Is Elon Musk’s strategy, betting the future of Tesla on the success of their efforts to make a working self-driving system with a camera-only machine learning car, similar to that of the Spanish conquistador Hernán Cortés, who legend says burned his ships so that his crew could not return home, leaving them no choice but to conquer Mexico?
Tesla’s plan has been very different from the strategies of companies like Alphabet/Waymo, GM/Cruise, Amazon/Zoox, Baidu Apollo and others. These teams have taken a make-it-work first approach, using whatever hardware necessary, expecting the cost of the hardware to plummet with time and scale as is generally the case for digital electronics. Tesla chose a low-cost hardware suite (cameras and a custom processor) already present in their cars with the hope that enough software effort, particularly machine learning, would make the system work safely enough with that hardware.
So far, the more popular approach has been a success. Waymo put their first no-driver car on the road in 2017, was running general vehicles empty in 2019 and today is giving about 350,000 autonomous rides per week to paying customers. Baidu is doing about 250,000 rides/week. Cruise was operating but was shut down because of a cover-up around a key safety issue. Tesla, on the other hand, does not have a working system that can drive unsupervised many years after the other companies pulled it off.
This week, however, they claimed they had made it work, and deployed a small number of robotaxis with no supervisor in the car, stating that they will have large fleets deployed soon in many cities. Their demonstrations have included chase cars, and much evidence points to Tesla having built a remote supervision and even remote driving system, which would allow them to do these demos without having a fully unsupervised system working.
Making things work with cheaper hardware is harder, as the results show. It offers some potential to be better at scaling, but mostly it offers the ability to convert Tesla’s already sold cars, which number in the millions, into self-driving vehicles with a software update.
To give that harder path a chance, Musk gave his team no other choice. He even removed the radars, another tool all other teams use, though today they use more advanced “imaging” radars which do much more than the radars Tesla removed. (Tesla experimented with an imaging radar but has not gone forward.) In this way, Musk increases the chances his team, given no other path, can pull it off.
Cortés didn’t actually burn his ships, he just sunk them, but the metaphor remains with us. It worked. His small band of Spaniards pulled off the impossible, and conquered the Mexica empire with alliances, superior technology and diseases.
It worked, but this strategy doesn’t always work. Indeed, while Tesla declared in 2016 that their cars had all the hardware needed for self-driving, they released a new hardware suite called HW3 in 2019, and another called HW4 in 2023. HW5 has been pushed to 2027. Tesla has effectively admitted they won’t make FSD work on HW3 and no longer releases new versions for it. (I own a Tesla which was upgraded to HW3.) Musk has stated that some advanced reasoning features they have planned for FSD will need HW5, but he has also stated that owners of older cars who earlier purchased the FSD beta package (now known as FSD Supervised) will somehow be upgraded, though at present the HW3 cars don’t have the wattage necessary for the upgrade in their computer cabinet, so it’s not an easy thing to do.
Tesla’s strategy is to use end-to-end machine learning to solve the problem. The tools used today did not exist in 2016, so when crafted, the strategy was more vague. LIDARs were too expensive to add to production cars then, though as expected their prices continue to drop to be much more affordable, though LIDAR+radar+cameras will always be more then cameras alone.
Many (including myself) have been skeptical about whether it can be done with just cameras, though a large number agree it’s probably possible some year, but the year is unknown. Musk now declares he will deploy it very soon in Austin and in several cities in the first half of 2026. But strangely, they have announced some demonstrations that give the appearance they have actually made it work unsupervised, while other clues say that they have not. The first demos had obvious chase cars supervising the operations. In the Jan 18 earnings call, Musk declared they were now going rides without chase cars (admitted that they had them, though they had previously declared the chase car rides as unsupervised) but there continues to be a strong case (including photos of Tesla Robotaxi HQ with consoles with steering wheels) that they have built a remote supervision system with even remote driving ability.
What would be the point of them doing obvious fakery? Once again, it may be a burning of the ships. By appearing to have unsupervised robotaxis, Musk forces his team to go all out to make them real. Musk believes they can do that, and if they do, he gets a big win. He is offering them little other choice. But you can only fake for so long. Using remote supervisors is expensive. It’s perfectly fine for small fleets (and all companies do it, I am confident Zoox is doing it today, though official confirmation of such things is rare.) You can test and demo, but you don’t make a fleet of thousands of cars easily that way.
Things like chase cars and remote supervision consoles don’t have to be fakery. You would have those just to be safe on first deployment anyway. But when they are there, people can’t tell the difference between caution and games. By far the biggest leap in developing self-driving is the move to be truly unsupervised. Before that point, it’s hard to judge if the player is really in the game, or just covering up flaws with the supervision.
Unfortunately, for Tesla, as powerful as the machine learning systems they are using have been in many areas, nobody has yet built a complex system with with pure ML “bet your life” reliability. Should Tesla do it they would be the first, and Musk might feel the best way to bet on making that happen is to leave no other choice for his team.
