Mustafa Suleyman: Empathy in AI

0:00Scaling a business in the US can feel like a maze. Each state has its own payroll, benefits, and compliance rules, pulling your focus away from growth, which is why founders use Deal. Deal is the professional employer organization, or PEO, that gives you a dedicated HR expert, plus Fortune 500-level benefits for your team. The National Association of PEOs says businesses can grow twice as fast if they use one. So, if you're scaling, make it simple with Deal. Go to deel.com. slash MOS and get up to three months free. One of the patterns I've seen over and over and speaking with successful founders is this.

0:40They don't wait for HR problems to show up. They build solid systems early on so their company can grow without tripping over itself. JustWorks is one of those systems. JustWorks is the human resources platform that helps you scale without sacrificing your team's needs. It handles automated payroll, benefits, compliance, even international hiring, so you can scale your team without slowing your momentum. JustWorks secures your hiring systems so they don't cost you time, money, or people, which is how you scale with confidence. JustWorks for your people. It takes a little bit of naivety to declare that your mission is to build artificial general intelligence ethically and safely.

1:30The goal of Pi is not to take a lot of time and distract you from human beings. The goal of Pi is to help you in your interactions with human beings. How do I learn more? How do I plan better? How do I get more efficient? And now you're going to have a personal AI that will be working for you in the background. people go like well i'm just going to sit down and watch television have my pie deal with everything it's like no no that's not actually in fact the design the world gets built the way we intentionally design and build it as creators of technology we suddenly have a new way of creating very very magical experiences for people

2:17That was Mustafa Suleiman and Reid Hoffman, co-founders of Inflection AI. I'm executive producer Chris McLeod. Reid recently sat down with Mustafa to discuss the ever-changing landscape of artificial intelligence, as well as the ideals that were essential in creating the AI assistant pie. And we're so excited to share this interview with you today, because it's the perfect prologue to our upcoming miniseries, AI and You, where Reid will talk with an array of AI leaders, including Mustafa, to explore how you can harness AI to scale your productivity, your business, and yourself, while staying safe in the process.

3:01Mustafa Suleiman has been an AI leader for decades. Before he co-founded Inflection AI with Reed, he was co-founder of the AI company DeepMind, which was acquired by Google in 2014. Mustafa's recent book, The Coming Wave, explores the seismic changes that are on the horizon for humanity and what we should be doing to build a future that is safe and secure for everyone. The interview you're about to hear is a peek inside the minds of two of the foremost AI leaders. They'll give a clear-eyed but optimistic view on how AI will revolutionize our lives. Now, Reed's interview with Mustafa Suleiman.

3:56Well, I have been looking forward to this Masters of Scale interview for years. Mustafa Suleiman, my co-founder of Inflection, a friend who I have learned from in multiple vectors ranging from poker to actually, in fact, really serious things that matter. So, Mustafa Suleiman, welcome to Masters of Scale. Thank you, Reid. It is super exciting to be here. I have obviously learned infinite amounts from you over the last decade, and so I'm super excited to be here having this conversation. It's been a long time coming. And so let's start with when you dropped out of Oxford and we're thinking, OK, how do I make a contribution?

4:34You're like, OK, I'm going to create what in the U.S. is called a 501c3, a nonprofit, but a kind of equivalent of an NGO for helping society. You worked the mayor's office. What got you into artificial intelligence? What made you suddenly realize, oh, gosh, this is something that I should be turning my attention to? Yeah, I guess I've always been a systems thinker. I love thinking about the relationship between complex ideas over time. And when I dropped out of Oxford, it was to start a nonprofit charity, a telephone counseling service. And I was really interested in the question, how do we have the biggest possible impact at scale to make the world a better place?

5:19And that was genuinely back in 2002, you know, my main motivation. And for the first, you know, five years of my career, I worked in nonprofits. I worked in local government. I ended up co-founding a conflict resolution firm and working all over the world as a facilitator and a negotiator, you know, at the UN for big companies, for local governments. And all of that experience led me to realize that actually it's technology that really has an outsized impact in the world. And technology was accelerating at an incredible pace, even in those days. I think it was 2008 or so when I really started to pay attention to Facebook's growth.

