Andrew Ross Sorkin (3:24)

For the last year and a half, there's been this battle playing out between OpenAI and Anthropic and you. And I think there was a view that you thought you needed to catch up. I think behind the scenes you, you thought you were going to get there. But something happened this year. What do you think it was? Because I think at this moment your model may very well be at the top of the, the charts, if you will.

Demis Hassabis (3:45)

I think it's. Look, it's been a, it's been a, it's been a kind of long journey last couple of years to kind of corral together all of the assets that, that we have as Google and DeepMind. Incredible research bench, our TPUs and all of the kind of research we've been doing over the last decade plus really that underpins a lot of the industry. So we always had all the ingredients, amazing product surfaces to plug AI into. And I think in the last year what people are feeling is we've kind of organize that all in a really efficient, efficient way.

Andrew Ross Sorkin (4:14)

Okay, so where, where are you in the journey though, in terms of what comes next about sort of productizing all of this, really getting it into all of these different services? When am I going to see it in Gmail, for example?

Demis Hassabis (4:27)

Yeah. So look, everything starts first of all with the research and the model quality. So as you mentioned, we're very happy with how our latest Gemini model, Gemini 3 is working. We think it's, you know, it's topping most of the leaderboards on most of the benchmarks and, and the other thing we worked hard on in the last year is accelerating the infrastructure and actually doing a lot of rewriting of the infrastructure to get the model quality as quickly as possible into our product services. So you're starting to see that in. We're simultaneously shipping in search, obviously in the Gemini app. And then what you'll see this year is appearing in more and more places across all of our product services, especially things like Gmail.

Andrew Ross Sorkin (5:02)

So Apple has now chosen to work with you. How transformational is that for Gemini and for your. What your work is?

Demis Hassabis (5:08)

Well, I think it's a massive sort of vote of confidence in the quality of our model. So we're very pleased with that partnership. Obviously it's an amazing partnership for us. Very important one. Apple ran a very rigorous evaluation process and you know, Gemini came top of that. And I think that's, that's testament to the work that we've.

Andrew Ross Sorkin (5:26)

Were you surprised that they decided they wanted to work with you as opposed to, did you think that they would ultimately try to create their own model?

Demis Hassabis (5:33)

Well, you know, creating models costs a huge amount of resources and money and also a certain type of research base in terms of like the research teams and so on. So I think it makes sense now. These models are mature. There's a kind of ferocious battle going on on the frontier to make, if you're a product company, to make use of those models and then downstream do amazing things with it.

Andrew Ross Sorkin (5:58)

You used to use the word interesting word, mature, that the models are mature. When people talk about getting to AGI and sort of how quickly all of this can scale, there has been, there's sort of two views. One is if you just throw more compute processing power at this, we can get there. There's another view that there needs to be some kind of really breakthrough scientific research shift. Which is it?

Demis Hassabis (6:24)

Well, actually I have a kind of in between view which is that I think it's an empirical question. So I feel like we're still getting lots of amazing gains out pushing the existing paradigms. These LLMs, these foundation models like Gemini, they're getting better and better with each iteration and we see no end to that. But on the other hand, to get to full AGI, there's some missing capabilities still. They don't do continual learning, they can't, they don't have true creativity yet. They don't do long term planning and reasoning. So it's a question mark about are some, a couple more, maybe a handful, two or three new big breakthroughs needed in order to get to all the way to AGI or will scaling up the existing techniques be enough? And we're doing both. So we're in where we have the luxury of basically having the deepest and broadest research bench and we can push both to the maximum.

Andrew Ross Sorkin (7:10)

How much do you worry that all these large language models ultimately converge? Meaning they become very similar in terms of what they can all do and then what differentiates them?

Demis Hassabis (7:20)

Yeah, I don't think we're seeing that. I think if you look at the last year, models are actually getting good. Different models are getting good at different things. You know, maybe like claw first co. Specifically things like Gemini. Amazing for multimodal Understanding and you know, image, image generation, things like that. So actually I think you're seeing quite a lot of differentiation at the frontier and my, my expectation is that will actually increase the gaps in that and the differences between the models.

Andrew Ross Sorkin (7:44)

This year there's a big question about whether there's an AI bubble. You've heard that over and over again. Is there a bubble?

Demis Hassabis (7:50)

Well, I don't think it's a binary yes or no. I think some, the industry is big now and there's many parts to it. I think some parts of the, of the AI field may be in a bubble. I mean if you look at things like the new hot startups that raising billions of dollars in a seed round with, you know, no product or technology yet, that seems a little bit frothy to me and perhaps unsustainable. But on the other hand, you know, there are lots of amazing use cases going on in product speed use.

