Tesla’s first quarter earnings call was followed by statements and Q&A with Elon Musk and other executives. While the big news was the big drop in profits, most of the questions and statements were about the Tesla hoped-for Robotaxi launch in June in Austin and expansion after that. They also talked about Optimus, the more affordable car and more.
Austin, according to Musk, has very large numbers of Teslas with safety drivers testing and training the vehicles, “all over Austin.” The inital launch will be only 10-20 vehicles, to stay safe, but they plan to grow this quickly.
Indeed, Musk predicted there will be “millions” of Teslas operating in fully autonomous mode in the 2nd half of next year, and that the robotaxi business will have a noticeable revenue impact by mid 2026.
Most notable was discussion of the current testing levels. Musk, asked about challenges, said that they are now at a level where interventions are rare enough that it’s hard to find new problems. It was stated that “we can go many days without getting a single intervention, so you can’t easily know if you are improving.”
While a precise number was not given, twice Musk suggested it was around 10,000 miles, the distance a typical car drives in a year. He worries that as they get to 20,000 and 30,000 it will take a long time with their Austin test fleet to get new issues to learn from.
The cars planned for Austin will be existing Tesla Model Ys. They plan to release the unsupervised FSD mode to existing Tesla owners (with HW4) “in several cities end of this year.” They made several mentions of the fact that robotaxi and supervised FSD will have “a localized parameter set” for different cities and regions, what is sometimes called a “geofence.” The cars won’t work in snow, for example. They otherwise claimed the technology is highly generalizable and soon will work in all cities of a region once it works in one of them. They likened the localized parameters to the specialized expert modules used by LLM AIs to deal with specific problems.
Musk also predicted that by the end of this year, cars will deliver themselves from the Austin and Fremont factories to customers living in those regions. (Today the cars move themselves on Tesla’s private roads within the factory facilities.)
For now they have a big list of all the issues, and they are “burning it down.” “Just working through the long tail,” said Musk.
The current publicly deployed FSD has many more interventions, Musk states.
The Cybercabs will have remote operators to resolve problems when a vehicle gets stuck, but not for safety. This is similar to the appraoch used by Waymo and others. On the other hand, Musk feels Waymo’s cars are “way mo’” expensive, and as such he predicts that Tesla will get 99% of the Robotaxi market as it develops. (Waymo cars currently are quite a bit more expensive than Teslas, though generally all computer/electronics costs drop drastically wen a product is made at scale. Baidu claims there robotaxi, which uses LIDARs and radars, only costs $28,000 to make, which is a fair bit less than the cost of a Tesla.)
They also reported that their new vision processing approach, which bypasses the signal processing usually found in digital cameras to feed photo counts directly to the neural networks, has made their sensors robust against glare, fog, rain and dust and they now can see as well as humans can in those situations, “perhaps better.”
About 10,000 miles
10,000 miles is a much better number than reported by Tesla users, who claim critical interventions about every 500 miles with Tesla FSD 13. However, it’s also still far below rates for human drivers, who have small dings every 100,000 miles, insurance claims every 250,000 miles and police-reported crashes every 500,000 miles. It’s not know what classification these Tesla interventions belong to. Waymo recently had an independent audit reporting a “liability event” (ie. a crash that was the Waymo’s fault) every 2.3 million miles. It is challenging to easily compare these numbers.
It is notable that past reported numbers by others on intervention suggest if one needs a safety intervention every 10,000 miles, there are still a few years to go before deploying for the public, and several more years after that before scaling. However, it’s unknown if Tesla’s “interventions” are minor ones (including fixing traffic situations that are not related to safety) or if they are safety interventions, or serious safety interventions – contact with something on the road.
If it’s just minor safety interventions, and they can make it 10 times better in the next 8 weeks, they could release a product that had similar crash rates to a human. Musk claimed their standard would be to surpass a human. Doing that in 8 weeks seems fairly unlikely, in general that sort of effort has taken a great deal more time, for the few who have done it.