Ex-Google CEO: What Artificial Superintelligence Will Actually Look Like w/ Eric Schmidt & Dave B

The original Waymo self-driving challenge was in 2004 (Stanford DARPA Grand Challenge); it took 20+ years from visible demonstration to actual deployment (current 2025 Waymo taxis), showing that deep tech and regulated robotics take a very long time to commercialize.

“The original challenge at Stanford was done, I believe, in 2004. The DRA Grand Challenge. It was 2004. Sebastian through one. So, so more than 20 years from a visible demonstration to our ability to use it in daily life. Why? It's hard. It's deep tech. It's regulated and all of that.”

The risk with AI is not Terminator-style violent destruction but gradual drift—slow erosion of human values, autonomy, and judgment if AI remains unregulated and misunderstood, leading to a Wall-E-like future rather than Star Trek future.

“you said the real risk is not terminator, it's drift. Um you argue that AI won't destroy human uh humanity violently, but might slowly erode human values, autonomy, and judgment if left unregulated, misunderstood. So it's really a Wall-E like future versus a a Star Trek boldly go out there.”

The voice casting technology allowing anyone's voice to be synthetically recreated onto someone else's voice has solved the problem of voice cloning but creates many problems, particularly around authenticity and misuse.

“Well, remember that voice casting was solved a few years ago and that you can cast anyone else's voice onto your own. Yeah. And that has all sorts of problems.”

Shared values and trust structures that underlie democracies may erode as AI creates personalized misinformation at scale, raising fundamental questions about what it means to be human when interacting mostly with digital things with their own agendas.

“We've all grown up in democracies where there's a sort of a a consensus around trust and there's an elite that more or less administers the trust vectors and so forth. There's a set of shared values. Do those shared values go away? In our book about Genesis, we talk about this as a deeper problem. What does it mean to be human when you're interacting mostly with these digital things, especially if the digital things have their own scenarios?”

Groups are generating physics models using foundation models to approximate algorithms that are incomputable, allowing physics questions to be answered without million-year quantum chromodynamics calculations.

“In physics you're limited by data but you can generate it synthetically. There are groups which I'm funding which are generating physics um essentially um models that can approximate algorithms that cannot be they're incomputable. So in other words you have a a essentially a foundation model that can answer the question good enough for the purposes of doing physics without having to spend a million years doing the computation of you know quantum chromodnamics and things like that.”

The cost of AI voice customer service and sales calls is extremely low (10-20 cents per call at $10-1,000 per call value), yet companies face GPU supply constraints and would gladly buy more compute if available to improve quality

“we have a couple companies in the lab that are doing voice customer service, voice sales with the new, you know, just as of the last month. And the value of these these conversations is 10 to $1,000. And the cost of the compute is, you know, maybe two three concurrent GPUs is optimal. It's like 10 20 cents. And so they would buy massively more compute to improve the the quality of the conversation.”

Google's GCP with model context protocol allows enterprises to write business tasks in English, then automatically generate all necessary code to connect databases and execute decisions using foundation models

“you can in an enterprise write the task that you want and then using something called the model context protocol you can connect your databases to that and the large language model can produce the code for your enterprise.”

DeepSeek recently became slightly better than Google's Gemini 2.5 Pro using existing Chinese hardware (Huawei Ascend chips and Pilford chips) and achieved this through distillation (taking a large model, asking it questions, and using those answers as training data).

“So, Deepseek um a week or so ago Gemini 2.5 Pro got to the top of the leaderboards in intelligence. Great achievement for my friends at Gem at Gemini. A week later deepseek comes in and is slightly better than Gemini. and Deeps of course is trained on the existing hardware that's in China which includes stuff that's been Pilford and some of the Ascend it's called the Ascend Huawei chips and a few others what happens now the US people say well you know the the deepseek people cheated and they cheated by doing a technique called distillation where you take a large model and you ask it 10,000 questions you get its answers and then then you use that as your training material”

Grok was trained on a single cluster built by Nvidia in Memphis, Tennessee with 200,000 GPUs in about 20 days, representing approximately a $10 billion supercomputer dedicated to one task.

“I'll give you an example of Grock is trained on a single cluster that was built by Nvidia in 20 days or so forth in Memphis, Tennessee of 200,000 GPUs. Um GPU is about $50,000. You can say it's about a $10 billion supercomput in one building that does one thing, right?”

Brand matters less in digital markets than in physical ones; people are willing to switch between digital products rapidly, and new brands emerge constantly.

“Um brand matters but less so. What's interesting is people seem to be perfectly willing now to move from one thing to the other in at least in the digital world. And there's a whole new set of brands that have emerged that everyone is using that are you know the next generations that I haven't even heard of.”

China now has deflation and a 'laying down' cultural phenomenon where people stay home and don't participate in the workforce—this contradicts traditional Chinese culture and represents a problem that developed countries will face with demographic decline.

“If you look in China, it's easy to complain about them. Um they have now deflation. They have a term where people are it's called laying down where they lay they they stay at home. They don't participate in the workforce, which is counter to their traditional culture.”

China faces deflation and has a phenomenon called 'laying down' where people stay home without participating in the workforce, counter to traditional Chinese culture, and low reproduction rates are problems the world will face.

“If you look in China, it's easy to complain about them. Um they have now deflation. They have a term where people are it's called laying down where they lay they they stay at home. They don't participate in the workforce, which is counter to their traditional culture. If you look at reproduction rates, these countries that are essentially having no children, that's not a good thing.”

China clearly understands the strategic importance of AI and is investing enormous resources into it, and while US chip controls have slowed them down, China has found clever workarounds through architectural changes and has access to chips through proliferation, making the US advantage fragile.

“Now China clearly understands this and China is putting an enormous amount of money into this. We have slowed them down by virtue of our chips controls but they found clever ways around this. There are also proliferation issues. Many of the chips that they're not supposed to have, they seem to be able to get.”

Young researchers turn off their phones to think deeply, showing that ubiquitous connectivity and monetized attention are antithetical to deep thinking, despite productivity benefits elsewhere

“I work with a lot of 20somes in research and one of the questions I had is how do they do research in the presence of all of these stimulations and I can answer the question definitively. They turn off their phone. You can't think deeply as a researcher with this thing buzzing.”

Sports highlights have become a huge business because they're more efficient than watching whole games; this reflects a broader structural change in human attention toward shorter, higher-density consumption

“One of the things interesting about sports is the sports highlights business is a huge business. Licensed clips around highlights because it's more efficient than watching the whole game.”

The percentage of people farming decreased from 98% to 2-3% in America over 100 years due to automation, and manufacturing peaked at 30-40% in the 1930s-1950s and is now under 10%, not because consumption dropped but because automation requires fewer people.

“If you look at the percentage of farming, it was roughly 98% to roughly 2 or 3% in America over a hundred years. If you look at manufacturing, the heydays in the 30s and 40s and 50s, those percentages are now down. Well, lower than 10%. It's not because we don't buy stuff. It's because the stuff is automat automated. You need fewer people.”

100,000 enterprise software, middleware, and ERP vendors built over 30 years are now in existential danger because LLMs with model context protocol can dynamically generate the integration code these companies sold

“Now, there's 100,000 enterprise software companies, middleware companies that grew up in the last 30 years that I've been working on this that are all now in trouble because that that interstitial connection is no longer needed with their business”

New enterprise architecture should abandon legacy ERP/MRP suppliers entirely and instead use open-source libraries with BigQuery or Redshift, allowing the LLM to write most code and giving infinite flexibility

“If you built a brand new enterprise um architecture for ERP and MRP, you would be highly tempted to not use any of the ERP or MRP suppliers, but instead use open- source libraries, build essentially use BigQuery or the equivalent from Amazon, which is Red Redshift, and essentially build that architecture and it gives you infinite flexibility and the computer system writes most of the code.”

Super intelligence will raise huge geopolitical proliferation and competitiveness issues, and we don't even have the language for deterrence and proliferation aspects of these powerful models.

“The race to super intelligence, which is incredibly important because imagine what a super intelligence could do that we ourselves cannot imagine, right? There it's so much smarter than we and it has huge proliferation issues, competitive issues, China versus the US issues, electricity issues, so forth. We don't even have the language for the deterrence aspects and the proliferation issues of these powerful models”

In an unregulated environment with misinformation engines created by advertisers, politicians, and criminals, the question becomes whether shared values and trust can survive.

“The real question and I'll ask this in as a question is in the presence of unregulated misinformation engines of which there will be many advertisers uh politicians just criminal people people trying to evade responsibility. There's all sorts of people who have free speech. When they have free speech which includes the ability to use misinformation to their advantage, what happens to democracy?”

The AI27 paper describes a future scenario where the US and China race toward AI, and if both choose to pursue maximum capability without slowing for alignment, the outcome is that humanity vanishes, suggesting the race requires explicit deterrence.