6:03And it was that realization that this product had gone from a complete non-entity two or three years earlier to then having, I think it was like 100 million monthly active users at that time. I was like, wow, technology really is the thing that is going to transform our world. And I want to try and participate and try and steer it to, you know, deliver the best possible impact that it can for our species. So say a little bit about the conversations that you started with Demis and Shane. Yeah. So, I mean, the best way to, you know, get involved in a completely new alien paradigm is to throw yourself in at the deep end.

6:43And so I ended up hanging out with Shane Legg and Demis Asabis at the Gatsby Computational Neuroscience Unit at University College London. And they very kindly snuck me in through the back door to listen in to the Lunchtime Lecture Series, which was about how neuroscience could be a tool for inspiring new machine learning algorithms. And I have to say, whilst I didn't understand all of it, I found it deeply inspiring. And, you know, we started going for lunch together, the three of us at Carluccio's, which is an Italian restaurant, a couple minutes around the corner from Russell Square in London.

7:21And we started thinking big and planning, you know, what it would be like if we put together an AGI company. And over the course of about six months, started reading papers, reading textbooks, you know, really teaching myself to understand at a high level, you know, how these machine learning systems work, what their objective function is, you know, how they're trained. I think the really cool thing that we had spotted at the time was that models that had the ability to learn were in fact starting to gain traction. And artificial neural networks had been around since the 80s, you know, proposed in various different formats.

8:00And traditionally, a lot of the AI models had been handcrafted. So they were kind of just complex sets of rules that interact with one another. And everyone realized that that wasn't how the brain worked. What everyone was looking for was sort of more brain-inspired type algorithms. And the idea, of course, was that a reward signal from an environment could update a set of weights in a network to ensure that the network was a better fit for the learning objective, eventually now building the kind of models that we see with large language models. And I think looking back, it takes a little bit of naivety to be bold enough to declare that your mission is to build artificial general intelligence ethically and safely.

8:51That was the strapline of our 25-page business plan, which we took to Silicon Valley in the summer of 2010 and ended up successfully pitching Peter Thiel at the time for a minuscule 1.5 million pounds, which we were over the moon about. And that then became our series A. Yep. And then what were some of the kind of the key insights that you were getting from the deep mind time that kind of shed a light forward to where we are now? What was the lens, you know, kind of deep mind forward from the early days? What the team were trying to do was use these models to generate a novel example of a black and white handwritten digit, right?

9:34So the AI would read lots and lots of examples of images. I think they were sort of like 300 pixels by 300 pixels, black and white, handwritten digits. And then it would try to generate a brand new seven that didn't exist in the data set, for example, that wasn't a match. It was a novel style of handwriting. and you know we we sort of saw this almost like a small video like a short 20 seconds as the video goes from complete black and white pixelated fuzziness and slowly resolves to you know a clear black background with a very distinct seven you know emerging out of the darkness and it sounds really small but it was a really mind-blowing moment for me and you know in many ways, all of the subsequent breakthroughs, whether they were learning to play Atari games or AlphaGo or AlphaFold or many of the health applications that we did that produced best in the world class of x-ray recognition or ophthalmology, eye disease, diagnosa, each of these was really a continuation of that very early work back in 2011, 2012.

10:49And when did you start seeing that scale compute was going to be an important part of the equation? What was the beginning of that realization? Well, we certainly didn't see scale back in 2013 in the way that we think of scale now. Even though looking back, they were really large amounts of compute that were used to train the Atari model, for example. So the Atari model used two petaflops of compute, A flop is a floating point operation. It's a unit of computation. And a peta flop, peta, P-E-T-A, refers to a million billion operations. So at the time, it used two million billion operations, which sounds like an enormous number.

11:37But to put that into perspective, every year since then, the cutting edge of AI models has used 10 times more compute. Over the last 10 years, the amount of compute used to train the best and the biggest models in the world has 10xed, 10 orders of magnitude in 10 years. So that gives you a sense of the trajectory that we've been on over the last decade. It's kind of hard to comprehend. And so when did you kind of shift to kind of the large language model focus? And what brought you to kind of Silicon Valley as a way of kind of building that? Well, I moved to work at Google full time in 2020. And I was lucky enough to be able to work on an earlier version of Lambda, which was called Mina at the time.

