Elon Musk has once again emphasised what it will take to have actual safe, unsupervised Full Self-Driving (FSD). To Musk, the solution is not flashy demos or limited pilots but a huge scale. The Tesla head now believes that it will take around 10 billion miles of real-world driving information to actually make FSD unsupervised.
Musk explains this massive requirement with what he calls reality’s “super long tail of complexity”, the infinite number of rare, unpredictable, and edge-case situations that autonomous systems must master perfectly before they can be put on the roads.
The comments were made by Musk in reaction to a X-post by Paul Beisel, a former Apple employee and Rivian executive. Beisel published an opinion piece pointing out the very high contrast between supervised technology demonstrations and the scale deployment of real-world autonomous driving systems. He said that the autonomy could not be easily caught up by the competitors with the help of simulation or restricted road testing.
The idea that anyone can be expected to keep pace with this problem by means of simulation and the restricted exposure to the real road process sounds terribly naive to me, Beisel wrote. “This is not a demo problem. It is a problem of scale, data, and iteration.
Musk replied that 10 billion miles of training data would be needed to make unsupervised self-driving safe.
A Shift From Earlier Estimates
Interestingly, Musk’s latest figure represents an increase from his earlier projections. In Master Plan Part Deux, Tesla proposed that there could be some 6 billion miles of driving data that would persuade regulators across the world to accept autonomous driving systems. The revised estimate represents an increasing appreciation of the sheer fact that it is hard to get the last few percentage points of autonomy solved.
Roughly 10 billion miles of training data is needed to achieve safe unsupervised self-driving. Reality has a super long tail of complexity.
— Elon Musk (@elonmusk) January 8, 2026
As reliable numbers of 90% or even 99% are relatively easy, the other portion, which addresses the cases of the edge cases, is exponentially more difficult. These are such odd road patterns, unpredictability of human behaviour, extreme weather, and unforeseeable traffic situations that are rare but have high risks.
Tesla’s Current FSD Data Advantage
At the end of 2025, Tesla seems to be on the path to this ambitious goal set by Musk. The community trackers put FSD at approximately 7 billion miles of real-life driving data. It is important to note that the total of over 2.5 billion miles is based on complex inner-city environments, in which autonomous systems are most challenged. This large amount of data is systematically fed back into Tesla’s neural networks so that it can quickly iterate and enhance.
Musk has also warned that the final journey toward full autonomy will be the hardest. He recently repeated this while commenting on Nvidia’s autonomous driving efforts, stating that it is “easy to get to 99% and then super hard to solve the long tail of the distribution.”
Moreover, Tesla’s Vice President of AI Software, Ashok Elluswamy, confirmed Musk’s claims by posting on X that “the long tail is sooo long, that most people can’t grasp it.” This long tail represents millions of rare scenarios that must be encountered, learned from, and solved, something only large-scale real-world deployment can provide.
