This interview covers George Hotz and his company Comma.ai's approach to self-driving cars. He argues that the current method used by Tesla and Waymo—breaking driving into many small tasks—is outdated and will be replaced by 'end-to-end learning' like AlphaZero. He is bearish on Waymo, calling its product logic flawed, and criticizes Nvidia for overpricing its chips. Key holdings: Comma.ai (sells a $1,200 device, ~2,000 daily users), Tesla (may win long-term but needs fixes), and Waymo (explicitly bearish, predicts failure).
George Hotz (Geohot), founder of Comma.ai, hacker, and former iPhone jailbreak developer. The core theme of this episode is Comma.ai's end-to-end autonomous driving technical approach, business model, and competitive landscape. The most significant takeaway: George Hotz believes that the ultimate solution for autonomous driving is end-to-end learning similar to AlphaZero / MuZero, rather than the "task engineering" (multi-task perception + rule-based stitching) approach currently adopted by Tesla and Waymo; the latter is essentially a higher-order form of feature engineering and will ultimately be replaced by end-to-end methods.
George Hotz argues that Tesla's multi-task perception (segmentation, detection, prediction) approach is a form of "task engineering," a higher-level feature engineering that will eventually be replaced by end-to-end learning.
He cites the AlphaZero case, noting that the modern chess engine (Stockfish) has a much larger codebase than AlphaZero, but the latter is more elegant and concise through learning. He calls MuZero "the foundational paper of the entire deep learning era" and believes it is the "solution" for autonomous driving—namely, performing "murder" and "rollback" (referring to extensive trial and error in simulation) within a learned dynamic model (simulator).
Argument Process:
Hotz also acknowledges uncertainty: "I cannot make promises on timelines. But AI history shows that feature engineering approaches are always replaced by end-to-end approaches."
Comma.ai的使命是“解决自动驾驶汽车，同时交付可发货的中间件”。 Hotz强调，“交付可发货的中间件”不仅是融资手段，更是让公司保持诚实的约束，避免像Waymo那样陷入“设定任意里程碑，仍然离目标无限远”的陷阱。
论证过程：
Hotz levels sharp criticism at major industry players while maintaining a nuanced stance toward Tesla.
| Position | Guest Attitude | Key Data |
|---|---|---|
| Comma.ai | Bullish (self-described) | ~2,000 daily active users; 70% D1 retention; Comma.2 priced at $1,200; projected revenue of $10 million next year |
| Tesla / Autopilot | Bullish (long-term), but needs correction | Hotz believes it will ultimately win, but needs to add driver monitoring and shift to an end-to-end approach |
| Waymo | Clearly Bearish | Hotz thinks its product logic is flawed, predicts it will fail, and says "it has already spent $10 billion, but the model is unsustainable" |
| NVIDIA | Critically Bearish | Hotz criticizes its pricing strategy (A100 vs. RTX 3080) and believes it will lose the ecosystem |
| Optimism (Ethereum L2) | Neutral (technical partner) | No position action, but Hotz modified its compiler (300 lines of code) to resolve its technical difficulties |
1. Endpoint of the technical route (George Hotz): “MuZero wrote the solution for autonomous driving on paper.” — End-to-end learning plus a learned dynamic model (simulator) is the only path; segmented task engineering (feature engineering) will eventually be replaced.
2. Business model anchor (George Hotz): “Real capitalism is based on consent.” — Users buying products with real money is the most honest proof of a company's value, and also a corrective mechanism to avoid “infinitely approaching but never reaching the goal.”
3. Fatal critique of Waymo (George Hotz): “Their product makes no sense. Users can already choose to trade time for money, but defaulting to a slower and more expensive option is illogical.” — The core argument is that the business model of L4 Robotaxi is less competitive than existing Uber/Lyft, and the market will fall into a race to the bottom like the scooter market.
4. Warning to NVIDIA (George Hotz): “NVIDIA is killing the goose that lays the golden eggs. Sell chips, not dreams.” — Its high pricing (A100 vs RTX 3080) will drive customers (like Tesla) to develop their own chips, ultimately losing the market ecosystem.
5. Mild criticism of Tesla (George Hotz): “Elon will install driver monitoring before he reaches L5.” — He believes Tesla's vision solution (no infrared) is a major flaw, but it is a correctable “software problem” that does not affect its long-term success.
6. Product design philosophy: Adaptive driver monitoring (George Hotz): “Driver monitoring strategy must be scene-adaptive. The loosest on highways, the strictest in urban areas.” — The goal is to prevent fatigue and false alarms, and to prevent users from “learning to ignore it.” This is achieved through machine learning, not hand-written rules.
7. Key falsification condition (George Hotz): “When our unplanned disengagement rate improves from every 100 miles to every 100,000 miles, we will be approaching L5.” — It clearly sets an order-of-magnitude target, acknowledging the huge gap between current levels and human driving, but believes in exponential improvement.
8. Unique insight on AI (George Hotz): “Compression is intelligence. The Hutter Prize framework made me understand modern AI.” — He proposes that reducing the problem of intelligence to a lossless compression problem is the core framework for understanding artificial intelligence, driving his exploration in this field.