12:34And I joined a team that was maybe five or six people strong at the time. And it was really just a small research project. And the model was, you know, significantly smaller than GPT-2. It was, you know, sort of almost incoherent. But you could see glimmers of really impressive text. Like every now and then it would string together a sentence or two sentences. And, you know, it seemed really, really impressive. And over the course of my time there, once we renamed it Lambda and we scaled it up and trained a much larger model, it was just unbelievable to see how good it had become so quickly. And I think what we built with Lambda was an interactive back and forth agent.

13:19So in many ways, it was ChatGPT way before ChatGPT. And we were completely blown away with how good your seventh and tenth turn of conversation was with the model, because it obviously had the prior turns of interaction in its working memory. And I think that was the key insight that made us realize that actually agents, you know, were really going to be the future of this next wave of technology. Yep. So, and this is where at the very earliest glimmers, the inflection and pie story begins, because that conversation about like, wait, the agent interaction is part of what unlocks all of this capability into human experience and creates this productive amplifier in a number of different ways.

14:10and obviously one of the things that you were trying to do at the time was kind of say okay i'm trying to get this productized and launched to google you know it's part of my job here and you know google at the time until you know chat gbt kind of kicked them in the rear end and got them moving was kind of like no no no no we don't want to launch anything that threatens our search business and and we don't know what how people respond to this so let's just keep this as an r &d project. And you're like, well, but wait, this could be so important for humanity. We tried really hard to get that launched at the time, but there just wasn't the appetite for taking the kind of risk.

14:47And it was pretty clear, I think, to a lot of people at Google that this was potentially going to unseat Google's existing search business. And so it's super hard for a company to try to compete with itself and upend itself from within. But I think the interactive component of it was just so obvious. I mean, once we had hooked it up to search, we were trying to sort of ground the generations that the AI produced in the context of the search results and try and make it more factual by sort of training Lambda to reference a search result when it produced an output. And actually, that's what you now see in BARD, you know, Google's new AI, it's pretty much exactly the same model.

15:30So it's very clear to me that actually it was going to be conversation that is the new user interface, right? I think that the future of the web and future of digital interactions in general will be that you just ask your AI. I think everybody is going to have a personal AI. And what you really want is to make sure that it is aligned with your interests, on your team, in your corner, because you're going to turn to your AI for all sorts of important and exciting and entertaining and even sensitive moments in your life. It's not just going to be about directions or getting factual information. It's also going to be about, you know, venting your frustrations.

16:13It's going to be, you know, about asking a stupid question that you're sort of too embarrassed to ask a friend or a colleague, or, you know, it's going to be sharing in a quiet, private moment when you sort of want to reflect. And so I think that that new user interface gives us a kind of clay. It's like a new design material. As creators of technology, we suddenly have a new way of creating very, very magical experiences for people. And as a result, these sort of more relationship-based, interactive-based AI agents are what I think is going to be the future. And so say a little bit about, we have multiple theories of scale that we're going to actually get to in this discussion, because there's scale and a number of different interesting vectors, which may surprise some people.

17:01But let's start with a micro, which is the design of Pi. What did you think were the kind of the key things in an agent? What are some of the places where people can understand Pi is different from ChatGPT and other things they may have heard of? What's the, as it were, the one-on-one or the tactile kind of design of how Pi currently operates? Yeah, so we've designed Pi, which by the way stands for personal intelligence, to be really sensitive and kind and supportive. Our thesis was what makes for great conversation. And that really is going to be the backbone of the new set of surfaces in web, in apps, and in all of your devices.

17:47It's going to be conversational. So Pi needed to be very respectful, very patient, very kind, always curious, right? It's super important that Pi seeks to understand your intent, right? It doesn't make assumptions, And, you know, it doesn't stay firm when it's wrong. So it was really important to us that Pi is able to back down and seek feedback and ask clarifying questions. And I think that's slightly different to the other AIs on the market, which try to be, you know, sort of more of a sort of conversational Wikipedia, giving you facts and lists and assuming that it knows the right answer straight away.

