Zuckerberg showed up as Mark Elliott — then argued for many AIs, not one

Zuckerberg had just fought his first jiu-jitsu tournament, and Fridman, a fellow competitor, wanted to know what that felt like with a company of billions of users still in his pocket. The Meta CEO said it was fun. He is competitive; sports that demand full attention are how he stays mentally healthy enough to do the rest of the job. Friends trained with him in a mini-academy in his garage, one of them suggested a tournament, and he was not going to shy away from a challenge like that. He was pretty sure he would do okay.

Getting there without making it a scene took a costume. He rolled up in a hat, sunglasses, and a COVID mask, and he registered under his first and middle name, Mark Elliott. People did not know it was him until he pulled that kit off right before he stepped on the mat. He did not want to lose. Fridman pressed on the rawer fear: not just losing, but being embarrassed in a sport that is just you and another human being, with nowhere to hide. Zuckerberg did not look scared. He looked excited.

"I just think part of learning is failing," he said. People who train jiu-jitsu cannot afford pride, because you only stay in a bad situation if you refuse to tap after you have already lost, and a beginner is going to lose a lot. He has been embarrassed enough times in his life that it has become a core competence. As people grow up, they build an adult identity and stop wanting to look stupid. His view is that the ability to keep doing interesting things is the willingness to go back to step one, get kicked around, and look foolish. The moment you decide you are too embarrassed to try something new, you are not going to learn anything anymore.

He draws a straight line from the mat to Meta. The big things the company is on are ten-plus-year missions. Early on, people doubt them, the first work looks silly, and the ethos is not to wait until something is perfect to put it out. Get it out, get feedback, change it. Fridman notes that the fear of looking like a beginner only gets worse with age, once you have expertise to protect. Zuckerberg agrees. Growth lives in the places where you are willing to say something stupid again.

Fridman asked how he handles fear on the big-decision days at work. Zuckerberg's answer was immediate: the thing that stresses him most is always the people challenges. Strategy questions do not keep him up much. He has enough conviction about the values of what they are trying to do, what matters, and what he wants the company to stand for. He does not get everything right — they make a lot of mistakes — but he has a strong sense of where he wants them to go.

After almost twenty years, one lesson is clear to him. When a team is cohesive, you can get almost anything done: run through hard challenges, make hard decisions, push the work. When there is tension in that group, things get really tough. Talking to friends who run other companies, he thinks he spends a disproportionate amount of time fostering a tight core of people who run Meta with him. That group is what makes the job fun. It is also what weighs on him when it frays. Facebook, Instagram, WhatsApp, the metaverse, AI — the people piece is the part he is most anxious to get right.

He builds that circle by spending a lot of time with it, grappling openly with the biggest problems. Every Monday morning, about the top thirty people get together. People rotate in and out, but the group has worked together a long time, and the meeting often runs a couple of hours, somewhat unstructured. He brings a few topics that are on his mind and asks others to bring theirs: a policy issue in some country, a new technology, a partner problem, a WhatsApp design trade-off between two values they care about deeply. Over time, working through those issues in the open is how people develop intuition for each other, and a bond. It is a fairly combative group. He says that like a compliment. Combat, in that room, is how design, engineering, and policy actually get decided.

Fridman pulled the conversation back to the mat and asked for a favorite submission. Zuckerberg prefers no-gi, because it sits closer to MMA, and his style sits closer to MMA too. A lot of jiu-jitsu players are fine on their back; a good guard is core to the sport. In MMA you do not want to be on your back, because even with control you are taking punches. He would rather be the top player, applying pressure. He is not a heavyweight, though, and a lot of the friends he trains with are bigger than he is, so back takes matter. That is where the leverage advantage is largest, because people's arms are weak behind them.

The right strategy, he said, is not to be too committed to any single finish. That said, he does not like hurting people, so he thinks chokes are a somewhat more humane way to go than joint locks. Fridman compared the style to Khabib Nurmagomedov. Zuckerberg went with the clean version: back take to a rear-naked choke. Advice for people who want to start, given how busy he is, was not a curriculum. You have to be willing to get beaten up a lot. Over time there is a flow to it. The lesson he keeps is simpler still: once you are locked, you have already lost. The only question is whether you will admit it.

