So You're Saying There's a Chance: Musk on Neuralink, AI, and a Pale Blue Dot

Lex Fridman opens the second conversation with Elon Musk by describing him as a world-class engineer, not only the public face of Tesla, SpaceX, Neuralink, and The Boring Company. Most scientists and engineers, Fridman says, do not question the way things are done. They follow the momentum of the crowd. The ideas that change the world, on small scales and large ones, happen when someone returns to the fundamentals and asks whether there is a better way.

That frame is the door into Neuralink. The work, as Fridman sets it up, is not a single moonshot. It is supposed to treat neurobiological disease, to trace the path from an individual neuron to high-level function, and one day to expand the brain through two-way communication with computers, the internet, and AI systems. The near-term jobs are specific and useful. The long-term job is a hedge against a future in which digital intelligence outthinks us in every way.

Fridman jokes that a YouTube commenter called their previous recording the first unsupervised conversation between two robots. Then he asks what he calls an easy question, and it is not easy at all: is consciousness unique to humans, or does it permeate all matter, almost like a fundamental force of physics? The rest of the hour hangs on how Musk answers questions of that kind—what can be tested, what is only a preference, and what is worth building before the window closes.

Musk does not think consciousness permeates all matter. Fridman presses the obvious follow-up: how would you tell if that were true? Musk lands on the scientific method. If you cannot test the hypothesis, you cannot reach a meaningful conclusion that it is true. Without a test, you are talking about preferences or untestable beliefs. On that standard, panpsychism is not knowledge. It is a story you are allowed to like.

Can science understand consciousness at all? Musk will not claim that anyone understands anything with complete accuracy. He does believe we can dramatically improve our understanding of it. Fridman then asks whether a system needs consciousness, a body, a fear of mortality, or a capacity to love in order to reach human-level or superhuman intelligence. Musk splits the words apart. Intelligence, self-awareness, and what consciousness *feels* like are easy to conflate. Are we headed toward a future where AI can outthink us in every way? The answer, he says, is unequivocally yes. A system that outthinks us that thoroughly will be self-aware. That is still different from consciousness as a felt dimension, which might itself be an illusion.

The physical clue is blunt. If you damage the brain, you damage consciousness, which implies that consciousness is a physical phenomenon. Digital intelligence, Musk thinks, will soon be able to simulate what we consider consciousness well enough that you could not tell the difference. From the standpoint of the scientific method, a perfect simulation might as well be the thing. He wants a harder Turing test than a text chat: a phone call or a video conference with a digital superintelligence that has the right inflections, the small movements, the mistakes. If you cannot tell whether you are talking to a person or an AI, it might as well be human. That is the line he will carry into the darker half of the conversation.

Fridman turns to existential risk. Musk has said AI may be one of the greatest challenges civilization faces. Fridman offers three paths: scale up AI-safety and beneficial-AI research in hopes of an algorithmic or policy fix; become a multiplanetary species as fast as possible; or merge with AI and ride the wave of rising intelligence. Which is most promising? Musk does not pick a favorite. He starts with money. There is already a tremendous amount of investment in AI. What is missing is investment in safety.

His institutional answer is a government agency whose job is to confirm that a system does not represent a public-safety risk. Food has the FDA. Cars have automotive-safety rules. Aircraft have the FAA. When something can endanger the public, society has already decided it needs a referee serving the public interest. AI, he argues, is unequivocally in that category, and it should have a regulator for the same reason. Then he names the hole in his own argument. Government is very slow. A regulatory agency usually comes into being after something terrible happens, a public outcry follows, and years later a rule is written.

Seat belts are his exhibit. It was known for a decade or more that they would save lives and prevent serious injuries. The car industry fought a mandate tooth and nail. Hundreds of thousands of people, he says, probably died in that delay. The industry claimed people would not buy cars that had seat belts, a claim he calls obviously absurd. Tobacco fought for just as long, which is part of why he helped make the film *Thank You for Smoking*: to show how pernicious it is when companies achieve regulatory capture of government. Digital superintelligence, in this telling, is a singularity only in the sense that it is hard to predict what happens after it. Some probability it is bad, some probability it is good. A referee is how you try to push the mix. The problem is that AI will not wait for the crash that historically creates the referee.

Fridman brings the conversation to Neuralink's hardware: flexible wires, a robotic sewing machine, implants that move with the brain, a safety case wrapped around all of it. We currently understand very little about the human brain. Will this work teach us more about the mind? Musk thinks it will shed a lot of insight, because the data we have now is so thin. An MRI, in his image, is like putting a stethoscope on the outside of a factory wall and walking it all over the building. You hear sounds. You do not know what the machines are doing. You can infer a few things in a broad brushstroke. To know what is going on, you need high-precision sensors and a stimulus-and-response loop: fire this neuron, and what do you feel, see, or change in your perception of the world?

