The Exchanges

Q of the largest kind of media publication owners in the world, um, and he, he said, I'm worried, Harry, I don't think that I will have a business in a couple of years. I think, bluntly, we're getting killed on our advertising because everything's getting scraped, uh, and they're not coming to our websites, and that's where we get paid. We get pay-per-clicks. Uh, is he right to be worried?

A I think he is right to be worried. Like, again, you look at Google's announcements yesterday, they were talking about this day after Palm Two, You suddenly look at the new Google page where they've got the language model, and it's just text, and where are the clicks? It was like when Google introduced AMP. You know, this is where, rather than looking at the New York Times page, you have this formatted thing with no New York Times kind of stuff there. Like, these search entities that aggregate are just intermediating more and more, and people are going to become used to just having synthesized input. So what does search look like? What does it look like when you're GPT-IV can write you an article about any news that's happening in a way that's customized to you and your context and everything like that. This is massively disruptive for media and information, and so they have to think, where am I in the future where, again, the worst way to think about the impact of this is they're really talented grads that occasionally get off their meds, and we push a button and we can get a thousand of them. Those grads include journalists, and you can have your own journalist army, your own writer army, your own coder army, your own designer army.

AI assessment note: “I think he is right to be worried. Like, again, you look at Google's”

Q another podcast and I was astounded and inspired by it. We mentioned before, you know, my mother's got MS and I hate the doomsday only version of kind of AI and the future of, you know, GBT. You said to me before about its impact on health and MS in particular and other conditions. How can it be so transformatively to solve some of the world's most challenging chronic conditions?

A So I think a large part of our problem is that we can't scale because information flow is so limited as we write these things down. Like you can never capture all of that. So anyone who's had a loved one that has one of these conditions, Notice how difficult it is, because you go from specialist to specialist to specialist, and you try to build that mental map, and we're so lucky that we have so much access. But why isn't it that we can't just push a button and see every clinical trial, and a deconstruction of all those and things? What if you had a thousand GPT-IVs organizing all that knowledge, and then make it available to everyone, so you can see the exact potential mechanisms that way which MS works, and all the potential food, Other things that work with that. So as you try different things with your family member, you can see, well, she reacted this way to the food or this way to this medicine, and it is a more holistic thing because you can have personalized medicine versus one specialist for a thousand people. You can have a thousand GPT-IVs or equivalents or MedPalm-II for you. So we need to organize all this knowledge and then use these language models and others to make it accessible to you.

AI assessment note: “use these language models and others to make it accessible to you.”

Q Can you just help me understand AI integrated versus AI first? What is, what is.

A AI integrated means that I have an existing newsroom and I bring in AI to write faster drafts and things like that. AI first is saying I have an army of things I can spin up instantly that can help me achieve these certain things to create news that is valuable for this reason with this feedback loop. And so you build the system kind of from the start thinking AI at the core versus AI being integrated into improve existing systems. Because so much of news is what? We find information, we have drafting, we have this, we have that, we do these checks. A lot of that can be simplified, just like we move from the analog to the digital age, to the internet age, the next age is the AI age.

AI assessment note: “AI integrated means that I have an existing newsroom and I bring in AI”

Q another podcast and I was astounded and inspired by it. We mentioned before, you know, my mother's got MS and I hate the doomsday only version of kind of AI and the future of, you know, GBT. You said to me before about its impact on health and MS in particular and other conditions. How can it be so transformatively to solve some of the world's most challenging chronic conditions?

A So I think a large part of our problem is that we can't scale because information flow is so limited as we write these things down. Like you can never capture all of that. So anyone who's had a loved one that has one of these conditions, Notice how difficult it is, because you go from specialist to specialist to specialist, and you try to build that mental map, and we're so lucky that we have so much access. But why isn't it that we can't just push a button and see every clinical trial, and a deconstruction of all those and things? What if you had a thousand GPT-IVs organizing all that knowledge, and then make it available to everyone, so you can see the exact potential mechanisms that way which MS works, and all the potential food, Other things that work with that. So as you try different things with your family member, you can see, well, she reacted this way to the food or this way to this medicine, and it is a more holistic thing because you can have personalized medicine versus one specialist for a thousand people. You can have a thousand GPT-IVs or equivalents or MedPalm-II for you. So we need to organize all this knowledge and then use these language models and others to make it accessible to you.

