This Issue at a Glance
Elon Musk (CEO of SpaceX, Tesla, Neuralink, and The Boring Company) in a deep discussion on the Lex Fridman podcast about SpaceX's Mars plan, Tesla's full self-driving technology roadmap, artificial intelligence, and robotics. The most significant judgment in the entire episode: Musk believes that the current cost (approximately $1 billion per ton to the Mars surface) must be reduced by at least 1,000 times to the million-dollar level in order to achieve a self-sufficient Mars city, and the key lies in "fully and rapidly reusable orbital rockets" — this is the Holy Grail of rocket technology, which has never been achieved by anyone.
Theme 1: Starship – The Hardest Engineering Problem Is Not Design, But Mass Production
Musk believes that the biggest bottleneck facing Starship is "engine production," not the design itself.
- Technical core: The Raptor engine is the world's first full-flow staged combustion engine, operating at a pressure of 300 bar, exceeding the 267 bar of the Russian RD series, the highest ever. For every 10% increase in pressure, the difficulty rises by about 50%. It requires specially developed alloys (multiple new materials), and the simultaneous flow of multiple propellants and high-temperature gases creates complex feedback loops.
- Production cost logic: Musk proposes a "magic wand number" — break the rocket down into raw materials (aluminum, steel, titanium, Invar, copper, etc.), calculate the raw material cost, and that is the theoretical minimum manufacturing cost. The actual cost is far higher, indicating that the problem lies in "how to arrange the atoms into the final shape," not the raw materials themselves.
- Mass production vs. design: Musk emphasizes that "prototypes are easy, mass production is hard." He criticizes the common engineering excuse that "costs are high because production volume is low." His diagnostic method: assuming an annual production volume of 1 million units, would it still be expensive? If the answer is "yes," then the cost problem is not about scale, but about the design itself.
- Reusability is key: The Falcon 9 currently only recovers the first stage and fairing, not the second stage, with a marginal cost of about $15–20 million per launch. If Starship achieves full reusability, the estimated cost per launch could drop to $1–2 million, delivering over 100 tons to orbit — a cost reduction of roughly 100 times.
Musk's view on this: Physics does not forbid it, but "success belongs to the set of possible outcomes, not a certainty." He admits, "There was a time when I was not convinced that success was in the set of possible outcomes." Currently, the team works day and night, and the key challenge is "rapid reusability" — which requires multiple test flights to achieve.
Theme 2: Mars — A Thousandfold Cost Leap from "Building Spaceships" to "Building Cities"
Musk proposes: A sustainable Mars city must pass the "great filter" — the city must be self-sufficient even if the Earth supply line is cut off.
- Current cost: Approximately $1 billion per ton delivered to the Martian surface (including rockets, launch, heat shielding, navigation, communications, and landing systems).
- Target cost: At least reduced to $1 million per ton, ideally lower. He believes that "without this cost reduction, you can't buy a ticket to Mars even with a trillion dollars."
- Requirements for self-sufficiency: Approximately 1 million tons of supplies (including semiconductor factories, iron smelting, and various infrastructure). A Mars city cannot lack any critical material, just as "a ship on a long voyage lacking vitamin C will eventually die."
- Timeline: Best case 5 years, worst case 10 years; the core bottleneck is Starship engineering progress.
- Civilization perspective: Musk cites Stephen Hawking's estimate of a 1% probability per century of a civilization-ending event. He believes becoming a multi-planetary species is "buying insurance for life itself." The Sun will expand and engulf Earth in about 500 million years — "this is the first time in 4.5 billion years that it is possible to extend life beyond Earth, and the window may be narrow."
Extrapolation and uncertainty: Musk is not sure he will see a self-sufficient Mars city in his lifetime, but hopes to see "at least a lot of momentum." He finds it worrying that "Americans haven't been back to the moon in 49 years" and questions whether civilization has already passed its peak.