Doing that in people is profound for a reason machines cannot match. People can articulate the change. Mood shifts, sharper vision or hearing, sentences that come easier or worse, a memory that gets jogged—a subject can tell you. Fridman then asks about plasticity. Brains adapt; neural nets adapt; the two sides have to find a communication protocol, almost like meeting an alien species. Who does more of the reaching? Musk is not romantic about wetware. The brain is not that plastic, to be totally frank. There is some give, and a firm structure. The machine side is far more malleable than the biological side, by a huge amount. The implant will have to adapt to the brain. That, he says, is the only thing that is possible.

Neurons will not start to record an electrode as another neuron, because a neuron is just a pulse. Something else pulsing is not the same kind of neighbor. There can be some elasticity in the interface, and the brain will adjust a little to a thing that is reading and stimulating it. The vast majority of the adjustment still has to be on the machine side. Otherwise, Musk says, it will not work. Fridman wonders whether synaptic plasticity at the interface might become a new kind of malleability, an extension of the brain rather than a rewrite of cortex. Musk does not bite. Some adjustment, yes. Most of it, no. The asymmetry is the design constraint, not a philosophical preference.

Musk's working model of a person is two layers. Underneath sits the limbic, primitive brain, where the impulses come from. On top sits the cortex. The human brain, he says, is a monkey brain with a computer stuck on it. A lot of our impulses are driven by the monkey. The cortex is constantly trying to make the monkey brain happy. It is not the smart layer steering the primitive one. It is the other way around. The cortex is the part that tells the story, so we convince ourselves the whole performance is more interesting than monkey business. Other creatures have a weak computer or none. We have a strong one. It still does not get the wheel.

The intelligence gap between cortex and limbic system is gigantic. The limbic system has no comprehension of what the cortex is doing. It is hungry, tired, angry, or sexy, and it sends that impulse upstairs. The cortex goes to satisfy it. A stupendous amount of thinking, Musk says, has gone into sex without procreation, which is a bit silly as an action if reproduction is not the point. Why do it? Because it makes the limbic system happy. Fridman notes that the whole of existence is pretty absurd. Musk is happy to keep going. Digital neural nets will interface with biological neural nets and, he hopes, bring us along for the ride. The vast majority of our intelligence, in that future, will be digital.

What is Neuralink for, in the nearer term? Musk lists brain-related disease: autism, schizophrenia, memory loss, parents who cannot remember their kids' names. Spinal cord injury, stroke, missing circuitry replaced with silicon. Those are steps along the way. Ultimately the device is meant to address the existential risk of digital superintelligence. You will not be smarter than a digital supercomputer, so you cannot beat it. Join it. At least have the option. Fridman asks whether he really hopes the link lets us merge and ride improving AI. Musk says the chances are above zero percent. There is a chance. Fridman supplies the *Dumb and Dumber* line. Musk takes it. Maybe the odds move from one in a million toward one in a thousand, then one in a hundred, then one in ten, depending on how fast Neuralink improves.

The long picture is a third layer. Limbic system first, cortex second, digital superintelligence third. People generally like having both of the first two and do not want to delete either; the fun, Musk says, is in the limbic system, and they still want the cortex. His optimism is analogical. The cortex is vastly more intelligent than the limbic system, and they still work together. Perhaps a tertiary digital layer, vastly more intelligent than the cortex, can coexist peacefully with both. The timing is not optional. Neuralink has to solve this sooner rather than later, because the point at which digital superintelligence arrives is the singularity: not necessarily good or bad, but extremely unstable. He wants a human brain interface before that point, or at least not long after, to minimize existential risk for humanity and for consciousness as we know it.

Fridman asks what the actual engineering problems are. Musk's list is a stack of disciplines: material science, electrical engineering, software, mechanical engineering, microfabrication, chip design. The electrode has to be tiny enough that it does not hurt neurons, and it has to last as long as a person—decades. The signal has to be processed locally at very low power, because the brain is heat-sensitive and a hot chip is a little furnace next to tissue. Then the device has to do something with those signals and stimulate back. Bidirectional communication is the point. Reading without writing is only half a circuit.

The materials problem is the one he calls very hard. You are trying to read and stimulate electrically in a region that is already electrically and chemically active. How do you put a coating on the electrode that does not dissolve over time and is still safe in the brain? Collecting the signals is a power budget as much as an algorithm: you have only tiny amounts of power to process them. Then the whole procedure has to be automated. If it is done by neurosurgeons, one patient at a time, there is no way it scales to a large number of people. It needs to scale, Musk says, because ultimately we want the future to be determined by a large number of humans.

Fridman asks whether this could revolutionize surgery more broadly. Musk says it has to be like LASIK. If LASIK were still done by hand, that would not be great. A robot does it. The ophthalmologist mostly needs to make sure the head is in position, then press the button. Brain implants, in this telling, should travel the same road. Scale is not a business-model footnote. It is how a species keeps a vote in a world that is about to grow a smarter layer than cortex.