AI assessment note: “What if you had a thousand GPT-IVs organizing all that knowledge”

Q The deep mind kind of desegregation or kind of unification was supposed to, of course, a lot of friction and be a negative press reported. Do you disagree with that?

A Of course. It is a lot of kind of replicated jobs. There was kind of brain and mind, and now they're kind of brought together. And it's a very different management style and other things. These things are never easy. But this is why, like, you saw Palm, five hundred and forty billion parameters, and that you had DeepMind with sixty seven billion parameter chinchilla, chinchilla, which is like, just train more as opposed to more parameters. You look at Palm two as a combination of both. And so it's trained for far more on far better data. And then that means it's only a fraction of the size, like fourteen billion parameters is one of the test comparative models versus the 540 and 67. So you can start to see this fusion of ideas, even if the teams, you cannot Integrate two big teams like that instantly.

AI assessment note: “you can start to see this fusion of ideas, even if the teams”

Q You kind of unpacked so much for me that I want to kind of go one by one. You said about kind of national data sets. Why national data sets versus supranational data sets?

A Because, like, I'll give you an example. There was a team that did Japan Diffusion, including some of our staff. Uh, so we took Stable Diffusion, and then changed the language model. Because when you typed in salaryman and Stable Diffusion, it was a very happy man. Whereas in Japan, a salaryman is a very sad man. You know, um, local context is important in these models, because we're going to outsource more and more of our thinking and minds to it. And so, do we want to have a British model, or do we want all the models to be Palo Alto? You know, like, a sparkling wine has to be from the Champagne region, like, Is the only real foundation model AI from Palo Alto?

AI assessment note: “local context is important in these models, because we're going to outsource”

Q Can you just help me understand AI integrated versus AI first? What is, what is.

A AI integrated means that I have an existing newsroom and I bring in AI to write faster drafts and things like that. AI first is saying I have an army of things I can spin up instantly that can help me achieve these certain things to create news that is valuable for this reason with this feedback loop. And so you build the system kind of from the start thinking AI at the core versus AI being integrated into improve existing systems. Because so much of news is what? We find information, we have drafting, we have this, we have that, we do these checks. A lot of that can be simplified, just like we move from the analog to the digital age, to the internet age, the next age is the AI age.

AI assessment note: “AI integrated means that I have an existing newsroom and I bring in AI”

Q I'm really naive and, and, uh, basic in terms of my thinking, which is why I'm a venture capitalist. Uh, my question is, what do we need to do to get to that stage? When we look at the data needed from the individuals, the data, the data that GPTs need, how we make the models work most efficiently?

A What do we need? First, we don't have to have that data from individuals. So we had Galactica as a scientific language model, but now we have MedPOM-II that exceeds doctor levels. So that was a Google announcement yesterday. We have AIs that can understand Articles better, as good as doctors, shall we say now. So we can scale that, because why do you need one when you have a thousand? So we take the existing generalized knowledge and all the hypotheticals, and we bring that together into an integrated common system available to everyone, because the building blocks are nearly here for that. Then you can personalize it later, and again, there are regulations and things around that, to how your, again, how we treat our loved ones and other things like that. The first thing is, let's get all the knowledge in one place and make it organized and useful. And so I think we're at that point now where the language models have just hit that point, that we can organize all of the world's Alzheimer's knowledge, longevity knowledge, autism knowledge, MS knowledge, and you can just type, and it can say this is the source, this is what it looks like, these are some hypotheticals, this is what we know, what we know we don't know, what we think we might know, etc. And then it can learn about you and your queries, because this is the other thing about lots of the language model things we've seen…

AI assessment note: “What do we need? First, we don't have to have that data from individuals.”

Q The joys of doing what I do is going on schedule, find a woman, and then we will kind of retain some form of normality of schedule. What is the future of Healthcare systems that, like, do you think with GPT models operating in this way?

