AI Labs: Google DeepMind plans its comeback

0:00Before we start this episode of Tectonic, we want to know what Tectonic listeners think of the show, so we've put together a short survey. It's your chance to give us feedback and help shape future episodes. And doing the survey will also give you the chance to win a high-end pair of Bose headphones. There's a link in the show notes of this episode, plus all the T's and C's. Okay, on with the show.

0:27Back in 2017, researchers at Google came up with a new way of building AI systems called transformers. They allowed AI to process language better than ever before. It was a breakthrough that would dramatically change the AI landscape. But at the time, Demis Hesibis, the head of Google's AI lab DeepMind, didn't think transformers were the future of AI. DeepMind hadn't bet on the LLM approach. You know, Demis didn't really believe that you could get to AGI by, you know, ingesting the internet and figuring out patterns in language. He just didn't believe that that represented human intelligence. Madhami Tamurgia is the FT's AI editor.

1:10It turned out, actually, you know, LLMs were a much better way of sort of mimicking intelligence than he had previously expected. In 2022, a startup called OpenAI released ChatGPT, a chatbot using transformer technology, and sparked a new AI boom that left Google, DeepMind and Demis Hesibis playing catch-up. I heard him talking about this publicly recently. Somebody asked him, you know, how did you feel? It's been your life's work to develop AGI and to kind of lead in this area. And yet OpenAI, you know, this little lab came and kind of ate your lunch and they put chat GPT out there. How do you feel about it?

1:51And he said, you know, for me, that was war. They came and parked their tanks on my lawn and I'm going to war.

2:02This is Tectonic from the Financial Times. I'm Murad Ahmed, the FT's technology news editor. A handful of Silicon Valley companies are vying to lead the world in artificial intelligence. In the last few years, startups like OpenAI and Anthropic have surged ahead with sophisticated chatbots and advanced coding models. But now there's a feeling in the valley that Google is back in a big way. It has the money, the scale and the talent. So is Google destined to win the AI race?

2:49To find out if Google is really on the AI combat trail, I spoke to Madame Ita Mergier, who's followed the career of Demis Hesibis for the best part of a decade, and to Stephen Morris, the FT's bureau chief in San Francisco, who covers, amongst other things, Google. Okay, guys, so I want the bull and the bear case for why Google and DeepMind are going to win the AI race. Stephen, give me the bull case. Well, when I first arrived here in San Francisco a couple of years ago, the narrative around Google and AI was very negative. You know, there was widespread allegations they'd fumbled their early lead in this technology, allowed two now trillion, potential trillion dollar startups heading for IPO to eat their own lunch after taking their staff and technology they'd incubated.

3:39But over the last, I would say, 12 to 18 months, that has turned around and Google really has started to show its strengths and started to back up their rhetoric that they have been building towards this AI moment for the last decade plus. What we are seeing is both a resurgence in their actual technical capabilities powered by their deep mine research lab. Gemini, which is the name of their chatbot and of their larger family of models, is starting to be adopted by both consumers and pushed to enterprises more aggressively. But more than that, Google has this whole ecosystem around the actual models, which they say gives them both a cost advantage and a technology advantage.

4:20I'm thinking here of their giant cloud business. I'm talking about their custom TPUs, which are specialized AI chips. And then you have this vast reservoir of money that they are sitting on, which they're able to pump and recycle into AI. I use that in the widest possible sense for attracting and retaining talent, for building more data centers, powering research. So we are really seeing Google completely change the narrative and like really underline their credentials. And I think every aspect of AI here. Now that doesn't mean we shouldn't criticize them for being a bit slow to start and allowing Anthropic and OpenAI to grow into these massive competitors right under their noses.

5:02But if I'm going to gamble like long-term, who's still going to be there at the frontier monetizing AI, making the best technology, my money's probably going to be on Google. Okay, this is a rhetorical exercise, Madhu, but I want you to engage in it. What's the bear case here? No, I think it's a legitimate question, right? Because as Stephen has laid out, there is no reason why Google isn't winning right now. They had everything in place, not just now with their cash reserves and the stack and the chips and the brains and the talent and Demis Hassabis, you know, their secret weapon. The fact that they also invented the transformer back in 2017, which is the technology at the heart of this entire revolution that we're seeing, these large language models.