Theme 3: Tesla Full Self-Driving — From "Number One Problem" to "Neural Network Eating Software"
Musk made it clear: To solve full self-driving, the human driving mechanism must be rebuilt — replacing human eyes + biological neural network with cameras + digital neural network.
- Core idea: From photon stream (image space) compressed into "vector space" (vehicles, pedestrians, lane lines, traffic lights, etc.). Once the vector space is accurate, the control problem becomes like a video game. Musk said: "Accurate vector space is very difficult, the control problem is difficult but not insurmountable."
- Latest engineering progress:
- Remove image signal post-processing, directly use raw photon count, performance far exceeds humans in low light, while saving 13 milliseconds of latency.
- Self-developed C compiler, compiles directly to FSD hardware, tracks latency from photon to actuator, and focuses on eliminating "jitter", because fixed latency is predictable, jitter is random and uncontrollable.
- Transitioning from a "bag of points + C++ code assembly" model to "neural network directly outputs vectors" (such as lane lines, drivable space, etc.), which Musk calls "neural network eating software, replacing heuristic rules".
- Introduce surround video and spatiotemporal memory to solve occluded object tracking (e.g., predicting a pedestrian's position after they walk behind a bus). Musk compares it to human "object permanence" — before a baby learns it, an object disappearing is a new thing, so "peekaboo is always fresh".
- Key data: FSD Beta's driver intervention rate (disengagements per million miles) is rapidly declining. Musk believes "it is likely achievable next year (2022) that the FSD accident rate will be lower than the human average," but the company's standard is "at least 2-3 times safer than humans" before declaring completion.
- Version 11: Includes a fundamental neural network architecture rewrite, will remove a large amount of C++ code, replaced by neural networks. Musk describes it as "one stack to rule them all" — a single stack handling all driving scenarios.
Musk's self-reflection: "I originally thought autonomous driving would be very difficult, but it is much harder than I imagined." He admitted that Tesla started based on Mobileye, then decided to develop it from scratch, which was a "bold decision."
Theme 4: Tesla Bot (Optimus) — From Factory Tool to Potential Companion
Musk initially focused on replacing "dangerous, boring, and repetitive labor," but acknowledged that robots may eventually become companions.
- Technical foundation: Tesla possesses "the most advanced real-world AI" (developed for autonomous driving) and "the best manufacturing capabilities," which are the two pillars for building humanoid robots. The robot will use Tesla's autonomous driving inference computer and a large amount of training data.
- Cost mindset: Musk believes that at scale, the cost of any product can approach the cost of raw materials. He cited an example: designing from the outset for the "cost limit at an annual production of a million units" can avoid inefficiencies caused by "inertia from existing tools and processes."
- Timeline: A "decent prototype" might be shown by the end of 2022.
- Companion potential: Musk had not previously thought deeply about this, but acknowledged that "robots could develop unique personalities and match the owner's character like a friend." He cited the Japanese aesthetic "wabi-sabi" (the beauty of imperfection) — a robot's flaws and quirks can make it feel relatable. He compared to C-3PO and R2-D2: "They have flaws, they argue, they make mistakes — but that's what makes them endearing."
- Neutral stance: Musk is open to "robot companionship" but emphasized that "the initial goal is to make it useful."
Theme 5: Currency, Institutions, and Information Theory — Musk's "Database" Perspective
Musk argues that currency should be viewed as a "resource allocation database," with governments holding editing rights, and cryptocurrencies attempting to reduce the "error" introduced by government dilution of the money supply.
- Real-world problem: The current banking system runs on "ancient COBOL mainframes." Governments have editing rights to the currency database and can increase the money supply at any time, which introduces "error." Musk believes this is "a hidden tax."
- Cryptocurrency assessment:
- Bitcoin: Limited transaction volume, excessive latency, unsuitable for daily transactions, but potentially suitable for "store of value."
