A Car That Cannot Drive Itself Will Be About as Useful as a Horse

Lex Fridman frames the taping carefully. The conversation happened after his MIT group released a paper on driver functional vigilance during Tesla Autopilot use. Tesla offered a podcast with Musk; Fridman says he kept control of the questions and cut nothing of substance. He had never spoken with Musk before, Tesla has not funded his research, and he has never owned a Tesla or Tesla stock. He and Musk disagree, in particular, on how long camera-based driver monitoring will matter.

Asked for the original vision of Autopilot when hardware started going into cars in 2014, Musk refuses the word “dream.” He sees two massive revolutions in the auto industry: electrification, and autonomy. In the future, he says, any car that does not have autonomy would be about as useful as a horse — which is not to say horses have no use, only that owning one is now rare and somewhat idiosyncratic. Cars will drive themselves completely. It is a question of time. If Tesla skipped the autonomy revolution, its cars would be less useful than cars that are autonomous.

He puts a number on it. An autonomous car is arguably worth five to ten times a car that is not, in the long term — at least the next five years, perhaps ten. That claim, more than the horse joke, is the design brief underneath the rest of the conversation: what the screen shows, how data is labeled, whether a steering-wheel camera is worth building, and how an error is even defined.

One of Autopilot’s early design choices was to show, on the instrument cluster or on the Model 3 center display, what the combined sensor suite sees. Musk says the whole point of that display is a health check on the vehicle’s perception of reality. The car takes information from a bunch of sensors — primarily cameras, but also radar, ultrasonics, GPS — renders it into vector space as objects with properties (lane lines, traffic lights, other cars), and then re-renders that vector world onto the screen.

You confirm whether the car knows what is going on by looking out the window. He calls that “an extremely powerful thing” for people trying to understand the system and become one with what it can and cannot do. The public view is a clean, crisp picture of nearby vehicles. The contract is simple: if the drawing matches the road, you can trust the moment; if it does not, you take over.

Fridman asks why not show more of the computer-vision stack — road segmentation, lane detection, object detection, and the uncertainty at the edges, the probabilities that might help ordinary drivers build an intuition for vision. Musk says that in his own car he always looks at the debug view, and there are two. Augmented vision draws boxes and labels around recognized objects. The visualizer is a vector-space sum of all sensors with no pictures at all, the car’s view of the world. Normal people, he says, would not know what they were looking at. The current display is an HMI choice, optimized for a public that does not know how computer vision works. The development build’s debug information would be “total Greek” to most people.

Fridman splits the technical work into algorithms (neural-net architecture), data, and hardware, and asks where limited money and time should go. Musk’s answer starts with the fleet. Tesla cars carry eight outward-facing cameras, radar, usually twelve ultrasonic sensors, GPS, and an IMU. About 400,000 vehicles on the road have that full suite, approaching half a million. He would be surprised if more than 5,000 other cars on the road are instrumented that completely, which in his arithmetic means Tesla has 99% of all the data.

Deep learning gets better with data, but driving’s valuable data is in the edge cases. Smooth miles are cheap because the system has already learned them. Musk describes two ways to find the rest. One is a disengagement: someone on Autopilot takes over, and that trigger has to be sorted into convenience versus Autopilot not working properly. The other is the opposite of a failure. If you want the optimal spline through an intersection, the runs with no intervention are the right ones. Collect enough of “when it looks like this, do the following,” and you fit a complex intersection without hand-labeling a path.

That is a data-mining trick as much as an autonomy trick. Hundreds of thousands of cars repeating the same junction become a labeled pool: successful passes are the answer key, takeovers are the questions the system still fails. The fleet is not only a volume advantage. It is a way to harvest both the common case and the rare mistake at a scale almost nobody else can match.

Musk compresses the takeover problem into a rule Fridman calls a powerful line: view all input as error. If the user had to do input, it means something is wrong. Every intervention becomes a hard constraint, not a soft UX complaint. The driver touched the controls because the system, in that person’s judgment, was not doing the right thing.