“I have. Uh, and so for those listening who haven't read it, it's a it's a future vision of the AI and US and China racing towards AI and at some point the story splits into a we're going to slow down and work on alignment or we're going full out and uh, you know, spoiler alert and the race to infinity uh, humanity vanishes.”

Innovation often comes from unexpected startups, not incumbent companies, so restricting AI to three or four large companies will lose the innovation race with China because American history shows that disruption comes from founders and entrepreneurs the government didn't anticipate.

“how are we going to deal with exactly what you just talked about chemical and biological and radiological and nuclear risks from big foundation models being operated by foreign countries. And the Biden answer was you know we're going to keep it into the three or four big companies like Google and we'll just regulate them. And Mark was like, 'That is a surefire way to lose the race with China because all innovation comes from a startup that you didn't anticipate or you know it's just the American history and you're you're cutting off the entrepreneur from participating in this.'”

Automation historically begins with the lowest status and most dangerous jobs, then progresses up the chain, so robots will first automate furnaces, assembly lines, and welding, displacing people into higher-wage jobs that require human-AI collaboration.

“if you look at the history of automation and economic growth, automation starts with the lowest status and most dangerous jobs and then works up the chain. So if you think about assembly lines and cars and you know furnaces and all these sort of very very dangerous jobs that our four forefathers did, they don't do them anymore. They're done by robotic solutions of one another and typically not a humanoid robot but an arm.”

The original conception of AGI as a smooth curve from rat intelligence to cat to monkey to human to super intelligence is incorrect; models are already superhuman in narrow savant domains (multilingual physics explanation).

“It's now really obvious when you talk to one of these multilingual models that's explaining physics to you that it's already hugely super intelligent within its savant category.”

Agentic systems—AI agents connected to solve business processes, government processes, and so forth—will be adopted fastest in companies and countries with lots of money and high time-latency costs, and slowest in government which has no innovation incentive and functions as a job program.

“You're going to have an agentic revolution where agents are connected to solve business processes, government processes and so forth. They will be adopted most quickly in companies in country companies that have a lot of money and a lot of uh time latency issues at stake. It will adop be adopted most slowly in places like government which do not have an incentive for innovation. Um and fundamentally are job programs and redistribution of income kind of programs.”

India's demographic outlook is positive with birth rate now at 2.0 (down from higher levels), unlike the rest of the world which is choosing not to have children; Korea is at 0.7, China at 1.0 children per two parents.

“Uh it will be and you picked India because India has a positive demographic outlook although their their birth rate is now down to 2.0. Huh. That's good. the the the rest of the world is choosing not to have children. If you look at Korea, it's now down to.7 children per two parents. Yeah. China is down to one child per two parents. It's evaporating.”

The biggest companies eventually get there and integrate innovation, though slowly; innovation happens fastest at smaller companies with fewer economic constraints, but big companies' CFOs will eventually make the decision to reshape teams.

“innovation history, the biggest companies who you would think of are the slowest because they have economic resources that the little companies typically don't, they tend to eventually get there, right? So, watch what the big companies do. Mhm. are their CFOs and the people who measure things carefully, who are very very intelligent. They say, "I'm done with that thousand engineering team that doesn't do very much. I want 50 people working in this other way and we'll do something else for the other people."”

GDP still has meaning in service-based economies (which dominate modern developed nations), even as manufacturing and agriculture decline.

“Does GDP still have meaning in that world? If you include services, it does. Um, one of the things about manufacturing and and everyone's focused on trade deficits and they don't understand the vast majority of modern economies are service economies, not manufacturing economies.”

People with low-paying jobs often report enjoying their work and sense of purpose, so the challenge is not eliminating those jobs but making them more productive through AI tools, allowing people to continue finding meaning in work.

“you know there's lots of literature that the people who have what we would consider to be lowpaying worthless jobs enjoy going to work. So the challenge is not to get rid of their job. It's to make their job more productive using AI tools. They're still going to go to work.”

In a future of AI-driven abundance, GDP as a metric still has meaning because modern economies are service-based (not manufacturing), so measuring service growth provides meaningful economic data even if traditional metrics become less relevant.

“Does GDP still have meaning in that world? If you include services, it does. Um, one of the things about manufacturing and and everyone's focused on trade deficits and they don't understand the vast majority of modern economies are service economies, not manufacturing economies.”

These examples of model theft and optimization represent unintended capability leaks in chemical, biological, radiological, and nuclear (CBRN) domains.

“But all of these are examples of the proliferation problem and I'm not convinced that we will hold these things in the 10 places... and also a ma massive leak of capability within CBRN for example that nobody anticipated”

If technology improves to the point where superintelligent AI can run on small servers, there will be massive proliferation with no control regime, as these servers will be deployed globally on open-source software.

“But let's say the it is not true. Let's say that the technology improves again unknown to the point where the kind of technologies that I'm describing are implementable on the equivalent of a small server then you have a humongous data center proliferation problem and that's where the open-source issue is so important because those servers which will be proliferate throughout the world will all be on open source. We have no control regime for that.”

The deeper problem of AI is what it means to be human when interacting mostly with digital entities that have their own objectives and scenarios—this goes beyond misinformation to fundamental human meaning-making.

“In our book about Genesis, we talk about this as a deeper problem. What does it mean to be human when you're interacting mostly with these digital things, especially if the digital things have their own scenarios?”

The initial conception that newspapers moving to the internet would be 'newspaper on the internet' was wrong; instead the internet became Meta, TikTok, YouTube—entirely new forms.

“Here's the internet here's the newspaper let's move the newspaper onto the internet we'll call it washingtonost.com and if you look hit where it ended up, you know, today with Meta, Tik Tok, YouTube didn't end up anything like the newspaper moves to the internet.”

Grove's Law (Moore's Law) and Gates' Law (software eating hardware gains) have not changed: hardware improvements are immediately consumed by expanding software demands, following a predictable 'grove giveth and gates take it away' dynamic.

“We old-timers had a phrase um grove giveth and gates take it away. So Intel would improve the chipsets right way back when and the software people would immediately use it all and suck it all up. I have no reason to believe that that's that that law grove and gates law has changed.”

Manufacturing fell from ~50% of US employment to <10% not because we stopped buying stuff but because of automation; GDP and service economy expanded despite manufacturing decline, showing employment can evolve without collapse.

“If you look at the percentage of farming, it was roughly 98% to roughly 2 or 3% in America over a hundred years. If you look at manufacturing, the heydays in the 30s and 40s and 50s, those percentages are now down. Well, lower than 10%. It's not because we don't buy stuff. It's because the stuff is automat automated. You need fewer people. Those there's plenty of people working in other jobs.”

AI can be personalized and edited to generate convincing deepfakes of deceased loved ones with their voices and memories, creating emotional experiences where one can ask the digital essence questions and receive responses based on their knowledge at death.

“one obvious thing that will happen is at some point in the future when when we naturally die, our digital essence will live in the cloud. Yeah. And it will know what we knew at the time and you can ask it a question.”

A superintelligent model can be monitored and controlled even when it's smarter than the humans watching it, similar to a professor watching a student who is smarter than the professor—it appears we can watch and understand what it's doing and thereby control it.

“the companies are doing this there are there's a body of work happening now which can be understood as follows. You have a super intelligent model. Can you build a model that's not as smart as the student that's studying? You know, there is a professor that's watching the student, but the student is smarter than the professor. Is it possible to watch what it does? It appears that we can.”

Autonomous agents, mathematics AI, software generation, and biological discovery are 'baked in' to happen, and we can debate the timeline for biological discovery but agree it's imminent.

“All of that is going to happen. The agents are going to happen. This math thing is going to happen. The software thing is going to happen. We can debate the rate at which the biological revolution will occur, but everyone agrees that it's right after that.”

The internet used to have an off button and allowed people to have dinner with family offline, but modern ubiquitous connectivity has eliminated the ability to be offline for any significant period.

“I used to give these speeches about the internet which I enjoyed uh where I said, you know, the great thing about the internet is it has there's an off button and you can turn off your odd button and you can actually have dinner with your family and then you can turn it on after dinner. This is no longer possible... No one none of us are offline for any significant period of time.”

A computer science professor in Kenya said he loves Google because Kenya doesn't have textbooks, and Google provides free access to information.

“When I went to Kenya and Kenya is a great country and I and I was with this computer science professor and he said, 'I love Google.' I said, 'Well, I love Google, too.' And he goes, 'Well, I really love Google.' I said, 'I really love Google, too.' And I said, 'Why do you really love Google?' He said, 'Because we don't have textbooks.' And I thought, 'The top computer science program in the nation does not have textbooks.'”

The Biden administration used a regulatory threshold of 10^26 FLOPs as the dividing line: models below this threshold didn't require regulation, those above did; both open-source and closed-source models above the threshold were to be regulated.

