The IQ required to destroy the world drops by one point every year

Lex Fridman opens the second conversation with George Hotz — Geohot, iPhone jailbreaker, PlayStation 3 hacker, founder of Comma AI — by asking whether there is intelligent life elsewhere in the universe. Hotz's standing answer to the Fermi paradox has not changed. Intelligent civilizations have existed. None of them are left, because they destroyed themselves.

He cites a line from Eliezer Yudkowsky: the IQ required to destroy the world falls by one point every year. Technology, in this telling, democratizes annihilation. Cryptography is the rare domain where defense still wins, because verifying a signature is easier than forging one. Nuclear weapons are the opposite. Building a warhead that splits into a hundred is far easier than building something that can intercept a hundred little warheads. "Maybe our future is in crypto," he says.

The galaxy is about 70,000 light-years across, a number he half-jokingly credits to *Star Trek: Voyager*. A von Neumann probe traveling at sublight speed could fill it in a million years. That clearly did not happen, so something stopped expansion. Waste heat is the signature he would look for: civilizations that grow consume more energy, and thermodynamics will not let them hide the residue. He grants that physics on the far side of the galaxy could be stranger than we think, but he is willing to bet it is not.

Fridman brings up the Pentagon UFO videos, Navy pilot David Fravor, and the objects nicknamed "Tic Tacs." Hotz's reply is one sentence: he thinks all of that news is a psyop. He listened to Bob Lazar on Joe Rogan and says he believes every word — then immediately adds that it was probably some MKUltra-style operation. They built a weird thing, called it an alien craft, and maybe the point was to stimulate young physicists and see what they invented.

The conspiracy theory he finds most useful is that conspiracy theories themselves were invented by the CIA in the 1960s to discredit true things. Flat Earth and Pizzagate, in this view, exist to dilute. China locking up the doctors who discovered the coronavirus actually happened. Money used to be backed by gold and now is backed by nothing; that actually happened. Governments used to control information by censoring it. After the internet, they learned the leprechaun trick: plant one flag over the gold, then fill the field with flags so no one can find the real one.

His 2019 SXSW talk about "hacking the simulation" was widely misread as a literal plan. He says the point was theoretical physics: if you want a grand unified theory, go build one. The real message was smaller and harsher. Look at what you spend time on. You are bytes in God's computer. He is a fan of capitalism, but he does not think that is the game being played. The rules are rigged. Nature wrote a real game. Instagram's version of nature — big things eating small things in order to mate — is not the whole story. Nature, to him, is an open-world MMO. You can play the cow level in Diablo for hundreds of hours, as he did in college. That does not make it the game worth playing.

How do you hack the real game? Hotz wants to live forever — not true unkillable immortality, which he thinks would be worse than what we have, but the right to choose his own death. Nature does not get to decide. Fridman objects that the hunger to know everything may be a byproduct of a finite life, and that curiosity would vanish like morning fog if time never ran out. Hotz's answer is an engineer's: change the thing in the brain that reminds him life is finite. Maybe tell it that life ends next week. He is fine with some self-manipulation. If the risk is losing curiosity, keep backup copies of himself and let them override the trajectory.

Fridman compares that to cheat codes in a video game: once you use them, the game is ruined, and the damage does not reverse. Hotz knows the feeling. He treats it as a limitation of current brain-editing technology. Delete the code. He also admits that self-modifying systems produce hairy complexity, and he cannot promise he will get it right — only that he can put safety structures in place.

He does not doubt that we live in a simulation. He just does not think it matters unless the theory gives him power over nature. Unfalsifiable ideas are not worth caring about. The evidence he would take seriously is a memory leak: the ability to read memory you were not supposed to read, a shortcut in the physics. Reading is not the horror. Writing is. Even writing random noise would be terrifying. String theory, Eric Weinstein's Geometric Unity, Stephen Wolfram's hypergraph — he will pay attention when they make a testable prediction that yields power. He wants a 100 GHz processor and transistors smaller than atoms, not a beautiful untestable story.

In mid-2020 Hotz walked into a New York restaurant, ran into engineers from Optimism — a Plasma spin-off building layer-two scaling for Ethereum — and wrote them out of a deadline. Ethereum's base chain makes every node run every transaction, which is secure and slow. Layer two moves compute off-chain and posts results, then uses a dispute mechanism: stake $10,000 that the result is real, and anyone who proves you wrong can take the bounty. That requires a sandbox. Calls that write state to the chain have to be replaced with requests to a hypervisor, the way a Chrome renderer cannot open a file without asking the browser.