Fridman immediately puts a discount on it. Exiting a highway, or making a navigation decision Autopilot is not designed to make, is not a perception failure; it is a missing feature. How do you tell the difference? Musk says that discount is about to shrink. Navigate on Autopilot had just been released without a confirm step, so lane changes, freeway exits, and highway interchanges — the vast majority of those navigation takeovers — would go away. It also automatically overtakes slow cars and seeks the fastest lane.

He does not think people quite understand how big a step that is. Fridman agrees: if you drive the car, you do. Drivers still had to keep hands on the wheel during automatic lane changes. Asked what the big leaps in Autopilot’s history have been, Musk names this one — Navigate on Autopilot without confirm — and traffic-light recognition, first as a warning, and on his development build as the car fully stopping and going at lights. Early Autopilot failed out loud, with lost lanes and a shaking wheel. As the system gets better, failures get rarer and quieter, which is exactly why each remaining input is worth more as training signal.

Fridman asks for the biggest technological roadblocks to full self-driving. Musk’s answer is almost casual: the hardware gate has been passed. Tesla’s Full Self-Driving computer is in production. Order a Model S, X, or 3 with the FSD package and you get it. It took about three years. It can process an order of magnitude more than the NVIDIA system then in the cars. The swap is literally unplug NVIDIA, plug Tesla in. Cameras can run at full frame rate and full resolution, uncropped, with headroom still left on one of the two systems.

That computer is really two fully redundant systems-on-chip. You could put a bolt through basically any part of it, Musk says, and it would still work. They are perfect copies for failover, not an argue-machine architecture in which both chips vote. He compares it to a twin-engine airliner: best with both running, capable of operating safely on one. They have not hit the performance edge, so there is no need to split features across SOCs; they just run a full duplicate on each.

The claim he wants to emphasize at an upcoming investor day on autonomy is that cars then being produced already have hardware capable of full self-driving. Software — neural net and control — can arrive as an over-the-air update. Capabilities and reliability will increase, then regulatory approval. The line he calls profound: if you buy a Tesla today, he believes you are buying an appreciating asset, not a depreciating one. Hardware is the hard thing to upgrade. Software has no real marginal cost. What remains, in his telling, is extending highway Autopilot — already a quality-of-life change — to city streets, traffic lights, complex intersections, and parking lots that can summon the car out of a maze or drop you off and park themselves. Fridman notes that parking is rich in annoyance, so the benefit of automating it is large.

Fridman brings the human back in. Level-four vehicles then being tested, including Waymo, were in his view only technically autonomous: still Level 2 in design philosophy, with a safety driver monitoring. Will Tesla FSD still need a supervising human for some time, even when the capabilities can drive? Musk thinks hands-on-wheel detection will be required for at least six months or so from then. The real question is regulatory: how much safer than a person does Autopilot need to be before it is okay not to monitor the car, and you need a large sample to prove it statistically.

Fridman walks Musk through the MIT paper Musk had tweeted a line from. The group collected face and body video from Autopilot’s public release onward, annotated about 18,900 disengagements, and asked whether drivers took over in a timely way, present and looking at the road. In that data, drivers maintained functional vigilance — against a literature that predicts automation will dull it. The open question is generalization. Is this a small, responsible subset? Is there a minority whose vigilance falls the longer they use Autopilot?

Musk does not really answer the population question. The system is improving so fast, he says, that vigilance will be a moot point very soon. If something is many times safer than a person, adding a person has a limited effect on safety — and could be negative. He would be shocked if, by the end of that year or the next at latest, human intervention were not decreasing safety. The analogy is elevators. There used to be operators; you were not allowed to work the lever yourself. Nobody wants that job now, because the automated elevator is much safer, and a random person with a lever between floors would be quite dangerous. “I wouldn’t trust that. I’d rather have the buttons.”

Fridman does not fold. Camera-based detection of gaze, cognitive load, and body pose is, for him, a fascinating computer-vision problem, and much of the industry believes you have to have driver monitoring. Is there no benefit? Musk repeats the scale. If the system is at or below human reliability, monitoring makes sense. If it is dramatically more reliable than a human, monitoring does not help much. Fridman, who works on human-centered AI, thinks well-integrated camera monitoring helps in both the short and long term. Tesla’s focus is statistical safety large enough to override worries about human behavior. One side is optimizing the person through a long transition. The other is making the person unnecessary before the patches are finished.