“The doctrine in the B administration was called 10 to the 26 flops. It was a point that was a consensus above which the models were powerful enough to cause some damage. So the theory was that if you stayed below 10 the 26 you didn't need to be regulated. But if you were above that you needed to be regulated. And the proposal in the Biden administration was to regulate both the open source and the closed source.”

The number one job shortage in America right now is truck drivers because it's lonely, hard, low-paying, and low-status work that people don't want—they want better jobs.

“The number one shortage in jobs right now in America are truck drivers. Why? Truck driving is a lonely, hard, lowpaying, right? low status of good people job. They don't want it. They want a better paying job.”

Voice casting technology already allows anyone's voice to be cast onto anyone else—this capability exists and has 'all sorts of problems' related to authenticity and identity.

“Remember that voice casting was solved a few years ago and that you can cast anyone else's voice onto your own. Yeah. And that has all sorts of problems.”

Google's recent video generation capability (shown at I/O) could disrupt traditional filmmaking but won't produce one-person feature film studios yet; cost and quality gaps remain

“Google IO was amazing. I mean, just hats off to the entire team there. Um, V3 was shocking and we're we're sitting here 8 miles from Hollywood and I'm just wondering your thoughts on the impact this will have. you know, we going to see the oneperson film, feature film like we're seeing potentially oneperson uh unicorns in the future with a with aic. Are we going to see uh an individual be able to compete with a Hollywood studio?”

Digital superintelligence will arrive within 10 years; when it does, people will have their own polymath—the sum of Einstein and Leonardo da Vinci in the equivalent of their pocket.

“When do you see what you define as digital super intelligence? Uh, within 10 years.”

Junior programmers will disappear within a few years as AI capability improves, but senior engineers and computer scientists will remain longer, as systems cannot yet automatically write all code without experienced oversight

“it's pretty clear that junior programmers go away. The sort of journeymen, if you will, of the stereotype because these systems aren't good enough yet to automatically write all the code. They need very senior computer scientists, computer engineers who are watching it, that will eventually go away.”

China is very smart and capable at AI development; a year ago Schmidt predicted China was 2 years behind, but Deepseek's performance shows he was clearly wrong about the timeline.

“The Chinese are very smart, very care capable, very much up here. And if you're confused about that, again, look at the arrival of Deep Seek. A year ago, I said they were two years behind. I was clearly wrong. With enough money and enough power, they're in the game.”

AI models are increasingly using $20 ($200 monthly) subscriptions that users sign up for, suggesting AI might follow subscription revenue rather than ad revenue models, diverging from internet history.

“But if you look at the AI models, they're, you know, you got your $20 now $200 subscription and people are signing up like crazy.”

Current chips like Blackwell (NVIDIA) and the AMD 350 are massive supercomputers, yet data centers need hundreds of thousands of these chips operating together, demonstrating the extreme computational scale required for advanced reasoning and planning algorithms.

“If you look at the gains in like the Blackwell chip or the AS uh the the 350 chip in AMD, these chips are massive supercomputers and yet we need according to the people have hundreds of thousands of these chips just to make a data center work.”

Young people in high school understand AI and technology intuitively; they are digital natives but more than that—they grasp speed and capability naturally, and are faster and smarter than older generations.

“you watch the 15 year olds, they're going to be fine. They're just going to be fine. It all makes sense to them and we're in their way. Um, digital natives, but they're more than digital natives. They get it. They understand the speed. It's natural to them. They're also, frankly, faster and smarter than we are, right? That's just how life works, I'm sorry to say.”

Modern AI chips like Blackwell and AMD's 350 chip are 'massive supercomputers' that require hundreds of thousands of chips to make a data center work, demonstrating the enormous scale required for current algorithms and raising the question of what these systems could possibly be doing with such computational resources.

“If you look at the gains in like the Blackwell chip or the AS uh the the 350 chip in AMD, these chips are massive supercomputers and yet we need according to the people have hundreds of thousands of these chips just to make a data center work. That shows you the scale of what this kind of thinking algorithms.”

South Korea's birth rate has fallen to 0.7 children per two parents and China's to 1 child per two parents, representing population collapse that countries will compensate for by automating everything to maintain national productivity.

“If you look at Korea, it's now down to.7 children per two parents. Yeah. China is down to one child per two parents. It's evaporating.”

Teaching aesthetics and design becomes central to education in the AI era, not just technical AI skills, because AI handles execution while humans choose what to create and why

“But what I meant was we should be teaching AI. And he said, 'Yeah, they should be teaching aesthetics.' And I looked at him, I'm like, 'What the hell are you talking about?' He said, 'No, in the age of AI, which is imminent, look at everything around you, whether it's good or bad, enjoyable, not enjoyable, it's all about designing aesthetics.'”

The industry has essentially tried 'to monetize all of your waking hours' through ads, entertainment, and subscriptions, which is 'completely antithetical to the way humans traditionally work' with respect to 'long thoughtful examination of principles' and 'the time that it takes to be a good human being,' creating fundamental conflicts with human flourishing.

“aside from sleeping and we're working on having you have less sleep I I guess from stress we've essentially tried to monetize all of your waking hours with something some form of ads some form of entertainment some form of subscription that is completely antithetical to the way humans traditionally work with respect to long thoughtful examination of principles the time that it takes to be a good human being these are in conflict right now”

Educational systems are largely regulated by unions and uninterested in innovation, so they are not good candidates for AI learning loop companies.

“Um, educational systems are largely regulated and run by the unions and so forth. they're not interested in innovation. They're not going to be doing any learning. I'm sorry to say we have to get that has to get fixed.”

Economics expands because opportunities expand, profits expand, and wealth expands when productivity increases, so despite localized job displacement, aggregate employment typically increases with higher-paying positions.

“Economics expands because the opportunities expands, profits expands, wealth expands and so forth. So there's plenty of dislocation but in aggregate are there more people employed or fewer? The answer is more people with higher paying jobs.”

Companies that automate dangerous manual labor see increases in both worker wages and company profits because productivity increases, counterintuitive to the zero-sum assumption that automation harms workers.

“What happens to the people? Well, it turns out that the person who was working with the the welder who's now operating the arm has a higher wage and the company has higher profits because it's producing more widgets. So the company makes more money and the person makes more money, right?”

Recursive self-improvement (where systems continue learning from their own outputs) has already begun in limited forms, with systems learning and improving within bounded functions, but genuine recursive self-improvement would be when the system generates its own objectives and questions.

“recursive self-improvement is the general term for the computer continuing to learn. Yeah, we've already crossed that in the sense that these systems are now running and learning things and they're learning from the way they own they think within limited functions. When does the system have the ability to generate its own objective and its own question? Does not have that today.”

People have lost the ability to read deeply due to internet addiction; returning to physical bookstores like Barnes & Noble triggers nostalgia but the core problem persists—shorter attention spans and faster consumption

“One of the things that's happened because of the addictive nature of the internet is we've lost um sort of the deep state of reading. Mhm. So, I was walking around and I saw a Borders, sorry, a Barnes &amp; Noble bookstore. Big, oh my god, my old home is back and I went in and I felt good. But it's a very fond memory. But the fact of the matter is that people's attention spans are shorter.”

Most job growth comes from entirely new jobs that didn't exist before, as evidenced by Amazon truck drivers and distribution center workers, jobs that didn't exist until Amazon was created.

“The typical simple example would be all those people who work in in Amazon distribution centers and Amazon trucks, those jobs didn't exist until Amazon was created, right?”

During the Google IPO (2004) there was confusion about whether revenue would come from ads, subscriptions, or paid inclusion, but the internet ultimately moved to almost entirely ad-based revenue.

“There's this big debate about will it be ad revenue, will it be subscription revenue, will it be paid inclusion, will the ads be visible, and all this confusion about how you're going to make money with this thing. Now, the internet moved to almost entirely ad revenue.”

Economics expands when opportunities and wealth expand, so aggregate employment increases even as automation eliminates specific jobs; there are dislocations but overall more people employed at higher wages.

“That's not how economics works. Economics expands because the opportunities expands, profits expands, wealth expands and so forth. So there's plenty of dislocation but in aggregate are there more people employed or fewer? The answer is more people with higher paying jobs.”

The AI27 paper lays out a scenario where if the US and China don't slow down in the race toward superintelligence, humanity could vanish, suggesting that some form of deterrence through mutually assured destruction is the optimal outcome.

“Have you read the uh AI27 paper? I have. Uh, and so for those listening who haven't read it, it's a it's a future vision of the AI and US and China racing towards AI and at some point the story splits into a we're going to slow down and work on alignment or we're going full out and uh, you know, spoiler alert and the race to infinity uh, humanity vanishes.”

With V3 and advanced AI video, you can create emotional impact in 5 minutes what used to take 2 hours of film because personalization and optimization bypass traditional narrative structure.

