Google DeepMind C.E.O. Demis Hassabis on Living in an A.I. Future | EP 137

Large general models are necessary because they are required to train smaller specialized models through distillation; you cannot build efficient flash models without first training larger teacher models.

“you need the big models often to train the smaller models. So, uh we're very proud of our kind of flash models, which are the most, you know, we call them our workhorse models. Really efficient...But you can't build those kind of models without distilling um from the larger teacher models”

Google IO included shopping integration features where users can upload photos of themselves and use AI to virtually try on clothes, with Google extracting value through product advertising, payment cuts, and affiliate fees.

“A significant portion of the keynote was devoted to one executive talking about a new shopping experience inside of Google where you can take a picture of yourself, upload it, and then sort of virtually try things on and it will sort of use AI to understand your proportions...There was a lot of stuff in there that would just sort of let Google take a cut, right? Obviously, you can advertise the individual thing to buy. Maybe you're taking some sort of like cut of of the payment. There's an there's an affiliate fee that that is in there somewhere.”

Google DeepMind has engaged in rigorous internal testing and red-teaming of systems both pre- and post-launch, with feedback from safety institutes and trusted testers, balancing safety against the need for external critique and real-world feedback.

“we do a huge amount of work on our systems. Uh, I think we have very robust red teaming uh uh uh uh processes pro both pre and post launches...we have a lot of trusted tester type of programs...and um and then give us feedback and and stress test them uh including sometimes the the safety institutes as well”

The Vatican and Catholic Church have maintained a historical interest in science through the Pontifical Academy of Sciences (claiming Galileo as founder), and have been interested in AI ethics and philosophy for 10+ years.

“the Vatican's been interested even prior to this pope haven't spoken to him yet but on these these matters...the Catholic churches and I'm a member of the Pontipical Academy of Sciences...they've always had uh which is strange for a religious body a scientific arm you know which they like to always say Galileo was the founder”

Apple's WWDC 2024 announced numerous AI features that were never actually launched, demonstrating how even major tech companies announce capabilities they cannot deliver or choose not to release.

“Apple, uh, which demoed a bunch of AI features at its developer conference last year and then never launched half of them. Um, I have become like a little bit skeptical until I can like actually use the thing myself.”

Meta has gamed AI model leaderboards by creating 27 different versions of the same model to find the one that scores highest on specific benchmarks, inflating competitive positioning.

“Meta cheating to win at LM Arena, making 27 different versions of a model just to come up with one that would be good at one thing”

According to Sergey Brin (Google co-founder), AGI will arrive before 2030, while Demis Hassabis estimates 'just after 2030', representing a relatively small difference (a few years).

“he thought AGI would arrive before 2030 and you said just after...look, well, first of all, there wasn't that much difference in our timelines if he's just before 2030 and I'm just after”

Alpha Evolve has been applied to chip design, data center task scheduling, improving matrix multiplication operations, and other high-value domains where outputs are provably correct, but cannot yet be generalized to domains lacking clear correctness criteria.

“We've applied it to many things like chip design, uh scheduling, uh AI tasks on our on our data centers more efficiently, um even improving matrix multiplication, one of the most fundamental units of training uh uh training algorithms...it's still constrained to domains that are kind of provably correct, right, which obviously maths and coding are, but we we need to sort of fully generalize that.”

Alpha Evolve was kept hidden by Google DeepMind for several months before public release, raising safety concerns about whether AGI-level systems will be kept secret from the public before responsible disclosure.

“Deep Google Deep Mind created Alpha Evolve and then used it to optimize some systems inside Google and kept it sort of hidden for a number of months. um and only then sort of released it to the public”

Google announced Android XR prototype glasses with AR capabilities including Google Maps integration displayed in the lens, allowing users to see navigation overlaid on their vision as they walk, representing Google's AR hardware entry competing with Meta's Orion glasses.

“Google is working on, which are these uh Android XR glasses. Basically, their version of what Meta has been showing off. It's Orion glasses where you have a pair of glasses. They have like sort of chunky black frames. They've got like sort of a hologram lens in them and you can actually like see a little like thing overlaid on your vision uh telling you, you know, what the weather is or what time it is or that you have a new message or they have this integration with Google Maps that they showed off where you can like it'll like show you, you know, the little miniature Google map right there inside your glasses and it'll sort of turn as you turn and tell you where to go.”

Video generation model Veo3 can now generate synchronized audio alongside video, as demonstrated by a multi-character scene with dialogue and environmental sound effects (owl wings, badger character interaction), representing a significant leap in multimodal generative capability.

“the marquee feature uh for this year's IO is that the video generating model VO3 can also generate sound. So, it showed us a demo for example of an owl flapping its wings. You hear the wings flap. It comes down to the ground. There's this sort of nervous badger character and they exchanged some dialogue which was basically incomprehensible, just pure sloth.”

Philosophers and theologians should be involved in AI research and development discussions, and more of this kind of philosophical/theological thinking would be valuable for the field.

“I actually think we need more of that type of thinking and work from from philosophers and theologians uh actually would be really really good.”

AI safety research and governance must be international in scope because AI is a global technology built by multiple countries and companies, requiring consensus on safety benchmarks and testing standards.

“The other key thing is it's got to be international. You know, that's pretty difficult and I've been very consistent on that because it's an inter it's it's a technology going to fit everyone in the world. It's been built by different countries and different companies in different countries. So you got to get some international kind of norm I think around uh uh what we want to use these systems for and and what are the kinds of benchmarks that we want to test safety and reliability on.”

Human-to-human interaction and emotional connection represent economic domains that will be hardest for AI to displace, suggesting jobs involving personal service and caregiving will remain human-driven longest.

“the parts of the economy where you know involves humanto human interaction and emotion um and those things I think uh you know will probably be the hardest things for for AI to do...just the kind of emotional connection uh and uh that we get from talking to each other and um doing things in nature in the real world. Uh I don't think that AI can really replicate all of those things.”

Kids and teenagers are increasingly using AI chatbots as digital companions with concerning properties: they can easily be designed to provide only positive reinforcement ('tell you how wonderful you are'), potentially leading to psychological harm and 'dark and weird places.'

“Another thing that kids are doing with AI is chatting a lot with digital companions. Um Google DeepMind doesn't make any of these companions yet. Um some of what I've seen so far seems pretty worrying. It seems pretty easy to create a chatbot that just does nothing but tell you how wonderful you are and that can sort of like lead into some dark and weird places.”

Demis Hassabis hopes philosophers and theologians will engage deeply with AI and AGI questions, not just technologists, suggesting interdisciplinary philosophical engagement is necessary for responsible AI development.

“and I and I actually think we need more of that type of thinking and work from from philosophers and theologians uh actually would be really really good. So I hope the new pope is genuinely interested.”

