't get that internship, or if I didn't get that 15-minute meeting, life might have gone a very different way.

Yeah. The other memorable part of my PhD was my time here in London at DeepMind. I told Demis this when he named all the rooms after scientists. So every single room was named after like Faraday or Ramanujan or Einstein. And so that day I felt as an intern that if I ever have a company and if it ever gets to be good enough to have an office, I'm not going to name it after different cities or neighborhood towns. I'm going to name it after scientists, and that would be like a celebration of knowledge. And that's what I ended up doing at Perplexity as well.

And so my time at DeepMind was very motivating for a lot of reasons. I felt like you could actually create a successful company that's in service of like knowledge discovery and curiosity. And they had a library, and there was a book called 'In the Plex' that I got a chance to read. And as every other intern would, you'd spend most of your time in the office, even the night. So that's what I did. And I would launch training jobs on the cluster during the day, have it be running, and be reading a book in the library during the night.

And that was the book that truly convinced me that I had to be an entrepreneur, where there was a foreword in that book by Larry Page which said there were only two career paths that I saw possible for me. One is to be a professor, or the other is to be an entrepreneur. And every other path I would have to be executing on somebody else's vision. In these two paths I could execute on my own vision for the future. And I was like, 'Wow, that was pretty brilliantly put.' Like that perfectly resonates with me. And I think I don't want to be in academia for too long, so I know what I need to do. So that was how my time at DeepMind influenced me a lot.

I think we often think about like how hard it looks when you start a company, but for us we were so focused on solving hard problems and staying curious. Like most of the things that we had to do, we just had to learn ourselves, and we wanted to build a tool that allowed us to learn. Literally, that was the birthing of the product. When we started, we were like, okay, AI is getting really cool. This was pre-GPT-3.5 and ChatGPT days. But because we were researchers, we got to see the, you know, the scaling law advances and the capability advances. So we knew that incredibly new products that could never be built before were completely AI-first, rather than AI being bolted on.

I think that was the moment when we felt like we had to start a company. But we still did not know what was the right product. We were searching, and we were first-time entrepreneurs, all three of our co-founders, and nobody knew anything. Like nobody knew how to set up servers, nobody knew how to hire people, nobody knew how to set up insurance for other employees. I didn't even set up insurance for all three of us, and I didn't set up any HR tool, like I didn't have an office space. Like there's a lot of things that I didn't know, how to raise money. I never made a pitch deck ever for the company.

Jeff Dean, one of the most famous Google engineers of all time, I got connected — I had known him partially through my time at Google, but I got formally connected to him through a professor of mine, and he asked me for a pitch deck to angel. I was like, 'Hey Jeff, I don't have a pitch deck. Here's the link to the product. You can just try it out, and if you like using it, invest in it.' And he actually loved using it, and he's like, 'Okay, like I want to put in $250,000, take it.' And it's like, 'No, I don't have that much space for you. I will take $25,000.' So that's how I used to raise capital.

And so a lot of problems we just figured out incrementally over time. And whenever the product couldn't answer our own questions when we were problem solving, we would just go and make the product better. And that philosophy still continues today. Like even today I have a lot of problems to solve at a different scale, and I want to use my own product to help me solve those problems. And when it doesn't work, I would push our team to make those problems go away in the product. And when the product gets a lot better, it really helps me to solve my own problems.

And I feel like this is the fundamental crux of Perplexity. It's meant for anybody who's curious, always wants to be a problem solver, push the edge, push the boundaries, always wants to research new things. And as the scope and scale of these problems gets more complex, our product also evolves to become better and better. And we kind of want to write the trajectory of our own scope and scale of problems and the company's magnitude as well as the quality of the product in the same exponential.

Like I said in my talk, the goalposts always keep shifting. People often talk about the word AGI. It's a very theoretical concept. The reality is that what we are as humans today is far different from what we were in 2021 or 2022. With the power of these tools, we've gotten a lot smarter, too. We've been able to pose questions we never even had the opportunity to find answers to. And now we're posing questions for which there are no answers, even where we're using the help of these tools to figure out small pieces that'll help us further and further advance questions about the unknown, right? And so AI is no match to a human who's using AI.

standalone because AI is still not good at posing the right questions. You have to shape and steer the AI in directions that allow you to solve hard problems. To even want to be solving these problems is a complete human agency and driven by your own curiosity. And so those are the things that AIs are not going to take away. And so AIs are going to help us solve problems like there are a group of mathematicians who are trying to use AIs to solve like Millennium Prize problems, and there are a group of physicists who are trying to solve unknown problems, there are a group of biologists who are trying to cure cancer. I think it's amazing that they have the power of these tools to go and you know try to solve these problems that will truly advance humanity. And what it's not taking away from us is our own intrinsic curiosity to go even like approach new problems.