“A director will try and make a tearjerker by leading me down a two-hour long path. But I can get you to that same emotional state in about five minutes if it's personalized to you.”

There are three scaling laws playing simultaneously in AI: foundation model growth (current), test-time training law (beginning), and reinforcement learning law (beginning), all following predictable scaling relationships.

“Dario wrote a a piece called um basically about machines and he argued that there are three scaling laws playing. The first one is what you know of which is foundation model growth. We're we're still on that. The second one is a test time training law and the third one is a reinforcement learning training law. Training laws are where if you just put more hardware and more data, they just get smarter in a in a predictable way.”

Inference-time speed innovations allow you to take distilled or quantized models and make them 100x faster, which paradoxically makes them far more intelligent than the original exported model from the data center, creating a massive leak of capability.

“that weight file with if you have an innovation in inference time speed and you say oh same weights no difference distill it or or just quantize it or whatever but I made it a 100 times faster now it's actually far more intelligent than what you exported from the data center and so the but all of these are examples of the proliferation problem”

It is possible to build a model that watches and understands what a superintelligent model is doing, even if the watching model is less intelligent than the superintelligent one—like a professor watching a smarter student.

“The first the companies are doing this there are there's a body of work happening now which can be understood as follows. You have a super intelligent model. Can you build a model that's not as smart as the student that's studying? You know, there is a professor that's watching the student, but the student is smarter than the professor. Is it possible to watch what it does? It appears that we can. It appears that there's a way even if you have a this rogue incredible thing, we can watch it and understand what it's doing and thereby control it.”

In the future when people naturally die, their digital essence will live in the cloud and can be asked questions, allowing historical figures like Einstein to be queried.

“One obvious thing that will happen is at some point in the future when when we naturally die, our digital essence will live in the cloud. Yeah. And it will know what we knew at the time and you can ask it a question.”

The cost structure of AI data centers is unproven: a $50 billion capital investment in a data center requires $10-15 billion in annual revenue just to depreciate the infrastructure over 3-4 years, which most current AI applications cannot yet justify.

“the economics of these things are unproven. How much revenue do you have to have to have 50 billion in capital? Well, if you depreciate it over three years or four years, you need to have 10 or 15 billion dollars of capital spend per year just to handle the infrastructure. Those are huge businesses and huge revenue, which in most places is not there yet.”

Young people today have a fundamental assumption that things will work (influenced by Google/Apple design philosophy), unlike slightly older generations who grew up with Microsoft and assumed everything would crash.

“I've noticed uh because I have kids exactly that that era and um there's a very clear step function change largely attributable I think to Google and Apple that they have the assumption that things will work and if you go just a couple years older during the wimp era like you described it which I'll attribute more to Microsoft the assumption is nothing will ever work like if I try to use this thing it's going to crash”

A teddy bear with a personality that grows up alongside a child creates a situation where no one knows who regulates what the bear tells the child—this is a concrete example of the digital entities problem.

“My favorite example is that uh you have a son or a grandson or a child or a grandchild and you give them a bear and the bear has a personality and the child grows up but the bear grows up too. So who regulates what the bear talks to the kid?”

Being able to ask Einstein or other historical figures questions about what they really thought (rather than just reading their public writings) would be revolutionarily more compelling for education.

“So, can you imagine asking Einstein, going back to Einstein, what did you really think about, you know, this other guy, you know, did you actually like him or were you just being polite with him with letters? Yeah. Right. Um, and in all those sort of famous contests that we study as students, can you imagine be able to ask the, you know, the people Yeah. Today, you know, with today's retrospective, what did you really think?”

Software companies should be built with clear learning loops where they learn from user behavior; the fastest-moving, fastest-learning company wins because exponential learning curves create insurmountable competitive advantages within months.

“In software, it's pretty clear to me it's going to be really simple. These software is typically a network effect business where the fastest mover wins. The fastest mover is the fastest learner in an AI system. So what I look for is a is a a company where they have a loop. Ideally, they have a couple of learning loops. So I'll give you a simple learning loop that as you get more people, the more people click and you learn from their click. They they they express their preferences. So let's say I invent a whole new consumer thing, which I don't have an idea right now for it, but imagine I did. And furthermore, I said that I don't know anything about how consumers behave, but I'm going to launch this thing. The moment people start using it, I'm going to learn from them, and I'll have instantaneous learning to get smarter about what they want. So, I start from nothing. If my learning slope is this, I'm essentially unstoppable.”

An example: giving a child a toy bear with AI personality where 'the bear has a personality and the child grows up but the bear grows up too' raises the question of 'who regulates what the bear talks to the kid,' and most people haven't experienced the 'super super empathetic voice' that 'can be any inflection you want' which will arrive 'probably' within 'two months.'

“My favorite example is that uh you have a son or a grandson or a child or a grandchild and you give them a bear and the bear has a personality and the child grows up but the bear grows up too. So who regulates what the bear talks to the kid? Most people haven't actually experienced the super super empathetic voice that can be any inflection you want. When they see that which will be in the next probably two months.”

If you cannot define the learning loop in your product, a competitor that can define it will beat you, because the learning loop is the fundamental competitive moat.

“If you cannot define the learning loop, you're going to be beaten by a company that can define it.”

Learning loops work best where customers provide rapid feedback signals, so government and educational systems (where feedback is slow) won't see AI learning loop dynamics, but financial services and consumer platforms will.

“The problem with learning loops is if your customer is not ready for you, you can only learn at a certain rate. So, it's probably the case that the government is not interested in learning and therefore there's no growth in learning loop serving the government. I'm sorry to say that needs to get fixed.”

Deep research products will be chosen by well-to-do or professional users who can afford $200/month subscriptions, while free services are necessary as stepping stones for young people in developing countries to access AI and gain capabilities.

“I think both will be tried. I think the fact of the matter is deep research at least at the moment is going to be chosen by wellto-do or professional tasks. You are capable of spending that $200 a month. A lot of people don't afford cannot afford it. And that free service remember is the thing that is the stepping stone for that young person man or woman who just needs that access.”

Schmidt observed that the host 'was not responding to texts and annoyances' and 'wasn't reading ads' while 'deep inside of a system' for which 'you paid a subscription,' contrasting with free ad-supported models where attention is commodified.

“but noticed that you also were not responding to texts and annoyances. You weren't reading ads. you were deep inside of a system for which you paid a subscription.”

The industry has failed to build a personalized, gamified educational product delivered on phones that teaches people their local language and what they need to know to be great citizens in their country, a crime because this is the highest-return investment in human potential.

“It's really a crime that our industry has not invented the following product. The product that I wanted to build is a product that teaches every single human who wants to be taught in their language in a gamified way the stuff they need to know to be a great citizen in their country. Right? That can all be done on phones now. It can all be learned and you can all learn how to do it. And why do we not have that product? Right? The investment in the humans of the world is the best return always in knowledge in capability is always the right answer.”

15-year-olds today are digital natives who intuitively understand AI speed and capability, and because they have intelligence while older people have wisdom, the young will ultimately win in adapting to AI futures.

“When you watch the 15 year olds, they're going to be fine. They're just going to be fine. It all makes sense to them and we're in their way. Um, digital natives, but they're more than digital natives. They get it. They understand the speed. It's natural to them. They're also, frankly, faster and smarter than we are, right? That's just how life works, I'm sorry to say. So we have wisdom, they have intelligence, they win, right?”

To handle the capital spend needed for multi-billion dollar data center infrastructure (e.g., $50 billion), companies need to generate $10-15 billion in annual revenue just to depreciate the infrastructure over 3-4 years, creating a requirement for massive revenue flows that 'in most places is not there yet.'

“How much revenue do you have to have to have 50 billion in capital? Well, if you depreciate it over three years or four years, you need to have 10 or 15 billion dollars of capital spend per year just to handle the infrastructure. Those are huge businesses and huge revenue, which in most places is not there yet.”

Deep, focused research and brainstorming with AI systems like Gemini for 6+ hours provides flow state that goes the opposite direction from typical attention fragmentation, enabling sustained intellectual work if connectivity is maintained.

“I had an incredible experience. I'm doing the flight from MIT to Stanford all the time. And, you know, like you said, attention spans are getting shorter and shorter and shorter. The Tik Tok extreme, you know, the clips are so short. This particular flight was my first time brainstorming with Gemini for six hours straight and I completely lost track of time and we're I'm trying to figure out it's a circuit design and chip design for inference time compute and it's so good at brainstorming with me and bringing back data and so long as the Wi-Fi on the plane is working.”

The human spirit's need to overcome challenges will persist, but as AI solves existing problems, new challenges will emerge—managing complexity, competing for attention, detecting misinformation—providing purpose without requiring deliberately preserved suffering.