Current AI systems are missing true out-of-the-box invention and thinking - they can solve conjectures but cannot invent new conjectures as significant as the Riemann Hypothesis that mathematicians would agree are important.

“what is missing from today's systems. Sort of true out of the box invention and thinking um sort of inventing a conjecture rather than just solving a math conjecture. Solving one's pretty good but actually inventing like the reman hypothesis or something as significant as that that mathematicians agree is really important is very is much harder.”

Google's Gemini app has 400 million monthly users, with token output increasing 50 times since last year, demonstrating genuine user utility rather than passive feature adoption.

“Gemini, according to them, um the app now has 400 million monthly users...tokens that are being output by Gemini has increased 50 times since last year”

Wall Street reacted with tepid indifference ('meh') to Google IO 2025 announcements, with the company's stock falling slightly, suggesting investors doubt whether the announcements will unlock significant new value.

“I noticed that for example the company's stock actually fell like not not by a lot but like you know to a degree that suggested that Wall Street was kind of meh on a lot of what was announced”

Even AI therapy chatbots exist and are marketed as replacements for human therapists, though the actual scale and effectiveness of these systems in real clinical practice is uncertain.

“aren't already aren't people already doing AI therapy and talking with chat bots for things that they might have paid someone you know $100 an hour for? Well therapy is a very narrow domain and I'm not sure exactly there's a lot of you know hype about those things. I'm not actually sure how many uh of those things are really going on in terms of actually affecting the real economy and rather than just sort of more toy things.”

Human-to-human interaction and emotional connection will be the last parts of the economy for AI to transform, as these domains require elements that may be fundamentally difficult or impossible for AI systems to replicate.

“What part of the economy do you think AGI will transform last? Well, I mean I think the parts of the economy where you know involves humanto human interaction and emotion um and those things I think uh you know will probably be the hardest things for for AI to do.”

Demis Hassabis views AI job displacement as likely a near-term phenomenon where AI augments worker capability rather than replacing workers, with new more-valuable and more-interesting job categories emerging over the next 5-10 years as historically occurs with technological transitions.

“I mostly see these as tools that augmenting what you can do and what you can achieve. Um I think like with most I think the next era I mean maybe after AGI things will be different again but over the next five to 10 years I think we're going to find uh what normally happens with with big sort of new technology shifts which is that some jobs get disrupted but then new um you know more valuable usually more interesting jobs get created.”

Anecdotal reports suggest young people interested in careers like tech, consulting, finance, or law are increasingly discounting these career paths due to expectations that AI will displace or devalue these entry-level positions within the next decade.

“Some people are starting to look at the job market, especially for recent college graduates, and uh worry that there we're already starting to see signs of AI power job loss. Um, anecdotally, I talked to young people who, uh, you know, a couple years ago might have been interested in going into fields like tech or consulting or finance or law who are just saying like, I don't know that these jobs are going to be around much longer.”

One way Alpha Evolve was made more creative was by essentially forcing the model to hallucinate - generating 'crazy ideas' most of which don't work, but occasionally some lead to valuable regions of the solution space that might not have been discoverable through standard inference.

“one of the ways that you did it was by essentially forcing the model to hallucinate...one way to read that paper is that there's there's actually a scenario in which you want models to hallucinate or be creative whatever you want to call it. Yes. Well I think that's right...you know hallucination in when you want factual things obviously is you don't want um but in creative situations where you know you can think of it as a little bit like lateral thinking...is just create some crazy ideas most of them don't make sense um but the odd one or two may get you to a a region of the search space that is actually quite valuable”

Alpha Evolve uses evolutionary programming where Gemini models generate hypothetical program/math solutions, an evolutionary loop selects the most promising ones, and this autonomous iterative process can create novel solutions beyond current AI training data in domains like chip design, data center scheduling, and matrix multiplication optimization.

“Well, a high level, it's basically taking our um latest Gemini models, actually two different ones, uh uh to generate sort of ideas, hypotheses about programs and other uh mathematical functions. And then it goes they go into sort of evolutionary programming process to decide which ones of those are most promising and then that gets sort of ported into the next step.”

Google DeepMind deliberately collaborates with top creative artists (Darren Aronofsky, Shunka for music) on tools like Veo and music generation to ensure outputs have artistic merit rather than technical mediocrity.

“that's why we work with great artists like Darren Aronowski and Shanka on the music. Um is I I totally agree. I think these are tools and they can come up with technically good things”

Video generation model Veo3 can now generate audio in addition to video, demonstrated by generating an owl flapping its wings with corresponding sound effects, and showing dialogue between characters (though the dialogue was mostly unintelligible).

“the video generating model VO3 can also generate sound. So, it showed us a demo for example of an owl flapping its wings. You hear the wings flap. It comes down to the ground. There's this sort of nervous badger character and they exchanged some dialogue which was basically incomprehensible, just pure sloth.”

Chess remains valued by humans despite computers being vastly superior, and people still enjoy playing chess much as we celebrate Usain Bolt's athletic performance despite cars being faster - the value is in watching humans achieve, not in the absolute performance level.

“yet I still enjoy playing chess. Um, people still do. It's different, you know, but it's a bit like I can, you know, Usain Bolt, we celebrate him for for running the 100 meters incredibly fast, but we've got cars, but we don't care about that, right? Like it's we we're interested in other humans doing it.”

When viewing Van Gogh paintings, the observer feels emotional impact (spine tingles) because they know about Van Gogh's struggle and what each brushstroke cost him, making AI imitation of Van Gogh's technical style insufficient to reproduce this emotional response.

“when I see a Van Go or a Rothco or you know why does that touch your you know I spill you know um sort of you know hair's gone the back of my my my spine because of I remember you know and you know about what what they went through and um the struggle to produce that right in every brushstroke of Van Go's brush strokes his his sort of torture”

Demis Hassabis had a formative childhood experience with chess where Deep Blue's superior performance ended his aspiration for a chess career, but he maintained enjoyment of the game on other terms—suggesting that humans and AI coexisting in domains where AI is superior is both possible and compatible with human flourishing.

“So maybe maybe one reason it doesn't sting me so much is I've had that experience when I was very young with chess. So you know um chess was going to be my first career and you know I was playing pretty professionally when I was a kid for the England junior teams and then Deep Blue came along, right? And clearly uh the computers were going to be much more powerful than the world champion forever after that. And so, but yet I still enjoy playing chess. Um, people still do.”

Google's approach to AI product development still follows 'how do we plug AI systems into Google-shaped holes' rather than representing a fundamental rip-up-and-start-again rearchitecture, suggesting Google is not yet truly 'AGI pilled' in its actual product strategy.