He believes that accomplishing anything is, in large part, being willing to push through, and that grit is often the difference between the people who get the most done and the people who do not. Almost everyone who starts a successful company, in his view, is working extremely hard. Sports taught him the other half. You cannot push through everything. In MMA a bad counter is a punch in the face; in surfing you cannot go the other direction on the wave; on a foil you can pump in several directions until the momentum against you is strong enough that it simply will not work. Company cycle times are months, and hardware can be years, before you get the next set of decisions. Short-turnaround sports give the feedback in seconds. The humble lesson for people who run things is knowing where finesse matters more than force.

Fridman turned to AI, calling the year an interesting one for the whole of human civilization, and asked for the story of LLaMA — Meta's 65-billion-parameter language model — and the decision of how to release it. Zuckerberg agreed that the prior year had seen a bunch of advances scaling large transformer models: large language models on the language side, large diffusion models on the image side. Meta had taken a quite open and academic approach. Part of that is talent. A lot of the best people in the world want to know they will be able to share their work. The deal is industry-scale infrastructure plus an ethos of sharing what gets invented.

LLaMA was the language model the research team made. They did a limited open-source release intended for researchers. Responsibility and safety mattered enough that they did not think they should release the first version commercially, so they punted on that question for v1. By releasing it for research it was out there, but companies knew they were not supposed to put it into commercial products. Follow-up models were already in the works, with more time spent on safety. His broader view is that it would be good if a lot of different people could build state-of-the-art technology here, not just a small number of big companies.

Training a state-of-the-art model, he said, takes hundreds of millions of dollars of infrastructure. Not that many organizations in the world can do that at the biggest scale today, even if it gets more efficient every day. He wants to see what students, hackers, and startups build, and he points to a precedent inside Meta's own infrastructure. The company put its data-center and server designs into the Open Compute Project. The theory was that if more people could use the server design, that would enable more innovation, make the design more efficient, and make Meta's business more efficient too. That had worked. He wanted to try the same logic on models.

The limited February release got, in Fridman's phrase, leaked: it escaped the research circle. Zuckerberg said they shared it with researchers, which is a slow release in all but name. What happened next was a vibrant open-source scene. llama.cpp re-implemented the model to run more efficiently on smaller computers. People wired up reinforcement learning from human feedback. Projects like Alpaca, Colossal AI, and GPT4All sprang up on top of it. Zuckerberg thought it was neat. Individuals who do not have a large cloud budget could run it on a laptop. Even Meta, running its own versions, could do it on way less compute. That saved a lot of money for everyone who used it.

He was careful to place the model. Because it was a relatively early release, LLaMA was not quite as on the frontier as the biggest OpenAI models or the biggest Google models. The largest LLaMA they released had 65 billion parameters. Nobody outside OpenAI knew GPT-4's specs, but his understanding was that it was about ten times bigger, and that Google's PaLM had about ten times as many parameters too. The LLaMA models were very efficient, and performed well for something around 65 billion. That efficiency was part of the safety argument. There is a real debate about whether everyone should have the most frontier models as they approach superhuman intelligence. Nobody, he said, looks at that generation of LLaMA and thinks it is anywhere near superintelligent.

Fridman put the sharp version of the worry: people terrified of existential risk will say you have to be careful about the open-sourcing step, and open source also puts a powerful tool in the hands of bad actors. Zuckerberg's reply was staged. At the current stage, the equities balance strongly toward doing this more openly. If you got something closer to superintelligence, you would have to discuss that more; Meta had not decided what it would do in that position, and he thinks there is a good chance they are pretty far off. He does not claim the current stance applies to every future thing. Open-source software, in his view, tends to be more secure than software developed in a silo, because more people look at it and find holes. Security through obscurity is the worse pattern. Meta still does not open-source the core app code for WhatsApp or Instagram; that code is not general enough to be useful to a lot of people. Memcache, a caching tool he coded and led — one of the last things he led directly — is the other kind of software: broadly useful, like the language models now, and like the translation stack Fridman had just praised.