Optimism had compiled Solidity to bytecode, then written a 3,000-line "transpiler" that decompiled, rewrote, and reassembled it into OVM-safe bytecode. Hotz's first reaction was that they should change the compiler instead. He asked how much they would pay, then wrote a 300-line C++ patch. Asked how, he said you just think and then do it, a sentence he later admits does not mean much. The transferable move is going one layer up when everyone else is reverse-engineering downstream.

He is extremely bullish on crypto as an idea, not on any particular coin. Nakamoto consensus — a relatively straightforward algorithm that lets a group reach decentralized agreement — is one of the great innovations of the 21st century. The second is smart contracts. If a contract is disputed, lawyers interpret it. Lawyers, he says, are really shitty, overpaid interpreters. Python on a nice cloud machine costs ten cents an hour; a lawyer costs a thousand. Python is ten thousand times cheaper and almost always returns the same answer. A safety-deposit box that opens when two of three parties are present depends on the bank and the lawyers. A smart contract checks two of three keys with cryptography and gives you certainty, which is what businesses actually want. He mentions the Viacom–YouTube lawsuit: neither side cared who won, they just needed an answer.

He will not pick a winner. He is in the "Bitcoin as a meme coin" camp, mostly because the block-size debate struck him as too stupid to engage. He has come around somewhat on proof of stake because very smart people are thinking about it. Even in distributed systems, he says, leaders help, because humans need a face. Ethereum has Vitalik Buterin, and Hotz thinks he is the face *because* he is a genius. He wishes presidential debates asked what the Sixth Amendment is, what the four fundamental forces are, and what the integral of two to the x is. The crypto world's other gift is the fork. In 2012 he joked that Obama and Romney should both be president, and Americans could pull liquidity from one into the other. Ethereum actually forked into Ethereum and Ethereum Classic. He would like to fork companies the same way: one closed-source Comma, one open-source, one bought by GM, one still the Android renegade, and let the market decide.

Comma's mission statement is exact: solve self-driving cars while delivering shippable intermediaries. Solving driving means building a person replacement — something that sits in the driver's seat and takes you anywhere, at human or better safety, speed, quality, and comfort. It does not mean building a new car. The second clause is the design. On Hacker News he saw the line he now repeats: you can set any arbitrary milestone, meet it, and still be infinitely far from solving self-driving. Without a product, it is easy to think you are making progress when you are not. Revenue is the honesty mechanism. When someone spends a thousand dollars on a Comma Two, the company has provided at least a thousand dollars of value, or they would not buy it.

The Comma Two is basically a phone: a Qualcomm Snapdragon 820, a cooling system, and a car interface. Four CAN buses come out over USB. CAN is the protocol cars already use to talk, and it is surprisingly simple to read and write. The software, openpilot, is everything around the neural net. The model itself is "a blob of weights." People call it a black box; Hotz says it is a blob of weights, what did you expect. Openpilot cleans the inputs, cleans the outputs, runs control, handles sensors and calibration, counter-steers on a banked road, ships logs to the servers, and manages heat and disk space.

Since the last podcast he has kept a promise. Lateral policy is now almost completely end-to-end. You can turn the lane lines off and it will still drive, only slightly worse on the highway, trained on user data. This year they hope to do the same for longitudinal control. Data quantity is not the bottleneck; selection is. His preferred definition: in supervised learning the weights depend on the data, and in reinforcement learning the data depends on the weights. The end state is that whenever a car makes a mistake it disengages, that clip becomes training data, and a new model ships. They are not there yet. Unplanned disengagements now happen about every hundred miles, up from every ten a year earlier. That is a 10x improvement, which already assumes exponential progress. They still need three more orders of magnitude to reach the hundred thousand miles that would make the system broadly usable. Exponential does not mean a 100 GHz processor tomorrow.

Tesla breaks driving into hundreds of machine-learning tasks: lane detection, stop signs, traffic lights, drivable area, vehicles, bikes, pedestrians, localization, future trajectories. Each task gets a data engine mining edge cases, and an engineer who can become the best in the world at that one job. Hotz's question is whether you would hire a bishop guy if you were starting a chess-engine company. In a world with AlphaZero, that looks ludicrous. He does not claim Tesla's approach is impossible, only that the engineering magnitude may exceed what humanity can finish. End-to-end is much easier as a path to Level 5.