5:47But they've lost every single person of the seven co-authors who wrote that paper to other companies, including OpenAI and elsewhere. And now we're at a point where there's an entirely new market that's been created by these upstart competitors. You've got Anthropic and OpenAI who've opened up entirely new applications. And I think the bare case there is, you know, so far they've shown that they've sort of incubated talent and then lost all of that talent to build these new companies. Right. And these companies like OpenAI and Anthropic, as well as taking AI talent from Google, they're able to be more nimble in a way that's harder for a tech giant like Google to do.

6:26But what Google also has in its favor is that its rivals are, in addition to being nimble and able to pivot, they are also unique and often bonkers organizations. For example, the CEO of Anthropic picking a fight with Donald Trump and Defense Secretary Pete Hegseth over the appropriate military use of their technology, resulting in them being thrown into a tailspin after essentially being banned from all US government servers. And good for their reputation in some ways. Good for their reputation in some ways, in particular among very worthy, very moral AI researchers. Bad for their income. And of course, you never really want to make an enemy of Donald Trump in this current environment.

7:06Equally, you have OpenAI CEO Sam Altman and President Greg Brockman being absolutely massacred in court in a lawsuit with Elon Musk at the moment, accusing them of essentially stealing the assets and the IP of a charity, which could derail their for-profit conversion. Remember, they were initially founded as a nonprofit to develop AGI for the benefit of all humanity. They're now an$850 billion company with lots of outside investors, very keen to see them IPO this year. Google, compared to those two situations, is a relative oasis of calm. And I will say, before we go to DeepMind specifically, that, you know, there is something to be said about Google being the grown-up in the room.

7:47Stephen's pointed out some of the really interesting personality quirks around the leaders. they didn't want to found a company. They aren't necessarily sort of business operators. And now they're trying to shepherd these nearly trillion dollar companies and these huge powerful technologies and, you know, fielding calls from CEOs of banks and trying to fight the US president. It's like way beyond the ken of like probably what these guys as postdocs at universities were thinking would happen with their lives, right? Whereas, you know, Google has been through this for decades. It knows how to navigate the legal system.

8:22It knows how to deal with politicians moving through the White House over many, many administrations, you know, staying the course and sort of shepherding this technology into the economy in a way that the others, they aren't sort of making us feel very confident in their ability to do that. So let's talk about DeepMind specifically. This is the British AI startup bought by Google in 2014. now the research lab at the center of Google's AI ambitions and headed by Demis Hesibis. Maddy, you said Demis is Google's secret weapon. Why? Yeah, it's so interesting because I think here in the UK, we are so proud of the achievements of Demis.

9:03And I'm not sure it's reflected in the same way in Silicon Valley. But here, you know, his story is legend, really, because he was a, you know, a child chess genius, a video game designer and then went on to study computer science and then completely pivot and got PhD in the neuroscience of imagination like this is a polymath like the word was invented for him right he seemed like he could turn his mind to anything and master it and the one through line his sort of passion his mission has had always been AI but what's really interesting is like even sort of 10, 15 years ago, this was a fringe thing.

9:41And for him to even found a company and get investment from, you know, he's talked to Elon Musk about it. He eventually got some money from Founders Fund, which Peter Thiel runs. He had to really pitch for it and convince people. And the money he did get was people who were sort of doing counterintuitive bets because, you know, they wanted to be contrarian, not because they believed this was going to be big. So it's just to look at his story shows you a little bit about the trajectory of AI in such a short time, over a decade. And when he sold to Google, from Demis' perspective, he's said this many times, he felt that he needed the compute power in order to do his work in peace.

