16 Questions About Self Driving Cars

• Core Thesis: Andreessen Horizons operates on the belief that "everything that moves will eventually go autonomous" as costs collapse and utility increases, with cars representing the most significant near-term market opportunity due to their massive revenue scale despite lower unit volume compared to smartphones.
• Market Comparison: The automotive market is described as having "much higher revenue" than Apple (the most valuable tech company) but involves fewer units sold.
• Timeline Consensus: While predictions vary widely, all public estimates indicate widespread autonomy will occur within the current decade or two:
  • Waymo: Deployed in Singapore in 2018; targeting 10 top cities by 2020.
  • Delphi/Mobileye: Systems for manufacturers by 2019.
  • Volkswagen/Baidu: Systems already deployed.
  • GM: 2020.
  • Ford/BMW: Level 5 availability or iNext shipment by 2021.
  • Tesla: 2023.
  • Uber: Full fleet autonomy by 2030.
  • IEEE: 40% of road cars autonomous by 2040.

Technological Decisions and Architectures

• Autonomy Levels (SAE Standard):
  • Level 0: Fully human-driven.
  • Level 1-2: Driver assistance (e.g., anti-lock brakes, cruise control, lane keeping).
  • Level 5: No steering wheel; the vehicle is solely self-driving mode.
  • Strategic Debate: Incumbents (e.g., GM) favor an incremental approach adding features one by one, whereas disruptors (e.g., Google) argue for "jumping" straight to Level 5 to avoid complex human-machine interface (UX) failures and mixed-mode accidents.
• Sensor Configuration (LiDAR vs. Cameras):
  • LiDAR currently costs ~$75,000 but is projected to drop to $250 with solid-state (no moving parts) technology, offering superior 3D accuracy over stereo cameras.
  • Counter-argument: Some entities (e.g., Tesla) argue that stereo cameras are sufficient to compute 3D space and that LiDAR introduces unnecessary cost and data fusion complexity.
  • Industry Trend: The consensus among most industry observers leans toward the inclusion of LiDAR once costs stabilize.
• Mapping Requirements:
  • High-Definition (HD) Maps: Standard navigation (Google/Apple Maps) lacks granular details like curb positions or specific glare times required for autonomous navigation.
  • Dependency Risk: Reliance on HD maps could limit vehicle range based on map availability, similar to electric vehicle range anxiety.
  • On-Board Compute Trade-off: Eliminating map dependency requires massive on-board supercomputing power (50W–500W), which negatively impacts EV range or fuel efficiency.
• Software Architecture:
  • Deep Learning: NVIDIA advocates for end-to-end deep learning connecting sensor inputs directly to control outputs (steering, braking, throttle).
  • Control Systems: Alternative approach utilizes traditional robotic control algorithms (search, directed pathing) that guarantee vehicle control and safety without neural net training.
  • Predicted Outcome: A hybrid blend of deep learning and traditional control systems is expected.
• Training Data Sources:
  • Simulation: Significant debate exists on whether neural nets can converge to real-world performance weights using virtual environments (e.g., gaming engines) versus real-road data.
  • Advantage: Simulation allows for infinite iteration on edge cases (rain, snow, fog) without physical risk.
• V2X Communication (Vehicle-to-Everything):
  • Capabilities: Enables Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure (V2I) communication to prevent T-boning accidents and optimize traffic light timing.
  • Deployment Barriers: Requires widespread protocol compatibility, low-latency connectivity (sub-second), and security against spoofing; currently limited to specific fleets (e.g., Mercedes-Benz S-Class).
  • Future Vision: Potential elimination of traffic lights entirely if intersections can algorithmically manage car "packets" with zero accidents, though this remains a long-term theoretical goal.
• Localization of Driving Culture:
  • Challenge: Safe driving conventions vary significantly by city (e.g., Boston vs. Bangalore).
  • Solution: Moving away from pre-programmed city-specific versions toward vehicles that "learn" local conventions (pedestrian flow, tailgating norms) during their first month of operation via machine learning.

Business Model Shifts

• Competitive Landscape:
  • Incumbents: Traditional manufacturers (Ford, GM, Toyota, etc.) are aggressively hiring software talent in Silicon Valley to integrate autonomous features into existing vehicles.
  • Disruptors: Silicon Valley-native companies (e.g., Tesla) and Chinese firms (Baidu, etc.) are viewed as potential winners in inventing the car from scratch.
• Ownership vs. Service:
  • Shift: Consumers may transition from buying cars to renting transportation (MaaS), shifting brand loyalty from manufacturers (e.g., BMW) to fleet operators (e.g., Uber/Lyft).
  • Consequence: Automakers become B2B providers to fleet managers; consumer-facing marketing (e.g., Super Bowl ads) and engineered sensory details (e.g., door slam sounds) would become obsolete.
• Insurance and Liability:
  • Pricing: Premiums will likely shift from driver demographics to vehicle algorithm effectiveness (safety scores of the car itself).
  • Liability: New legal questions arise regarding hacking incidents (e.g., hacked garage door leading to a stolen car), creating shared liability between hardware makers, software providers, and homeowners.
  • Repair Costs: Accident frequency is expected to drop, but repair costs may rise due to the expense of replacing onboard supercomputers.

Social Implications and Future Trends

• Accident Rates:
  • Long-term: Projected to approach zero as algorithms eliminate the 24 out of 25 primary causes of accidents (distraction, speeding, alcohol, etc.).
  • Short-term: Rates may temporarily increase due to the "mixed fleet" problem where human drivers misinterpret the cautious behavior of autonomous vehicles.
• Regulation:
  • Future Prohibition: If algorithms are statistically superior, it is expected that driving will eventually be made illegal for humans, relegating recreational driving to controlled environments like Legoland.
• Commute Dynamics:
  • Induced Demand: Historically, new transport modalities (highways) increase total commute times due to latent demand.
  • Urban Optimization: Autonomous fleets could free up parking and repair spaces, potentially allowing cities to restructure density so people live closer to work, offsetting induced demand.
• Unpredictable Second-Order Effects:
  • The shift from car ownership to autonomy may enable unforeseen social shifts similar to the emergence of Walmart, driven by the radical reorganization of logistics and city infrastructure.