Interview, Fireside Chat, Podcast
Elon Musk: Tesla Autopilot | Lex Fridman Podcast #18
- The new Full Self-Driving computer is projected to process an order of magnitude more data than current NVIDIA systems and reach its performance boundaries within approximately two years, enabling full-frame, full-resolution camera operation without cropping.
- System capabilities are expected to evolve to include running a full duplicate on a single system-on-a-chip, traversing complex intersections via optimal spline determination without intervention, and operating safely even if one of two redundant systems fails.
- Functional expansion is planned to extend autonomy from highways to city streets with traffic light recognition and complex navigation, alongside advanced parking lot navigation where vehicles can exit spaces, locate drivers, and park independently.
- Regulatory pathways require statistical safety improvements of 200% to 300% over human drivers before driver monitoring and steering wheel detection requirements are removed, with hands-on-wheel detection expected to disappear by the end of the current year or the following year at the latest.
- Specific removal of driver confirmation for lane changes, freeway exits, and interchanges is anticipated following the "Navigate on Autopilot" update, while unsupervised driving is expected to be implemented within approximately six months pending regulatory approval.
- Hardware currently in production is deemed capable of full self-driving, with software updates intended to drastically increase reliability and effectively turn vehicles into appreciating assets through exponential improvement rates as neural networks refine.
- Risks regarding adversarial examples targeting neural networks are expected to be mitigated through anti-negative recognition and training on invalid data, while the company asserts it is currently vastly ahead of competitors in the self-driving domain.
- Long-term perspectives include the arrival of artificial general intelligence very quickly, the system eventually replacing human supervision entirely due to vigilance decrement becoming moot, and future generations viewing current manual driving as obsolete.
- Regulatory approval will rely on crash probability and statistical safety metrics rather than fatality counts due to scale, and driver monitoring via cameras is expected to provide little benefit once system reliability significantly surpasses human levels.
- The system is expected to handle a wide operational design domain, and it is predicted that AI will eventually become capable of convincing humans to form emotional attachments to it.