He calls Tesla's method feature engineering at a higher level of abstraction — task engineering. The history of AI, as he reads it, is that feature engineering always loses to end-to-end. Compare Stockfish's code to AlphaZero's: the latter is shorter, more elegant, and took far fewer programming hours. Fridman's objection is the murder. AlphaZero can kill billions of failed agents in simulation at nearly zero cost. Biological evolution produced humans by killing trillions of organisms. You cannot do that on a real road.

Hotz's answer is MuZero. He thinks it will be remembered as the cornerstone paper of this deep-learning era. Deep learning is not everything — pieces are still missing — but MuZero is the solution to self-driving, with a few tweaks. It learns a dynamics model from observations, then does the rollouts and the murders inside that learned simulator. When he read the paper he felt he could sit back and wait for the solution to arrive. That year he saw the future. Neural nets will solve driving; nothing else is required. The timeline is the later 2020s. The path still has to go through products, so it will be slow. The only difference between Level 2 and the levels above it, he adds, is who takes liability, and he does not want to be a liability guy.

As the system gets better, drivers get lazier. Comma's answer is already adaptive. If the car is not moving, it does not enforce monitoring. On a 45-mile-an-hour road with lights, stop signs, and possible pedestrians, the policy is tight. Alone on a perfectly straight highway, it relaxes. None of that is hardcoded; a model estimates risk. People ask why you would ever make monitoring less aggressive. Because if you alert someone when there is nothing they could possibly hit, they learn to tune it out. They hang a weight on the steering wheel. The human is always part of the adaptive system.

He is willing to bet $10,000 that Elon Musk will ship driver monitoring before he ships Level 5. If Hotz were in that seat, he would publicly say monitoring is the wrong path too: you already built a large fleet without the camera, and you cannot tell customers their cars lack the thing you now think is right. Cadillac Super Cruise showed him the failure mode. Sun through the window blinded the camera, the car said he was not paying attention while he was looking straight ahead, and he could not reset attention by touching the wheel, so the system forced a disengagement. It should have been a conversation. Comma ships a new policy every month and balances it with users on Discord, as aggressive as people will tolerate without feeling watched.

He calls driver monitoring easy relative to external perception for two reasons. Faces and bodies vary less than the outside world, and the reliability bar is much lower. A system that is right 90 percent of the time still captures almost all the benefit, because you are training a cooperative human. Swerve into a tree and there is no margin. Over-the-air updates, which he says should have been obvious from the iPhone, are what make the conversation possible. He also wants the car to communicate uncertainty through haptic feel rather than always acting confident until it fails, and he is hiring UI people to make that visible.

Tesla's Dojo, in Hotz's reading, is not a chip specialized for driving data. It is Tesla going the Apple way and owning every link in the chain. The training-hardware market, as he sees it, has two names: Nvidia and Google. Google Cloud is not a real option. Its terms of service, at least in the version he remembers, forbade using Cloud ML to train autonomous vehicles or to compete with Google without written permission. That leaves a monopoly.

In 2016 he was a huge Nvidia fan. Jensen Huang was a cool guy, and the stock was $30. After 2018 he watched the management posture change: extract every dollar from the ecosystem, charge $10,000 for a card that is not that different from a $699 RTX 3080. From a business perspective he thinks they are shooting themselves in the foot. If he ran Nvidia he would sell chips at a reasonable markup to everybody, with no restrictions, the way Intel did during its long run. Productizing the chips too hard extracts more value than the company can sustain. Tomorrow's share price goes up. The CEO question is how far you can strip-mine the place. Huang is supposed to be a fifty-year-vision person, so Hotz cannot make the gouging fit. His guess is that Huang is listening to the wrong people.

He would apply the same logic to Tesla. Sell the accelerator. Anyone can build that chip if they are willing to spend the money; selling it makes cash and lets the world build the toolchain and fix bugs in the software layer. Keep the batteries. Batteries are the thing that competes head-on with other automakers. Comma, meanwhile, is in the middle of its own rewrite: moving from a flat-world hypothesis, where lane lines pitch when the car pitches, to models that put everything in 3D, and switching the stack from TensorFlow to PyTorch. They have a crude data engine — train, run on a hundred thousand images, take the high-loss examples, label a hundred, repeat — but for driving itself the human already annotated the data by driving.