Andrew Ross Sorkin (8:18)

Clearly Alphabet and Google have remarkable financial backing, but there's a lot of even big independents and I'm thinking of anthropic, I'm thinking of OpenAI that have to continue to get new capital to really meet some of the spending that that's necessary to get to the next place. Do you worry about whether they'll be able to get there?

Demis Hassabis (8:42)

Well, I think, I mean I mostly worry about what we need to do and, and going back to your bubble question, I think my job is as head of Google deep mind is to make sure whatever happens, whether there is a bubble and it bursts or whether, you know, the trajectory continues of, you know, the bull case continues from here, we're in the, you know, a winning position and doing really well. And I think, I think because of our balance sheet and all of the existing amazing products we have that are natural fits for AI. You mentioned, you know, email and there's Chrome and there's, you know, search of course. So there's all these amazing products billions of people use every day that I think AI can enhance and you know, we're only scratching the surface, I think of what we can do then there's going to be a lot more I think amazing kind of product features this.

Andrew Ross Sorkin (9:22)

Year in terms of spending. People talk a lot about spending on data centers and the need for new chips. When you think about sort of what comes next, do you think we could have a deep seek like mine, deep seek like moment rather where there's some kind of major technological revolution where people say oh, maybe we actually don't need all this processing power?

Demis Hassabis (9:42)

Yeah, I think it's very unlikely. And I also think the deep seq moment was, was a little bit overblown in my opinion. Like they had to utilize some of the western models in order to kind of fine tune against and to train from. So it wasn't, you know, what they reported. This very small training number wasn't quite the full picture. It is possible that one of you know these, there could be a left field breakthrough that does increase the efficiency of things. Maybe things like self improvement where you could imagine there's some cycle which requires, requires less compute power. But for now, you know, we think that more compute, you need compute for lots of things for training, for serving and for exploring new ideas and but of course we're also, all of us are trying to make our models, our leading models as efficient as possible. So in fact of the Gemini 3 models, our flash model, which is sort of the workhorse model, may be the most important because we can deploy that everywhere.

Andrew Ross Sorkin (10:30)

There's also a big question about the depreciation schedule of these chips. Meaning do they really last for four years, do they last for seven years? And it's not so much do they last? Are they useful in the same way that you probably buy a new phone every year or two, you're going to want, people are going to want to buy and need the next cutting edge chip. And so are we laying down, you know, railroad tracks or are we laying down railroad tracks that need to be replaced every couple of years?

Demis Hassabis (10:55)

Yeah, well I think that's one of the advantage we have is we're full stack. So I think we're the only organization really that has a frontier lab and our own chips and TPUs and our cloud business. So we have a lot of sort of ways of utilizing any spare compute anywhere on our, on our, you know, on our, on our systems and data centers and maybe as compute gets older, the older generations you start moving them towards serving or maybe like labeling data for you. So you can always utilize, you know, even quite older sets of generations of chips for useful, useful work.

Andrew Ross Sorkin (11:28)

We've mentioned OpenAI, we've mentioned Anthropic, we have not mentioned Metta and we have not mentioned Elon, Musk and xi. Where do they sit in this sort of competitive stack for you right now?

Demis Hassabis (11:41)

Well, obviously, you know, they're extraordinary companies and they're led by very ambitious and you know, aggressive leaders and I think they're working on very interesting things and we'll see how what comes out of that. This year especially on matter, I guess they've rebooted their, their sort of research division and you know, we're all waiting to see what they're going to come up with next.

Andrew Ross Sorkin (12:00)

But do you think that ultimately there's going to be one or two or three model big models and that's how it's, I mean, if we're sitting here together five years from now, what does this look like to you?

Demis Hassabis (12:08)

I think there's two pictures from here. I think there could be room. It's such, you know, is going to be so transformative and I think it's going to create so much new opportunities. I think there is plenty of room, especially on the enterprise side for, you know, maybe two, three, four windows winners. But I think each year it's getting harder because the pace is ferocious. These are very capable companies and organizations and research groups that are pushing the frontier and we're all pushing that as hard and as fast as possible. You know, I work 100 hour weeks. I've been doing that for the last three years and I see no end to that. And I think all the leading labs are doing that. And so for anyone to catch up to the frontier, it's progressively a harder, harder problem.

Andrew Ross Sorkin (12:46)

What's the defensive moat around your business? And the reason I ask is I always thought that persistent memory, the idea that you would know me potentially better than somebody else would be the thing that would keep me using one model from switching to another. But I just recently learned that actually all of this is actually quite portable. So you could actually take whatever persistent memory OpenAI has about me and actually hand it to Gemini.

