From a Five-Year Phone Queue to 480 Trillion Tokens a Month

Fridman opens with the arc that made this episode feel personal: a childhood in a two-room apartment in Chennai, almost no access to technology, and a career that now runs a company Lex puts at about two trillion dollars. Pichai says the twelve-year-old version of himself, living in 1984, would have laughed the whole story off. He had started reading about computers by then. He had not yet seen one.

The city he describes is crowded, energetic, and, in his memory, simple. Neighborhood kids played cricket in the street until it was too dark to see the ball, barefoot, pausing whenever traffic needed the road. Before computers, he says, there was an almost shocking amount of free time; now solitude is something you have to seek. Newspapers and books were how he reached the world’s information. His grandfather, who worked in the post office, was the deepest influence — so good with language that Pichai still calls his English handwriting the most beautiful he has ever seen.

The household had books of every grade, from Ayn Rand and philosophy to “stupid crime novels.” That hunger for knowledge, he says, is why Google’s mission later felt like home. Technology, though, always arrived late. There was a five-year waiting list for a rotary telephone. When it finally came, neighbors walked over to call people they loved, and a hospital trip that used to cost two hours each way — often to be told the blood-test results were not ready — collapsed into five minutes. “This light bulb went in my head,” he says: technology can change a life in a discrete jump.

The jumps kept coming. During a drought there was no running water; trucks delivered maybe eight buckets per household, and he, his brother, and sometimes his mother waited in line to carry them home. Years later there was a tap, then a water heater, then a hot shower. A Panasonic VCR, bought in twelfth grade from a shop that had “kind of smuggled it in,” let them record World Cup matches and watch bootlegged tapes after a childhood of one TV channel. Fridman notes that humans are bad at feeling how good they have it. When Pichai goes back to India now, the thing that still shocks him is what mobile phones have done.

Asked what he would tell young people who want the next version of his story, Pichai does not lead with tactics. Luck matters, he says, and you still have to make smart choices. The more important move is to listen to your heart enough to find work you actually enjoy, because love of the work is how you meet the best version of yourself. That is hard, he admits. It is still the first piece of advice he would give.

The second is to put yourself in rooms with people who feel better than you. He wants a few moments in a life when you are sitting across from someone and a quiet “wow” goes off in your head. Uncomfortable company, he argues, is what grows ability; people who stay open enough to walk into those rooms often surprise themselves. Fridman then presses on a different puzzle: Pichai climbed a cutthroat hierarchy while remaining, from the outside, humble and kind.

Pichai does not romanticize his temperament. He gets angry and frustrated like anyone else. What he has learned is that mission-oriented people with an inner drive to excellence rarely need to be frightened into doing the right thing; they often feel a miss more sharply than the boss does. Losing his temper, he says, has become less frequent over the years because it is not required to get the outcome. Sometimes the unspoken message is stronger. People can see you are unhappy without a speech. Silence, on those days, delivers it more cleanly than volume.

Fridman asks him to rank AI among 12,000 years of productivity multipliers — agriculture, industrial mechanization, electricity. Pichai notes that around 2017 or 2018 he called AI the most profound technology humanity will ever work on, more profound than fire or electricity, and he is not taking it back. He does worry about recency bias, the sports habit of crowning whoever just left the field. Two years before the taping he had surgery, processed the fact that people once went through such procedures without anesthesia, and in that moment decided anesthesia might be the greatest invention of all. We cannot fairly score eras we did not live through.

Even so, he argues from first principles that AI is in a different league. Electricity and the internet were general-purpose and touched everything, but their pace and their ceiling were visible. AI is progressing so fast that the ceiling is unclear, and it is recursively self-improving — the first technology that can dramatically accelerate creation itself. If models reach superhuman performance on AI research, the system starts improving the thing that improves the system. He points to AlphaGo starting clueless and becoming superhuman over the course of a day, and to sampling Veo 3 at 30 percent and 60 percent of training: inspiring, and “a little bit unsettling.”

Fridman borrows the historian’s idea of a Neolithic package: the agricultural revolution was not just planting food, but pottery, surplus, hierarchy, early government, trade, and division of labor — second- and third-order effects no nomad could have forecast. What, then, is in the early AI package? Pichai says most of it is still unknown, but one piece is already tangible. Thoughts in your head will translate into things that exist. Software, games, films, tools: people who once could only write sentences will build objects. We are in the early days of vibe coding, he says, stitching prompts together, but “this is the worst it’ll ever be.”

