Jensen Huang and Elon Musk on self-driving cars and security

In 2015, the NVIDIA and Tesla CEOs discuss difficult driving scenarios, software updates and the risk of a fleet-wide hack.

EMElon MuskInvité · Tesla
Chaîne
Economic Archive
Publié
16 nov. 2023
Vidéo
19:42
Lecture
4 min
Original
anglais
Ouvrir sur YouTube
En faisant défiler, la vidéo se réduit en bas à droite et continue. Les sous-titres suivent votre langue de lecture.
Langue de lectureDisponible en 2 langues
Cet article n’est pas encore disponible en français ; il s’affiche dans l’original en anglais.
RÉSUMÉ

En 30 secondes

In a 2015 conversation with NVIDIA CEO Jensen Huang, Elon Musk predicts that autonomous driving will eventually become routine, but identifies suburban streets—with children, bicycles and unexpected roadworks—as harder than highways or low-speed maneuvers. He says regulators would need substantial safety evidence and that connected vehicles need protection against attacks affecting many cars. These were forecasts and design claims made at the time. Tesla’s current [Full Self-Driving (Supervised) guidance](https://www.tesla.com/support/fsd) requires active driver supervision and says the feature does not make the vehicle autonomous.

  1. 1Musk predicted autonomous cars would eventually feel ordinaryHe compared them with automatic elevators, while acknowledging that widespread replacement of existing vehicles would take years.
  2. 2Software improvements depend on a hardware foundationMusk describes using cameras, radar, ultrasonic sensors and onboard computing, then adding capabilities through connected software updates.
  3. 3Unpredictable streets are harder than controlled highway drivingHe points to road closures, open manholes, children and bicycles as situations that complicate perception and decisions.
  4. 4Regulators would need evidence, not just a claimMusk suggests comparing automated decisions in shadow mode with actual outcomes across a large population before seeking approval.
  5. 5A fleet-wide hack was a central security concernMusk says Tesla focused on resisting attacks against many cars and protecting safety-critical components with additional layers.
  6. 6Musk distinguished narrow driving AI from broader AI risksHe says his concern about more general intelligence is different from the narrow use of AI in a car.

À retenir

  1. Musk identified unpredictable suburban situations as a difficult problem for autonomous driving in 2015.
  2. He argued that even if all new vehicles became autonomous immediately, replacing the existing fleet would take many years.
  3. Musk proposed collecting large amounts of evidence, including simulated decisions, before seeking regulatory approval.
  4. He said Tesla prioritized making a fleet-wide remote attack difficult and adding protection around safety-critical controls.
  5. A software update can add features, but its capabilities remain constrained by sensors, computing power and safety validation.
TEXTE INTÉGRAL

Lire le texte remanié, chapitre par chapitre

Transcription seule
01

The promise and pace of autonomous cars

Huang asks about AI risk and Musk’s expectations for self-driving. Musk predicts broad adoption but notes the time needed to replace an existing vehicle fleet.

At NVIDIA’s 2015 GPU Technology Conference, CEO Jensen Huang welcomes Tesla CEO Elon Musk and asks how Musk reconciles concern about powerful AI with AI used in self-driving cars.

Musk distinguishes driving as a narrow AI application from broader intelligence risks. He predicts autonomous driving could eventually be safer than human driving and become ordinary, comparing it to elevators that no longer need operators. These are his predictions in 2015, not demonstrated results in the conversation.

Huang wonders whether vehicles could someday be lighter if collisions became rare. Musk says that would depend on being able to count on avoiding crashes, and adds that older vehicles would remain on the road for a long time.

Musk estimates that replacing a global fleet of billions of cars and trucks would take many years even if every new vehicle immediately became autonomous or electric. He uses the comparison to distinguish a new technology’s arrival from the time needed for a complete fleet transition.

02

Sensors, software and difficult streets

Musk describes the sensor and computing platform in 2015 vehicles. He identifies the variety of mid-speed suburban situations as a difficult technical problem.

Huang asks about the path from existing driver-assistance features to self-driving. Musk says Tesla starts with sensors and computing hardware, then adds functions through connected software updates. He describes the 2015 vehicle’s camera, radar and ultrasonic sensor suite.

Musk says the existing hardware could support improvements such as highway steering, but would not provide the reliability he wants in a complex urban setting. He says broader sensing and more computing power would be needed to handle more difficult situations.

Huang asks the engineers in the audience what problems deserve attention. Musk says very low-speed maneuvers and relatively structured freeway driving are different from intermediate-speed streets, where unexpected events are more varied.

He lists road closures, open manholes, children and bicycles as examples. The difficult task, he says, is recognizing what is happening and making the right decision in a suburban environment. He predicts a solution in a few years; that is his assessment at the time.

03

Evidence, regulation and connected software

Musk discusses safety evidence and regulatory review, while Huang asks about over-the-air software updates. The two describe cars as evolving computing platforms.

Huang asks how regulations might evolve when cars can drive themselves. Musk says regulators would need a large amount of evidence that a system is safer, and might take time to assess it after a technical capability exists.

Musk proposes using shadow mode: record what the automated system would have done, compare it with actual outcomes and examine errors over a large population. He says regulators could then evaluate the statistical argument. He also acknowledges the need for caution when public safety is involved.

Huang describes the car as a software platform whose capabilities change through updates. Musk says Tesla tries to establish the sensor and computing foundation first, then make the car more capable through software. They discuss an upcoming 2015 software release and the growing demand for computing performance.

04

Protecting connected cars from attacks

Musk describes concern about a remote attack on multiple vehicles and additional protection for safety-critical systems. The conversation ends with AI and future driving.

Huang asks how a connected car can resist hacking. Musk says protecting against a remote attack on many vehicles is a particular concern. He says Tesla uses outside parties to test security and adds protection to safety-critical components such as steering and drive controls.

Musk says the intention is to make an attack on an entertainment function different from gaining control of vehicle movement. He describes this as a layered security design, not a guarantee that attacks are impossible.

Huang says Tesla is both an electric-car maker and a computing-platform builder. Musk says he wants Tesla to make autonomous cars available to customers and invites engineers interested in the problem to join the company.

Musk returns to his distinction between narrow driving AI and the broader AI systems he believes deserve caution. Huang says progress in AI can improve safety and convenience through intermediate features, before fully autonomous vehicles exist.

The two close with a lighter exchange about how they like to drive their own cars. Tesla’s current [Full Self-Driving (Supervised) guidance](https://www.tesla.com/support/fsd) requires active driver supervision and says the feature does not make the vehicle autonomous.

Sur la même chaîne