18:28And, you know, that was the core hypothesis is that if we can design a personal AI that, you know, really gets to know you over time, that remembers the conversations that you've had in the past, then increasingly it will be able to personalize its style and tone to your style and tone and therefore give you, you know, a much better quality experience. So let's now broaden that out some to the kind of the view of the universe in which, you know, Pi is a stepping stone to coming, the kind of the personal intelligence universe and what that means for amplifying human beings. When this vision gets to scale, what does that vision look like?

19:08Well, I think that everybody over the next decade is going to have access to a personal intelligence, a Pi. You know, I think there will be many, many different types of AI in the world. Some of those AIs will represent brands. Some of them will represent businesses or digital influencers. Some of them would be trying to sell you stuff. Maybe you'll have a healthcare AI or a lawyer AI. Even governments will have their own AIs, you know, that help you with government services and your tax return. And so what you want as a consumer is an AI that is really on your side, that can represent you and that can interact with other AIs that are trying to sort of sell you something or persuade you of something.

19:55And really, that's what a personal AI is. I think over time, it will start to feel like a browser for your life, right? So if you imagine what a browser is today, it represents the sum total of your digital curiosities, right? Some of your tabs are going to be clusters, you know, where you're researching a, I don't know, a new camera that you want to buy. Another set of tabs will be like you, you know, trying to book a new holiday. Another set of tabs will be like you doing some research for work, or maybe you're looking for a new job. Each of these are threads or lines of inquiry in your life that require you to maintain state.

20:35You sort of have to remember where am I at on this little mini journey and how do I learn more? How do I plan better? How do I get more efficient? How do I keep up the rhythm of this little inquiry or investigation in my life? And now you're going to have a personal AI to help you maintain state, to help you dig deeper and learn more that will be working for you in the backgrounds. You know, your pie is going to go off and find, you know, new articles, give you summaries, find you nice how-tos and little videos, little snippets to help you keep progressing on each of these lines of inquiry. And that's kind of how I see it, preserving state across all these different sort of areas of your interest and helping make you smarter and save you time.

21:24And, you know, it naturally brings a kind of a comedian kind of perspective is like, well, my pie is going to talk to your pie and we'll sort it out. And intersect that with the kind of the human amplification, the humanism that's at the kind of the core of the design of this? Because obviously, to some degree, people go like, well, I'm just going to sit down and watch television, have my pie deal with everything. It's like, no, no, that's not actually, in fact, the design. The world gets built the way we intentionally design and build it. How does that play in the human amplification side? Well, pie is clearly, in the end, in the next few years, going to save people vast amounts of time, right?

22:02It is going to give you back time. And I think it'll be a question for you as to how you want to spend that time. I hope that it means that it will free us up to spend more time with our kids and our friends and our family and our loved ones and be out in the world because, you know, you're going to spend less time doing mundane administrative tasks online. You know, in many ways, Pi is going to be like a chief of staff for you or like a PA or a secretary organizing, planning, booking, buying, arranging, the amount of time that we all spend just processing payments online or just like ordering the groceries when it's a run of the mill order and we know what we need.

22:43There's a lot of wastage in front of our screens. So I think trying to make your life more efficient and more productive hopefully frees you up to spend more time interacting with other people. Equally, it could be that, you know, you now have more time to pursue your hobbies and your passions and your new learning interests, or it might be that you're trying to upskill yourself to change jobs, switch careers, move cities. Each of those things will be something that your personal AI can help you with too. Yeah, and one of the ways that I put a different lens on the same point you've just made is the goal of Pi is not to take a lot of time and distract you from human beings.

23:23The goal of Pi is to help you in your interactions with human beings. So if you come in, like, you know, I have this difficult discussion with a friend and you come and start talking to Pi, it's not, oh, just talk to me and I'll make you feel totally better and you can ignore your friend totally. It's like, no, let me help you. Like kind of like how you engage with your friend and understand each other and have a great conversation as part of the design ethos. It's to help you being your best self and navigate the world you're in, whether it's life or work or anything else, but from you out is, I think, you know, another lens into what you just said.