A I think that you can change the nature of a doctor because a lot of this stuff is kind of very basic. I think, you know, you had Babylon Health and others trying that chatbot. It wasn't ready. Now you've got this. Everyone should have their own AIs looking out for their own health with that objective function, you know, and then the nature of a doctor becomes different in terms of they have more rich information about an individual while it being preserved in a private manner. I think what you have is you have things like processes and procedures improving, like, uh, wound care, for example, in the NHS. Um, if you are injured as an elderly person, and your wounds aren't treated properly, you're more likely to die by a factor of eight times. Being able to monitor those types of things with this information set means you're eight times less likely, and then you have far more efficiency around that. So the information density around healthcare improves, which means that then our own healthcare improves. We all have access to as much knowledge as we want to within our own context, and so do our providers and the people that help us.

AI assessment note: “you can change the nature of a doctor because a lot of this stuff”

Q You kind of unpacked so much for me that I want to kind of go one by one. You said about kind of national data sets. Why national data sets versus supranational data sets?

A Because, like, I'll give you an example. There was a team that did Japan Diffusion, including some of our staff. Uh, so we took Stable Diffusion, and then changed the language model. Because when you typed in salaryman and Stable Diffusion, it was a very happy man. Whereas in Japan, a salaryman is a very sad man. You know, um, local context is important in these models, because we're going to outsource more and more of our thinking and minds to it. And so, do we want to have a British model, or do we want all the models to be Palo Alto? You know, like, a sparkling wine has to be from the Champagne region, like, Is the only real foundation model AI from Palo Alto?

AI assessment note: “local context is important in these models, because we're going to outsource more”

Q of the largest kind of media publication owners in the world, um, and he, he said, I'm worried, Harry, I don't think that I will have a business in a couple of years. I think, bluntly, we're getting killed on our advertising because everything's getting scraped, uh, and they're not coming to our websites, and that's where we get paid. We get pay-per-clicks. Uh, is he right to be worried?

A I think he is right to be worried. Like, again, you look at Google's announcements yesterday, they were talking about this day after Palm Two, You suddenly look at the new Google page where they've got the language model, and it's just text, and where are the clicks? It was like when Google introduced AMP. You know, this is where, rather than looking at the New York Times page, you have this formatted thing with no New York Times kind of stuff there. Like, these search entities that aggregate are just intermediating more and more, and people are going to become used to just having synthesized input. So what does search look like? What does it look like when you're GPT-IV can write you an article about any news that's happening in a way that's customized to you and your context and everything like that. This is massively disruptive for media and information, and so they have to think, where am I in the future where, again, the worst way to think about the impact of this is they're really talented grads that occasionally get off their meds, and we push a button and we can get a thousand of them. Those grads include journalists, and you can have your own journalist army, your own writer army, your own coder army, your own designer army.

AI assessment note: “I think he is right to be worried.”

Q I'm really naive and, and, uh, basic in terms of my thinking, which is why I'm a venture capitalist. Uh, my question is, what do we need to do to get to that stage? When we look at the data needed from the individuals, the data, the data that GPTs need, how we make the models work most efficiently?

A What do we need? First, we don't have to have that data from individuals. So we had Galactica as a scientific language model, but now we have MedPOM-II that exceeds doctor levels. So that was a Google announcement yesterday. We have AIs that can understand Articles better, as good as doctors, shall we say now. So we can scale that, because why do you need one when you have a thousand? So we take the existing generalized knowledge and all the hypotheticals, and we bring that together into an integrated common system available to everyone, because the building blocks are nearly here for that. Then you can personalize it later, and again, there are regulations and things around that, to how your, again, how we treat our loved ones and other things like that. The first thing is, let's get all the knowledge in one place and make it organized and useful. And so I think we're at that point now where the language models have just hit that point, that we can organize all of the world's Alzheimer's knowledge, longevity knowledge, autism knowledge, MS knowledge, and you can just type, and it can say this is the source, this is what it looks like, these are some hypotheticals, this is what we know, what we know we don't know, what we think we might know, etc. And then it can learn about you and your queries, because this is the other thing about lots of the language model things we've seen…

AI assessment note: “First, we don't have to have that data from individuals.”