The deeper claim is access. Forty years ago, he guesses, maybe five people in the United States could make a living doing interviews the way Fridman does; YouTube multiplied that. Blogs did something similar with writing. AI, he thinks, will not stop at hundreds of thousands of new makers. It could be tens of millions, maybe a billion people putting work into the world in a deeper way. That is the expansionary effect he believes people keep underestimating. Bring someone from the 1940s to YouTube and their mind would break. He expects the next ten to twenty years to do the same to us.

Fridman jokes that cable networks are nervous a person in a suit can do what they do, and that a Notebook LM podcast might one day be five or ten times better than his. If two AIs ever make a show he prefers, he says, he will stop hosting and go listen. Pichai’s counter is chess. You would not watch Stockfish play AlphaGo; it would be boring. Magnus Carlsen versus Gukesh is a different object. AI-generated content will be consumed because it is efficient. The premium experiences people protect may still be the ones where human essence comes through — the way watching Messi dribble still moves you even if a machine eventually dribbles better.

That is also how he thinks about artists and Veo. Darren Aronofsky was an early collaborator because he could see the future of the tool and get inspired by it. Artistic free expression, Pichai says, is one of the most important values in a society; artists have always pushed the frontier of thought. He wants to put APIs in their hands the way you provide electricity: as infrastructure, a paintbrush, not a list of approved canvases. Society still has to decide, at a fundamental level, what is not okay. Responsibility sits underneath the freedom. The freedom is still the point.

Fridman had noticed a shift in the models themselves. Earlier Geminis, when he researched Genghis Khan, the Aztecs, or the world wars, tended to flinch: are you sure you want to learn this? Gemini 2.5 Pro answered with facts, nuance, and depth. Pichai’s explanation is capability. A year ago the models were weaker and did more stupid things, so the patches on edge cases compounded. Once a model crosses a threshold of intelligence, it can reason through messy human material on its own. Users, he says, want as much access to the raw model as possible, including custom prompts. The long-term method is first-principles reasoning built from the ground up, “not like some subset of humans kind of hard coding things on top of it.”

Gemini, given Fridman’s notes, invented a question of its own: Pichai had told the world the system was churning out 480 trillion tokens a month — what is the most life-changing five-word sentence hiding in that haystack? Pichai confirms the number. Twelve months earlier it was 9.7 trillion tokens a month, a fifty-fold rise, and he had the same feeling he once had about Search: there is no limit to human curiosity. Maybe one day a five-word phrase will say what the universe actually is. He does not think we are there yet. The tokens, though, are already delivering small aha moments at planetary scale.

Will scaling hit a wall? Pichai says that question is a cherished micro-kitchen conversation when Demis Hassabis is visiting, or when Koray, Jeff Dean, Noam, and Sergey Brin are around. They see a lot of headroom: pre-training, post-training, test-time compute, tool use, more agentic systems, more general world models. Veo 3’s grasp of physics, he says, is dramatically better than Veo 1. He is lucky to work with researchers who also think there is more room. Each year he asks whether throwing ten times more compute at the problem will still buy progress. Sitting in that room, he thought the year ahead would.

The bottleneck he feels is compute. That is why each generation ships Flash, Nano, and Pro, but not Ultra. The Pro model, in his telling, reaches something like 80 or 90 percent of Ultra’s capability while Ultra would be slow and expensive to serve. The trick is to make the next Pro as good as the previous Ultra, and fast enough to actually use. Scaling laws are working. The models most people live with are still a few months behind the maximum the labs can deliver, because the maximum is not the most usable.

Fridman adds that “performance” itself is getting harder to measure. Gemini Flash can be more impactful than Pro simply because of latency; a slightly less intelligent model that answers now may beat a smarter one that makes you wait. Benchmarks, he argues, capture less and less of real-world usefulness. Pichai does not disagree. Impact is migrating off the leaderboard and into whether the thing is actually in someone’s hands.

On AGI timelines he first offers another acronym, one he thinks may have come from Andrej Karpathy: AJI, artificial jagged intelligence. Models do astonishing things and then trip on counting the r’s in strawberry. We are in that phase, he says, including in a Waymo working a crowded San Francisco street, or in Gemini Live answering a question about the world and getting it right. Will AGI arrive by 2030? He thinks we will just fall short, and that the definition will keep moving anyway. What he is sure of is that by 2030 the progress will be dramatic enough that society is dealing with both the positive and negative externalities in a serious way. DeepMind talked about a twenty-year path in 2010. In 2012, near where they were sitting, Jeff Dean showed a neural net recognizing a cat. People said a couple of decades. His sense is slightly after 2030, and that the label will matter less than the consequences.

