Comma.ai Drive and Tour with George Hotz and Lex Fridman

Product Demo & Performance

• George Hotz demonstrated the "Kama 2" (identified as Comma 2 in context) system on a drive in San Diego, highlighting its end-to-end neural network capabilities.
• The system successfully handled a curvy, banked road section on Highway 163 without disengagement, a maneuver noted as difficult for competitors like GM Super Cruise.
• Testing revealed the system can perform lane changes on roads lacking lane markings, relying on visual context rather than explicit line detection.
• Longitudinal control (braking/acceleration) currently leverages Hyundai's existing Advanced Cruise Control (ACC) system, with plans to transition to Comma's own end-to-end longitudinal policy once finalized.
• Lateral control (steering) required five years of engineering to achieve smoothness and handle subtle maneuvers, such as navigating split lanes.
• Safety validation includes "injection tests" where a safety driver uses a joystick to simulate erratic inputs while the vehicle maintains stability; the system's torque limit is set lower than Tesla's to prevent aggressive over-correction.
• The system communicates disengagement clearly via distinct auditory and visual cues to alert the driver of handoff points.

Operational Scale & Data

• The Comma 2 platform has accumulated nearly 30 million miles on OpenPilot, marking a significant product breakout.
• Active device count ("botnet") breakdown:
  • Daily active devices: ~1,500.
  • Weekly active devices: ~2,500.
  • Monthly active devices: ~3,000.
• Engagement metrics:
  • Percent of miles engaged: ~50%.
  • Percent of time engaged: ~30%.
  • Disengagement causes are evenly split between cancel, gas override, and brake override.
• Geographic distribution covers the U.S., Korea, Canada, Taiwan, and Japan, with the U.S. accounting for 2,200 devices in the last 30 days.
• The majority of segments and data now originate from Comma 2 hardware units.

Manufacturing & Infrastructure

• Comma AI operates as a multinational corporation with offices in both the U.S. and Europe.
• The San Diego facility includes a hardware production line where units undergo 24-hour stress testing for heat, light sensors, and cosmetic checks.
• Assembly involves retrofitting original smartphones with custom circuit boards and heat sinks before casing.
• The company recently expanded 3D printing capabilities to approximately 20 printers sourced globally from Poland, Germany, Peru, and France to produce device cases.
• Compute infrastructure consists of:
  • Three racks of CPU machines totaling 2 petabytes of spinning disk storage.
  • GPU farms containing NVIDIA 3080s (for desktop use) and V100s (for server use), referred to as a "21st-century farm."

Strategic Decisions & Roadmap

• Software Release Strategy: OpenPilot 1.0 remains in development, while OpenPilot 1.1 is slated for over-the-air (OTA) deployment to Comma 2 devices.
• Technical Migration: The team is evaluating a move from PyTorch to "Pi Torch" (likely a transcription error for a specific framework or internal codename) for model ingestion and processing.
• Product Focus: The primary value proposition and current usability of OpenPilot remain concentrated on highways and two-lane roads rather than dense city streets, which require mature end-to-end longitudinal control.
• Safety Philosophy: Future updates aim to increase the wheel touch timeout and improve the holistic interaction between lateral and longitudinal policies to reduce "freak out" moments.
• Hiring & Talent: The company recently hired "Virtually Chris," a prominent YouTube creator known for OpenPilot content, to its machine learning team.

Market Positioning

• The company views its success as symbiotic with Tesla's, stating, "If iOS succeeds, Android succeeds."
• George Hotz expressed reluctance to ship the system until it achieves a "holistic" end-to-end capability, including full stop sign and red light handling.