Q The joys of doing what I do is going on schedule, find a woman, and then we will kind of retain some form of normality of schedule. What is the future of Healthcare systems that, like, do you think with GPT models operating in this way?

A I think that you can change the nature of a doctor because a lot of this stuff is kind of very basic. I think, you know, you had Babylon Health and others trying that chatbot. It wasn't ready. Now you've got this. Everyone should have their own AIs looking out for their own health with that objective function, you know, and then the nature of a doctor becomes different in terms of they have more rich information about an individual while it being preserved in a private manner. I think what you have is you have things like processes and procedures improving, like, uh, wound care, for example, in the NHS. Um, if you are injured as an elderly person, and your wounds aren't treated properly, you're more likely to die by a factor of eight times. Being able to monitor those types of things with this information set means you're eight times less likely, and then you have far more efficiency around that. So the information density around healthcare improves, which means that then our own healthcare improves. We all have access to as much knowledge as we want to within our own context, and so do our providers and the people that help us.

AI assessment note: “I think that you can change the nature of a doctor because a lot of”

Q What did you mean? What did you mean? Yeah, exactly. What did you mean by, like, the biggest bubble ever and the biggest shit show?

A Oh, I mean, like, the dot-com bubble, we've seen all these bubbles happen, you know, you had hundreds of billions into web three, and then developers got paid millions. Already the, uh, there's certain Chinese companies paying 1.2 million dollar salaries for PhDs. Uh, it's already getting a bit insane. They're in remembrance of that. The amount of money relative to the amount of opportunity within the sector is just completely misaligned. Like, my TAM analysis is that A thousand companies spend ten million in the next year, a hundred companies spend a 110 companies spend a billion. Like PwC just announcing they'll spend a billion over the next three years. And that's a accountancy firm, you know? Where's that gonna go? They don't know, nobody knows. And so a multiple of that will be allocated to this as the only growth theme in the entire market against the backdrop of rising rates, real estate crashing, etc. So the amount of capacity versus the amount and whale and wall of money Into something that's growing faster than anything we've ever seen is completely mismatched. And what will that cause? Like already you're seeing GitHub stars leading to a hundred million dollar funding rounds with zero traction and zero business model. Like stability, we actually have a business model and it's a good business model because I designed it. Um, but other things like money will go everywh…

AI assessment note: “amount of money relative to the amount of opportunity within the sector is just completely misaligned”

Q If you were an investor today, how, if you and me came, what would you do? I'm an early stage investor, I invest globally. What would you, what would you do?

A I would, again, I think it comes down to, there's going to be this tailwind of beta. And then you have an alpha play on top of that, right? So the beta play is that you just invest in any good founder. And if you get in, you figure out what can I offer as kind of a value add there? Am I offering distribution? Am I offering people? Am I offering this? And you emphasize kind of your value set. I think right now what people need is people. There are a few people that are like coming out here, but then what you see is you see good companies with like good ideas, but not businesses. They're building surface level things, these wrapper layers and others, and they're not thinking about distribution and data. It's like, if you want to have distribution, what do we do? We went to Amazon and said Bedrock, because then it gives us a 100,000 SageMaker SMEs, and we just have to give them the models that they can then take to the private data, and we get a share of all of that. This is how we saw it, like, rather than being responsible for that. So if you can bring that distribution to that, it was important. This is part of that Google memo that went out. We don't have an edge, neither does OpenAI. OpenAI used Microsoft for distribution and that flywheel. You know, if you have a business that's focused on innovation at the core, that's not actually a business. It becomes a business when tha…

AI assessment note: “So the beta play is that you just invest in any good founder.”

Q bluntly seeing this in action in society, I'm sure it's very aware of technology cycles taking so much longer than one anticipates. How do you think about that in actual, there's kind of twofold. One is adoption on enterprise and another is adoption on consumer. Say if we do the adoption on the consumer side, which is impacting freelancer jobs and impacting education, what do you think that looks like?

