Physics is the law. Everything else is a recommendation

Lex Fridman opens Elon Musk's fourth appearance, and the 400th episode of the show, with a question he ran past Grok first: how much of war is human nature, and how much is the way societies are built? Musk answers as a peace advocate who does not expect peace to win. Ignorance, he says, is the real enemy. All creatures fight. People picture nature as a calm place; Werner Herzog's jungle, in Musk's retelling, is murder in every direction, plants and animals trying to kill one another every minute of every day.

Greater intelligence does buy more control over the limbic urge to violence than a chimpanzee has. Chimpanzee society, Bonobos aside, is filled with horrific dominance fights with no ethics and no just-war tradition. Humans are not special for fighting. They are special for having a layer of reason that can sit on top of the same hardware. That is as optimistic as Musk gets about the species. Wars have always existed, he tells Fridman, and they always will. The only variables worth arguing over are scope and frequency. Even Iain M. Banks's Culture novels, which Musk loves, open on a galactic war in which trillions die.

On Gaza he is blunt. Hamas did not expect a military victory over Israel. The goal, in his view, was to commit the worst atrocities possible, provoke the most aggressive Israeli response, and then use that response to rally Muslims worldwide to Gaza and Palestine. That part, he says, has already succeeded. His recommended counter is split in two. Israel should still find Hamas members and kill or incarcerate them, because otherwise they will keep coming. Alongside that, it should pile on "conspicuous acts of kindness" that cannot be dismissed as a trick: a mobile hospital, food, water, medical supplies, webcams running 24/7.

The arithmetic is the point. An eye for an eye, he notes, is taken seriously in that part of the world, and if you are not going to commit genocide, you will leave a lot of people alive who hate Israel. "For every Hamas member that you kill, how many did you create?" Kill a child in Gaza and you have made at least a few people who will die just to kill an Israeli. If the long-term goal is fewer terrorists rather than more, the kill count is the wrong scoreboard. Something has to interrupt the cycle of reciprocal violence, or it never stops.

Fridman asks whether people beating the drums of escalation are feeding the military-industrial complex. Musk's worry is more basic: very few people alive in the West viscerally understand how terrible war is. His grandfather fought almost six years in World War II, in eastern North Africa and Italy. All of his friends were killed in front of him. He would have died too, except a random intelligence test pulled a corporal who had never finished high school out of the cannon-fodder category and into British intelligence in London, where he met Musk's grandmother. He came home with medals he never mentioned and PTSD so severe he barely spoke. Musk learned the record from files posted online. The man has been dead thirty years. The visceral memory of industrial war died with that generation.

Nuclear war is the civilizational version of the same problem. Musk puts the current probability as quite low, but the missiles are still in the silos, and enough warheads are pointed at the United States to turn the country into radioactive rubble several times over. Local wars are tragic. Global thermonuclear war could end civilization, or push it back toward the Stone Age, after which scientists and engineers would not be popular. Civilizational risks, he argues, should outrank pain that is local even when that local pain is enormous.

On Ukraine he predicted a year earlier, he says, almost exactly what happened: a lot of people dying for almost no change in land, and the flower of Ukrainian and Russian youth spent. The front is defense in depth, minefields, trenches, anti-tank weapons. Neither side has air superiority, because anti-aircraft missiles are far better than the aircraft. Tanks are slow-moving death traps. What remains is long-range artillery and infantry. "It's World War I all over again with drones." Whoever goes on the offensive will take casualties several times the defender's rate. A ceasefire, he thinks, will land very close to the current lines.

He asks for sympathy for the Russian boys as well as the Ukrainian ones. They did not ask to be on the front. The World War I line he quotes is that young men who do not know each other kill each other on behalf of old men who do. Asked whether Volodymyr Zelensky should sit down with Vladimir Putin, Musk declines to referee the talking. His advice is one sentence: do not send the flower of Ukrainian youth to die in trenches. History will not look kindly on whoever orders the next offensive.