That translation work was its own sci-fi delivery. Meta had expanded text-to-speech and speech-to-text from around 100 languages to more than 1,100, and the model could identify more than 4,000 spoken languages, which Fridman said was forty times more than any known previous technology. Zuckerberg framed the product as the old idea of an earbud or a pair of glasses that translates in real time. He thinks technology is basically delivering that now. For languages that never had enough data to train a traditional model, the coverage is the point: a lot of people the stack used to ignore are finally in it.

On the next version, Zuckerberg would not announce a date. A lot of the work was taking the research-first v1 and building a version with the latest safety precautions, trained on more data from across Meta's services, with internal effort focused on making it as aligned and responsible as possible. The product picture was specific. There will be an assistant you can talk to in WhatsApp. Every creator, he thinks, will have an AI agent that can act on their behalf, that fans can talk to. He wants every small business to have an AI people can talk to for commerce and customer support. The language model underneath is the engine. Open-sourcing it, as with v1, unlocks ecosystem innovation that makes Meta's products better and yields feedback on security that you do not get behind a closed door. The work is taking a research project toward core infrastructure, for Meta and hopefully for a lot of other things too. The internal debate, he said, is not whether to open it so much as the right way. They had been clear that v1's code and weights had no commercial license, and they had generally seen people respect that; no reputable companies, in his telling, were stuffing it into products. By sharing it with so many researchers, it did leave the building. He had no news to announce on v2.

Fridman passed along a question from Yann LeCun: will LLM fine-tuning need to be crowdsourced, Wikipedia-style? Zuckerberg has talked with LeCun about it a bunch and finds it compelling. Different groups test different data recipes; instead of one lab picking a winner, it would be nice to get to a collaborative way for a broader community to fine-tune as well. Infrastructure, community management, and product design make that hard, and he has not worked it out. As an idea it fits the open-source ethos. How to produce aligned models is very much a research area, and one that needs as much progress as the core intelligence of the models themselves. He does not think simply scaling current LLMs to higher parameter counts will, by itself, get to general intelligence. There are probably more fundamental steps along the way. Even so, the current level of technology can already create a tremendous amount of product value. He would rather the world had some time to build those breakthroughs into things people actually use than stack surprise on surprise.

More than three billion people, Fridman noted, use WhatsApp, Facebook, and Instagram, so any LLM-fueled product in those apps is a civilization-scale event. Zuckerberg sketched three categories of work. The most interesting, and the biggest distinction from OpenAI, Google, and Microsoft, is that everyone else is building the one singular AI — ChatGPT, Bard, Bing — and his view is that people will want to engage with a lot of different AIs, the way they use different apps and have different people filling different roles. Creators have 24 hours in a day; bottling up a personality so fans can get concert info is valuable, and it is not one model. Businesses will want an AI that only sells their products, not a general assistant that might recommend a competitor. The second category is simpler and already close: tell the system to edit a photo, or tell it you are a dog walker and let it generate the ad unit and find the right people. The third is that every Meta product, including metaverse worlds, evolves in this direction: describe the world and it writes the code. Natural language becomes more of the interface, in any language, with translation in the stack.

Fridman asked whether some of the people you are friends with on these networks will be AI systems. Zuckerberg thinks we will get there, and he pointed to evidence that is already old. Microsoft's Xiaoice in China, a pre-LLM chatbot, had tens of millions of people who became quite attached. Services like Replika exist now. Most people, he suspects, do not have as many friends as they would like; demographic work he cited says the average person has fewer close friends than fifteen years ago. The primary job is still tools that help people connect with each other. An AI that reminds you it is a friend's birthday, tells you what is going on in their life, and helps you reach out as more than a copy-paste emoji is a social assistant he finds powerful. Beyond that he can imagine a mentor, a cheerleader, an education tutor. His kids, five and seven, love learning to code and get stuck waiting for a parent. A coding assistant designed for that age would be extremely patient and never annoyed. He does not think all of that buckets into one singular AI.