I actually think the dangers of AI are like way more boring than most people talk about. I think there's a lot of like fake content on social media at this point that it's hard to even know like what is real and fake at this point when you scroll through it. And fake content gets a lot more engagement. And so a lot of AI compute capacity is being put into making you know like content to increase social media engagement. And so that's like having a far more negative effect on our own attention spans and like the amount of time we consume on these platforms and like using all that data to feed us more ads. You know like those kind of things like or sycophantic AIs that allow us you know kind of help make people feel better in that moment in time but not truly solving their problems. These are actually the dangers in my mind about AIs rather than them taking over jobs. There's still like a lot of evidence that companies like whatever layoffs have happened have been unrelated to AI. It's just been overhiring during the COVID era and what's happening at least in our company and companies that I'm seeing elsewhere is that people who are already there are using a lot of AIs that every single person is functioning at like 20 to 30% more than what they could without AI that the overall output of a company is far higher and companies are advancing their business, their revenue, their topline revenue or the bottom line profits far more efficiently than they could have ever done before. And people are just way better problem solvers. Everybody's working even harder than they used to. And not because they're forced to, they just are finally able to have that agency.

I really think the idea of having like independent evaluations of AI models in terms of where they are capabilities or like auditing the products regularly and seeing where the companies might not have like you know had sufficient guardrails enough for rolling out those models in the form of products. I think that's very good. Pointing these out and like helping the companies like solve them in a data-driven scientific manner is very important and I think we should do more of that. Every single company should have an eval team, red teaming team and that focuses on like you know the right kind of guardrails for agents running in sandboxes or especially social media and like personal AI companion companies should focus a lot on like the non-engineering aspects of it too like you know making sure that AIs are not sycophantic or trying to overengage the user just to increase the amount of minutes spent on the app. And so I do think a good amount of regulation on the personal AI side, the personal AI companions or AIs for social apps is probably worth it. And as far as like core capability advances itself, like how far are they getting into like you know real super intelligence paradigm, we should do that with independent evals and audits of like where the security standards are, the sandboxes, the guardrails built and like how it's being rolled out. I think all that stuff was very good.

So I'll preface that by saying that five years ago I was actually like during my PhD, nobody, I had access to all the best researchers and I can guarantee you that nobody saw any of this coming you know so you can take whatever I say with a grain of salt. I can only say one thing is I feel like we'll be more curious than we are right now and I feel like everyone will have a problem solver that is quite affordable and we'll be able to throw it at whatever hard problems we want to solve and who can make the best use of it will be the people who are like most curious to even want to go and attack these problems and I hope like that gives us all sufficient agency to go wanting to solve really hard problems that the solutions of which can actually move humanity forward by like multiple decades. That's my hope.

I think we need independent audits for sure. And that's what the recent like super intelligence accord was trying to get at where there would be an independent group whether it's a committee or a nonprofit research organization. That's yet to be decided. But it's good for an independent group to constantly evaluate the security standards of the most frontier agent products and work closely with each company to share those findings and allow those findings to be incorporated in their post training or the way the infrastructure is built. And so that's what I'm excited about and there are already some good companies in the space. So I'm hoping to see more of that.

So fundamentally it just changes the abstraction space in which we are doing our thinking. So you know like when nobody writes assembly code. So we're not actually dealing with how memory is managed, how different processes are run, how different registers are being used. And so you know like when you're writing code in Python it's even garbage collection is handled for you for free. So every single time like something is taken care of for you, you're just trying to solve a problem in a higher abstraction space. Let me give you some examples like there are a lot of engineers, research engineers at our company who are now beginning to post-train our own models who feel so free having to not think about how to do multi-GPU training jobs because AI is taking care of all those details and they just have to design the experiments. They have to look at the loss curves. They have to look at the evals. They have to design the evals. And then they have to decide what experiments to run next. Okay, maybe you can have an AI help you do that too. Then you decide where to even allocate compute. What are the ideas even worth working on? So every single time you're just moving to a higher abstraction space that you never even had the opportunity to work in because you were like so focused on the mundane details. So that's what is very exciting.

So I feel like what is really lacking today in the models as far as continual learning is concerned is when you give an AI perfect context it requires to solve a problem. I feel like we already have an AI in today's frontier we already have that taken care of. Like if you give a model perfect context and all access to all infrastructure it needs to solve the problem, it absolutely can do it. But what continual learning should address is what happens when you don't have perfect context. How do you elicit whatever information is missing? How do you even know what is missing first and then use that knowledge to figure out how to get the remaining information either by asking other human beings or either by going and learning some things from scratch again or if you're asked to predict the price of Bitcoin in 2030. What's all the set of questions that you need to pose? Maybe you don't even have all the knowledge that you need to know to make that prediction today. But what if I allowed you like two years worth of compute? Would you be able to do that? And in those two years, what are all the experiments that you want to run in the world to gather more information? That's what continual learning should address. And it's possible that it could still be a coding agent similar to what current frontier models look today that are just able to run real experiments on their own and can run like you know long horizon tasks over like multiple years worth of like tokens and compact those tokens efficiently. But where this thing breaks today is that long horizon. So you need long horizon capability even way longer than what is needed for a single task and you need the ability to keep posing new questions and you need the ability to run real experiments in the world. So you could argue that like if you could create just like how we've created coding agents, if you could create a research agent that can actually be like a human scientist and run experiments, then that could allow you to solve continual learning.