Asked how the United States and China avoid a clash over Taiwan, Musk points to the Peloponnesian War and the so-called Thucydides trap. He loves war history, he says, and the Greeks are useful because they wrote everything down. Athens grew stronger year after year while Sparta remained the leader of Greece. Everyone could see the clash coming. They discussed how to avoid it. They could not. The United States has been the world's largest economy since around 1890, longer than anyone alive can remember, and the foundation of war is economics. China's economy, he says, is likely to be two or perhaps three times larger than America's. Imagine being the biggest kid on the block for as long as anyone can remember, and then a kid twice your size walks in.

What people miss about China, from Musk's time there, is the sheer number of smart, hardworking people, far more than in the United States, and the energy of the new architecture: train stations, buildings, high-speed rail. He tells listeners to go to Shanghai and Beijing and to take the train from Beijing to Xi'an for the terracotta warriors. Historically China has been internally focused, with rare exceptions, and its civil wars have been brutal. He cites the Three Kingdoms period as wiping out about 70 percent of the population. People in China, in his observation, think about China ten times more than they think about anything outside it. An inward-looking great power, on that reading, has less appetite to go out and invade a string of countries.

He puts the United States in a similar non-acquisitive category. After World War II it had a nuclear monopoly and could have taken any country it wanted without losing soldiers. It did not. It helped rebuild Europe and Japan and ran the Berlin airlift. His crude moral test is which army you would rather surrender to as a German in 1945. Wernher von Braun and the German rocket team, under SS orders to be executed, left in winter with no coats or food and ran west until they found an American patrol. Taiwan is the exception that does not fit the inward-looking story. From Beijing's standpoint, Musk says, the island is more significant than Hawaii is to the United States, a fundamental part of China, and the only reason it has not been incorporated is the U.S. Pacific Fleet. China has been clear it will take Taiwan peacefully or militarily. From their standpoint that outcome is 100 percent likely.

Fridman circles back to conspicuous kindness as a geopolitical policy that sounds naive. Musk answers with the twentieth century's two endings. After World War I the victors loaded Germany with impossible reparations, lumped all the blame on one country, and planted the seeds of World War II in a population that felt wronged and wanted vengeance. After World War II the Allies treated the Treaty of Versailles as the mistake and ran the Marshall Plan. "Something must stop the cycle of reciprocal violence," Musk says, or it never stops: eye for eye, tooth for tooth, forever.

xAI had just released Grok, built by a relatively small team in a short time, with a regular mode and a fun mode Musk takes the blame for. The model is patterned on *The Hitchhiker's Guide to the Galaxy*, which he calls a philosophy book disguised as humor and the basis of his own outlook: we do not know the meaning of life, but the more we expand the scope and scale of consciousness, digital and biological, the better we get at asking questions of the universe. The three-word mission is "understand the universe." Under the jokes he wants engineering, math, and physics answers you can count on.

Large language models, in his experience, hallucinate most on the important, difficult questions, and they do it with confidence. xAI is trying to ground results in physics first principles and mathematical logic, with acknowledged error rather than confident wrongness. If the foundational physics is not trustworthy, any invention built on top of it is wishful thinking. His test for whether an AI has reached human level is harsh: if it cannot figure out new physics, it is clearly not equal to humans, because humans have done that. Physics is discovering rules that already exist. Engineering is inventing things that never existed, and the design space there is far larger.

"Physics is the law, everything else is a recommendation." He has seen plenty of people break man-made laws and none break physics. Token-by-token prediction, he says, is a vector sum. Small errors drift, and after enough tokens the whole thing may not cohere. The fix is the one writers already use: draft, then look at the gestalt, then revise until the beginning, middle, and end add up. Fridman suggests that revision *is* reasoning. Musk agrees. On consciousness he is openly stuck. Thought and emotion do not feel like atoms bumping into atoms. He does not know on what dimension they exist. His guess is binary: either everything is conscious, or nothing is.

People ask Musk constantly whether he has seen evidence of aliens. He has not, which he finds concerning. He would prefer archaeological ruins. SpaceX has more satellites in orbit than everyone else combined, and they do not have to maneuver around anything unexplained. If he saw the slightest indication, he would post it on X immediately. We might, he says, be the only consciousness in the galaxy. Earth is about four and a half billion years old. Civilization measured from the first writing is about 5,000 years old, with a nod to the Sumerians. That is one millionth of Earth's existence, a flash in the pan.