A Reddit question put the dystopia cleanly: if people talk through chatbots that polish their language, will a generation communicate by grunts? Zuckerberg thinks the long-term effects are mostly good. The question was framed negatively; language translation already helps you talk to people whose language you do not speak, and nobody serious says you should have just learned it. Language is an imperfect map of thought. A jiu-jitsu instructor Fridman knows runs emails to parents through GPT so he comes off less like a jerk. Creators on Meta's services tell the company that negotiating brand deals in DMs is one of the most stressful parts of the job; they want to connect with a community and freeze when a brand wants something. A tool that helps them push back respectfully is a simple product he has heard real demand for. He and Priscilla had just had their third daughter. Babies cry because they cannot tell you what they need; his five-year-old still cries for the same reason; a lot of adults get frustrated because pride and skill get in the way of saying what they mean. Technologies that help people navigate social complexity are, in his view, generally good. A generation later, nobody laments Google for ruining memory.

Fridman said he still writes maybe 80 or 90 percent of his code himself, then edits, and that models mean you no longer have to remember the specifics of every function. Zuckerberg was more interested in the large-company version of that problem. Before an engineer at Meta can sit down to code, they have to figure out libraries and dependencies that tens of thousands of people wrote. He is excited about tools that summarize the internal knowledge base, not only tools that write code. In public experiments, you ask a model for a script and it already knows which libraries to pull in — which was always the annoying part. Fridman raised the flip side: if you outsource the thinking, you miss subtle bugs in code that looks convincingly right, and you lose the skill to find them. Zuckerberg agreed it is a trade-off. It is still valuable, he said, to be able to do basic things yourself. Standing on taller giants, you forget how they climbed.

Fridman asked about human-like bots on social networks, phishing, and social engineering. Zuckerberg split the problem. For five or seven years Meta has been ramping classical AI systems — not generative ones — to classify harmful content. He listed the work as about eighteen categories: terrorism, child exploitation, incitement to violence, intellectual property violations, and the rest. Generative AI will test those systems. So far, and he is optimistic this continues, they can bring more computing power to bear with even stronger AIs on defense. Hate speech, in his sketch, does not get much more sophisticated. Nation-states do. Whether it is Russia or other countries standing up networks of bots, or real people masquerading as others in a coordinated way, the behavior is adversarial: each time Meta finds them, they evolve. The internal name is coordinated inauthentic behavior. If you told him twenty years ago that a college site would one day include a counterterrorism organization with AI, he would have thought that was pretty wild. That is where they are.

He wants to be specific about what is bad. People organically bouncing ideas off each other until a narrative appears is how culture and art get made. A network of bots and fake accounts with someone pulling the strings, acting as if the behavior were organic, is the problem, and he does not think people should run coordinated networks without disclosing them. His guess is that most of the AI challenge for the foreseeable future lives here, amplifying known harms, not in existential risk. Those later questions are hard to disprove; his own intuition is that current technology will not get to superintelligence without another set of significant advances. That does not mean there is no danger. Fridman noted you can already ask a model how to build a coordinated inauthentic network that Meta will not detect. Zuckerberg said a lot of the fine-tuning is training shipped AIs so you cannot ask them to help you commit a crime. None of the systems will be 100 percent perfect. The bar is making this not an easier path than the next best alternative — people still have Google — and making it expensive enough that sophisticated actors go somewhere else.

Then comes the part nobody can define. Fridman asked GPT about whether the virus leaked from a lab and got a nuanced, well-cited answer, and he would hate for that nuance to be sanded off by moderation. Zuckerberg agreed. Everyone agrees child sexual exploitation and terrorism should be off the service. Misinformation is trickier: being wrong is not always harmful, and you do not necessarily censor someone for it. Early COVID was worse, because there were real health implications, the science had not been fully vetted, and a lot of the establishment, in his account, waffled and asked for things to be censored that in retrospect were more debatable or true. That undermines trust. He tries to boil disputes down to harms people agree on, especially physical harm, and then let people choose how they want things flagged. He thinks Twitter's Community Notes implementation is quite good: adding context without taking the post down. Meta had recently introduced a control so fact-checking does not have to affect distribution for people who do not trust the fact-checkers. Incitement to violence is still not allowed. Science, he noted, never gets to 100 percent certainty. Edge cases are not a reason to stop trying to give people what they are looking for.