10:22He just didn't want to spend his brain energy pitching for money and thinking about how he was going to fund this work of building an artificial intelligence that was going to solve all humanity's problems. And Google felt like the answer to that. And then with Google's money and computing power, Demis and DeepMind had some big breakthroughs in AI, including one that won a Nobel Prize. Tell us about those. The first was called AlphaGo. And this was essentially an AI system that learned to play Go, which is this fiendishly complicated Chinese board game supposed to be, you know, exponentially harder than chess.

10:59And for a long time, it had been like a problem that was uncrackable by machines. And the crowning achievement of that was a machine which they kind of put to the test in Korea against the human Go champion called Lisa Dole. And I went out to Korea actually to watch this game. And it was kind of amazing to watch the human champion be defeated by this AI system for the first time. So that was the first breakthrough. And then the second one, which built on that, was AlphaFold, for which Demis and his team won a Nobel Prize for. And essentially, this was using the same techniques to sort of predict all the different ways in which proteins could fold.

11:38And that led to amazing breakthroughs in biology and biochemistry, allowing people to generate proteins much more quickly, proteins being the substrate of life. Stephen, this point about how Demis Hacibis graduated from a scientific lab to the front lines of the AI war. Talk us through how he's become so important within Google's corporate strategy. He's built up a lot of power around himself and is the key figure in their AI strategy. Well, I think if you look back, when was it Madhu? Spring 2024, maybe April or May, that was probably when I would identify their Nadia in particularly large language models and image generation.

12:20They had this series of embarrassing flops. And Demis, you know, according to our reporting and colleagues, took this very hard. And within a few months, he kind of resolved to take more control of the AI organization within Google. He demanded more power. There was a big restructuring. They merged their two rival research labs, Brain and DeepMind. And he kind of became, I guess, the second most important person in the company, maybe third, behind Sundar Pichai and then the two co-founders, Larry Page and Sergey Brin. And as Madhu has flagged out like child chess prodigy, computer game designer, everything you hear about Demis indicates an incredibly driven, incredibly competitive kind of guy.

13:01And like his initial focus was AI and science. He wanted to win the Nobel Prize. He wanted to solve unsolvable problems. But now I think he's kind of got the message, at least it seems, that he has to be competitive in terms of technology and earnings as well. He's out there on Twitter shilling for something they like to call Nano Banana 2, which generates short videos. I would categorize them personally as AI slop, but you can see that Google might be able to sell this technology into advertising, creating full custom video text advertising campaigns in the future. And Demis is certainly much more present as a booster of the commercial aspect of this technology than he has been before.

13:42And some of the changes he's made at DeepMind are trying to transform it from this more academically oriented lab into a more competitive one. For example, one story that we wrote about was the publication of its kind of cutting edge research. A lot of the scientists and researchers and engineers at DeepMind wanted to put their research out as soon as possible into the marketplace, upload them online, so that their fellow scientists and researchers could absorb it, praise them presumably, but then also build off the technology. Demis ended this. He said, no, there's going to be a six-month lag because we're essentially giving away some of our most novel research to free for our competition.

14:17And this was quite controversial internally because it seemed to many people who were kind of in AI for the more academic and research side, moving towards AGI for the benefit of humanity, as opposed to the benefit of Google's bottom line, seemed like a pretty clear shift. It's also dramatically different, isn't it, Stephen, to like five years ago, even when they evaluated their impact by their publications. I remember when I was covering it, you know, five, six years ago, the PR would be, we're in nature, we're in science. You know, they were so proud of their publications. And then that quietly went down.

14:51And the fact that that's changed so dramatically just goes to show what the sort of macro AI environment has done to kind of shape Demis' leadership and the company itself. Yeah, it's gone from like, he still tweets about, you know, these cutting edge like research that goes in Nature or the journal of this or that. But then next, it'll be followed by a tweet. Look at this picture of a Labrador and a party hat in a flying car that flights off the edge that Nano Banana has generated. Isn't this great? And you're like, do you really care about that? Or is this as part of your new expanded responsibilities as essentially an employee of Alphabet?