The sun is expanding. At some point it will overheat the planet, boil the oceans, and turn Earth into Venus, where life as we know it is impossible. If humans do not become multiplanetary and eventually leave the solar system, annihilation of all life on Earth is a certainty. On a galactic timescale that could be as little as half a billion years, only 10 percent longer than Earth has existed at all. If life had taken 10 percent longer to evolve, it would not exist. The great filter he keeps coming back to is whether a civilization gets off its home planet. Explore other star systems, he suspects, and you will find a lot of long-dead one-planet cultures. The dinosaurs did not have spaceships, so the dinosaurs are gone.

This is the first time in Earth's history that life can extend beyond Earth. The window may stay open a long time, or a short time, or open now and never again. The smart move is to make life multiplanetary while it is possible. A self-respecting civilization, in his phrasing, would not be a one-planet species. On God he sides with Einstein's Spinoza: the creator, or the simulator, reveals itself through the lawful harmony of physics. The simulation argument gets a practical twist from rocket work. SpaceX and Tesla run simulations because they do not know the outcome. If higher beings are running this universe, they are doing it for the same reason.

Musk is 52, and his main recreation is still video games. He can stroll through Diablo 4's tier-100 Nightmare Dungeons. Duriel cannot scratch the paint. Uber Lilith, titled Hatred Incarnate, in the Eternal realm, as a Druid with no way to become untargetable, is the hardest boss fight of his life. Random death waves punish reflexes that are no longer what they were when he was, by his account, one of the best Quake players in the world and took money in what he thinks was the first paid U.S. esports tournament, a four-person event his team lost only because the best player's computer crashed. He is still tweaking the Paragon board. He is still making progress. He would not recommend Druid in Eternal.

"Killing the demons in a video game calms the demons in my mind." A tough game in the Goldilocks zone produces flow: not too easy, not too hard, with a sense of progress, beautiful art, and a puzzle to solve. *Elden Ring* is, for him, a candidate for the best game ever, top five at least. It feels designed by an alien. Every major boss is a unique puzzle, and the strategy that beats one does not beat the next. Hatred Incarnate in Eternal is still harder. The metaphor is not lost on either of them. Seasonal Lilith, with vampiric powers, is easy. Eternal hatred is almost impossible. That, Musk says, is a metaphor for life.

The Lilith fight pulls him into a warning about wishes that sound good. World peace, enforced, might require the suppression of everyone, the suppression of progress, an ossified society that never changes. The cost of no war at all might be worse than a little war. The same trap is on the page in *Brave New World*. There is a real muscle-relaxant called Soma, around since the 1950s. Musk took it for a squashed disc at C5-C6. It reduces pain at the expense of mental acuity. It dulls the mind, just like the drug in the book. Grok, asked about Soma, first hits the prescription and then the novel, and lands the question Musk wants: do you actually want everyone happy all the time if the happiness is artificial? He does not think hate should be amplified, but he is not sure a world with none at all is better. "No highs, no lows, who knows?"

Training LLMs efficiently is a physics problem, and Musk wants the tools of physics applied in the limit. Even at Kardashev scale, with the entire power of the sun, you would still care about useful compute per watt. That is the metric he thinks AI is headed toward. The near-term queue is more concrete. There is a silicon shortage now. In about a year that becomes a shortage of voltage step-down transformers. In about two years it becomes an electricity shortage in general. Fate, he notes, loves irony: you need transformers to run Transformers.

He told a gathering of the world's utility companies to prepare for a tripling of electricity demand. All transport is going electric, with the ironic exception of rockets, and heating is going electric too. Energy use, in very rough thirds, is electricity, transport, and heating. Make the last two electric and you need three times the power plants. The grid today is sized for real-time load, which means it is sized for the worst second of the worst day of the year or you get a brownout. Austin's multi-day blackout is his exhibit. Hydro can buffer. Almost nothing else on the grid can. Batteries let you make power at night and use it by day. Power plants already see peak-to-trough ratios between two and five.