Fridman asked how Meta withstands pressure from governments and interest groups, including in a world of AI moderation. Zuckerberg said there is no one-size-fits-all answer. The principles are to let people express as much as possible, with clear categories of things they do not want on the services. When a government says it does not want something up, the hard cases split. A democratically elected government, with different values in a different place, raises the question of whether a California company should overrule what that country has decided is unacceptable — he thinks there is a certain amount of hubris in that. An autocrat cracking down on dissent, out of line with the people, is another case. In general they advocate and push back on takedown requests. They draw a much harder line on requests for access to information. Being told you cannot say something violates expression; a government getting data in a way that would be unlawful in the United States exposes people to real physical harm.

By the time you are litigating, it is late in the funnel. A lot of the protection starts years earlier, in where they put data centers. There are countries with many users where a local data center would cut latency, and they still will not put one there, because the government does not have a good track record of leaving people's data alone. If you put a data center in a country and the government wants the data, they have the option of showing up with guns and taking it by force. Architecture decisions in advance matter more than the press-conference confrontation. On speech norms, especially with a democratically elected government, he thinks there should be a certain amount of deference. On the Twitter Files, he said it is really hard from the outside to know exactly what happened in each case. Agencies flag threats they want security teams aware of; in his telling that is often information, not orders, and it remains Meta's call. There is so much pressure from all sides — elected officials, agencies, media, activist groups, in every country — that he is not sure any specific ask adds that much to the mix. You cannot make decisions by trying to make people happy. You do what you think is the right balance and accept that people will be upset wherever you land.

His view of freedom of speech has gotten more nuanced, not less pro-expression. You do not build these services unless you think people expressing themselves at scale is a good thing. It has become clear that some categories almost everyone accepts as bad, and that even in the United States, with a protective First Amendment, you are still not allowed to immediately incite violence or violate intellectual property. The practical lesson is often not "should this be allowed" but how to deal with someone being a jerk, especially as repeat behavior, without confusing a controversial viewpoint that might be good for a community with harassment. With AI, he thinks it is less about censorship and more about the most productive answer. One case study he reviewed with the team: someone asks how to 3D-print a gun. One proposed reply is a hard shutdown, the "as a large language model I cannot talk about that" style. Another is, I do not think that is a good idea, and here is the factual situation in a lot of countries, including the United States. He prefers the second: assume good intent, educate, stay informative. You will not be able to do that on everything. He thinks it is a better default.

Fridman asked about rumors of a text-based social network, a Twitter competitor internally known as P92. Zuckerberg said there is a project. He has always thought a text-based information utility is really important to society, and that for whatever reason Twitter has not lived up to the potential he would have expected. He thinks Elon Musk thinks that too, which is probably one reason Musk bought it. An alternative he finds interesting is the open, federated approach visible in Mastodon and a little in Bluesky, melded with the graph and identity people have already built on Instagram. He did not want to get ahead of himself; they explore a lot of things, and there was no launch date. Text, for him, is uniquely good for discussion because comments can be first-class, so conversations fork, unlike TikTok, where the video is the architecture and comments come after. He always thought Twitter, or whatever ends up in that space, should have a billion people using it. Building that is not automatic. Meta plateaued early, then got good at internationalization and at helping people spread a service with their friends — a competence that, once they had it, made him feel they could grow Instagram and WhatsApp after those companies joined. ByteDance did it with TikTok. Twitter, in his outside view, had a great idea and had not cracked that piece, which is why you are seeing other cuts at the same thing.

He is fascinated by running a larger experiment in more decentralized control and a more open stack. End-to-end encryption on WhatsApp is a version of that: it pushes the service to the edges. The server Meta runs for WhatsApp is relatively thin compared with Facebook or Instagram; more of the complexity is apps negotiating to pass information fully encrypted. That model, if you can make it work, puts more power in individuals' hands. He called the whole thread speculative. Fridman then asked, in the name of camaraderie after some harsh public words, what Musk was doing well. Zuckerberg declined a detailed critique from the outside. What he would say is that Musk led a push to make Twitter a lot leaner, more technical, with fewer layers of management and less distance between engineers and the person running the company. You can agree or disagree with the tactics. He thinks those principles were generally good changes, and good for the industry, because a lot of other people thought the same thing and had been shy about doing it. When Zuckerberg wrote a letter in March outlining organizational changes, and when people saw what Musk was doing, other founders told him it gave them room to think about how to shape their companies. He will not opine on whether Musk cut too much.