A So I think on the consumer side, um, You're freer with your information, so you can use a lot of these things, and the APIs of OpenAI and Cohere and others are fantastic, right, and Google Palm now kind of being out there. So it will be integrated to deliver better consumer experiences without it being creepy, like you've seen with some of the chat bots, etc. Because it's going into Word, it's going into workspace, you know, like, it helps already. Like, we will have a conversation that will be automatically logged by our Pixel phone, and then we'll get a transcript and remove bits that we don't want to share, that it goes into a global knowledge base that reminds us of things. That's inevitable. On enterprise it takes longer because you need to have auditable standardized models. If you're a financial services institute, you can't have a single piece of crawl data in there. And so that's what we're deliberately building with the largest companies in the world because we're building dedicated teams.

AI assessment note: “So I think on the consumer side, um, You're freer with your information”

Q What does that look like? Sorry, AI plus humans.

A AI plus humans means that you have information coming in and then the stories are, drafts are automatically written, reviewed by humans who then give their input to train it better. This is kind of the feedback flow, and then what happens is it comes out, and there's a factual anchors, and then it gets customized to Alabama, and then Alabama context, and all sorts of other things, because you can tell it, TLDR, too lazy didn't read, explain it like I'm five, make it more complex, and so you're going to see something very interesting here, which is the right news at the right time. The localization will return, but again, through AI first. I think this is the thing, we're seeing AI integrated, But the next wave is going to be, once we understand design patterns, AI first, everything and information flows, once these technologies are a bit more mature.

AI assessment note: “drafts are automatically written, reviewed by humans who then give their input”

Q What does that proper partnership mean, and who's that with? That's with IBM, that's with SAP, that's with Apple.

A So, we've announced Amazon, let's say we have lots of other announcements, the biggest companies in the world, where they have amazing teams, but they can only move so fast. And I'm building dedicated teams to help them move and understand the whole sector, Without trying to, like, sell them on services. I'm trying to say, I will build you a customized model if you want, but I'm only doing that with a dozen companies, you know, so I can kind of focus down. And I will tell you that GPT-IV is great, or Cohere is great, or all this stuff. All the latest research through the communities we support, I will make sure you're on top of, rather than into your sector, and you've got dedicated people helping you in this transition period.

AI assessment note: “we've announced Amazon... I'm building dedicated teams to help them move and understand”

Q And one thing I heard you talk about before, which I thought was fascinating, was your access to super compute, and you compared it to existing large incumbents. Why do you have more super compute than other people?

A Because I went and I did it. So we had articles coming out saying about our burn. I'm like, I have oil wells when everyone wants to build petrochemicals. Every day we have companies coming to us asking us for our super compute because it's not available on the market. You need these chips lined up with interconnect, and we've got 7000 A-one hundreds now. You know, we have TPUs, we have all these things, and we know how to use them, and we can share them with people because we're open. You know, whereas Anthropic and others cannot. So this is like, at the worst case, I'll build a foundation model as a service company, and I'll make a hundred million dollars in profit this year without having to charge even market rates, and I can retire, but I wouldn't do that. I wouldn't bring this to the world. So I think compute is misunderstood. It's not like Bird and, you know, all these scooter companies and others, they spent money on marketing. This is actually an asset right now that's scarce. And so there's no harm in scaling compute, and then with the top chip manufacturers, they're building us dedicated teams, and again, they're coming in and supporting us because our models drive demand for their chips. The more open models there are, the more open demand is, so it's a virtuous circle there as well, and so we get compute before everyone else.

AI assessment note: “they're coming in and supporting us because our models drive demand for their chips”

Q What do you make of, like, the open AI competitors? I've seen, Quite a few which are OpenAI for Europe. Uh, and we've seen three or four now. Like, is this a zero sum game and OpenAI's won that race, so to speak?

A I think it'd be difficult to compete against them because they're executing incredibly well. And I think, again, why would you use OpenAI for Europe versus Palm II? Versus GPT-IV? What can you bring? But you will have national champions and others. I think it's incredibly difficult to compete in proprietary. Um, I think in open, it's a bit different because the standardization element there, but again, my play is to be the benchmark across every modality, because there's no other company apart from me in open AI that does every modality. There's no company that's as aggressive as me in emerging markets. Um, and so they have to say, what is my edge? Because you can have an edge, like you can be the open AI for government or defense or for healthcare, and really get in and understand those, and then you can be sticky, you know?