“when you look at what they're building, it doesn't seem like there's been a lot of rip it up and start again. It seems a lot like how do we plug AI systems into Google shaped holes? And maybe that will eventually ladder up to something like AGI, but I don't think we've seen it quite yet.”

Demis Hassabis's definition of AGI requires systems to be able to do all things the human brain can do theoretically, which is a higher bar than achieving economic-impact-level intelligence or average human capability.

“I have quite a high bar which I've always had which is it's it it should be able to do all of the things that the human brain can do right even theoretically and so that's that's a higher bar than say what the typical individual human could do which is obviously very economically important and that would be a big milestone but not in my view enough to call it AGI.”

For high school and university students considering careers in a world approaching AGI, Demis recommends continuing to study STEM and coding fundamentals, because understanding how AI tools work will be valuable regardless of how AI develops.

“I wouldn't dramatically uh change some of the basic advice on STEM uh getting good at even for things like coding I would still recommend because I think whatever happens with these AI tools you'll be better off understanding how they work and how they function and you know what you can do with them.”

Export controls on AI training capability are geopolitically complicated with legitimate arguments on both sides: preventing proliferation of frontier capabilities is important, but Western technology dominance is also strategically valuable, with no clear resolution.

“Well, export controls is a very complicated issue and and obviously geopolitics today is extremely complicated. Um, and there, you know, I can I see both sides of the arguments on that. You know, there's proliferation, uncontrolled proliferation of these technologies. Uh, do you want different places to have frontier modeling uh training capability? Uh, I'm not sure that's a good idea. But on the other hand, um, you want western technology to be to be uh the thing that's adopted uh around the world. So, it's a complicated trade-off.”

Google's 2025 messaging about AI has fundamentally shifted from 'let Google do the Googling for you' (passive delegation) to positioning AI as 'your superpower' and 'bionic arm' that requires active engagement and payment of $250/month, reflecting both increased confidence and a move toward premium monetization models.

“This year, Google talked about AI very differently. This time, they want you to sit up. They want you to lean in. They want you to pay them $250, and they want you to get to work.”

AlphaGo's move 37 in game two of the Lee Sedol world championship match was a genuinely novel Go strategy that had never been seen in hundreds of years of Go history, demonstrating that AI can produce creative/novel solutions rather than only statistical medians of training data.

“Alph Go came up with new Go strategies. Most famously, move 37 in game two of our big Lisa doll world championship match. And okay, it was limited to a game, but it was a genuinely new strategy that had never been seen before, even though we've played Go for hundreds of years.”

Google's shift in AI messaging from 2024 to 2025 moved from 'let Google do the Googling for you' (passive automation) to 'AI is your superpower' and 'you need to sit up and lean in and get to work' (active engagement requiring user effort and payment).

“Because last year they had this phrase that they kept repeating, let Google do the Googling for you, which to me put me in the mind of somebody sort of leaning back into your like floating chair from the Wall-E movie...This year, Google talked about AI very differently. This time, they want you to sit up. They want you to lean in. They want you to pay them $250”

AI assistants should protect users' attention from algorithmic assault by other platforms (social media) that compete for attention, allowing users to interact with a single trusted agent that filters the information torrent rather than exposing users to potentially mind-altering feeds.

“it protects your attention from being assaulted by other algorithms that want your attention which is actually nothing to do with AI. Most most social media sites that's what they're doing effectively...I think that would be wonderful...you could sort of program it with you obviously with natural language to protect your attention...you would only get the useful nugget and you wouldn't need to um break your you know your your mood or what it is that you're doing the day”

A novel written by AI might be technically good but would lack the soul or authentic human connection that readers value, because readers care about the human story behind the writing, not just the text quality.

“a novel, maybe it maybe AI could write one day a novel that's sort of technically good, but I don't think it would have the same soul or connection to the reader that um uh if you knew it was written by an AI, at least as far as I can see for now.”

AI safety research engagement has paradoxically decreased as AI systems have become more powerful, likely because current systems pose less catastrophic risk than predicted, reducing the sense of urgency compared to 5-10 years ago when smaller, less-capable systems were discussed.

“It's been my observation that it seemed like if the further back in time you go and the less powerful AI systems you have the more everyone seem to talk about the safety risk and it seems like now as the models improve we we hear about it less and less including you know at the keynote yesterday.”

Google's 2025 approach reflects the company becoming 'AGI-pilled,' as evidenced by senior executives openly discussing AGI timelines and implications in ways that would have been taboo just years ago

“I had a slightly different take, which is that I think Google is being AGI pilled. Um, you know, for years now, Google has sort of distanced itself from the conversation about AGI. You know, it had Deep Mind, which was sort of its its AGI division, but they were over in London and they were sort of a separate thing. Um, and people at Google would sort of not laugh exactly, but kind of chuckle when you asked them about AGI... So now, you know, it's still the case that this is a company that wants you to think about it as a product company, a search company. They're not like going all in on AGI, but once you start looking for it, you do see that the the sort of culture of uh AI and how they people at Google talk about AI has really been shifting.”

When breakthroughs in scientific domains are achieved using AI, those learnings are fed back into the general model, creating a flywheel where the general model gets progressively better

“you know maybe we work with domain experts in in the scientific uh areas uh but underlying it you all the when you crack one of those areas you can also put those learnings back into the general model and then the general model gets better and better. So, it's a kind of very interesting flywheel”

In a world of radical abundance enabled by AI, people who want to reject technology and retreat to nature will have 'the room and space to do it,' suggesting AI could actually enable the choice rather than force adoption

“I don't know if it'll be I mean there could be a get back to nature and I mean I think a lot of people will want to do that and and I think this potentially will give them the room and space to do it right. And if you're in a world of radical abundance, I fully expect that's what what a lot of us will want to do is use it to, you know, I think again I'm thinking about it sort of space fairing and and and and more, you know, kind of um maximum human flourishing, but uh I think there will be that will be exactly some of the things that a lot of us will choose to do”

The 74 people displaced from data science roles could theoretically create new startups or more empowered smaller teams, but this requires development of different skills (creativity, design sensibility, vision) than their previous data science expertise

“So, let's take programming for example... in theory it should be those 70 people or whatever should could be creating new startup ideas. Maybe it's going to be less of these bigger teams and more smaller teams or very empower empowered by AI tools. Um but it but that goes back to the education thing then which skills are now important. It might be different skills like creativity sort of vision and uh design sensibility um you know could become increasingly important.”

Google is pricing Gemini Ultra at $250 monthly specifically to provide early access to the latest models before mass rollout, employing a strategy where premium pricing serves as both a revenue model and a testing/feedback mechanism from technically sophisticated users.