Meta's own layoffs, he said, were uniquely hard because a significant number of people were leaving for reasons unconnected to their performance or a culture mismatch. It was a strategy decision, and in the changes that year more cultural and strategic than purely financial: he wanted a stronger technology company, more focused on building higher-quality products faster, with a stable position to keep investing in AI and the metaverse after eighteen months of macroeconomic, competitive, and self-inflicted thrash. Then you have to execute it as compassionately as possible, and there is no good way. He wanted to empower engineers, including putting individual-contributor engineers at almost every level up the stack, so the people doing the work are not eight layers down. Every extra layer is latency on information. He thinks part of the art of running a company is constantly re-engineering it so feedback loops get shorter — the same lesson as jiu-jitsu, applied to an org chart. Support functions matter, in the right proportion; too little legal or marketing and you blunder, too much and you get slow and conservative. At the start of the process, the average manager had around three reports, closer to three than four, because the company had been hiring so fast that managers needed spare capacity. Best practice, he said, is more like seven or eight. Going forward he does not want to hire that many people that quickly. In a leaner world you can defragment toward a more compact structure without undervaluing management or coaching.

How do you hire great engineers once the company is no longer growing at that pace? Attracting them, Zuckerberg said, is working on cool stuff, having a vision, and having a track record people believe. Selecting them is the art. The heuristic he uses for his own hiring, and that he thinks works if you recurse it through the organization, is that you should only hire someone if you would be happy working for them in an alternate universe. He is not in a rush to stop running the company; in that universe he would still want to learn from those people, and he respects their judgment and values. For people out of college, internships are the sorting function. Meta has a much better sense of who the best people are after a couple of months on the job than from a resume or a short interview loop. In-person feel tells you more than paper; a lot of fit is cultural and even team-specific. That is why they have invested so much in internships. It is useful for the company and for people who want to try the place.

Remote work, he thinks, is here to stay, with an asterisk for the metaverse work: once presence is dialed in, physical location may matter less. Today it still does. He would not want to run a fully remote company yet. People with special skills who live far from an office are absolutely worth having. People who have already built internal relationships can go remote without a performance hit — they have done the studies. New people, and people earlier in their career, still need the hallway after the meeting. Fridman had tried Quest 3. Zuckerberg said more details would come at Connect, with the headset coming that fall, priced at $499. Two big additions over Quest 2: high-resolution mixed reality, where you see the physical world and place virtual objects in it — a movie screen, a desktop, zombies through the wall, a Dungeons & Dragons board on the table — and a better VR headset, with twice the graphics processing power, 40 percent sharper screens, and a device 40 percent thinner. Quest Pro had started at $1,500 and come down to $1,000; in a lot of ways the mixed reality on Quest 3 was ahead of what they had delivered on Pro. The mission constraint is the point. They want this accessible, not an elite gadget, which is an extremely hard technical problem because you cannot put unlimited hardware in the box.

Apple had just announced Vision Pro at $3,500, available early 2024. Zuckerberg had seen the launch materials, not the device. First, he took it as validation of the category Meta had been out there arguing for. Apple coming in would make a lot of its fans consider mixed reality, and the price meant a lot of people would not find it affordable, so there was a chance Apple increased demand for the space and Quest 3 was the primary beneficiary. Apple, in his view, builds high-end things; Meta has a more democratic ethos. They had sold tens of millions of Quest devices. His understanding, based on rumors, was that Apple was building about a million of its headset. The companies are also optimizing for different use cases. Meta is focused on social interaction, fitness, gaming, being active. A lot of Apple's launch material, as he read it, was people sitting and looking at screens, which is a real use, including replacing a laptop over time. Apple emphasized no controllers, elegance of eye gaze and hand tracking. Quest 3 can be navigated with hands too; one demo teaches piano with highlighted notes and no controllers. If you care about gaming, a controller still captures fine motor movement more precisely. Different design trade-offs, different costs, different rooms of the house.