AI assessment note: “I think it'd be difficult to compete against them because they're executing incredibly well.”

Q This is why I was surprised when I saw you sign the petition of Elon in terms of pausing for six months. Can you just unpack why you did that?

A Well, I mean, for six months, you're not getting H-one hundreds and TPU V-Fives anyway. So it's a natural pause, but then also because the shit show is coming next year. So I said, we have to self-regulate. Like for example, um, the adversaries already have GPT-IV. Why? Because you can just download it on a USB stick. You know, you don't have to train your own when you can just steal it. Let's have better OPSEC. Let's have better standards around data. Let's stop and move off web scrapes by next year. We had hundreds of millions of images opt out of stable diffusion because we were the only company in the world To offer opt out of data sets. You know, like, let's bring in some standards around this. Before, it's everywhere. Basically where we are now, you remember COVID, your mum is talking about this, and your aunt, and everyone's talking about generative AI, and they're asking you, Harry, what's going on? You know, but you haven't had the Tom Hanks moment yet. Because everyone was talking about COVID before Tom Hanks got it, and then when Tom Hanks got it, that's when global policy changed. Because if Tom Hanks can get it, anyone can get it. What is that moment for generative AI?

AI assessment note: “So it's a natural pause, but then also because the shit show is coming”

Q Two more that I have to ask, then we'll do a quick fire. When you look at the incumbents at your Microsoft, your Apple, your Amazon, your Google, uh, Who has been the worst? You said Google were actually incredibly impressive. Apple, Amazon, are they well placed?

A Well, Apple's a black box, right? So we'll see at WWDC, uh, next month, in a few weeks. And so they could surprise us all, but let's face it, Siri's crap, you know? Um, but they have all the ingredients in place, the identity architecture, the secure enclave, other things, neural engine, uh, stable diffusion was the first model ever optimized on the neural engine, et cetera. But let's see that one. Um, Amazon, again, Amazon have moved faster than I think they've moved before. Amazon's interesting because they're an engineering organization. So they have self-driving cars. They have satellite internet, and all these kind of things. Because once they've got it, and they can take it from research to engineering, it's there. One of the struggles they've had is that it's not moved from the research side yet. You're still evolving on research. They're like, what do we do now? But they are inclusive, like Jeff Bezos said, for his first hundred billion in revenue, he envisioned half of it being proprietary, and half of it being marketplace, and they're having the same approach with Bedrock and things. Microsoft had a winning bet, Sati did amazing with the OpenAI thing. It's been mutually beneficial, even if there are clashes there, right? And Google's kind of something that's moving slowly. Meta, I think, is the dark horse. I think Mark's probably pissed off that OpenAI bought AI.com, …

AI assessment note: “Siri's crap, you know? Um, but they have all the ingredients in place”

Q Two more that I have to ask, then we'll do a quick fire. When you look at the incumbents at your Microsoft, your Apple, your Amazon, your Google, uh, Who has been the worst? You said Google were actually incredibly impressive. Apple, Amazon, are they well placed?

A Well, Apple's a black box, right? So we'll see at WWDC, uh, next month, in a few weeks. And so they could surprise us all, but let's face it, Siri's crap, you know? Um, but they have all the ingredients in place, the identity architecture, the secure enclave, other things, neural engine, uh, stable diffusion was the first model ever optimized on the neural engine, et cetera. But let's see that one. Um, Amazon, again, Amazon have moved faster than I think they've moved before. Amazon's interesting because they're an engineering organization. So they have self-driving cars. They have satellite internet, and all these kind of things. Because once they've got it, and they can take it from research to engineering, it's there. One of the struggles they've had is that it's not moved from the research side yet. You're still evolving on research. They're like, what do we do now? But they are inclusive, like Jeff Bezos said, for his first hundred billion in revenue, he envisioned half of it being proprietary, and half of it being marketplace, and they're having the same approach with Bedrock and things. Microsoft had a winning bet, Sati did amazing with the OpenAI thing. It's been mutually beneficial, even if there are clashes there, right? And Google's kind of something that's moving slowly. Meta, I think, is the dark horse. I think Mark's probably pissed off that OpenAI bought AI.com, …

AI assessment note: “Apple's a black box... Siri's crap, you know? Um, but they have all the ingredients”

Q If you were an investor today, how, if you and me came, what would you do? I'm an early stage investor, I invest globally. What would you, what would you do?