15:26that. Okay, let's take a quick break. And then I want to talk more about Demis Hasebis and his shift from scientist to tech CEO, and whether he might even be in line for the top job at Google.

15:44Right now, someone is asking AI a question about your industry. They're looking for the perfect recommendation or an accurate solution, and they're ready to act on it. When AI gives an answer, is your brand a part of it or not? Most companies have no idea which. Gist changes that. We show you where you're visible, where you're not, and give you the tools to do something about it. It's time to make AI work for your brand with Gist. Learn more at gist.ai.

16:17So Stephen, Google CEO Sundar Pichai is getting a lot of credit for helping to refocus the company on AI over the last 10 years. But given how important Demis Hesibis is to their AI efforts, could he one day be in charge of Google itself? So Dennis sits within this like interesting constellation of very capable, very competitive people. But I do think he can lay claim to being Google's secret weapon. Like Elon Musk is very fond of saying that Google didn't buy DeepMind. This is a long-term reverse takeover of Google by an AI research lab, in addition to being probably one of the cheapest tech deals of all time in retrospect.

16:58So I think we need to set where Demis, you know, sits within the Alphabet organization in its appropriate context. But I'm saying like, if you put Sundar Pichai or Sergey Brin on the spot and say, what's the one executive? What's the one piece of talent internally that you couldn't use? I think they'd be hard pressed to come up with another answer that isn't Dennis Osalbus. I mean, I guess the question is like, would, yes, you know, I think from our reporting, there is this leadership struggle and not necessarily the obvious sort of who's going to be the next CEO, which is also part of it, you know, in due course, someone will have to replace Sundar and is that going to be Dennis?

17:35But more broadly, the question of like, who's going to lead in AI? who's going to be the person that takes us over the line. And if we're going to buy into the AGI hype, who's going to be the person that gets there first, you know, who creates a super intelligent technology. And I think it's been clear, and I don't think Demis has wavered from this all through the time that I've covered AI, that his goal is to create that technology, the kind of technology that surpasses the most intelligent human that's able to solve problems that we would never be able to do without it. And for him, I think everything else is a means to that end.

18:11And so I guess the question is like, you know, if he believes that being the CEO of Google will help to get to that end, or whether it will be a distraction from that mission, you know, having to do quarterly calls and be answerable to your shareholders and think about, you know, your massive advertising machine. I think he could do that job. And, you know, he could win out in that power struggle but it's I think only useful to him if it feeds into this his goal that he's never lost sight of which is creating you know a very powerful AI and I think that makes him in that way very unique amongst leaders of AI technologies who are all motivated in slightly different ways to do slightly different things maybe he's closest I would say to Dario Amadai in that sense.

19:01And the pair of them have a bit of a bromance, don't they? Yeah. They're Miss Sindario. Yeah, it's coming out a bit more lately. They're like, look at what the companies that are run by like real AI researchers and scientists are doing. And then look at what the companies who are run by marketing guys or advertising people are doing. They said this on stage when they appeared together at Davos in January, very thinly veiled jibe at Sam Altman and Mark Zuckerberg there. And their point is, I guess potentially Elon Musk as well, their point was that we are the ones that can be trusted to deploy AI safely, to really deeply think about what an AGI future looks like from everything from the economy, employment, to safety and military applications.

19:47Whereas if you were just using AI to improve your advertising margins, or indeed line your own pockets, perhaps you can't be trusted as much for the technology. Yeah, and this is a strongly held belief by Demis. You know, he, when asked, you know, why did you do a PhD in neuroscience and go into video games and play chess and all of these things, he has said he needed his credentials to be unimpeachable when he came to running a company like DeepMind. And what he meant was that he wanted to surround himself with PhDs and for them to know that he probably was the smartest guy in the room and that he could talk to them at that level as well as running a company.