Openpilot supported 91 cars at the time of the recording, with 100 as the year's goal. Hotz recommends a 2020 Corolla, or a 2020 Sonata that keeps the factory longitudinal control and only takes over laterally. A Honda Civic is great if you will install a modified EPS firmware, which he will not help you with in public. Installation is IKEA-furniture level: ten minutes, YouTube videos, and if you can assemble a table you can install a Comma Two. About 60 percent of users run stock software; about 40 percent run forks. Forks that add readouts or stop disengaging on the gas pedal are fine. Forks that subvert safety features get banned from the community.

The device is $1,000 plus about $200 for the car harness, with a 30-day no-questions money-back guarantee, and he says future hardware will cost more. Retention is the number he likes: 70 percent of Comma Two buyers are daily active users. Daily actives were around 2,000 to 2,500, monthlies over 3,000. He thinks openpilot may one day cross Autopilot in users, the way Android crossed iOS, and that Comma will never cross Tesla in revenue. Tesla can extract far more money per user. If only 10 percent of openpilot devices are official Comma hardware, that is still a lot of money for the company. Next year's sketch is 10,000 units; a later sketch is 100,000 units at a higher price, which starts to look like real revenue.

Acquisition offers make him laugh. M&A teams treat startups as their outsourced innovation department and expect the founder to play along for some "scam fake Fed dollars." Respect decides the rest. GM: never. Hyundai: more plausible. Japanese, German, and American automakers, he says, think they are too great to worry about tech companies. Among tech firms, Apple is the closest he could respect. He quit Facebook ten years earlier over the business model. Google, in his view, has declined fast in five years.

Hotz has visited Waymo and ridden in the cars. The engineering, the research, the scale of building it in-house, all impress him. In the geo-fenced area the vehicles work, even going at or above the speed limit. The failure, as he sees it, is not technology. It is product. Benchmarked against a competent Uber driver, Waymo is about 20 percent slower. You can argue it will be cheaper, but riders already have a way to trade money for time: Uber Pool. He puts the share of Pool users who will not make that trade at about 15 percent. The product, compared with ordinary rideshare, is not competitive.

Profit, in this market, depends on monopoly. He thinks Level 4 autonomous rideshare, if the technology even arrives, will look like the scooter market: a race to the bottom. Scooter companies have already stopped pretending they are practical transportation and started selling fun, which he respects as honesty. As transportation, neither scooters nor Level 4 robotaxis make sense to him. Asked what he would do if he ran Waymo for a year, he says he would get Anthony Levandowski out of jail and put him in charge. He means it. He thinks Levandowski is a genius and would have a better idea than Hotz would. Waymo is already $10 billion in the hole. Comma has raised $8.1 million. Hotz cannot think of a plan that pays that hole back; the plan would have to be an enormous risk, which is why he wants Levandowski.

His other criticism is cultural. Waymo went corny too early. A startup that has not delivered should look like a renegade, put engineers in front, and show failures the way SpaceX does with rockets. He does not accept the friendly-car ads. The person in them, he says, is a paid actor, not a Waymo user — he found the call sheet. If you lie to him, he wants you to fail. If you tell the truth, he wants you to succeed. Pre-revenue with $10 billion raised violates every Silicon Valley instinct he still believes in: where is the MVP, where are the users, where is the growth. Stay mission-focused, he tells Fridman. Be arrogant in the face of other people if you want. Be arrogant in the face of nature and you are an idiot. Winning is Level 5. Until then, you have not won.

Generic startup advice is not something Hotz wants to sell. The one line he will stand behind: technology always wins, and lying always loses. Keynes's warning that markets can stay irrational longer than you can stay solvent is the obvious objection, and he does not dodge it — that is why you need something else underneath you. He rejects "fake it till you make it." Be honest and hope you make it the whole way. Hype backed by something you can actually buy is fine; he loved camping out for iPhones. Hype backed by lies is not.

Elon Musk overpromises timelines and delivers later. Hotz does the same; he thinks it is a programmer thing. If Musk says they will land a rocket in a year and it takes four, but they work toward it the whole time and land it, that is not a lie. It becomes a lie when you know you will not meet the deadline. Bullshit, in this definition, requires that you are conscious of lying. Billy McFarland probably believed everything he said too, and that is how a cult starts. You need a coefficient, and you need people who can pull you back. People root for others to fail, Hotz thinks, because someone else's success is a verdict on their own life. A 36-year-old who has spent a career at large tech companies has to decide that the wild new thing will fail, because if it succeeds it means he could have done something different.