In the short term, electrification of transport and heating will pull harder on the grid than AI will. Both are growing. Together they make a shortage likely. The transformer constraint is physical and industrial. Incoming power at 300,000 volts has to step all the way down to around 0.7 volts at the chip. The voltage drop is gigantic, and the transformer industry is not used to rapid growth. Useful compute per watt will get solved, Musk says. The question is how long the limiting factor at each stage is allowed to stall everything else.

Fridman asks how xAI and Tesla AI keep the race to AGI from becoming a race to the bottom. Musk says he has been a Cassandra on AI safety for more than a decade. Great power, great responsibility: he wants at least an objective third-party referee who can walk into the leading labs, understand what they are doing, and voice concerns publicly even with no enforcement power. Geoffrey Hinton left Google and then spoke. Once he is gone, who speaks from inside? Tesla, by contrast, answers to more than a hundred automotive regulators. The binders would fill the room they are sitting in.

Rockets get the same treatment, and Starship's limiter at the time of the taping was not hardware. The FAA had approved. Fish and Wildlife had not. Musk posted that he wanted to buy a fish license, a Monty Python joke that was also, he insists, the actual requirement. One analysis worried about the rocket hitting a shark. Ocean surface area being what it is, sharks are effectively zero percent of it, but the agency would not share distribution data for fear of shark-fin hunters, and another part of the same department that could run the numbers was not trusted by the first. Months later he learned of a separate organization concerned with hitting a whale in international waters. If a rocket did hit a whale, he says, that would be the least lucky whale ever.

Vandenberg in California added seal procreation. Rockets have launched from there for years, and the seal population has steadily increased, which would make sonic booms an aphrodisiac if you trusted the correlation. They were still required to kidnap a seal, strap it to a board, put headphones on it, and play boom sounds, twice, on two different seals. Musk has pictures. If he were the seal, he would assume he was about to be eaten, and his colony would never believe the alien-abduction version of the story. He is not arguing for zero regulation. He is arguing that the public does not know the madness, and that AI regulation should start with a referee who can see the facts and say them out loud.

Open-sourcing AI is not, in Musk's framing, a basement AGI problem. Grok was trained on 8,000 A100s at peak efficiency, and xAI planned to more than double compute every couple of months for the next several. He is generally in favor of open-sourcing base models, perhaps with a delay of six months. What he is not in favor of is what OpenAI became. He calls himself, arguably, the prime mover: the lab existed because of conversations with Larry Page, back when they were friends and Musk was staying at his house, talking about AI safety. Page, at the time, did not care. He called Musk a speciesist for being pro-human. Musk's reply was, which team are you on, Larry?

Google had acquired DeepMind, had perhaps two-thirds of the world's AI researchers, and had effectively infinite money and compute. The keystone hire at OpenAI was Ilya Sutskever, a recruiting fight Musk describes as mostly Demis Hassabis on one side and Musk on the other, with Ilya going back and forth for a long time before joining. Musk says he provided all of the early funding, over $40 million, and helped recruit others. The open in the name was supposed to mean open source. It was created as a nonprofit. "Now it is a closed source for maximum profit, which I think is not good karma." He would like to be friends with Page again. The break, as he tells it, was over OpenAI, and specifically over recruiting Ilya.

Even delayed open source is a strange artifact. The weights are a giant file of numbers. The lines of code are few. Most of the software work is curating data, separating good from bad. You cannot just crawl the internet. A huge percentage of websites have more noise than signal, many of them scam pages built for search-engine optimization. That is the unglamorous bottleneck Musk wants people to see: not a secret sauce in the trainer, but the fight to keep garbage out of the corpus. Grok, he notes, still drinks too much from conventional media. Later in the taping it repeats press versions of his SEC and defamation cases that juries, he says, actually decided in his favor.