20:27I think we've given a lot of airtime to the Bull case. I am going to insist that we'd be good skeptical journalists and also really talk through the Bear case as well. I have a Bear case for you, Murad. Please, Stephen. You said Kremis' credentials were unimpeachable or impeccable. I think he said that he needed his credentials to be unimpeachable. You know, the culture, certainly DeepMind has always been very aligned to this don't be evil initial ethos. Now, now Google's obviously alphabets abandoned that and it's organized the world's information to make it useful or something, which is much more commercial strapline.

21:03There has been a bit of a spat, as we mentioned before, between Anthropic and the Pentagon about military use of technology, specifically autonomous lethal weapons, basically allowing an AI to decide whether to kill someone or groups of people or not, and also mass domestic surveillance. Anthropic felt strongly enough about these red lines that it tried to basically pull its models from use at the Pentagon, and it kicked off this huge spat, which saw them designated as a supply chain risk. A lot of staff at Google kind of backed Dario Amadei's position and said we should be doing the same thing and we should be asking for the same red lines.

21:37Obviously, Elon Musk and Sam Altman signed on the dotted line like literally the next day to try and get their AI systems into these potentially lucrative government contracts. And then last week, Google signed its own deal that has, you know, language which says it shouldn't be used for mass domestic, for surveillance and autonomous weapons, but no actual legal safeguard. The government wrote the contract in a way that it can just disregard that should it feel the need to. Hundreds of DeepMind staff and Google Cloud staff wrote to Sundar Pichai trying to stop this, and they didn't. Demis has been silent on the issue.

22:10He hasn't said anything, although he has been on the record before saying, I don't think either of these things should ever be used for AI, nor do I think AI technology is sufficiently controllable or advanced to be allowed to use for this. If you're looking for a bear case, if Alphabet continues down that path, it could potentially lose some of its staff and more idealistic rivals. And it does represent a bit of a sticky issue. Demis has doomer tendencies in his view of AI if deployed irresponsibly. And if you do allow the commercial imperative to override these concerns, that could be a potential hit to its lab who would favor more people along the lines of anthropic.

22:53I mean, what do you think, Maddy? Yeah, I mean, I was going to say you've hit on, I think, the, and I didn't expect you to be the one to bring that up. I thought you'd have a commercial bear case, but it's the soft thing. It's people, it's talent, which is a huge, huge commodity. And sort of, you know, there's only a limited pool of people who really are the best in the world at building AI and deploying it. And yes, we're going to train up more. But these companies, they're fighting over this limited pool of people. You can see that from the sort of salaries that they command. And a lot of these researchers, they are scientists.

23:27Many come from physics or chemistry. And they are driven by something more than a commercial imperative, just to get rich. I mean, we do know from the case this week that people like Greg Brockman might feel differently and are maybe driven by how can I get to my billion dollars. I want to give you a bit of background about that, Maddie. But this is OpenAI's president whose personal diary was read out? Yeah, in court in the case of Elon Musk versus OpenAI. And he had been musing about how he could hit a billion dollars. But I think like in contrast, many of the researchers who are sort of key to building AI are here because they want to invent something amazing that has a positive beneficial impact on society that solves huge unsolvable problems.

24:10and having spoken to some of them this past week, there is this real sense of disillusionment. And it's funny because like when you look at it from the outside as a journalist following it, you wonder why didn't they know this already? You can see how AI has been accelerating in terms of its rollout and how it's being used in military and other contexts. So why now? But one person I spoke to said, you know, for the first time they could see directly how something that they contributed to building could result in real loss of innocent lives. And there was nothing they could do to stop it. And the way this person described it is there's a feeling, a widespread feeling amongst this talent of just helplessness and lack of agency to the point where people are just quitting and saying, what is the point of anything really?