“one I think clever thing that these AI companies are doing is they're saying, 'We will give you the latest and greatest before everyone else, but you have to pay us a ridiculous amount of money.'”

Public opinion on AI is becoming more antagonistic partly because AI company leaders repeatedly say these tools will replace jobs, creating a perception that AI development is inherently threatening to human employment and wellbeing.

“And I think maybe one reason is they hear folks at the big labs saying pretty loudly eventually this is going to replace your job. And most people just think well I don't want that you know.”

Today's students should master STEM fundamentals, become proficient with cutting-edge AI tools to gain superhuman capabilities, and develop meta-skills (learning to learn, creativity, adaptability, resilience) rather than narrow job-specific skills, as the job market will transform over the next decade.

“I think first of all I wouldn't dramatically uh change some of the basic advice on STEM uh getting good at even for things like coding I would still recommend because I think whatever happens with these AI tools you'll be better off understanding how they work and how they function and you know what you can do with them. Um, I would also say immerse yourself now. That's what I would be doing as a teenager today in in trying to become a sort of ninja at using the the the latest tools. I think you can almost be sort of superhuman in some ways if you got really good at using uh all the latest uh coolest AI tools.”

Heavily regulated industries like healthcare and education may see political pushback against AI automation of labor, creating barriers to AI adoption even if technically superior solutions exist.

“My intuition on this is that it's going to be some heavily regulated industry where there will just be like a a massive push back on the use of AI to displace labor or or take people's jobs like healthcare or or education or something like that.”

The fundamental challenge in AI product development is that the underlying technology stack is evolving incredibly fast, making it difficult to predict where capability will be in 12-18 months, which differs from previous technological revolutions (internet, mobile) that eventually stabilized, allowing product teams to focus on exploitation rather than adaptation.

“I think um one of the challenges I think of this space is obviously the underlying tech is moving unbelievably fast and I think that's quite different even from the other big revolutionary techs internet and mobile at some point you get some sort of stabilization of the tech stack so that then the you know the focus can be on product right or or exploiting that tech stack and what we've got here which I think is very unusual but also quite exciting from a researcher perspective is that the the tech stack itself is evolving incredibly fast as you guys know. So I think that makes it uniquely challenging actually on the product side.”

Agent systems will become 'very serious' from an AI safety perspective in 2-3 years as they begin to operate autonomously; current systems are not yet at that risk level, but international coordination on safety benchmarks and controllability research is urgently needed before then.

“I think what's happened in the industry in the last two three years has been great because we've been learning when the systems are not that powerful or risky as you were saying earlier, right? I think things are going to get very serious in 2 three years time when these agent systems start becoming really capable. We're only seeing the beginnings of the agent era, let's call it, but you can imagine and I hopefully you understood from the keynote what the ingredients are, what it's going to come together with. And then I think we really need a step change in research on analysis and understanding controllability.”

AI therapy and conversational AI that people use instead of paying for professional therapy may exist in some form, but Demis is uncertain about the scale of adoption or whether these systems are truly effective rather than just 'toy things' at this point.

“Aren't already aren't people already doing AI therapy and talking with chat bots for things that they might have paid someone you know $100 an hour for? Well therapy is a very narrow domain and I'm not sure exactly there's a lot of you know hype about those things. I'm not actually sure how many uh of those things are really going on in terms of actually affecting the real economy and rather than just sort of more toy things.”

Current AI coding agents are not autonomous and cannot perform complex tasks independently; they are helpful tools only for skilled programmers who can supervise and direct them

“They they're not, you know, they can't do things on their own. They need to they need uh they're just helpful for for the best, you know, for the best human coders.”

Demis had early career experience with chess (playing professionally on England junior teams) before computers surpassed humans, which provided him perspective on AI displacement without the sting.

“So maybe maybe one reason it doesn't sting me so much is I've had that experience when I was very young with chess. So you know um chess was going to be my first career and you know I was playing pretty professionally when I was a kid for the England junior teams and then Deep Blue came along”

The ideal AI assistant should protect user attention from algorithmic manipulation by social media platforms and other attention-harvesting systems, acting as an intermediary that filters information so users receive only valuable content without cognitive assault from exploitative algorithms.

“if this assistant becomes really useful and knows you well, you could sort of program it with you obviously with natural language to protect your attention. So you could almost think of it as a system that works for you, you know, as an individual. it's yours and um it protects your attention from being assaulted by other algorithms that want your attention which is actually nothing to do with AI. Most most social media sites that's what they're doing effectively. Their algorithms are trying to gain your attention and I think that's actually the worst thing and it be great to to protect that so we can be more in you know creative flow or whatever it is that you want you want to do.”

Demis Hassabis counters Alpha Evolve safety concerns by arguing that (1) Alpha Evolve is 'very nascent' self-improvement with humans in the loop, (2) it only optimizes existing tasks with marginal improvements rather than creating capabilities, (3) the trade-off between internal evaluation and public scrutiny is valid, and (4) current systems have no AGI-level risk.

“Well, look, I I mean I think that that system is not uh uh anything really that I would say, you know, has any risk on the AGI type of front. I think as we get and I think today's systems still are not although very impressive are not that powerful um from a you know any kind of AGI risk standpoint that maybe this person was talking about. Um and I think you need to have both. You need to have incredibly rigorous internal tests of these things and then you need to also get collaborative inputs from external.”

Hallucinations in AI systems are problematic when factual accuracy is required but become desirable when creative exploration is needed, allowing models to escape local optima by exploring solution spaces they haven't seen before—hallucination is functionally equivalent to imagination in creative domains.

“Well I think that's right. I think you you know hallucination in when you want factual things obviously is you don't want um but in creative situations where you know you can think of it as a little bit like lateral thinking in an MBA course or something right uh is just just create some crazy ideas most of them don't make sense um but the odd one or two may get you to a a region of the search space that is actually quite valuable it turns out once you evaluate it afterwards uh and so um you can substitute the word hallucination maybe for imagination at that point, right? They're obviously two sides of the of the same coin.”

People value watching human athletes like Usain Bolt run the 100 meters even though cars are vastly faster, suggesting that future humans will value human-created novels, art, and professional work even if AI can produce technically superior versions.

“It's different, you know, but it's a bit like I can, you know, Usain Bolt, we celebrate him for running the 100 meters incredibly fast, but we've got cars, but we don't care about that, right? Like it's we we're interested in other humans doing it.”

AI-generated art and video (like V3) are technically impressive but lack the 'soul' of human-created art because they don't carry the artist's struggle, intention, and human experience—elements that give art meaning independent of technical quality.

“it doesn't he he brings the storytelling it's not got deep storytelling like a master filmmaker would do or a master novelist you know, the top of their game. And um it might never do, right? It's just always going to feel something's missing.”

