Tesla’s Full Self-Driving (FSD) software has been in the spotlight recently as the company prepares for its upcoming robotaxi launch and expansion outside of North America. CEO Elon Musk commented on the limited usefulness of miles driven with the system engaged for training the neural net-based model. He emphasized the need for more training miles to improve the effectiveness of the FSD system. In response to an investor discussion about FSD, Musk stated that only one out of every 10,000 miles driven with the system engaged is actually useful for training the neural network.
This may come as a surprise to some. Musk mentioned that the limiting factor for FSD progress was previously AI training compute, but it is now validation as interventions are becoming increasingly rare. Musk highlighted that interventions are becoming rarer as the software improves from camera- and video-based training. This underscores the need for increased validation, especially with handling unique road scenarios.
He pointed out the importance of accurate analysis and the need for as many training miles as possible for effective FSD development.