Fridman puts his own probability that AGI and then ASI destroy civilization at about 10 percent, and asks for Pichai’s. Pichai has been excited about AI while also insisting the risks have to be thought about actively and harnessed hard. The p(doom) question, he says, is another kitchen-table conversation. Running a large organization has taught him that if you can align incentives and get people marching toward one mission, you can achieve almost anything. Organizing all of humanity that way is a different problem.

The twist is the one he keeps coming back to. If p(doom) is actually high, then at some point all of humanity is aligned on making sure that is not the case, and more progress gets made against the risk. “There is a self-modulating aspect there.” Because of that irony he is optimistic on the doom scenarios, even though he thinks the underlying risk is pretty high. He has faith that humanity will rise to meet the moment once the threat is concrete.

Fridman adds a comparison people skip: the probability of doom without AI. Humans already have many ways to wreck themselves. AI might make the world smarter, kinder, and less resource-constrained, and resource stress is a classic source of conflict. In that frame, AI could be the thing that saves civilization from everything else. Pichai agrees. On some of the hardest problems, he would rather have AI in the room helping.

After a break, the conversation snaps to a demo Fridman had just taken: Google Beam. Pichai asks him to imagine the U.S. and Chinese presidents sitting in that system with live translation and making a little more progress than they would on a flat call. Fridman, who has been reading about how much World War II might have changed if Chamberlain had met Stalin in person, says Beam produced the same hard-to-name feeling as being in a room. Pichai’s shorter version is that people who go through the trenches together form bonds. The product motivation he names is not a board meeting. It is a grandmother far from a grandchild, or a deployed soldier trying to talk to family — nothing substitutes for being there, but being there is not always possible.

A year before the interview, Fridman recalls, analysts were writing that Pichai should step down, that Google had lost the AI race and its magic. A year later Gemini Pro was winning on a spread of benchmarks, and a plot of what had shipped looked, to Fridman, incredible. Pichai says his main bet as CEO was to make the company AI-first and to set it up to develop AGI responsibly, then put that capability into products people actually use. Even in the loud months he had a clear view of what was being built inside.

He is, by his own account, good at separating signal from noise. The image he uses is scuba diving: jump into a choppy ocean, drop a foot under, and it is “the calmest thing in the entire universe.” Running Google, he adds, can feel like coaching Barcelona or Real Madrid; you will have a bad season. Outside voices sometimes make real points, and he wants those. Most of what you read is weather. Under the surface he could see decisions already made: merging Brain and DeepMind into Google DeepMind, a ten-year-old bet on TPUs, an AI infrastructure team assembled to scale compute. TPUs are excellent, he notes, but ramping enough of them took time — a curve outsiders could not see.

The merger itself he compares to being told to take Stanford and MIT and make one great department. Brain was bottoms-up and diverse; a lot of important research came out of that. DeepMind had a sharp vision of how to build AGI. Jeff Dean had already said he wanted to return to scientific individual-contributor work, because management was eating his time. Demis was the natural person to run the combined lab. It still took patience, a few sleepless nights, and credit he gives to Demis, Jeff, Koray, and the people who did the stitching. Early in 2025 he got to celebrate, on back-to-back days, a Nobel Prize for Geoff Hinton and then one for Demis Hassabis and John Jumper.

On the decisions people hated, his method is simple. Hear everyone when you can; sometimes it changes your mind. Then make a clear call and ask the room to disagree and commit. You can be calm and still be firm. Clarity, he finds, is what people respect over time. Some days you already have conviction, you say so, you place the bet, and you move. One of his favorite rituals now is walking, several times a week, into the Gradient Canopy building where the Mountain View DeepMind researchers work. Sergey is often among them. They look at the model. They look at the loss curves. That cultural reconnection, he thinks, mattered as much as the org chart.

Google Search is still, in Fridman’s phrase, the front page of the internet: a legendarily minimal page of ten blue links. Now there is an AI Overviews layer on the results, and a separate AI Mode tab. Pichai says Search has been evolving since mobile made people ask more questions, but this jump is different because the underlying models can supply real context. The design goal he refuses to drop is that you still go out onto the web and find things. AI is a layer for summary, context, and, in AI Mode, a back-and-forth dialogue. The links stay.

The mechanism that most surprised Fridman, after listening to Search lead Liz Reid, is query fan-out. For each question, the best models treat Search as a deep tool, explode the query into multiple searches, and assemble the knowledge so you can choose where to go next. The other unlock is language. For people who do not speak English, the usable web in their own language is much smaller. AI Mode can reason across English pages during the hunt, not only after you arrive and hit translate. Multimodality, translation, reasoning, and better tool use, Pichai says, put that power in the flow of Search — the part of the AI package he is most excited about.