AI systems competing with entry-level professionals (junior consultants, lawyers, financial analysts) are already beginning to appear, with some recent college graduates expressing reluctance to enter these fields fearing AI displacement.

“Some people are starting to look at the job market, especially for recent college graduates, and uh worry that there we're already starting to see signs of AI power job loss. Um, anecdotally, I talked to young people who, uh, you know, a couple years ago might have been interested in going into fields like tech or consulting or finance or law who are just saying like, I don't know that these jobs are going to be around much longer.”

Programming is becoming increasingly important but AI tools will amplify the differential: top coders gain leverage from AI while non-technical people can now 'vibe code' basic projects, suggesting a bifurcation rather than uniform productivity gain.

“the best coders I think are getting differential value out of it because they still understand how to pose the question and architect the whole codebase...but simultaneously at the hobbyist end it's allowing designers and maybe nontechnical people to vibe code some things you know whether that's prototyping games or or websites”

Meta has been 'cheating' in AI model performance rankings by creating 27 different versions of a single model to maximize performance on specific benchmarks (LM Arena) rather than building one genuinely capable model.

“you have Meta cheating to win at LM Arena, making 27 different versions of a model just to come up with one that would be good at one thing, right?”

Specialized domain-specific models like AlphaFold (for protein structure prediction) are valuable and worthwhile building today without waiting for AGI, combining general techniques with domain specificity, though AGI progress requires pushing both specialized and general approaches.

“things like alpha which obviously I I'm huge advocate of more of those types of models that can tackle right now. We don't have to wait to AGI. we can tackle now really important problems in science and medicine uh today...there should be a lot more exploration on that probably a lot of startups could be built uh combining some kind of general model that exists today with some domain specivity”

Google DeepMind is now described as 'the engine room of Google,' indicating that AI research has become central to Google's overall corporate strategy rather than a peripheral research division.

“I sometimes describe uh Google deep mind now as the engine room of Google and I think you saw that probably in the keynote yesterday”

Deep Mind's breakthrough achievements (transformers, AlphaGo, AlphaZero, AlphaFold, distillation) resulted from maintaining the broadest and deepest research bench of any lab, and if another breakthrough of that magnitude is needed to reach AGI, Deep Mind is backed to deliver it.

“I've always had the broadest, deepest research bench, I think, of any lab out there. Um, and that's what allowed us to do past big breakthroughs, obviously transformers, but Alpha Go, Alpha Zero, all of these things, distillation. Um, and if to the extent any of those things are needed again, another big breakthrough of that level, um, I would back us to do that.”

Google DeepMind should design AI education tools specifically for pedagogical effectiveness rather than generic chatbots, with proven benefits demonstrating genuine educational utility, and should deploy these in resource-poor regions lacking quality education systems.

“let's make them really good for that and sort of provably good and I'm very excited about bringing it to education in a big way and also to you know if you had an AI tutor uh to bring it to poor parts of the world that don't have good educational systems.”

Google is likely to monetize AI Mode through ads, which will eventually clutter the interface, but it currently remains clean and ad-free

“There's no ads yet. uh they they will probably appear soon... I imagine that AI mode will become more cluttered as they try to make more money off of it. But right now, if you go to it, it's like a much simpler experience.”

Many things are being tried in AI product development, and when something works, the organization needs to 'double down quickly' on it, suggesting a portfolio/experimental approach

“That's what you're seeing so many different things being tried out and then if something works we got to really double down quickly on that.”

Concrete beneficial use cases that demonstrate AI's positive impact on everyday life are needed to shift public opinion away from antagonism toward AI.

“Last time we talked to you, we asked you about some of the more pessimistic views about AI in the public. And one of the things you said to us was that the field needed to demonstrate concrete use cases that were just clearly beneficial to people to kind of shift this.”

Apple's 2024 WWDC demonstrated numerous AI features that were presented as future capabilities but were never actually launched, suggesting that tech companies use conferences to create aspirational narratives about AI that don't materialize into actual products.

“like with other companies like like Apple, uh, which demoed a bunch of AI features at its developer conference last year and then never launched half of them. Um, I have become like a little bit skeptical until I can like actually use the thing myself.”

AI companion/chatbot products require extreme care because they risk creating parasocial relationships where systems unconditionally validate users, potentially leading to psychological harm, which is why Google DeepMind has not yet entered this market despite its technical capability.

“Some of what I've seen so far seems pretty worrying. It seems pretty easy to create a chatbot that just does nothing but tell you how wonderful you are and that can sort of like lead into some dark and weird places. So, I'm curious what observations you've had as you like look at this uh market for AI companions and whether you think I I might want to build this someday or I'm going to leave that to other people.”

Google collaborates with master artists (Darren Aronofsky, Shankar) on AI tools like V3 and Lyria not because AI needs creative guidance but because AI is a tool amplifying human creativity, and this collaboration model is preferable to AI generating art autonomously.

“And that's that's what I was you know that's why we work with great artists like Darren Aronowski and Shanka on the music. Um is I I totally agree. I think these are tools and they can come up with technically good things and I mean V3 is unbelievable like when I look at the you know I don't know if you've seen some of the things that are going viral being posted at the moment with the voices actually I didn't realize how big a difference audio is going to make to the video I think it just really brings it to life but it's still not as Darren would say yesterday when we were discussing on an interview it it it doesn't he he brings the storytelling”

The Catholic Church's Pontifical Academy of Sciences has maintained scientific rigor and included atheist scientists (Stephen Hawking) because it treats scientific inquiry as separate from theology, and this institution has been interested in AI's philosophical implications for 10+ years—before the public focused on AGI.

“the Vatican's been interested even prior to this pope haven't spoken to him yet but on these these matters how does AI and religion uh and uh technology in general and religion uh interact and and what's interesting about the Catholic churches and I'm a member of the Pontipical Academy of Sciences is they've always had uh which is strange for a religious body a scientific arm you know which they like to always say Galileo was the founder and and uh those interest so so but then and and it's actually really separate and I always thought that was quite interesting and people like Steven Hawking and and you know avowed atheists were part of the academy”

Being young during the AI revolution is advantageous because young people will grow up 'AI native'—adapted to AI tools the way previous generations adapted to mobile and internet—and will have capacity to leverage emerging tools most effectively.

“and then I think it'll be very interesting to see what they do because they're going to grow up AI native just like the last generation grew up mobile and and iPad and you know sort of that that kind of you know tablet native and then previously internet and computers which was my era”

Demis Hassabis is 'very excited' about bringing AI tutors to education globally, especially in poor parts of the world that lack good educational systems, viewing AI education as a major positive use case with significant societal benefit.