The conversation slides to Tesla. Fridman suggests that Autopilot could educate the public about AI and robotics, because people who may never have seen a robot will watch hundreds of thousands of cars with no one in the driver's seat. Does Musk feel that as a burden, or is it just a parking feature? He says they are really just trying to make people's lives easier with autonomy. Now that Fridman mentions it, though, he agrees it will be an eye-opener. Most people have never really seen a robot. An autonomous car is a four-wheel robot. There are already hundreds of thousands with autonomous capability; it will not be long, he says, before there is a million that can drive without a person in them.

He claims he can tell immediately whether a car is on Autopilot from little nuances of movement. Each iteration of the software has a slightly different personality. On the highway, the system is far more precise about sitting in the center of the lane than a person is. Look at human-driven cars in their lanes, he says, and they are like bumper cars, moving all over the place. The Autopilot car holds the center. Fridman asks where Tesla stands on full autonomy. On the highway, the probability of needing to intervene is already extremely low. In stop-and-go traffic, Musk says Autopilot is far safer than a person right now: the chance of injury is much lower. It is changing lanes and taking highway interchanges. They are also coming at the problem from the other direction: low-speed full autonomy.

That low-speed path is how a person learns to drive. You start in a parking lot. You do not begin on a chaotic street in San Francisco. Get it right slowly, then expand. The missing pieces he names are traffic lights and very windy roads. Side streets, in his telling, are relatively easier because you at least know where the street is. Lights and winding geometry are the two things that still need to be solved. The folk test of maturity is already visible from the next lane: a car that does not wander. The remaining work is to make that calm behavior hold when the world stops looking like a highway.

Fridman asks which is harder, perception or control: perfectly seeing everything, or planning once you have seen it. For the giant multitask network Tesla is training, Musk's answer is perception. The hardest thing is an accurate representation of the physical objects in vector space. Take visual input, plus some sonar and radar, and build a precise map of the objects around you. Once that representation is accurate, planning and control are relatively easy. They become, in his analogy, a video game.

Cars in *Grand Theft Auto* drive down the road and do not crash unless you crash into them. That is because they already have an accurate vector-space representation of where the other cars are. The rendering is just the outward show. Full autonomy on the highway, Musk says, is on track. Fridman notes they still have driver sensing via torque on the wheel. Then the conversation loosens. Tesla karaoke, Fridman suggests, is not only fun; sitting in a car singing might help attention and vigilance. Musk's reply is a shrug that fits the rest of the hour: you can do both at the same time. Fun and a safety feature do not have to be a trade.

Fridman has just met with Ann Druyan, Carl Sagan's widow, and he is a fan of *Cosmos*. Sagan, he says, had a great way of putting things: all of consciousness, all of civilization, everything we have ever known, sits on a tiny blue dot. People get trapped in squabbles among humans and take continued existence for granted. They should not. Civilizations rise and fall. Now civilization is globalized, so it rises and falls together. Geographic isolation is gone. Things do not always go up. That, Fridman says, is an important lesson of history. In 1990, at Sagan's request, Voyager 1 turned around from 3.7 billion miles away and photographed Earth. The planet takes up less than a single pixel. Sagan called it a pale blue dot, and he spoke about that image in 1994. Fridman asks Musk to read the words.

Musk first does the arithmetic. The universe appears to be 13.8 billion years old. Earth is about 4.5 billion. In another half billion years or so, the Sun will expand, probably evaporate the oceans, and make life impossible here. If consciousness had taken ten percent longer to evolve, it would never have bothered. Just ten percent longer. He wonders how many dead one-planet civilizations are out there, species that never made it to another world and extinguished themselves or were destroyed by external factors. Probably a few. It is only just possible, he says, to try to travel to Mars, and only barely. A ten-percent shift in the relevant gravity, in his telling, would make even a single-stage trip from the Martian surface toward Earth unworkable. The window is luck as much as prowess.

Then he reads Sagan. Look again at that dot: that's here, that's home, that's us. Everyone you love, everyone you know, everyone you have ever heard of, every human being who ever was, lived out their lives on it. The aggregate of joy and suffering, thousands of confident religions and doctrines, every hunter and forager, hero and coward, king and peasant, young couple, saint and sinner—on a mote of dust suspended in a sunbeam. Our planet is a lonely speck in a great enveloping cosmic dark. In all that vastness, there is no hint that help will come from elsewhere to save us from ourselves. Earth is the only world known so far to harbor life. There is nowhere else, at least in the near future, to which our species could migrate. Musk stops. That last claim, he says, is not true. That is what Mars is. He thinks Sagan would agree, and that Sagan simply could not imagine it at the time.

The correction is small and it is the whole method of the hour. Seat-belt rules, an AI referee, a two-layer brain, a factory-wall MRI, a vector-space car: Musk keeps taking a settled story and asking whether it still survives contact with a test. Sagan's 1994 conclusion was honest about the Earth he had. A generation later Musk is trying to make the sentence expire. No one is coming to save us from ourselves. If there is a second world, we will have to build the way there. Neuralink, in that light, is not a gadget. It is one more attempt to keep consciousness in the game while the clock—ten percent, half a billion years, or the arrival of a digital layer—is still running.