On the business, he reports that AI Overviews improved the product on the user metrics they watch and helped growth, while AI Mode was already in millions of hands with encouraging early numbers. Ads in AI Mode are not the first job. The early work is getting the organic experience right. He still treats ads as commercial information, held to the same quality bar as anything else, and as the thing that lets the service reach billions of people for free. Over time he expects AI itself to help figure out how to present that information — the way a podcaster learns where a read belongs. YouTube is already a mix of subscriptions and ads; Google is adding subscriptions across products, and the optimization point will move.

Will AI Mode eat the ten blue links? Current plan: keep it as a separate tab for people who want the bleeding edge, and let features that work overflow into AI Overviews and the main page. The core principle remains taking you to the human-created web, with higher-quality referrals because intent is better understood. People are asking more questions, longer questions, different questions; Search is becoming a machine for curiosity. Demis has talked about a slice of the web built for agents. Pichai’s analogy is the grocery store versus delivery: both will coexist. Sites may actually get better for humans because AI will help design them. An agentic web will grow too, if the incentives can be solved. Agents will not need human UI paradigms. He does not think they replace the human web.

Fridman calls Chrome maybe his top-three piece of software, possibly number one, and asks for the origin story. Around 2004 and 2005, AJAX made the web suddenly dynamic; Flickr, Gmail, and Google Maps arrived in a matter of months. Browsers had never been designed for that world. JavaScript was painfully slow. The Chrome team’s bet was to bring operating-system principles into the browser: a secure sandbox, a process per tab, things that now feel obvious and then felt strange. In Aarhus, Denmark, they found a team whose V8 JavaScript virtual machine was, at the time, twenty-five times faster than anything else. They open-sourced it into Chromium. The name Chrome was a joke on the chrome of existing browsers getting clunkier; they wanted to minimize it. Pichai calls it the most fun he has had building a product from the ground up.

Then-CEO Eric Schmidt was not hostile so much as experienced: he knew how crazy it was to build a browser. Support followed once people could feel how much faster the thing was. The lesson Pichai internalized from Larry Page and Sergey Brin is a three-part case for moonshots. Ambitious work attracts the best people. Because it is so ambitious, almost nobody else is on the path, so you have it to yourself — Waymo is the example he names in the same breath. And even if you only hit 60 or 80 percent of the original goal, that is already a terrific success. People, he thinks, do not fully internalize those three advantages.

Waymo now has ten million paid robotaxi rides. The team always knew the last 20 percent would take 80 percent of the time, and Pichai says they knew they were in that phase. The counterintuitive move was to invest more at the moment outsiders were doubting the project. Google is a deep technology company, he says; Waymo is a version of building an AI robot that works. Safety and ambition were treated as equal commitments. They have scaled hard in the year of the interview and planned to keep scaling in 2026. On Tesla, he notes that Google was one of the earliest and biggest backers of SpaceX, assumes as a default that Elon Musk succeeds at what he attempts, and does not see a direct car-versus-car fight. Alphabet is building a general-purpose L4/L5 Waymo Driver for many settings. The space, in his view, is large enough for both to do well.

Robotics is the next node in the AI package — Waymo, as Fridman puts it, is already a robot that happens to have four wheels. Demis and Google DeepMind are focused on Gemini Robotics, building the underlying models and pushing the research. Applications are coming; Pichai says stay tuned, the outside story is not fully articulated yet. Hardware, in his view, has made extraordinary progress. Software was the holdup, and general models are the unlock. Getting those models to work in the real world, safely and generally, is the frontier. For the first time as a CEO, he adds, one deep horizontal investment can drive multiple businesses — Search, Gmail, Android, cars, robots — instead of a pile of separate bets. That is what makes the moment unique inside Google and Alphabet.

Programmers, Fridman says, are scared. Pichai answers with two Google numbers. About 30 percent of code now uses AI-generated suggestions. The metric he cares about more, because they try to measure it rigorously, is engineering velocity, and the estimate is a 10 percent increase across the company. Ten percent on tens of thousands of engineers is, in his words, phenomenal. It is also not a headcount story. They plan to hire more engineers the following year, because the opportunity space of what can be built is expanding too.