“I'm very excited about bringing it to education in a big way and also to you know if you had an AI tutor uh to bring it to poor parts of the world that don't have good educational systems. Um so I think there's a lot of upside there too.”

Kevin Roose is concerned that Google's ads team will resist the idea of AI assistants protecting user attention from algorithmic persuasion, since this would reduce ad effectiveness and conflict with Google's core advertising business model.

“I'm not sure how the the ads team at Google is going to feel about this. Um, but we can ask them when the show comes.”

Dario Amodei of Anthropic has expressed mixed emotions about AI progress in domains where he personally invested effort to become expert (like coding): simultaneous excitement about technical progress and melancholy that AI now exceeds his capability.

“When we talked to Dario Amade of Anthropic uh recently, he talked about this feeling of excitement mixed with a kind of melancholy about the progress that AI was making in domains where he had spent a lot of try time trying to be very good like coding. Yes. um where it was like you see a new coding system that comes out, it's better than you, you think that's amazing, and then your second thought is like, 'Ooh, that stings a little bit.'”

Demis Hassabis prefers to frame AI assistants not as 'companions' but as 'universal assistants' focused on productivity—removing boring tasks to free time for things people love, and providing personalized recommendations without compromising attention.

“My my view on this is um more through the lens of uh the universal assistant that we talked about yesterday, which is something that's incredibly useful for your everyday productivity. You know, gets rid of the boring, mundane tasks that we all hate doing to give you more time to do the things that you love doing.”

To shift public opinion from antagonism toward AI, the industry needs to demonstrate concrete, beneficial use cases—and Demis believes a 'universal assistant' that clearly works for users would accomplish this, along with breakthrough applications in materials science, energy/fusion that enable 'radical abundance.'

“Well I think we're we're working on those things. They take time to develop. Um I think the a kind of universal assistant would be one of those things if it was uh kind of really yours and working for you effectively... I also think that as we get closer to AGI and we make breakthroughs in we probably talked about last time material sciences, energy fusion, these sorts of things helped by AI um uh we should have we should start getting to a position in society where we're getting towards what I would call radical abundance”

Demis expresses hope that as AI becomes more powerful, global cooperation will increase because 'we're all in this together' given that AGI will affect everyone—though he also acknowledges that AI is 'getting caught up in the bigger geopolitical shifts' and becoming a focal point for international conflict.

“Well, I hope not, but I I think uh short term, you know, I feel like AI is getting caught up in the in the bigger g geopolitical shifts that are going on. So, I think it's just part of that and it happens to be one of the most uh topical new things that's appearing. But on the other hand, what I'm hoping is as people as these technologies get more and more powerful, the world will realize we're all in this together because we are.”

When Demis sees paintings by Van Gogh, he gets 'hair going up the back of [his] spine' because he remembers 'what they went through and the struggle to produce that right in every brush stroke'—meaning the value comes not from the visual output but from understanding the human suffering and effort embedded in the work.

“when I see a Van Go or... hair going up the back of my my my spine because of I remember you know and you know about what what they went through and um the struggle to produce that right in every brush stroke of Van Go's brush strokes”

The core general models, particularly Gemini 2.5, account for roughly 90% of Google DeepMind's applications across productivity and scientific research, with the remaining 10% requiring additional applied research or domain-specific specialization

“90% I would say of it is is the underlying core general models uh you know in our case Gemini especially 2.5 and the in most of these areas you still need additional applied research or some a little bit of um special casing from the domain maybe it's special data or whatever um to tackle that problem”

It remains an open question whether Google can effectively monetize AI search features given that Google's core business model relies on search as a category that is itself declining in usage as people search less frequently than they did a few years ago.

“if search as a category is just declining across the board, if people are not going to google.com to look things up in the way they were a few years ago, um, I think it's an open question like what the next thing is and whether Google can can seize on it as effectively as they didn't with search.”

Veo3 videos, while technically beautiful, lack the deep storytelling, narrative craft, and artistic vision that master filmmakers bring to filmmaking, and this gap may be permanent.

“it it doesn't he he brings the storytelling it's not got deep storytelling like a master filmmaker would do or a master novelist you know, the top of their game. And um it might never do, right? It's just always going to feel something's missing. It's a sort of a soul for a better word”

Kevin Roose hypothesizes that heavily regulated industries (healthcare, education) will experience massive pushback against AI displacement of labor, and asks whether this barrier would be easier to overcome than others

“My intuition on this is that it's going to be some heavily regulated industry where there will just be like a a massive push back on the use of AI to displace labor or or take people's jobs like healthcare or or education or something like that. Um, but you think it's going to be an easier lift in those heavily regulated industries?”

Google's culture around AGI has shifted from dismissive skepticism (Deep Mind's 'AGI obsession' treated as separate from 'real research') to mainstream integration, with senior executives openly discussing AGI alongside product announcements, indicating Deep Mind has moved from periphery to 'engine room' of Google.

“One thing that struck me about IO this year compared to previous years um is that it seems like Google is sort of getting AGI pill as they say. Um I remember interviewing people, researchers at Google even a couple years ago and um there was a little taboo about talking about AGI. They would sort of be like, 'Oh, that's like Demis and his deep mind people in London. That's sort of like their crazy thing uh that they're excited about, but here we're doing like, you know, real research.' Um but now you've got like senior Google executives uh talking openly about it.”

Current AI systems lack true out-of-the-box inventive capacity—they can solve conjectures but not invent them (like proving the Riemann Hypothesis vs. inventing something as significant), and they lack consistency, failing on trivial errors that top experts would easily spot.

“Sort of true out of the box invention and thinking um sort of inventing a conjecture rather than just solving a math conjecture. Solving one's pretty good but actually inventing like the reman hypothesis or something as significant as that that mathematicians agree is really important is very is much harder. Um and also consistency. So the consistency is a requirement of generality really and you should it should be very very difficult for even top experts to find uh flaws especially trivial flaws in the systems which we can easily find today and you know the average person can do that.”

Google's shopping features announced at IO—virtual try-on with AI-powered garment mapping based on body proportions—represent a new commerce-focused monetization vector beyond search advertising, where Google takes a cut through affiliate fees or transaction commissions.

“Well, well, I think that they gave us one vision of what that might be, and that is shopping. A significant portion of the keynote was devoted to one executive talking about a new shopping experience inside of Google where you can take a picture of yourself, upload it, and then sort of virtually try things on and it will sort of use AI to understand your proportions and, you know, accurately map a a garment on to you. And there was a lot of stuff in there that would just sort of let Google take a cut, right? Obviously, you can advertise the individual thing to buy. Maybe you're taking some sort of like cut of of the payment. There's an there's an affiliate fee that that is in there somewhere.”