A I would, again, I think it comes down to, there's going to be this tailwind of beta. And then you have an alpha play on top of that, right? So the beta play is that you just invest in any good founder. And if you get in, you figure out what can I offer as kind of a value add there? Am I offering distribution? Am I offering people? Am I offering this? And you emphasize kind of your value set. I think right now what people need is people. There are a few people that are like coming out here, but then what you see is you see good companies with like good ideas, but not businesses. They're building surface level things, these wrapper layers and others, and they're not thinking about distribution and data. It's like, if you want to have distribution, what do we do? We went to Amazon and said Bedrock, because then it gives us a 100,000 SageMaker SMEs, and we just have to give them the models that they can then take to the private data, and we get a share of all of that. This is how we saw it, like, rather than being responsible for that. So if you can bring that distribution to that, it was important. This is part of that Google memo that went out. We don't have an edge, neither does OpenAI. OpenAI used Microsoft for distribution and that flywheel. You know, if you have a business that's focused on innovation at the core, that's not actually a business. It becomes a business when tha…

AI assessment note: “the beta play is that you just invest in any good founder”

Q How did you think about distribution? You know, you've seen, you know, HuggingFace has partnered with, um, Amazon. You've seen, obviously, OpenAI with Microsoft. When you think about distribution and your competitive edge there, where did you land?

A So my business model is actually very simple. I haven't really talked about it much. Um, stimulate open. We're one of the biggest providers of, uh, grants to open source software, tens of millions already, and then take the best of open, which hopefully we build ourselves, and then an open base with an open data, and then commercial variants with licensed data, and then national variants. So you have Hindi insurance adjusted stable chat. Or Indonesian pharmaceutical worker stable chat that's available in every cloud on-prem on device with licensing fees, royalties, and revenue share. And the system integrators work with us as well. Lots of announcement to come. And so by standardizing and stabilizing all the complexity to these very sophisticated building blocks, these very intentionally built models, that really helps the world integrate this stuff by building playbooks and other things. That's the core business, because it doesn't require actual innovation. We are still innovative in the leaders and media in particular. Instead, it requires data and distribution. Data to the models, the models are open and interpretable, and models to the data via our partners. And that's valuable, because the private data in the world is far more valuable than the data that you will send to proprietary models. And it's not a race to the bottom either. So, that's what we are. We're a modeling…

AI assessment note: “models to the data via our partners.”

Q How do we think about open source versus closed source human healthcare data? Cause like obviously for us all to benefit as one, you know, MS sufferers around the world need to submit that data around, you know, responses to certain treatments.

A Yeah. So I think the wonderful thing about these models is they're few shot learners, so they don't need to have much information. And so isn't the classical big data problem. HDR UK has been one of the pioneers here with the UK Biobank, Federated Learning and others. And there are kind of, um, with FL seven, HLR and other standards being built around this to allow for full federated learning. If you have open source, Language models that are fully auditable, understand, I call them organic free-range models, the ones we're building with no web scrape data. Those can sit on device, like Google yesterday announced, um, POM-II. The smallest POM-II model is four hundred million parameters. It works on your Google Pixel phone. You don't need giant models anymore, and then that model can just share the specific information that preserves your privacy with the bigger thing, and then it can take from that global knowledge base as well. So you'll have big global models, on-device models, And I think open works for that, because you don't need to have all the data open. You just need to know that Harry is old enough to have a drink, not that all the details about Harry and his birthplace and everything like that.

AI assessment note: “open works for that, because you don't need to have all the data open”

Q The deep mind kind of desegregation or kind of unification was supposed to, of course, a lot of friction and be a negative press reported. Do you disagree with that?