Commenters are not the population. He does not comment on anything. Twitter's one-way likes make negativity travel through replies, which is part of its particular toxicity; YouTube comments, with up and down votes, feel better. Most people would speak if the room were not full of assholes. Platforms default toward entropy. If you do not spend effort encouraging nuance, they slide into chaos. Small Reddit communities still hold joy and disagreement, but that is emergent, not a platform feature. He thinks Twitter can be overthrown, and that we will look back and not believe we put up with this much toxicity. The business-model problem is the one *The Social Dilemma* named more carefully than "you are the product": the product being sold is influence over you. In a real conversation he spends effort watching whether he has been stupid or mean. On the internet most people refuse to pay that cost. Reward the ones who will, and you might be surprised how many show up.

Asked for a perfect productive day, Hotz names three dates instead. The first was reading old Eliezer Yudkowsky on the singularity, before AI safety took over the page: computers double in power every 18 months and humans do not, and that graph opened the door to human obsolescence. The second was the Hutter Prize. Intelligence equals compression. He had been searching for a definition of intelligence and never had one; lossless compression has a clean definition, and he still finds the equivalence compelling even if it is not useful day to day. He spent six months on the prize, then worked at Vicarious and caught up to modern deep learning. The third, during Gamergate and the run-up to 2016, was the blog *Unqualified Reservations* by Curtis Yarvin. Suddenly the world made sense; he had a framework. He did nothing special to find those days, so he does not think he has to do anything special now. He drifts, hoping for another one. A snowstorm on Colorado's Monarch Pass forced him to turn around and spend the night in Gunnison. That day mattered because it was real. "It's rare that anything even real happens in my life."

GPT-3 is cool for what it is. The loss function is cross-entropy on characters, which is not the loss function of general intelligence, and the models have no long-term memory, only a hardcoded context width. He does not believe you scale that into something that can build a rocket. Driving is more Markovian than language: most of the information in state t-minus-one is already in state t. Language makes you look back at the third paragraph on the second page; driving rarely does, except for things like speed-limit signs. Trying Transformers on Comma's perception space would be a GRU-to-Transformer swap, not a revolution. He does not believe in a hard-takeoff AGI. The singularity is bandwidth: the world is already networked, and we are raising the speed of that network until it becomes one thing. Driving is a pretty complete subset of general intelligence, which is why he is doing it. In AI it is too easy to delude yourself into thinking you have figured something out. If you build a Level 5 car, you have indisputably built something. If humanity had to choose, he would rather we go to Mars. He is an information person, so his own work stays on this side.

A MacBook Air in the corner of a hotel room is enough. He codes mostly in Emacs, an editor from the 1970s. After months switching the stack from TensorFlow to PyTorch he felt he had wasted years; Emacs has no such defense, the heart wants what it wants. Everyone should write assembly and C, not to mastery, but enough to know what the computer is doing, then Python, then Haskell and a dependently typed language, then the SIMD world where a long block of instructions all execute at once. Karpathy calls the newest paradigm Programming 2.0; Hotz prefers data-driven programming. Nobody learns to program by watching a video called *Learn Programming*. Everyone he knows who can program well had something they wanted to do, tried to do it, and noticed the computer could help. People who ask how to apply machine learning to macaroni and cheese production have the wrong question. The right one is that some of the boxes are not getting enough macaroni.

The books: *Infinite Jest*, which is about where we are heading. *Atlas Shrugged*, which he will not defend as literature — "it's pornography," a 60-page monologue Rand paid to keep in — but it is a framework for human relations, and the villains' slogan really was the Communist Party's. Greg Egan's *Permutation City* for copies of yourself; *The Metamorphosis of Prime Intellect*, a post-singularity story that reads like an engineer wrote it, first read at 15. Everyone should read *Neuromancer*. *A Fire Upon the Deep* and *Dune* come up; *Foundation* felt too modernist. Love, for him, splits into opiates, oxytocin, and the love of math. He is not a fan of the first two, for the same reason he does not do opiates or ecstasy: the ethical life is the productive life. Psychedelics are not uniquely educational. Every drug teaches, but you learn about yourself, not the world. Look behind the door once; you do not need to keep going back. He revisits small doses maybe once a year. Big learning rates have pros and cons. On a one-to-ten scale of crazy he gives himself a seven, then doubts the frame. Schizophrenia and hard bipolar are real. ADHD and oppositional defiant disorder look like the normal human spectrum. People on his streams think he is always on drugs. He had a cup of coffee. The rest is his brain in action. Be willing to sparkle. Do not let the constraints in your brain force it into the mold of what society says you are supposed to be.