23:57Yeah, totally. I mean, that's a great way of putting it. I mean, in many ways, Pi is kind of there to absorb your tiredness and help you, you know, translate your sort of your frustrations onto Pi rather than onto, you know, other people in your life. And I think, you know, that's just a, it's hard to wrap your head around, you know, it really is a completely different experience. And that's where I think we have to sort of really try to make the right choices, really think very, very carefully. I mean, one of the things that you and I have talked about a lot and that we're very clear on is that, you know, we don't want to let Pi be used for romantic relationships, right?

24:37So, you know, other people will build those things potentially. It's just not our bag. You know, so now if, you know, someone sort of approaches that kind of style of conversation with Pi, Pi will be really deliberate and clear that that's kind of off limits. It'll be super respectful. You know, that's the nice thing about Pi. It never judges you no matter what values or views you come with. It'll try to talk it through with you and, you know, present both sides of the argument. But it's never going to put you down or reject you. But it does have its own boundaries and it has its own values. And they may not work for everybody, but, you know, we certainly we hope that, you know, it'll be a healthy way to proceed for most people.

25:16Yeah, indeed. need. Real leaders don't back down when the stakes are high. They innovate, they push forward, and then they take the stage at the Masters of Scale Summit. Join us in San Francisco, October 7th to 9th, to hear from the CEO of the New York Times, scientists using cutting-edge technology to find cures, the leader of crypto powerhouse Coinbase, a retired four-star general, and many, many more. Apply now at mastersofscale.com slash apply25. That's mastersofscale.com slash apply25. If you're ready to take your startup from idea to impact, then AWS is your launchpad. From data storage to machine learning to secure app hosting, AWS gives you the tech trusted by the world's fastest scaling startups.

26:09And here's the best part. Join AWS Activate today and you could score up to$100 ,000 in AWS credits tailored to your stage and network. Go to aws.amazon.com slash activate and start building. Expanding your business in the U.S. can feel like a maze. Every state has its own payroll, benefits, and compliance rules, which can pull your focus away from growth. That's why founders use Deal. Deal is the professional employer organization, or PEO, that gives you a dedicated HR expert plus Fortune 500-level benefits for your team. The National Association of PEOs says businesses can grow twice as fast if they use one.

26:52So if you're scaling, make it simple with Deal. Go to deel.com slash mos and get up to three months free. Now let's go to another version of scale, which is the compute behind this. Say a little bit about what the scale of compute is, what inflection is. We just had an announcement recently about H100 clusters. Say a little bit about this. Right, yeah. I mean, so this entire story of large language models is a story of scale. It's really quite surreal. So the training data sets, for example, use hundreds and hundreds of billions of tokens, or you can think of them as words. You know, more texts than any one of us has ever read in our lifetimes or ever could, even if it was the only thing that we did.

27:41And you're right, in terms of compute, that is really the core engine powering the progress of this new revolution. and you know the workhorse of that engine is the nvidia gpu graphics processing unit these chips that were previously primarily used for gaming and turned out to be brilliant parallel processors when it comes to running computations for neural networks so we as you mentioned have managed to gather together a pretty impressive supercomputer recently we announced our new supercomputer of NVIDIA H100s. And on the open source ML performance benchmark, NVIDIA and our partner CoreWeave and us actually demonstrated that it was the fastest computer in the world, which is pretty remarkable in itself.

28:31And we also just announced our new funding round, where we've been lucky enough to be able to raise$1.3 billion. And we'll be using that to build the largest super cluster in the world, which we'll have up and running by the autumn of this year. And that's incredibly exciting. It was just quite surreal that us as a 40-person, one-year-old startup has been able to build the largest cluster in the world. Say a little bit about what that size of cluster is relative to how people talk about exoscale computers and supercomputers and so forth. Because it's another of the lens of scale here is on the amazing of the scale compute.

29:14Well, one way to think about it is that for each of the words being read by the AI, the AI needs to learn the relationship between that word and the previous words in the sentence. And there could be billions and billions of possible combinations of previous words in a sentence. So you can think of it as kind of like learning an all-to-all relationship between all the words that it has read in order to give you the likelihood that the next word in a sentence, because it's predicting or generating the next word in a sentence, it gives you the likelihood that one word will appear over another, right?