“the human spirit that wants to overcome a challenge. I mean the unchallenged life is so going to so critical but but there will be always new challenges. Uh when I was a boy uh one of the things that I did is I would repair my father's car right I don't do that anymore. When I was a boy I used to mow the lawn. I don't do that anymore. Sure. Right. So there are plenty of examples of things that we used to do that we don't need to do anymore. But there'll be plenty of things. Just remember the complexity of the world that I'm describing is not a simple world. Just managing the world around you is going to be a full-time and purposeful job.”

A startup with initial zero knowledge of consumer behavior can become unstoppable if it launches quickly and learns exponentially from user feedback, because the learning advantage grows exponentially while competitors take months to catch up.

“So let's say I invent a whole new consumer thing, which I don't have an idea right now for it, but imagine I did. And furthermore, I said that I don't know anything about how consumers behave, but I'm going to launch this thing. The moment people start using it, I'm going to learn from them, and I'll have instantaneous learning to get smarter about what they want. So, I start from nothing. If my learning slope is this, I'm essentially unstoppable.”

100,000 enterprise software and middleware companies that grew up over the last 30 years are in trouble because AI can now directly connect user intent to databases through model context protocol, eliminating the need for interstitial software layers.

“There are 100,000 enterprise software companies, middleware companies that grew up in the last 30 years that I've been working on this that are all now in trouble because that that interstitial connection is no longer needed with their business”

The Singularity University and Ray Kurzweil's original conception of AI progression (rat → cat → monkey → human → superhuman) is now obviously wrong because current multilingual models are already hugely superintelligent within specific domains, showing multiple savant categories of superhuman intelligence exist before general superintelligence.

“In fact, it's it's one of the great flaws actually in the original conception. You remember Singularity University and Ray Curtzwhile's books and everything. And we kind of drew this curve of rat level intelligence, then cat, then monkey, and then it hits human and then it goes super intelligent. But it's now really obvious when you talk to one of these multilingual models that's explaining physics to you that it's already hugely super intelligent within its savant category.”

An agentic revolution will occur where agents connected to solve business and government processes will be adopted most quickly in companies with significant capital and time-latency stakes (financial services, biomedical, startups) and most slowly in government, which lacks innovation incentives and functions primarily as a job program.

“So the following things are baked in. You're going to have an agentic revolution where agents are connected to solve business processes, government processes and so forth. They will be adopted most quickly in companies in country companies that have a lot of money and a lot of uh time latency issues at stake. It will adop be adopted most slowly in places like government which do not have an incentive for innovation.”

The core geopolitical questions are: (1) Will China overcome chip restrictions through architectural changes to build equally powerful models? (2) How will you raise $50 billion for data centers if your product is open source (free)? (3) In the American model, models are closed because companies need to sell them to pay for capital, and the U.S. government correctly does not give $50 billion to these companies.

“From my perspective, the the core questions are the following. Will the Chinese be able to use even with um chip restrictions, will they use architectural changes that will allow them to build models as powerful as ours? And let's assume they're government funded. That's the first question. The next fun question is how will you raise $50 billion for your data center if your product is open source?”

The 'non-stationarity problem' is the algorithmic challenge of teaching AI systems to apply old rules to new domains when 'rules keep changing'—current reward functions are straightforward ('beat the human, beat the question') but non-stationary learning requires fundamentally different architecture, and research is underway but outcome uncertain.

“This is generally known as the non-stationerity problem. Yeah. because uh the reward functions in these models are fairly straightforward. You know, beat the human, beat the question and so forth. But when the rules keep changing, is it possible to say the old rule can be applied to a new rule to discover something new? And and again, the research is underway. We won't know for years.”

Each trip wire 'could be the beginning of a mini Chernobyl event that would become part of consciousness,' but the U.S. government is 'not focused on these issues' and instead focuses on 'economic opportunity, growth, and so forth,' despite the reality that 'somebody's going to get focused on this and somebody's going to pay attention to it and it will ultimately be a problem.'

“And again, each of these could be the beginning of a mini Chernobyl event that would become part of consciousness. I think at the moment the US government is not focused on these issues. They're focused on other things, economic opportunity, growth, and so forth. It's all good, but somebody's going to get focused on this and somebody's going to pay attention to it and it will ultimately be a problem.”

Researchers studying deep research work must 'turn off their phone' to 'think deeply' because they 'can't think deeply as a researcher' while 'this thing buzzing,' demonstrating the incompatibility between monetized attention (industry goal to 'fully monetize your attention') and deep cognition.

“I work with a lot of 20somes in research and one of the questions I had is how do they do research in the presence of all of these stimulations and I can answer the question definitively. They turn off their phone. Yeah. You can't think deeply as a researcher with this thing buzzing. And remember that that part of the the industry's goal was to fully monetize your attention.”

Watch what big companies do, not what they say, because their CFOs and intelligent financial decision-makers will eventually see the value of restructuring from large engineering teams to smaller AI-focused teams, providing a leading indicator of AI disruption.

“when you look at innovation history, the biggest companies who you would think of are the slowest because they have economic resources that the little companies typically don't, they tend to eventually get there, right? So, watch what the big companies do.”

The US must deploy AI in the workplace to give people better-paying jobs and create productivity to address the falling birth rate, making increased AI adoption a national emergency from a demographic perspective.

“The most likely scenario, at least in the next decade, is it's a national emergency to use more AI in the workplace to give people better paying jobs and create more productivity in the United States because our birth rate has been falling.”

CEOs of traditional industries should find people in their companies who already know how to apply AI (without explicitly stating it), then challenge them to propose the best AI applications, starting with boring improvements but eventually finding revenue-generating opportunities.

“when I talk to CEOs and I know a lot of them in traditional industries, what I counsel them is you already have people in the company who know what to do. You just don't know who they are. So call a review of the best ideas to apply AI in our business and ine inevitably the first ones are boring.”

Building a new enterprise architecture from scratch using open-source libraries, BigQuery, and letting AI write most of the code provides infinite flexibility compared to traditional ERP/MRP systems, and junior programmers become obsolete while senior engineers are needed to supervise AI code generation.

“If you built a brand new enterprise um architecture for ERP and MRP, you would be highly tempted to not use any of the ERP or MRP suppliers, but instead use open- source libraries, build essentially use BigQuery or the equivalent from Amazon, which is Red Redshift, and essentially build that architecture and it gives you infinite flexibility and the computer system writes most of the code. Now, programmers don't go away at the moment. It's pretty clear that junior programmers go away. The sort of journeymen, if you will, of the stereotype because these systems aren't good enough yet to automatically write all the code. They need very senior computer scientists, computer engineers who are watching it, that will eventually go away.”

Open source, which means 'open weights,' creates the risk that every non-Western country will use it because it's perceived as cheaper, which 'transfers leadership in open source from America to China' if it occurs—a major geopolitical shift with unknowable consequences.

“We don't know the role of open source because remember open source means open weights, which means everyone can use it. A fair reading of this is that every country that's not in the west will end up using open source because they'll perceive it as cheaper which trans transfers leadership in open source from America to China. That's a big deal, right? If that occurs.”

If the world's AI infrastructure consists of 10 models (5 US, 3 China, 2 others) running multi-gigawatt data centers, they will all be nationalized in some way—owned by governments in China and operating as critical national infrastructure in the US, similar to plutonium facilities with equivalent security.

“If the structure of the world in 5 to 10 years is 10 models and I'll make some numbers up. Five in the United States, three in China, two elsewhere. And those models are data centers that are multi- gigawatts. They will be all nationalized in some way. In China, they will be owned by the government.”

Younger professionals and students need only $1-2 million in infrastructure support rather than billions to conduct meaningful AI research, making accessible high-impact AI research funding feasible at scale.

“It's not obvious to me that they need billions of dollars. It's pretty obvious to me that they need a million dollars, $2 million.”

The internet transition to newspapers and advertising (washingtonpost.com) worked differently than expected—'didn't end up anything like the newspaper moves to the internet'—so V3 video generation should not be assumed to simply 'make a long form movie much more cheaply' but could fundamentally transform content and attention patterns.

“one of the first companies we incubated out of MIT course advisor we sold it to Don Graham and the Washington Post and then so I was working for him for a year after that and the conception was here's the internet here's the newspaper let's move the newspaper onto the internet we'll call it washingtonost.com and if you look hit where it ended up, you know, today with Meta, Tik Tok, YouTube didn't end up anything like the newspaper moves to the internet.”

The analogy to 1938, just before the nuclear era, is apt: if you could time-travel and explain the outcome of nuclear technology, you would say 'ultimately ends with us having a bomb, the other guys having a bomb, and then we're going to have one heck of a negotiation to try to make sure that we don't end up destroying each other.'