- Dogecoin (DOGE): Due to "accident," it has higher transaction volume and lower fees than Bitcoin, and features a fixed annual inflation in quantity (rather than a fixed percentage), with the inflation rate decreasing over time. Musk argues that "this is the best foundation of any monetary system I have seen, even though it started as a joke."
- Mars currency: Due to the speed-of-light delay (4–20 minutes), Mars cannot run a blockchain synchronously with Earth and must issue a local currency. Musk says, "let the Martians decide for themselves."
- Institutional thinking: Musk supports direct democracy (people vote directly on legislation), with laws that must be short and understandable. He proposes that "laws should have sunset clauses" and that repealing a law should be easier than enacting one (e.g., 60% to pass, 40% to repeal). He criticizes "regulation accumulation without garbage collection" — "like C without garbage collection, Java has it" — and the result is "civilization's arteriosclerosis."
Targets Mentioned
| Target |
Guest Attitude |
Key Data |
| Tesla (TSLA) |
Optimistic about the future, but emphasizes that the difficulty exceeds expectations |
FSD Beta intervention rate declining rapidly; expects 2022 accident rate lower than human average; target safety 2-3 times that of humans |
| SpaceX (unlisted) |
Pushing forward with full effort |
Starship target cost $1-2 million per launch, carrying 100+ tons to orbit; Mars cost must be reduced to $1 million per ton |
| Dogecoin (DOGE) |
Slightly positive, believes "unexpectedly advantageous" |
Transaction volume higher than Bitcoin; low fees; fixed annual inflation in quantity |
| Bitcoin (BTC) |
Assessment is somewhat weak |
Limited transaction volume, high latency, high fees, not suitable for daily currency; Satoshi's anonymity assessed as neutral |
| British Museum (not an investment target) |
Neutral |
Points out the historical fact that "Britain took artifacts from around the world," but believes "on the whole, museums are a net positive" |
Judgments Worth Remembering
1. Musk: "Every time I see a rocket, I see all the things that could go wrong and need improvement, not the cool exterior." Support: As chief engineer, he signs off on nearly every design decision, and failure is "fundamentally my personal responsibility."
2. Musk: "Prototypes are easy, mass production is hard." Support: The most time-consuming part of Starship is engine production, not design; Raptor is the first full-flow staged combustion engine, requiring new materials, with a working pressure of 300 bar—the highest in history.
3. Musk: "A fully reusable rocket could reduce costs by 100 times." Support: Analogy to cars—buying a new car every time you drive is extremely expensive, while reuse (just refueling) is 1,000 times cheaper. Falcon 9 marginal cost is about $15–20 million; Starship targets $1–2 million.
4. Musk: "A Mars city must pass the Great Filter—even if Earth resupply is cut off, the city must be self-sufficient." Support: If any key resource (e.g., semiconductor factory, iron ore smelting) is missing, the city will eventually perish, like "a long voyage ship lacking vitamin C." Self-sufficiency may require 1 million tons of supplies.
5. Musk: "Money is a resource allocation database, and the government has edit permissions. This 'editing' is a hidden tax." Support: Current banks run on COBOL mainframes; governments can issue more at any time, increasing the "database error"; cryptocurrencies attempt to reduce this error.
6. Musk: "The difficulty of autonomous driving lies in accurate vector space, not control problems." Support: The control problem is like a video game, but compressing the stream of photons into vector representations of vehicles, pedestrians, lane lines, etc., is extremely difficult. Once solved, control becomes relatively simple.
7. Musk: "Laws should have a 'garbage collection' mechanism, like C++ or Java's GC; otherwise, civilization suffers from arteriosclerosis." Support: Regulations are added every year but never removed, eventually "everything is regulated, and nothing can be done." He suggests laws come with sunset clauses, making repeal easier than enactment.
8. Musk: "The laws of physics are the law; everything else is a recommendation." Support: First principles—starting from the most fundamental truths (e.g., conservation of energy), reasoning to conclusions, then cross-checking. Also use "extreme thinking" (scale to extreme) to test cost issues.