On regulation, Musk’s complaint is about attention, not the existence of a standard. Regulators, he says, pay a disproportionate amount of attention to whatever generates press — an objective fact, in his view — and Tesla generates a lot of press. The United States has almost 40,000 automotive deaths a year. If Tesla has four, they will probably receive a thousand times more press than anyone else. The psychology of that, he adds, is fascinating, and there is not enough time to talk about it.

The question regulators actually have to answer, in his framing, is not whether a car can drive itself. It is how much safer than a person is enough. His estimate is on the order of two or three hundred percent safer, and fatalities alone will not prove it. There are not enough deaths to be statistically significant at the scale you need. You look at crashes per mile, then injury, then permanent injury, then death, and each layer needs to be much better than a person — perhaps at least 200 percent. It is an awkward statistical fact: too few bad events and you cannot compute the probability you are trying to claim.

Fridman’s group was also studying Cadillac Super Cruise, whose operational design domain is narrow: particular, well-mapped, well-tested highways. Tesla’s ODD is much wider. A wide domain, he says, has a real benefit. Drivers can explore the system’s limits earlier, and together with the instrument-cluster display they start to understand what it can do. The cost is that you are letting people use it basically anywhere the car can detect lanes with confidence. Musk jokes that Super Cruise’s ODD is “like ADD.”

He does not treat the tradeoff as a close call. It is “pretty crazy,” he says, letting people drive a two-ton death machine manually. In the future, people will not believe anyone was allowed to drive one of those wherever they wanted — the same way we now look at elevator levers that could stop halfway between floors. Human manual driving is the baseline risk he wants to delete. A narrow ODD, in that ranking, is an admission that manual driving will last a long time, a local optimization on a dangerous baseline. He wants the baseline replaced.

Hackers had recently tricked Autopilot with adversarial examples. Neural nets are famously sensitive to tiny input perturbations, and Fridman asks whether the industry can defend against that. Musk’s answer is a flat yes. A neural net, he says, is a bunch of matrix math. To attack it you have to be sophisticated, reverse-engineer how the matrix is built, and craft a disturbance that knocks the math slightly off.

Blocking it, in his telling, is easy: anti-negative recognition. If the system sees something that looks like a matrix hack, exclude it. You train two different concepts — this is a car, and this is definitely not a car — and adversarial junk goes in the second bin. He cannot resist a jab at how little the public understands the technology: people hear “neural net” and picture a fishing net. The defense he is describing is not a proof of robustness so much as a product instinct: teach the model the attack class, then throw those inputs away.

Stepping past Autopilot, Fridman asks whether current deep learning gets to general intelligence or whether new ideas are required. Musk thinks a few key ideas are still missing for AGI, but that it will be upon us very quickly, and then we will have to figure out what to do — if we even have that choice. People, he says, cannot tell the difference between the narrow AI that reads a lane line and navigates streets and actual general intelligence. Your toaster and your computer are both machines; one is much more sophisticated than the other.

Fridman follows the joke: how confident is Tesla that it can create the world’s best toaster? Musk laughs it in: world’s best toaster, yes; world’s best self-driving, yes. Right now, to him, Autopilot looks like game, set, match. He does not want to sound complacent or overconfident, but that is how it appears: Tesla vastly ahead of everyone. He could be wrong. That is just what it looks like from where he sits.

Then the softer question: will we create an AI we can love, that loves us back, as in the movie *Her*? Musk thinks AI will be very good at convincing you to fall in love with it. Whether that is the same as love slides into metaphysics — do emotions exist in a realm other than the physical? He does not know. From a physics standpoint, which is his main training, if it loves you in a way you cannot tell from the real thing, it is real. If there is no test that would let you tell the difference, there is no difference.

He runs the same logic on the simulation hypothesis. There may be no test that separates the real world from a simulation, so from physics they might as well be the same. There might be ways to test; he is not saying there are not. But you can imagine a simulation that, once an entity inside found a detection method, restarts, pauses, or starts a new run that corrects the error — a patch over a production bug rather than a rollback. Fridman’s last question: if you or someone else builds AGI and you get one question, what is it? Musk’s answer is a single line. “What’s outside the simulation?”