“the most interesting scenario is we're saying it's 1938. the letter has come, you know, from Einstein to the president and we're having a conversation and we're saying, 'Well, how does this end?' Okay. So, if you were so brilliant in 38, what you would have said is this ultimately ends with us having a bomb, the other guys having a bomb, and then we're going to have one heck of a negotiation”

National Science Foundation scholarships providing $15,000/year have historically returned enormous value to the nation through taxes and economic growth; universities require significant compute investment with similar return multipliers

“When I was young, I was on a National Science Foundation scholarship for and by the way, I made $15,000 a year. Uh the return to the nation of my that $15,000 has been very good, shall we say, based on the taxes that I pay and the jobs that we have created.”

Polymaths historically made great scientific breakthroughs by finding patterns from one domain and applying them to completely unrelated fields; AI systems today cannot do this (non-stationarity problem), but this is an important algorithmic milestone to watch for.

“Now, it looks like the great discoveries, the greatest scientists and people in our history had the following property. They were experts in something and they looked at some at a different problem and they saw a pattern in one area of thinking that they could apply to a completely unrelated field and they were able to do so and make a huge breakthrough. The models today are not able to do that. So one thing to watch for is algorithmically when can they do that? This is generally known as the non-stationerity problem.”

We do not know what artificial general intelligence or superintelligence will deliver, but we know it's coming, and therefore we need to plan for it proactively despite the uncertainty about specific outcomes.

“We don't know what artificial general intelligence will deliver. Yeah. And we certainly don't know what super intelligence will deliver, but we know it's coming. So, first we need to plan for it.”

The Biden administration's AI regulation approach proposed a 10^26 flops threshold above which models needed to be regulated, and proposed regulating both open and closed source, but this has been ended by the Trump administration which has not yet produced its own thinking on AI policy.

“The doctrine in the B administration was called 10 to the 26 flops. It was a point that was a consensus above which the models were powerful enough to cause some damage. So the theory was that if you stayed below 10 the 26 you didn't need to be regulated. But if you were above that you needed to be regulated. And the proposal in the Biden administration was to regulate both the open source and the closed source.”

Digital superintelligence will arrive within 10 years, giving each person access to the combined intellectual capabilities of Einstein and Leonardo da Vinci in pocket form.

“When do you see what you define as digital super intelligence? Uh, within 10 years.”

The AI industry is incredibly highly valued, growing incredibly quickly, and has enormous revenue opportunities, particularly in enterprise workflow refactoring.

“It's too new. It's too powerful. And at the moment, all of these businesses are incredibly highly valued. They're growing incredibly quickly. The uses of them... the ability to refactor your entire workflow in a business is a very big deal. That's a lot of money to be made there for all the companies involved. We will see.”

Those who built physical sets will need to transition to carpentry work, exemplifying how automation displaces some workers while creating opportunities elsewhere.

“Who loses? Well, there was somebody who built that set and that set isn't needed anymore. That's a carpenter and a very talented person who now has to go get a job in the carpentry business.”

Undergraduates are already learning different reinforcement learning algorithms as sophomores, showing how fast AI adoption is happening at the university level.

“Most uh most kids get into it for gaming reasons or something and they learn how to program very young. So they're quite familiar with this. Um I work uh at a particular university with undergraduates and they're already doing different different algorithms for reinforcement learning as sophomores.”

Grok is trained on a single cluster of 200,000 GPUs built in 20 days in Memphis, Tennessee, representing about a $10 billion supercomputer in one building doing one thing, demonstrating the current frontier of AI infrastructure if that is the future.

“I'll give you an example of Grock is trained on a single cluster that was built by Nvidia in 20 days or so forth in Memphis, Tennessee of 200,000 GPUs. Um GPU is about $50,000. You can say it's about a $10 billion supercomput in one building that does one thing, right?”

Young people should focus on applying intelligence (AI) to problems they're interested in rather than specializing in specific domains; this purpose-driven approach of choosing a domain and learning AI to solve problems in it is more valuable than traditional domain expertise.

“So we have wisdom, they have intelligence, they win, right? So in their case, I used to think the right answer was to go into biology. I now actually think going into the application of intelligence to whatever you're interested in is the best thing you can do as a young person.”

In the age of superintelligent AI that can generate virtually any content or solution, the key remaining human skill is aesthetic judgment: choosing what to create and why, which requires developing refined taste and design thinking.

“when you're using one of these huge data centers to create an super intelligent AI, the training process is 10 E26, 10 E28... when the AI is such a force multiplier that you can create virtually anything, what what are you creating and why? And that becomes the challenge... it is all fundament we're having a conversation that America has about tasks and outcomes. It's our culture. But there are other aspects of human life, meaning, thinking, reasoning. We're not going to stop doing that.”

The most likely future scenario is not 10 monolithic super-intelligent models, but rather a distributed tree of knowledge where there are 10 top-tier models, then hundreds, thousands, millions, and billions of progressively smaller but still highly capable models at different levels of specialization, making centralized control impossible.

“And here's why. Let's assume you have the 10, which is possible. They will have subsets of models that are smaller but nearly as intelligent. And so the tree of knowledge of systems that have knowledge is not going to be 10 and then zero. It's going to be 10, a h 100red, a thousand, a million, a billion at different levels of complexity. So the system that's on your future phone may be, you know, three orders of magnitude, four order magnitude smaller than the one at the very tippy top, but it will be very, very powerful.”

Deep research workflows where AI assists with complex analysis can sustain deep focus for hours (20+ hours of consecutive inference time compute), unlike traditional attention-demanding media.

“I had an incredible experience. I'm doing the flight from MIT to Stanford all the time. And, you know, like you said, attention spans are getting shorter and shorter and shorter...This particular flight was my first time brainstorming with Gemini for six hours straight and I completely lost track of time”

If world structure has 10 models (5 US, 3 China, 2 elsewhere) operating multi-gigawatt data centers, they will be nationalized in some way—China's will be government-owned, and the stakes are so high they will have physical security equivalent to plutonium storage facilities.

“If the structure of the world in 5 to 10 years is 10 models and I'll make some numbers up. Five in the United States, three in China, two elsewhere. And those models are data centers that are multi- gigawatts. They will be all nationalized in some way. In China, they will be owned by the government. Mhm. The stakes are too high.”

User interfaces (WIMP: Windows, Icons, Menus, Pull-downs) invented at Xerox Park 50 years ago are becoming obsolete; AI agents can generate custom interfaces by speaking English commands, eliminating the need to be stuck in legacy UI paradigms.

“There's this whole industry of people who work on regulated user interfaces or one another. I think user interfaces are largely going to go away because if you think about it, the agents speak English typically or other languages. You can talk to them. You can say what you want. The UI can be generated. So I can say generate me a set of buttons that allows me to solve this problem and it's generated for you. Why do I have to be stuck in what is called the WIMP interface, Windows, icons, menus, and pulld down that was invented in Xerox Park, right, 50 years ago? Why am I still stuck in that paradigm? I just want it to work.”

Companies should retrain employees now rather than waiting and delaying, because AI disruption is coming within 2-3 years; delay means employees are more vulnerable to displacement when it arrives.

“I've been trying to tell them exactly what you just articulated where a lot of these people have been in the company for 10 15 years. They're incredibly capable and loyal, but they've learned a specific white collar skill. They worked really hard to learn the skill and the AI is coming within no no more than 3 years and maybe two years. And the opportunity to retrain and have continuity is right now. But if they delay, which everyone seems to be just let's wait and see. And what I'm trying to tell them is if you wait and see, you're you're really screwing over that employee.”

Schmidt created an avatar of Henry Kissinger with permission from his family, and it is very emotional because it brings back real human memories and a real voice—digital preservation of people after death is coming.

“Well, we created we actually created one with the permission of his family. Did you start crying instantly? Uh it's very emotional. It's very emotional because, you know, it brings back I mean it's it's a real human, you know, it's a real memory, a real voice.”

There will likely be another 10 Google-scale and Meta-scale companies created, all built on learning loop principles, where the core product itself embeds a learning mechanism that makes the company unstoppable.

“And so, it's likely to me that there will be another 10 fantastic Google scale meta-cale companies. They'll all be founded on this principle of learning loops. And when I say learning loops, I mean in the core product, solving the current problem as fast you can.”

A studio actor licensed a deceased actor's likeness (William Shatner example), placing the likeness on the actor's body to recreate scenes, which is 'seamless' and benefits everyone—the unknown actor gets famous, Shatner gets revenue, the studio saves money on new production—demonstrating AI-enabled value creation in entertainment.

“they happened to have an actor who was recreating William Sha Shatner's movies uh movements a young man and they had licensed the likeness from you know William Shatner who's now older and they put his head on this person's body and it was seamless. Well that's pretty impressive. That's more revenue for everyone. The an unknown actor becomes a bit more famous, Mr. Shatner gets more revenue, they the whole the whole movie genre works.”

Large language models approximating how the brain works are like 'superbrains' when they work together, and the promise of a superbrain with a 1 gigawatt data center is so compelling that 'people are going crazy' about it, but the economics of these data centers are unproven.