Near to mid term, he thinks AI takes away a lot of the grunt work and leaves more time for design, architecture, problem-solving, and brainstorming with colleagues. Creative power in more hands means more things get made — the chess analogy again: more people play chess now than ever. Fridman, coding on the side, feels roughly five times more productive and suspects a giant codebase has even more headroom. Pichai’s next unlock is robust agentic capability, including migrations and refactors of old code. On hiring, Google intends to keep at least one in-person interview so the fundamentals are still there. Using the tools to write better code is itself an asset. If you have a passion for computer science, he would still get the degree: the field is more than programming, and any education that trains first-principles thinking is good education, because AI will hit every field in ways that are hard to predict.

Android’s next act, he argues, is a paradigm shift in I/O, the way GUIs and multi-touch were. AR has been held back by brutal system integration and by the fact that, without AI, the interface is too complicated to feel natural. That is why Project Astra matters so much to Android XR. Apps and shortcuts will not vanish, but the OS needs to become more agentic: you describe what you want, or it learns your loops and adapts. Fridman had tried the glasses — glanceable time and weather, maps that show turn-by-turn when you look up and live translation when you look down, Gemini answering a question about a painting. Pichai calls it a moonshot that is closer than most. Developers were expected to get glasses later in 2025, consumers in 2026.

If Google actually built AGI, what would he ask it? Maybe by then it would be proactive and tell him a few things he should know. If he had to ask, he thinks it would help us understand ourselves in a way that surprises us — already visible, faintly, in today’s products, and much more so in an AGI. On a personal level, there is a chance it would understand you in a deep way. The other obvious use is expanding the frontier of how we understand the universe. He has not had access to anything that powerful. Those are the possibilities he can see.

Fridman’s first question is not practical. How many living and dead alien civilizations are there? How close are they, and are they dangerous? If there are none, or only bacterial life, why — is intelligence a trap that ends in self-destruction? Pichai smiles at the implication: that is an answer to the Fermi paradox. Fridman admits he is not sure what he would do with the information. It is the same curiosity Liz Reid was describing, the thing Search and now Gemini keep feeding.

Can you even measure the knowledge Google has unlocked? Years ago, Pichai recalls, an MIT study estimated Search’s value at a few thousand dollars per person per year. The number is almost immediately obsolete, because people take the new frontier for granted. AlphaFold is even harder to price. What makes humans special as AI beats us at more tasks? Pichai points at consciousness, the essence of being human, even as the line may blur. His hope is that if resources become more plentiful, the world feels less zero-sum — it never was, he says, but scarcity makes people perceive it that way — and then empathy and kindness surface more.

What gives him hope is a question he asks himself: would you rather have been born now, or at any other time? He almost always chooses now. Civilization has relentlessly made the world better, with ugly problems in every era, and he expects humanity to keep driving the frontier forward. Fridman is grateful to be alive for this work, lists a long string of Google inventions from Scholar to AlphaFold to quantum, and extracts a promise that Google Voice will not be killed. Pichai laughs and says they will not.

Fridman closes the episode with the two demos that, he says, blew his mind. Andrew, who leads the Google Beam team, walks him into a development unit: six ordinary color cameras, an AI video model, a light-field display, running in both directions. No headset, no pre-scan, no avatar. The person on the other side appears to lean out of a flat screen at real scale, with correct shadows and occlusion as you move your head. Spatial audio follows them across the room. When they try a high five, Fridman reports an almost tactile sense of contact. Andrew’s line is that the system is not trying to approximate you. If you nick yourself shaving or pin something on your jacket, that detail should travel. The window can seat a few people at a table; too many, and you fall back to familiar 2D tiles.

The second demo is Android XR glasses: light, styled like ordinary frames, with a camera, speakers, and multiple microphones for speaker isolation. Chris walks Fridman through a glanceable home screen — time, weather, calendar — designed to be used and then ignored. Gemini can be interrupted and followed up; asked about a painting on the wall, it names the work and refuses to invent a price. Look up while walking and Maps gives turn-by-turn; look down at your feet and the glasses show live subtitles. Fridman speaks Russian and watches English appear, with a slight delay when grammar needs the end of a sentence — “subtitles for the real world.”

Left alone with the microphone, Fridman returns to the Neolithic package. Rank inventions, he argues, by the size of the network that follows them, not by a single gadget. His current pick for the biggest leap is still the first agricultural revolution, the bootloader for ten thousand years of collective intelligence. AI has a real chance to take that number-one spot because of how large its follow-on network could be: translation that unlocks eight billion brains, autonomous vehicles that change how we move, science and energy, even a shift from industrial specialization toward generalists who integrate superhuman specialist AIs. He thinks the set of positive trajectories still outnumbers the negative ones, “but not by much.” He signs off with Jean de La Bruyère: “Out of difficulties grow miracles.”