Demis Hassabis (13:11)

Yeah, so look, I think that.

Andrew Ross Sorkin (13:12)

And vice versa.

Demis Hassabis (13:13)

Vice versa, absolutely. I think the key is going to be the quality, quality of the models and the capabilities. I think people can feel that when there's a differential there and it's suddenly more useful for whatever use case they have. So I think that's going to continue. Then there's the product features. But also I do think memory and personalization is going to be somewhat sticky. And we launched our first version of personal intelligence a couple of weeks ago and of course we have a whole ecosystem that you're already using Google workspace and email and, and of course search that we can kind of try and plug into obviously with the user's permission if they want a more useful personalized assistant.

Andrew Ross Sorkin (13:51)

Experience a lot of hand wringing around here about jobs and what ultimately happens 10 years from now. Maybe it's five years from now. I don't know. If you think even when you look at employment numbers today that to the extent that there's any softness, it's a function of AI and productivity being more successful.

Demis Hassabis (14:09)

Yeah, I think it's early days. I don't think there's any real evidence yet of any change in the job market. Perhaps a little bit on the entry level, maybe jobs and internships. But I think in the next five years at least it's going to be more than made up for. With extraordinary new opportunities these tools are going to are going to deliver, especially for individual creators, you know, artists and creators and game designers, for example, that can create whole apps almost on their own. So that should, I think, if I was to advise the youth of today.

Andrew Ross Sorkin (14:38)

What do you tell the kids is.

Demis Hassabis (14:40)

Get unbelievably proficient with the new tools, immerse yourself it, become native with it, and then leapfrog, you know, whatever professional ladder you're trying to get onto, leapfrogging the incumbent people on that with these new skills which are going to change the workspace, the workplace. And I think there's going to be plenty of new opportunities anytime.

Andrew Ross Sorkin (14:58)

There's a lot of disruption demos. Thank you.

Demis Hassabis (15:00)

Thank you very much.

Andrew Ross Sorkin (15:00)

That's great. Hey, Fidelity. Can I get a second opinion on stocks in the Fidelity app?

Cameron Costa (15:12)

With Fidelity, it's easy to get an outside opinion from independent experts in a single score. And then when you're ready, trade US stocks and ETFs with no commissions.

Demis Hassabis (15:22)

That's right.

Cameron Costa (15:23)

I am always right.

Demis Hassabis (15:25)

Investing involves risk, including risk of loss.

Andrew Ross Sorkin (15:27)

Online US equity trades and ETFs and.

Demis Hassabis (15:28)

Retail fidelity account sell order assessment fee not included, some account types and securities excluded. Details of fidelity.com commissions Fidelity Brokerage Services.

Andrew Ross Sorkin (15:34)

LLC member NYSE, SIPC ABC's David Muir, the most trusted anchor in America. The most watched anchor in America. Thank you for making World News Tonight with David Muir, the number one newscast in America. Most trusted Most Watched David Muir on.

Cameron Costa (15:51)

ABC not every sale happens at the register. Before AT&T business Wireless checking out customers on our mobile POS systems took too long. Basically a staring contest where everyone loses. It's crazy what people say during an awkward silence. Now transactions are done before the silence takes hold. That means I can focus on the task at hand and make an extra sail or two. Sometimes I do miss the bonding time.

Andrew Ross Sorkin (16:16)

Sometimes AT&T business Wireless connecting changes everything.

Cameron Costa (16:27)

Thank you for listening to this spot. Special Squawk pod reports from Davos. This is only one of many, many iconic interviews from the World Economic Forum this year. I promise it's worth following Squawkpod wherever you're listening. Now to catch each episode. Squawk Box is hosted by Joe Kernan, Becky Quick, and Andrew Ross Sorkin. Squawk Pod is produced by me, Cameron Costa and Zach Felici. Katie Kramer has been with the anchors in Davos the week. And stateside with us is Julie Tras, our editor. Have a great day.

Andrew Ross Sorkin (17:09)

Hey, Fidelity. How can I remember to invest every month?

Cameron Costa (17:14)

With the Fidelity app, you can choose a schedule and set up recurring investments in stocks and ETFs.

Andrew Ross Sorkin (17:21)

Oh, that sounds easier than I thought.

Cameron Costa (17:23)

You got this?

Andrew Ross Sorkin (17:24)

Yeah, I do. Now, where did I put my keys?

Cameron Costa (17:28)

You will find them where you left them.

Demis Hassabis (17:32)

Investing involves risk, including risk of loss. Fidelity Brokerage Services, LLC Member nyse, SIPC.