A Of course. It is a lot of kind of replicated jobs. There was kind of brain and mind, and now they're kind of brought together. And it's a very different management style and other things. These things are never easy. But this is why, like, you saw Palm, five hundred and forty billion parameters, and that you had DeepMind with sixty seven billion parameter chinchilla, chinchilla, which is like, just train more as opposed to more parameters. You look at Palm two as a combination of both. And so it's trained for far more on far better data. And then that means it's only a fraction of the size, like fourteen billion parameters is one of the test comparative models versus the 540 and 67. So you can start to see this fusion of ideas, even if the teams, you cannot Integrate two big teams like that instantly.

AI assessment note: “These things are never easy. But this is why, like, you saw Palm”

Q What did you mean? What did you mean? Yeah, exactly. What did you mean by, like, the biggest bubble ever and the biggest shit show?

A Oh, I mean, like, the dot-com bubble, we've seen all these bubbles happen, you know, you had hundreds of billions into web three, and then developers got paid millions. Already the, uh, there's certain Chinese companies paying 1.2 million dollar salaries for PhDs. Uh, it's already getting a bit insane. They're in remembrance of that. The amount of money relative to the amount of opportunity within the sector is just completely misaligned. Like, my TAM analysis is that A thousand companies spend ten million in the next year, a hundred companies spend a 110 companies spend a billion. Like PwC just announcing they'll spend a billion over the next three years. And that's a accountancy firm, you know? Where's that gonna go? They don't know, nobody knows. And so a multiple of that will be allocated to this as the only growth theme in the entire market against the backdrop of rising rates, real estate crashing, etc. So the amount of capacity versus the amount and whale and wall of money Into something that's growing faster than anything we've ever seen is completely mismatched. And what will that cause? Like already you're seeing GitHub stars leading to a hundred million dollar funding rounds with zero traction and zero business model. Like stability, we actually have a business model and it's a good business model because I designed it. Um, but other things like money will go everywh…

AI assessment note: “amount of money relative to the amount of opportunity within the sector is just completely misaligned”

Q bluntly seeing this in action in society, I'm sure it's very aware of technology cycles taking so much longer than one anticipates. How do you think about that in actual, there's kind of twofold. One is adoption on enterprise and another is adoption on consumer. Say if we do the adoption on the consumer side, which is impacting freelancer jobs and impacting education, what do you think that looks like?

A So I think on the consumer side, um, You're freer with your information, so you can use a lot of these things, and the APIs of OpenAI and Cohere and others are fantastic, right, and Google Palm now kind of being out there. So it will be integrated to deliver better consumer experiences without it being creepy, like you've seen with some of the chat bots, etc. Because it's going into Word, it's going into workspace, you know, like, it helps already. Like, we will have a conversation that will be automatically logged by our Pixel phone, and then we'll get a transcript and remove bits that we don't want to share, that it goes into a global knowledge base that reminds us of things. That's inevitable. On enterprise it takes longer because you need to have auditable standardized models. If you're a financial services institute, you can't have a single piece of crawl data in there. And so that's what we're deliberately building with the largest companies in the world because we're building dedicated teams.

AI assessment note: “we will have a conversation that will be automatically logged by our Pixel phone”

Q What do you make of, like, the open AI competitors? I've seen, Quite a few which are OpenAI for Europe. Uh, and we've seen three or four now. Like, is this a zero sum game and OpenAI's won that race, so to speak?

A I think it'd be difficult to compete against them because they're executing incredibly well. And I think, again, why would you use OpenAI for Europe versus Palm II? Versus GPT-IV? What can you bring? But you will have national champions and others. I think it's incredibly difficult to compete in proprietary. Um, I think in open, it's a bit different because the standardization element there, but again, my play is to be the benchmark across every modality, because there's no other company apart from me in open AI that does every modality. There's no company that's as aggressive as me in emerging markets. Um, and so they have to say, what is my edge? Because you can have an edge, like you can be the open AI for government or defense or for healthcare, and really get in and understand those, and then you can be sticky, you know?

AI assessment note: “I think it'd be difficult to compete against them because they're executing incredibly well.”