Fridman brought up Eliezer Yudkowsky and existential risk. Zuckerberg said that when a number of serious people raise an existential concern about something you are involved with, you have to think about it, and he has. Over time, as systems get closer to superintelligence, they will need to think about it even more. It is pretty clear to anyone working on these projects today that they are not there. His worry is that people get too focused on tail risk and then do not do as good a job on the harms that are almost certain to show up near term — fraud, scams, the known list. In some of the safety debates, he thinks intelligence and autonomy get conflated. You can scale intelligence quite far without the being of the thing. In human biology the neocortex is a prediction and reasoning engine; a more primitive old-brain structure is still calling the shots, and the neocortex does not have free will. We might build systems much more capable than a neocortex that still act as tools. Where we really need to be careful is autonomy. Relatively simple, unintelligent things with runaway autonomy already have a name: a virus, biological or computer. If he were a policymaker, he would want that distinction. Building intelligent systems can raise quality of life and productivity. The autonomy part is what we need to learn to govern before we give systems goals and let them make a lot of decisions on their own.

Fridman asked whether building AGI might make him the most powerful man in the world, and whether that power would corrupt him. Zuckerberg sent the question back to open source. He does not think the world is best served by a small number of organizations having this without it being more broadly available. History, in his telling, gets weird when there are unipolar advances and power imbalances. That is one reason he thinks open source is generally the right approach. He feels quite confident that open-sourcing models now is net positive. There is a good argument it remains right even closer to superintelligence, but he has certainly not decided that yet, and the debate will be informed by what happens between now and then. Asked what year we get superintelligence, he refused a date. The last year was a real breakthrough. One extreme is that one or two more leaps get you something truly crazy and close to general intelligence. The other, closer to what the industry has historically seen, is the Gartner trough: five, ten, fifteen years sitting with this one until the next big thing is figured out. Error bars are wide. He does not think breakthroughs stack every six months, and nobody knows how many are required. The piece he thinks most of the discourse has not grappled with is that we already have organizations and structures smarter than a single human: a company, if it is working, and the stock market, a distributed system that allocates capital. Those cybernetic organisms seem productive when we build them in balance with each other. AI, in that frame, can make the collective more informed and better at debating, on a human-time feedback loop, rather than racing off as a singular mind.

Fridman pulled him back to the body. Zuckerberg had recently done the Murph challenge — a mile, 100 pull-ups, 200 push-ups, 300 squats, another mile — in under forty minutes, not his best time, and not a win among friends that Memorial Day. He wanted it on the record that partitioned Murph, mixing the movements, is a different clock than doing each block in series. He does it each Memorial Day for Lieutenant Murphy. His older daughters did a quarter Murph with him; the seven-year-old wanted a weight vest and did not get one. He fights three or four times a week, cannot do it every day because of bruises, and fills the other days with cardio, strength, and mobility. He tries to do something physical every day. Meta's building has a park on the roof so one-on-ones can be walking meetings; he has a hard time just sitting. He does not believe a being is just a mind. Asked whether AGI should have a body, he said it depends on the goal. The stock market has no body and does not speak a language. There will be limits to what a pure intelligence understands about aging, having kids, designing a house, without modeling those sensations. If the goal is to understand people as much as possible, that probably matters. A lot of value can still be created by intelligence that is a separate thing. On AI replicas that persist after we die, he thinks we will have the capacity, and that it should be your choice — the same rule as pages: a Lex fan club is fine, a page that claims to be Lex is not.

Fridman closed on faith. Zuckerberg split it into a philosophical piece and a cultural one. At the beginning of Genesis, God creates the earth and creates people in God's image, and the only context you have for God at that point in the Old Testament is that God has created things. One lesson he takes is that there is a virtue in creating, artistic or useful. That drives a lot of how he thinks about a good life: helping the people around you, being a positive creative force, bringing new things into the world, including kids. Nothing makes him happier than coming home and seeing his daughters building Legos, the same spirit he hopes they do not lose. The cultural piece — community, values, tradition — has become a lot more important since he had children. He did not think about religion much in college. Kids make you choose what to impart. It is grounding, he said, to believe there is something much bigger than you guiding things. Among other things, that gives you a bit of humility while you keep trying to create.