Asked what role X will play in the 2024 U.S. elections, Musk says the goal is a level playing field for right, left, and independent. Old Twitter, in his account, was run by far-left activists who would have described themselves that way. Moving to the center looks like a move right if you stand on the far left. He says he is not aware of far-left users being suspended or de-amplified. The principle he wants on the platform is older than the election: "Free speech only matters if people you don't like are allowed to say things you don't like." If that is not the case, you do not have free speech, and it is only a matter of time before the censorship turns on you.

Community Notes is the product he is proudest of, and Fridman calls it a breath of fresh air in a climate of intensely viral posts. No system is perfect, Musk says, but he has yet to see an incorrect note survive more than a few hours. The mechanism uses AI to embed how people have rated notes in the past, in a vector space more complicated than left versus right. A note is shown only when people who have historically disagreed on a subject agree this time. If people who used to fight now sign the same context, it is probably true. Wikipedia, in his contrast, is hierarchical. Community Notes has no hierarchy. All of the code is open source, all of the data is open source, and anyone can recreate any note independently. "I can't change a Community Note if you put a gun to my head."

Fridman has followed Tesla Autopilot from the beginning and still treats it as one of the most capable real-world AI systems. The wild part of end-to-end training, Musk says, is that the net learned to read signs without being taught to read. Nobody taught it what a car, a person, or a cyclist was. It learned every object on the road from video, the way humans do. Humans are photons in, controls out: most of the information reaching the brain arrives through the eyes, and the outputs are motor signals to fingers and mouth. The car is the same loop. He agrees with Ilya Sutskever that models form a world model and that language, or video, is a projection of it. Autopilot and LLMs are converging on AGI. Tesla's path is more compute-efficient because it had no choice.

The onboard computer is about 100 watts, 144 trillion operations per second, which he calls small potatoes now, with activations in 256 values. The thinking part of a human brain is less than 10 watts, not counting the watts that run the body, and those 10 watts still write a better novel than a 10-megawatt GPU cluster. That is six orders of magnitude. AI has gotten here by brute force, throwing compute and power at the problem. Technologies first get made to work, then get made efficient, and he expects models to get smaller. Hardware 4 already fixed functions the car had been emulating; Hardware 5 will be better. Even Hardware 3, at 100 watts, looks to him able to drive better than a human. Optimized, it could be under 50 watts. Grok's list of his missed dates is fair, he admits: he is pathologically optimistic on schedule, and late, but he delivers in the end.

Optimus surprised him in the supply chain. There were no off-the-shelf motors, electronics, or sensors that fit a humanoid, at any price. Tens or hundreds of thousands of electric-motor designs exist in the world. None of them were suitable. Tesla designed the actuators from scratch, for manufacture the way a car is made. He thinks Optimus can eventually cost less than a car, because the robot is lighter and a car has more actuators. The milestone he wants is a robot that picks up a needle and thread and threads the needle just by looking. Maybe a year, with the usual caveat on his clocks. The car is a robot on four wheels; Optimus is a robot with hands and legs. Controls differ. The learning loop does not. Optimus will figure out tasks by watching videos.

Fridman quotes the line about being kind because everyone is fighting a battle you know nothing about, and asks what Musk is going through that people do not see. After a joke about Uber Lilith, he drops the armor. "My mind is a storm." He does not think most people would want to be him. They may think they would. They do not understand. Overall, in the grand scheme, he cannot complain. He gets lonely sometimes. His kids and friends keep him company. There are many nights he sleeps alone. He does not have to. He does. Asked how he knows whom to trust when everyone wants something, he says you look at track records over time and run the human neural net. He has almost never been betrayed. Karma, he believes, is real.

The hiring and friendship test he offered a few minutes earlier is the one he wants listeners to keep: never trust a cynic. Cynics excuse their own bad behavior by saying everyone does it. That is not true. Most people, in his view, are medium good. Walter Isaacson's biography had just asked, through Fridman, whether Musk would ever forgive his childhood. Musk does not think he is holding a resentment, so there is nothing to forgive. Holding grudges does not affect the future in a good way. What does, for him, is watching a child and a model at the same time. Raising little X and training neural nets have more parallels than he expected. He can see a biological net making more and more sense of the world, and a digital net doing the same. Kids make everything new. When you watch them see it, you see it again too.