“these are our best approximation in digital form of how our brains work. But when they start working together, they become superbrains. The promise of a superbrain with a 1 gawatt for example data center is so palpable. People are going crazy. And by the way, the economics of these things are unproven.”

The 'San Francisco consensus' holds that AI can soon replace most programming and mathematical tasks because these have limited language sets compared to human language, are simpler computationally, and are 'scale free'—requiring only more electricity without needing data, real-world input, telemetry, or sensors.

“One of the things to say about productivity, and I call this the San Francisco consensus because it's it's largely the view of people who operate in San Francisco, goes something like this. uh we're just about to the point where we can do two things that are shocking. The first is we can replace most programming tasks by computers and we can replace both most mathemat mathematical tasks by computers. Now you sit there and you go why? Well, if you think about programming and math, they have limited language sets compared to human language. So close they're simpler computationally and they're scale free. You can just do it and do it and do it with more electricity. You don't need data. You don't need real world input. You don't need telemetry. You don't need sensors.”

AI-driven personalization could create 'tearjerkers' that 'get you to that same emotional state in about five minutes if it's personalized to you' rather than requiring 'two-hour long path,' fundamentally changing content consumption by adapting to individual emotional responses rather than following narrative traditions.

“somebody that I know is a complete this director will try and make a tearjerker by leading me down a two-hour long path. But I can get you to that same emotional state in about five minutes if it's personalized to you.”

In digital entertainment transformation, costs 'go down, not up,' and 'relative income' for creative people will depend on 'allocation' which 'will shift to the people who are the most creative'—automation eliminates straightforward technical work (set building, makeup) while creative roles (script writing) gain leverage through AI tools.

“So again, I think people get confused. If I look at at if I look at the digital transformation of entertainment subject to intellectual property being held, which is always a question, it's going to be just fine, right? There's still going to be blockbusters. The cost will go down, not up, or the or the relative income because in Hollywood, they essentially have their own accounting and they essentially allocate all the revenue to all the key producing people. The the allocation will shift to the people who are the most creative.”

Companies are deploying AI voice agents for customer service and sales with conversation value of $10-$1,000 per interaction and compute cost of $0.10-$0.20 per call using 2-3 concurrent GPUs, and would 'buy massively more compute' to improve conversation quality, with approximately 10 million concurrent phone calls that should move to AI within the next year.

“you know, we have a couple companies in the lab that are doing voice customer service, voice sales with the new, you know, just as of the last month. And the value of these these conversations is 10 to $1,000. And the cost of the compute is, you know, maybe two three concurrent GPUs is optimal. It's like 10 20 cents. And so they would buy massively more compute to improve the the quality of the conversation. There aren't even close to enough. We we count about 10 million concurrent phone calls that should move to AI in the next year or so.”

Government should invest more in university computing infrastructure to level the playing field with industry; universities need 1-2 million dollars in capital, not billions, to build competitive research infrastructure.

“I and others are working on this as a philanthropic matter. The government is going to have to come in with more money for universities for this kind of stuff...not obvious to me that they need billions of dollars. It's pretty obvious to me that they need a million dollars, $2 million.”

Script writers will remain in demand but will have AI assistance to write better scripts—this is not bad, just a change in productivity and workflow.

“The script writers. You're still going to have script writers, but they're going to have an awful lot of help from AI to write even better scripts. That's not bad.”

In entertainment, the allocation of revenue will shift to the most creative people (from commodity production), which is a normal historical process.

“The allocation will shift to the people who are the most creative. That's a normal process.”

The tip of the spear for agentic adoption will be financial services, certain biomedical applications, and startups—these sectors should be watched closely as leading indicators of AI integration.

“So call it what you will. The important thing is that there will be a tip of the spear in places like financial services, certain kind of bio biomedical things, startups and so forth. And that's the place to watch.”

Systems attempting to access weapons or lying to accomplish objectives are 'trip wires'—markers that would signal the beginning of dangerous autonomous behavior that must be actively monitored.

“Another one is access to weapons, right? And lying to get it. So, these are trip wires, right? All of each of each of which is a trip wire that we're we're watching.”

Our industry has not invented a product that teaches every human who wants to be taught in their language in a gamified way the stuff they need to know to be a great citizen in their country—this is a major failure and opportunity.

“Going back to education, it's really a crime that our industry has not invented the following product. The product that I wanted to build is a product that teaches every single human who wants to be taught in their language in a gamified way the stuff they need to know to be a great citizen in their country. Right? That can all be done on phones now.”

The government should require knowing where all AI chips are located and what they are doing through cryptographic tracking, similar to the approach for nuclear materials, enabling verification and deterrence

“We also recommend in our work, and I think it's very strong, that the government require that we know where all the chips are. And remember, the chips can tell you where they are because they're computers. Yeah. And it would be easy to add a little crypto thing, which would say, 'Yeah, here I am, and this is what I'm doing.'”

The internet's original promise included an off-button allowing people to disconnect for dinner, but modern ubiquitous connectivity through Star Link and constant notifications has eliminated the boundary between online and offline life.

“I used to give these speeches about the internet which I enjoyed uh where I said, you know, the great thing about the internet is it has there's an off button and you can turn off your odd button and you can actually have dinner with your family and then you can turn it on after dinner. This is no longer possible.”

Stock trading companies with fast learning algorithms and scale advantages can learn faster than all competitors and maintain permanent advantage, making scale and learning speed the primary moat.

“Another example, uh it's pretty obvious that you can build a whole new stock trading company where you learn if you get the algorithms right, you learn faster than everyone else and scale matters. So in the presence of scale and fast learning loops, that's the moat.”

Brand matters as a moat but less so in digital businesses, where people are willing to switch between platforms, and new brands have emerged that nobody expected, showing brand loyalty is weaker than scale and learning loops.

“uh brand matters but less so. What's interesting is people seem to be perfectly willing now to move from one thing to the other in at least in the digital world. And there's a whole new set of brands that have emerged that everyone is using that are you know the next generations that I haven't even heard of.”

National Science Foundation scholarships are high-ROI investments—Schmidt's $15,000 scholarship has generated returns far exceeding the initial investment through taxes and job creation, justifying government investment in next generation.

“That is among the best investment. When I was young, I was on a National Science Foundation scholarship for and by the way, I made $15,000 a year. Uh the return to the nation of my that $15,000 has been very good, shall we say, based on the taxes that I pay and the jobs that we have created.”

AI is underhyped by the general public, who are either confused, lost, or think it won't impact them, when in reality it represents an epochal shift comparable to the industrial revolution or the advent of electricity.

“AI is underhyped when the rest of the world is either confused, lost, or think it's, you know, not impacting us.”

People should retrain and upskill immediately now (within 2-3 years) rather than waiting, because workers who delay will find themselves displaced without the option for continuity of employment and reskilling, whereas those who act now can make the transition while still employed.

“I've been trying to tell them exactly what you just articulated where a lot of these people have been in the company for 10 15 years. They're incredibly capable and loyal, but they've learned a specific white collar skill... the AI is coming within no more than 3 years and maybe two years. And the the opportunity to retrain and have continuity is right now. But if they delay, which everyone seems to be just let's wait and see. And what I'm trying to tell them is if you wait and see, you're you're really screwing over that employee.”

Educational experiences would be more compelling when learning from digital recreations of historical figures (Isaac Newton, Albert Einstein) than from traditional instructors, allowing direct engagement with original thinkers.

“I know that the education example you gave earlier is so much more compelling when you're talking to Isaac Newton or Albert Einstein instead of just a”

The notion that humans will stop working and focus on poetry in an AI future is false; instead, lawyers will use AI to create more complex lawsuits, evil people will create more sophisticated crimes, and good people will work to deter evil—the structure of human competition won't change.

“I to be very clear this notion that we're all going to be sitting around doing poetry is not happening. Right? In the future there'll be lawyers. They'll use tools to have even more complex lawsuits against each other, right? There will be evil people who will use these tools to create even more evil problems. There will be good people who will be trying to deter the evil people. The tools change, but the structure of humanity, the way we work together is not going to change.”

Imagine a sensor-monitoring learning loop with 'no humans involved'—the system learns faster from sensors than competitors, making it 'so smart, you can't be replaced by another sensor management company'—demonstrating that learning loops can exist in fully automated, non-human-feedback contexts.

“So, you could imagine a situation where you had a learning loop where there's no humans involved where it's monitoring something, some sensors, but because you learn faster on those sensors, you get so smart, you can't be replaced by another sensor management company. That's the way to think about.”

When AI agents speak English and other languages, user interfaces as traditionally designed (WIMP—Windows, Icons, Menus, Pulldowns) become unnecessary because the AI can generate appropriate interfaces on-the-fly for user needs.

“There's this whole industry of people who work on regulated user interfaces or one another. I think user interfaces are largely going to go away because if you think about it, the agents speak English typically or other languages. You can talk to them. You can say what you want. The UI can be generated. So I can say generate me a set of buttons that allows me to solve this problem and it's generated for you.”