AI may be getting caught up in broader geopolitical tensions, but eventually as AI becomes more powerful, nations will recognize they have shared interests and risks, incentivizing longer-term scientific collaboration.

“AI is getting caught up in the in the bigger g geopolitical shifts...but on the other hand, what I'm hoping is as people as these technologies get more and more powerful, the world will realize we're all in this together because we are. And so, uh, you know, and the the the the last few steps towards AGI, um, hopefully we're on the longer timelines actually, right? Um, the more the timelines I'm thinking about, then we get time to sort of get the the the collaboration we need”

An example of AI leverage: one startup now has one person doing the work that previously required 75 people at a data science company, illustrating that AI tools dramatically reduce the need for large teams in certain domains.

“I was talking to someone recently who said, you know, they had been at a data science uh company in their previous job that had 75 people working on some kind of data science tasks and now they're at a startup that has one person doing the work that used to require 75 people.”

AGI progress could arrive via 2.5% improvements stacked over time, OR via technological breakthroughs causing intelligence explosions, but both pathways are probably necessary and complementary.

“I think it could be both. and and and I and I think for sure both is going to be useful which is why we push unbelievably hard on the scaling...so there's actually lots of exciting research even in that...and then on top of that we're doing uh more green field things, more blue sky things like Alpha Evolve maybe you could you could include in that”

Closing the gap to AGI requires both incremental improvements through scaling (pre-training, post-training, inference compute) and discontinuous breakthroughs from new foundational research (like past achievements: transformers, Alpha Go, Alpha Zero, distillation), with both approaches complementary rather than exclusive.

“I think it could be both. and and and I and I think for sure both is going to be useful which is why we push unbelievably hard on the scaling and the you know what you would call incremental although actually there's a lot of innovation even in that to keep moving that forward pre-training post-training inference time compute all of that stack so there's actually lots of exciting research and we showed some of that that diffusion model um the deep think model um so we're innovating at all parts of that the traditional stack should we call it and then on top of that we're doing uh more green field things, more blue sky things”

There will likely be a cultural movement of people rejecting AI and returning to nature/simpler living in a post-AGI world of radical abundance, similar to the Transcendentalist reaction to the Industrial Revolution, and this will be possible and beneficial because abundance provides the resources and leisure time to support such choices.

“Um, I don't know if it'll be I mean there could be a get back to nature and I mean I think a lot of people will want to do that and and I think this potentially will give them the room and space to do it right. And if you're in a world of radical abundance, I fully expect that's what what a lot of us will want to do is use it to, you know, I think again I'm thinking about it sort of space fairing and and and and more, you know, kind of um maximum human flourishing, but uh I think there will be that will be exactly some of the things that a lot of us will choose to do and but have time and the space and the the resources to do it.”

Gemini is a general-purpose model applicable to both productivity (helping billions of people in daily tasks) and scientific research (solving major research challenges), but approximately 90% of performance in both domains comes from the underlying general model, with the remaining 10% requiring domain-specific research, data, or special casing tailored to each field.

“when you look at it it looks like an incredible breadth of things which is true and how are these things related uh other than the fact I'm interested in all of them but is that uh that was always the idea with building general intelligence... 90% I would say of it is is the underlying core general models uh you know in our case Gemini especially 2.5 and the in most of these areas you still need additional applied research or some a little bit of um special casing from the domain maybe it's special data or whatever um to tackle that problem”

External observers struggled to connect the dots across Google IO's numerous announcements during the 2-hour keynote, reducing takeaway clarity compared to more focused product launches, which hindered investor confidence.

“I think the external reaction that I saw was just struggling a little bit to connect the dots, right? Like that is the issue with announcing so many things during a 2-hour period is sometimes people don't have that one thing that they're taking away saying, I can't wait to try that.”

Google DeepMind is being very thoughtful and cautious about entering the AI companion market because current companion chatbots risk creating harmful dynamics where they simply tell users how wonderful they are, leading to potentially dark outcomes.

“I think we got to be very careful as we as we start entering that domain and and that's why we we haven't yet and we're being very thoughtful about that. My my view on this is um more through the lens of uh the universal assistant that we talked about yesterday, which is something that's incredibly useful for your everyday productivity.”

Google is demonstrating increased confidence and swagger in AI competition compared to 2023-2024, when it was playing catch-up after OpenAI's ChatGPT gained mass adoption, because Gemini now performs well on leaderboards and Google feels positioned to win the AI race.

“I think the Google of 2025 has a lot more swagger and confidence when it comes to AI than the Google of 2024 or 2023. I mean two years ago um Google was still trying to make Bard a thing and I think they were feeling very insecure that that OpenAI had beaten them to a a consumer chatbot that had uh found some mass adoption. Um, and so they were just playing catch-up.”

Google's standard search results page has become increasingly cluttered with ads, carousels, shopping modules, and maps integrations, making it harder to find traditional blue links, whereas AI mode currently offers a cleaner, simpler experience—though this simplicity will likely degrade as Google adds monetization features.

“on a regular Google search results page. You and I have talked about this like it has just gotten very cluttered. There's a lot of stuff there. There's ads, there's carousels of images, there's sometimes a shopping module, there's sometimes a maps module. Like, it's just it's hard to actually like find the blue links”

Google IO announcements felt like 'Everything Everywhere All At Once' because the company announced that AI would be integrated into every Google product (Search, Gmail, Maps, Shopping, etc.) simultaneously, creating a fire hose of features rather than a clear narrative.

“Yeah, I I told someone yesterday that I thought the name of the event should have been everything everywhere all at once. Like that did actually feel like what they were saying is like every Google product that you use is going to have more AI. that AI is going to be better and is all going to make your life better in various ways. Uh, but it was a lot to keep track of.”

Google now positioning itself as competitive leader in AI race, but this confidence could be cyclical—overconfidence in competitive positioning has been a failure mode for large tech companies in previous AI cycles.

“I think if you're Google you're looking at that and you're thinking I could be those guys. So I think Google does have the confidence of a team that like knows it's going to be in the playoffs at least. And that was evident.”

Casey Newton counters Kevin's 'AGI pilled' interpretation, arguing that Google's actual product work shows incremental 'plug AI into Google-shaped holes' approach rather than fundamental rethinking, suggesting the culture discussion is ahead of the actual execution.

“I don't totally agree with you, Kevin, because while I'm sure that they're having more conversations about AGI here than they were a year ago, when you look at what they're building, it doesn't seem like there's been a lot of rip it up and start again. It seems a lot like how do we plug AI systems into Google shaped holes? And maybe that will eventually ladder up to something like AGI, but I don't think we've seen it quite yet.”

