Interview
George Hotz: Comma.ai, OpenPilot, and Autonomous Vehicles | Lex Fridman Podcast #31
- Kama AI projects achieving profitability by mid-next year and plans to transition into a car insurance business model to monetize safety data, anticipating a market bifurcation where premiums vary significantly based on driver safety profiles.
- The company expects two additional years of development following four years of work to reach a consumer-ready OpenPilot version 1.0, which will focus exclusively on improved lane keeping without new feature additions, while version 0.6 is scheduled for release next month.
- OpenPilot development includes plans to disable safety features that currently trigger Automatic Emergency Braking (AEB) cut-offs prior to the 1.0 release and to utilize a temperature parameter set to zero on the live system to ensure safety while allowing an "explorer" mode for data gathering.
- The speaker forecasts that Level 4 autonomy will be feasible in defined areas like Chandler, Arizona, by 2020 or 2021, whereas Level 5 deployment in complex environments such as New York City is considered unlikely in the near term and may not exist as a practical target.
- Autonomous vehicle progress is categorized into three challenges: the "static" problem involving mapping and localization, which is deemed solvable but difficult without LiDAR; the "dynamic" problem requiring real-time modeling of other agents; and the "counterfactual" problem involving human interactions, which is identified as the hardest challenge requiring reinforcement learning.
- Kama AI expects to leverage 10 million miles of real-world data to verify system safety beyond manufacturer specifications and predicts that other companies promising autonomous fleets by 2020, 2021, or 2022 will fail to deliver.
- Waymo is expected to maintain a three to five-year lead over competitors with its current stack, potentially resulting in it being the sole major competitor if Level 5 autonomy takes decades to solve, while GM's acquisition of Cruise is viewed as a short-term insurance strategy against this potential monopoly.
- The speaker assesses that Cruise is three years behind Waymo, the Toyota Research Institute is significantly behind Kama AI, and Tesla utilizes a similar end-to-end learning approach, whereas Level 4 capabilities will require new hardware rather than software updates alone.
- Future business and technological outlooks include the prediction that the singularity will occur around 2038 when the majority of computing power is silicon-based, and the speaker anticipates merging with a machine in a long-term relationship while predicting the existence of AI relationships or "VR brothels."