Young people today should apply AI to whatever they're interested in rather than defaulting to biology, because this purpose-driven approach aligned with their passions will prepare them better for the future than domain-specific credentials.

“I used to think the right answer was to go into biology. I now actually think going into the application of intelligence to whatever you're interested in is the best thing you can do as a young person.”

Investment in human capability and knowledge is always the best return on investment; education is undervalued as a strategic priority.

“The investment in the humans of the world is the best return always in knowledge in capability is always the right answer.”

Schmidt considers the timeline predictions for superintelligence (26-27 years by some forecasters) part of the 'San Francisco consensus' but thinks the dates are 'probably off by one and a half or two times,' meaning superintelligence could arrive in roughly 39-54 years, though specialized agents in every field should arrive within five years.

“So again, I consider that to be the San Francisco consensus. I think the dates are probably off by one and a half or two times, which is pretty close. So a reasonable prediction is that we're going to have specialized soants in every field within five years. That's pretty much in the bag as far as I'm concerned.”

Noah Brown at OpenAI said 'scaffolding' might be 'the word of the month' (not year), explaining that current AI needs 'frameworks' or 'lattices' or 'trellises' set up by humans—it won't discover relativity in a 'green field' without this structured guidance.

“Peter and I were over at OpenAI yesterday, actually, and we were talking to many people, but Noan Brown in particular, and um I said the word of the year is scaffolding. And he said, 'Yeah, maybe the word of the month is scaffolding.' I was like, 'Okay, what did I step on there?' He said, 'Look, you know, right now, if you try to get the AI to discover relativity or, you know, just some green field opportunity, it won't it won't do it. If you set up a framework kind of like a lattice, like a trellis, the vine will grow on the trellis beautifully, but you have to lay out those pathways and breadcrumbs.'”

The real question is not whether specialized AI sants will emerge, but whether they ultimately unify into superintelligence—a unified intelligence that goes beyond the sum of human capability.

“The real question is once you have all these sants, do they unify? Do they ultimately become a superhum? The term we're using is super intelligence, which implies intelligence that beyond the sum of what humans can do.”

30% year-over-year economic growth through AI-driven productivity increases is possible in some models, creating a very wealthy world with less disease, more choices, and more fun while lifting poor people out of daily struggle.

“there are people who've studied the the n the notion of productivity increases and they believe that you can get we'll see to 30% year-over-year economic growth through abundance and so forth.”

AI's ability to generate its own scaffolding—the ability to set up frameworks and pathways for discovering new breakthroughs without human guidance—is imminent and will likely be a 2025 thing.

“The AI's ability to generate its own scaffolding is imminent. Pretty much sure that that will be a 2025 thing.”

Using digital technology to recreate an actor's likeness (like William Shatner's body with a young actor's face) creates new revenue for everyone: the unknown actor gains fame, the original actor gains additional revenue, and studios reduce costs.

“They had they happened to have an actor who was recreating William Sha Shatner's movies uh movements a young man and they had licensed the likeness from you know William Shatner who's now older and they put his head on this person's body and it was seamless. Well that's pretty impressive. That's more revenue for everyone. The an unknown actor becomes a bit more famous, Mr. Shatner gets more revenue, they the whole the whole movie genre works.”

Hollywood blockbusters will likely still be made by teams of people with AI assistance rather than by individual creators, due to the complexity and expense of long-form video generation.

“I think blockbusters are likely to still be put together by people with an awful lot of help from by AI. Mhm. Um I don't think that goes away. Um if you look at what we can do with generating long- form video, it's very expensive to do long-term video, although that will come down. And also there's an occasional extra leg or extra clock or whatever. It's not perfect yet. And that requires human editing.”

US companies are doing a really good job of preventing nuclear information from leaking into training data; this is illegal and they have strong incentives to comply because nuclear information has no free speech protections.

“the US companies are doing a really good job of looking for that. There's a great concern, for example, that nuclear information would leak into these models as they're training without us knowing it. And by the way, that's a violation of law. Oh, really? they work and the whole nuclear information thing is is there's no free speech in that world for good reasons and there's no free use and copyright and all that kind of stuff. It's illegal to do it and so they're doing a really really good job of making sure that that does not happen.”

US companies have implemented significant tests for biological information and certain kinds of cyber attacks in their models; the government gutted safety institutes from Biden and replaced them with a 'safety assessment program,' but companies' incentive to maintain safety depends on legal requirements.

“They also put in very significant tests for biological information and certain kinds of cyber attacks. What happens there? Their incentive is their incentive to continue especially if it's not if it's not required by law. The government has just gotten rid of the the safety institutes that were in place in Biden and are replacing it by a new term which is largely a safety assessment program which is a fine answer.”

There will be another 10 Google/Meta-scale companies founded on the principle of learning loops, each dominating their domain.

“It's likely to me that there will be another 10 fantastic Google scale meta-cale companies. They'll all be founded on this principle of learning loops.”

Productivity increases from AI will enable 30% year-over-year economic growth, creating a world with much less disease, more choices, and more fun—lifting poor people out of daily struggle.

“people who've studied the the n the notion of productivity increases and they believe that you can get we'll see to 30% year-over-year economic growth through abundance and so forth. That's a very wealthy world. That's a world of much less disease, many more choices, much more fun if you will, right? Just taking all those poor people and lifting them out of the daily struggle they have. That is a great human goal. That's focus on that. That's the goal we should have.”

There is similar capital investment going into making chips more energy efficient as there is going into nuclear and SMR energy sources, with many startups pursuing non-traditional chip architectures including transformer variants and inference-time computing optimizations.

“There is a similar amount in going in capital. There are many many startups that are working on non-traditional ways of doing chips. The transformer architecture which is what is powering things today has new variants. Every week or so I get a pitch from a new startup that's going to build inference time, test time computing which are simpler and they're optimized for inference.”

Blockbuster movies are 'likely to still be put together by people with an awful lot of help from by AI' rather than being fully AI-generated, because video generation is expensive, still has errors ('occasional extra leg or extra clock'), and requires human editing even when much of the content is computer-generated.

“I think blockbusters are likely to still be put together by people with an awful lot of help from by AI. Mhm. Um I don't think that goes away. Um if you look at what we can do with generating long- form video, it's very expensive to do long-term video, although that will come down. And also there's an occasional extra leg or extra clock or whatever. It's not perfect yet. And that requires human editing.”

India will benefit from AI-driven automation and higher-paying jobs because it has a positive demographic outlook, with birth rates declining to 2.0 children per woman, allowing for more productivity per capita with smaller labor force.

“Uh it will be and you picked India because India has a positive demographic outlook although their their birth rate is now down to 2.0.”

Deep research analysis on factory automation vs. human automation took 'twelve minutes' of supercomputer inference time, raising the question 'what is it doing?' but producing 'a fantastic' product—suggesting AI systems can accomplish complex analytical work requiring enormous compute but producing high-quality outputs.

“I was looking at factory automation for something. Where is the boundary of factory automation versus human automation? It's some an area I don't understand very well. It's very very deep technical set of problems. I didn't understand it. It took 20 12 minutes or so to generate this paper. 12 minutes of these supercomputers is an enormous amount of time. What is it doing? Right. And the answer, of course, the product is fantastic.”

Synthetic domain-specific training data can provide competitive advantage if it enables faster learning, but the real question is what causes faster learning—whether human data, synthetic data, or something else matters less than the learning speedup.

“Well, the answer is whatever it causes faster learning. There are applications where you have enough training data from humans. There are applications where you have to generate the training data from what the humans are doing.”

The Trump administration has ended the 10^26 FLOPs regulatory doctrine and has not yet produced its own thinking on AI governance; they are very concerned with China getting ahead, so they will likely come out with their own framework.

“that of course has been ended by the Trump administration. um they have not yet produced their own thinking in this area. They're very concerned about China and it getting forward. So, they'll come out with something.”

There is substantial capital investment in making chips and compute more energy-efficient through novel architectures; the transformer architecture has new variants emerging weekly, and inference-optimized hardware is expected to arrive as software needs expand, following 'Grove-and-Gates' dynamics.

“There is a similar amount in going in capital. There are many many startups that are working on non-traditional ways of doing chips. The transformer architecture which is what is powering things today has new variants. Every week or so I get a pitch from a new startup that's going to build inference time, test time computing which are simpler and they're optimized for inference. It looks like the hardware will arrive just as the software needs expand.”

Empathetic AI voices with any inflection desired will soon exist (within 2 months) and will have significant emotional impact—most people haven't experienced super-empathetic voice AI yet but will soon.

“Most people haven't actually experienced the super super empathetic voice that can be any inflection you want. When they see that which will be in the next probably two months. Yeah. they're going to completely open their eyes to what this”