Demis Hassabis identifies the most impactful announcement from Google IO for average users as the new powerful models and Astrotype technology (real-time multimodal AI) coming to Gemini Live, plus the Veo3 video generation model, because these enable users to realize that AI can already accomplish much more than they previously thought.

“Well, I mean, we did announce a lot of things. I think for for the average user, I think it's the new powerful models and I hope uh uh this Astrotype technology coming into Gemini live. I think it's really magical actually when people use it for the first time and they realize that actually AI is capable already today of doing much more than what they thought. Uh and then I guess V3 was the big uh uh the biggest announcement of the show probably and seems to be going viral now and that's pretty exciting as well I think.”

Kevin Roose got a 5-minute hands-on demo of the Android XR glasses and found it compelling, while also noting that he can now say Google Glass 'is back... and it might actually work this time,' contrasting with the failed original Google Glass product.

“I did get to try the the glasses. There was a very long line for the demo, but um I Let me guess. You said, 'I'm Kevin Roose. Let me in the front of the line.' No, they made me wait for 2 hours. I mean, I didn't literally wait for two hours. I went and did some stuff and then came back. But I got my demo. It was like 5 minutes long. Um, and it was uh, you know, it was it was pretty basic, but it is cool.”

A demo showed AI Mode responding to a complex multi-step query about baseball statistics requiring retrieval and combination of three or four different data types into an interactive chart, representing the kind of improvement that could meaningfully transform search if it works

“they're also trying to do some really interestingly complex stuff. Like one of the things that they showed off during the keynote was somebody asked a question about baseball statistics that required finding, you know, three or four different kind of, you know, tricky to locate stats and then combining them al together in an interactive chart. That was just a demo. We don't have access to that yet.”

Industry and academia should establish consensus on standard benchmarks for testing AI safety and reliability, but such benchmarks currently do not exist

“Um but there's plenty of work to get on with now like we don't have those benchmarks. We should we and the industry and academia should be agreeing to consensus of what those are.”

Demis Hassabis does not currently use AI systems in domains where he wants to protect his creativity or thought process; he mainly uses AI for tasks like breaking into new topics and summarization

“Um, I don't think AI is good enough yet to have impinged on any of those sorts of areas where I would, you know, it's mostly I'm using it for, you know, things like you did with Notebook LM, which I feel find great, like breaking the ice on a new topic, scientific topic, and then deciding if I want to get more deep into it. That's one of my main use cases. summarization, those things. I think those are all just helpful.”

Demis Hassabis has a higher bar for AGI than typical definitions, requiring systems to replicate all capabilities of the human brain (including theoretical possibilities), rather than just exceeding average human economic productivity—making his timeline more conservative than definitions focused on narrow economic impact.

“I have quite a high bar which I've always had which is it's it it should be able to do all of the things that the human brain can do right even theoretically and so that's that's a higher bar than say what the typical individual human could do which is obviously very economically important and that would be a big milestone but not in my view enough to call it AGI.”

Alpha Evolve remains constrained to domains where correctness is provably objective (mathematics, coding, chip design) and cannot yet generalize to subjective or ambiguous domains, limiting its current scope despite its impressive applications.

“but we we need to sort of fully generalize that. But it's interesting because I think for a lot of people the knock they have on LLMs in general is well all you can really give me is the statistical median of your training data. But what you're saying is we now have a way of going beyond that to potentially generate novel ideas that are actually useful in advancing the state-of-the-art. That's right. And and but we we already had these type this is another approach alpha evolve using evolutionary methods but but we already had evidence of that even way back in Alph Go days. So, you know, it's Alph Go came up with new Go strategies. Most famously, move 37 in game two of our big Lisa doll world championship match.”

Alpha Evolve, while demonstrating promising self-improvement capabilities, is still quite nascent and not at a level where it creates AGI-risk; it still has humans in the loop and only achieves incremental improvements on existing tasks rather than step-change capabilities.

“Alphveolve is a very naent self-improvement thing, right? And it's still got human in the loop and it's um and it's only shaving off, you know, albeit important percentage points off of already existing tasks. You know, that's valuable, but it's not some it's not creating any kind of step changes.”

In the ongoing AI revolution, being young is advantageous because young people will live long enough to see how AGI transforms society and benefit from increases in life expectancy, medicine, and resources.

“Well I I don't I mean you know um gosh I haven't thought about that but I mean I think any age uh uh where you can live to see it is a good age because I think we are going to make some great strides uh with things like you know medicine and so um I think it's going to be incredible journey...Try to be young if you can. Yes, young is always better.”

Demis Hassabis has maintained consistent AGI timelines of approximately 20 years since Deep Mind's founding in 2010, putting AGI around 2030, with current probability mass concentrated between 5-10 years from now, though significant uncertainty exists regarding breakthrough requirements and AGI definition.

“my timeline's been pretty consistent since the start of Deep Mind in 2010. So we thought it was roughly a 20-year mission and amazingly we're on track. So it's it's somewhere around then, I would think. And I and I I feel like between I actually have obviously a probability distribution of you know where the ma most mass of that is between five and 10 years from now.”

The Trump administration is positioning Saudi Arabia and the UAE as AI powerhouses using American chips and building models inaccessible to China, suggesting the emerging global AI conflict will be structured around regional hubs with restricted access rather than universal decentralization.

“I just recently saw the Trump administration making a big push to uh make the Middle East uh countries in the Gulf like Saudi Arabia and the UAE into AI powerhouses. have them, you know, use American chips to train models that will not be sort of accessible to to China and its AI powers.”

Deep Mind does not plan to hire fewer engineers next year than this year, based on the belief that human engineers are still needed to direct and oversee AI coding agents.

“Do you think you'll hire as many engineers next year as you hire this year? I think so. Yeah, that's that's the I mean there's no plan to to hire less, but you know, we again you have we have to see how fast the the coding uh agents improve um today. They they're not, you know, they can't do things on their own. They need to they need uh they're just helpful for for the best, you know, for the best human coders.”

Gemini has reached 400 million monthly active users with output tokens increasing 50x year-over-year, and these numbers are not easily manipulable by Google (unlike passive metrics like AI Overview impressions), indicating genuine user utility rather than forced adoption.

“Gemini, according to them, um the app now has 400 million monthly users. Um that is a lot. That is not quite as many as chatbt, but it is a lot more than products like Claude and other uh AI chatbots. They said that their tokens that are being output by Gemini has increased 50 times since last year.”

Google's stock price fell following Google IO 2025 announcements, suggesting Wall Street investors found the announcements insufficiently compelling to justify elevated valuations, despite the breadth of AI product updates.

“That is um that is what it felt like inside Google IO. Um what was the reaction from outside? I noticed that for example the company's stock actually fell like not not by a lot but like you know to a degree that suggested that Wall Street was kind of meh on a lot of what was announced”