24:56And one person, they said, whose Google Cloud technology had been used in Gaza has, you know, been on long-term leave, just knowing how their technology had been used in that context. So yeah, I think if the talent itself stops believing in the sort of mission here or sees ways in which this technology isn't being shepherded properly, that might be what sort of causes the downfall of the ambition of companies like Google. And this is part of the problem of being a big tech company. You go from a startup that was founded in a garage to becoming one of the biggest companies in the world. I've known people who've worked at Google and they get stifled they become cogs in a massive machine that's just an inevitability of being a big tech company that the likes of an open AI an anthropic is going to come along and be disruptive that's just the history of innovation right exactly and I think I remember talking to Google when I wrote my piece on the transformers and the co-authors of the transformers paper and the fact they'd all left and I didn't expect them to respond because, you know, it feels like they would be embarrassed maybe that they'd lost all this talent.

26:07But they sent me a response saying they were, you know, justifiably really proud. They were really proud that they brought together these people and like made this amazing sort of invention and that they'd then gone out and are building things with that and that they felt that they'd seeded this entirely new industry. And so I think, you know, I'm going back to being bullish again. Yes, there's this dilemma and they will probably lose talent for various reasons. They can't please everybody. And there's definitely going to be competitors that come out and sort of challenge Google. But I think, you know, they play a really important role in the ecosystem in that they have fostered research and innovation and kind of allowed this ecosystem to develop around them.

26:51I respect that. And I think it's going to result in overall probably better technology. Okay, so I think we've given a good airing to the bull and bear case, but now I'm going to ask you to put your colors to the mass. Google, are they going to win the AI race or where do you think they'll finish up in that particular race? Stephen? I think Google is in an incredibly strong position, both from a commercial standpoint, you know, it only takes one or two slips for open AI or Anthropic to fall behind a bad model, pointless fight with the government, a damaging court case. Whereas Google has about 10 safety nets below it.

27:29It can just afford to invest in new things, speculative products, fail in a few, and still keep coming back. Relative stability in leadership, the superpower in having the great AI mind, I think it's fair to say, and Demis Sasavis. I just think it sets them up to, if not win at all, to certainly be one of the very top players, not just in one aspect of this, but across the board. Maddy? I'm not going to disagree for the sake of it. If they can stay the course and ride out inevitably bumps around losing people due to political reasons or because they want to have a different kind of culture, all of that will happen.

28:07But if they can stay steady and like tune out the noise, I think they're in a position to lead. I think they need to figure out what their AI technology is for, because I feel they've not been very imaginative with that. Like, I guess it's an enterprise technology, but people only use it if it's sort of already on their systems or whatever. I don't think anybody's seeking it out in the way that people are seeking out clawed code, you know, because they know that it's amazing for a certain task. So I think they need to figure out like, what is their competitive edge? What is the product they're trying to sell and not rely on the fact they have a huge installed base of stuff.

28:43But those are things that they're set up to do. You know, they have the smartest people in the world there. And as Stephen said, they have a long rope to play with. So I bet on Google. And I think I would be, you know, not upset if they were the ones shepherding this technology, even from a trust and safety perspective. But Madhu, as somebody who's an expert in this, what happens if someone else gets to AGI first? I don't know. Am I allowed to say we're all screwed.

29:14That was Madhameet Mergier, the FT's AI editor, and Stephen Morris, the FT's San Francisco bureau chief, speaking to me about Demis Hesibis and Google DeepMind, thanks to them both. Next week, we'll be looking at the company that sparked the AI boom. Investors think OpenAI is worth nearly a trillion dollars. Others think it's in a crisis. The question is, is Sam Altman its greatest strength or its greatest liability?

29:51Tectonic was hosted by me, Murad Ahmed. The producer was Edwin Lane. Sound design by Breen Turner and Sam Giovinco. Original music by Metaphor Music. The FT's global head of audio is Cheryl Bromley.

30:37We'll see you next time. stand behind. Avalara, agentic AI tax and compliance with confidence. Right now, someone is asking AI a question about your industry. They're looking for the perfect recommendation or an accurate solution, and they're ready to act on it. When AI gives an answer, is your brand a part of it or not? Most companies have no idea which. GIST changes that. We show you where you're visible, where you're not and give you the tools to do something about it. It's time to make AI work for your brand with GIST. Learn more at gist.ai.