"Not Even Wrong" Podcast
Investing in fundamentally new concepts and engineering practices with large impact.
Robot models memorize action chunks. Bitter Lesson. Mac Schwager question: are they memorizing or generalizing? What does generalizing even mean? One token maps to more than one action chunk. Most don’t. Bitter Lesson. More data means more memorizing. Generalization? Adapt to unseen situations. Self and continuous learning. Sutton’s ‘Era of Experience’. Learn how to explore. Andrew Wagenmaker; teach robots how to explore. ❤️'Exploit Exploration' .Robinson Crusoe robot. Look at full stack AI cake (Energy, Chips, Cloud, Model, Application). Flexibility throughout stack. Company level, like Tesla is building FSD for flexible adaptability.
Robots will train world models. Not the other way around. Train robots on video. Then deploy them. Collect data from robot fleet to build real world embedded world models. Use those models to train robots. Data flywheel of real world AI. The key to world models is to create an orders of magnitude larger data set of state/actuator/transformation data.
The tables have turned in the chip business. Intel and AMD used to lag Nvidia. Now it’s reverse. Or is it? CPU demand is up because agentic AI requires sequential decision making. Nvidia is solving this problem with Vera Rubin and speculative decoding at scale. ‘Parallelize sequential compute’. The current AI infrastructure is stretched. Too many tokens pushed through legacy pipe (energy, CPU, cloud). Nvidia will solve the bottleneck of compute with massive scale up in inference and token/dollar. Cost structure of AI must change. Chinese want latest products because agentic AI requires cutting edge (Vera Rubin). Microsoft wants to raise token pricing . If Tesla sold the Model Y with FSD for 20k, demand would be through the roof. That’s what coding agents are doing currently. Token prices must go up and/or cost come down. Nvidia.
The West is at risk of succumbing to a cultural revolution reminiscent of China’s in the 1950s.
Woke ideology is the real threat — not the CCP.
Woke culture prioritizes redistribution and equality of outcome. These are little more than euphemisms for socialism. The greatest danger lies in the desire to halt progress, much like the Gang of Four did during China’s Cultural Revolution under Mao. Defending the core principles that built the West: freedom of knowledge creation, freedom of speech. China is optimizing for prosperity — and it is achieving that. Meanwhile, the West is optimizing for redistribution and equality of outcome — and we are getting exactly that.
A clear example is the education system. Engineering and science are often rewarded the same — or even less — than humanities, law, or business. Recipe for decline. Purpose of math is to describe physics. Physics explains nature and provides the foundation for engineering. Engineering, in turn, develops the technology that improves lives and protects us from harm. When we stop valuing these disciplines, we stop building the future.
Humanoid Robot pipeline starts with learning from humans. Then robots. Then world models. Flywheel. First step. Pre train: Internet scale video data. Mid train: Teach robots skills with goal oriented unsupervised RL. Robot uses pre training to learn how to explore skills to achieve goals. Fine tune: Make robot human friendly. Use robot data (state, actuator and transformation function) to feed world model. Create scenarios based on world model. Feed back to robot. Flywheel. World models learn from robots - not other way around. ❤️ The reason we have humanoid robots is not that the world is built for humans. It’s because robots can learn from humans and have similar morphology.
❤️ Robots will train World Models - not the other way around. World model companies are like Waymo or Yahoo. Good idea, but too early. True physics simulation must be embedded in realty. World models will succeed once we have enough robots to learn from. Waymo was too early because computer vision wasn’t around. Yahoo, because algorithmic search wasn’t around. Causes bad habits and Frankenstein tech.
❤️ Discussing presentation ‘Learning from Experience for Robots’. Andrew Wagenmaker ‘Posterior Behavioral Cloning’. Use imitation from humans to define action space for robots. Then let them explore themselves and solve problems. Instead of mimic human actions, use humans to define action space. Pre-training on video scale data builds a database of probabilistic physical interaction. Then mid-training, which is learning actual skills. Robot learns on its own by interacting with real world.
State of the Union in Robotics. 1/4 Large investments by corporates. 2/4 Science showing progress. 3/4 Following same model like digital agents. Pre-train on internet scale data from video. Mid-train and fine tune to push frontier. Pre-train from human video (humanoids make sense because training from humans). Mid-training specific skills. Fine tune so robots behave human like. Tesla FSD playbook. 4/4 Generative video to close the out of distribution gap in mid-training and fine tuning. World models enable self play.
Tesla Space X merger must happen but not at these levels of Space X. Technical risk with Space X is still high because wealth potential hinges on starship. Similar to FSD 5 years ago. As Tesla shareholders we run the risk of waiting again for wealth creation. Therefore, it’s better to deliver against our milestones (FSD unsupervised, Cybercab, Semi, Energy) and higher valuation when merging. This way the technical risk of startup weights less on TSLA shares. Starship will work, but might be delayed. Upside, like Space AI, Terafab, launch business, moon and competitive edge depend on rapid reusability of starship.
Book discussion ‘Stoner’ by John Williams. Literary masterpiece. A man goes through life without noise and fanfare. But all the little actions accumulate to a dramatic story of perseverance, love, rivalry and meaning. A man must have purpose and love. Stoner has both. Purity of academia versus dirt wars of real world. Accepting the mundane and insignificance elevates life to meaning and purpose. A novel from the American West. Answer to Nietzsche, Kant and strong repulsion of Foucault and other post modern idiots. Stoner is like water, ‘slips through fingers and holds up a ship’. Academia straddling between protective seclusion and utility of society.
Discussing presentation Robot learning from human data. Humanoid form factor is conducive to learning from human data. Similar morphology. Meta glasses as FSD for robots. Tesla pipeline in FSD same for robots. Learn from humans. Generated video to fine tune. Human video data is internet scale for robot learning. Morphology gap. Sim to real gap. Predict human action. Navigation, locomotion, manipulation. Diverse human actions yield emergent learning in latent space. Pay Dordash workers to carry camera.
Book impressions ‘Stoner’ by John Williams. Part 3. Stoner’s quiet battle is a defense of the university’s soul, often maddeningly self-absorbed yet possessing a fragile purity and beauty that enriches humanity and must be protected from opportunists like Walker. I agree to a degree: the same corruption of original spirit is visible in Silicon Valley, where figures like Sam Altman have diluted genuine curiosity and invention with hype, capital, and unleashed greed. The challenge is walking the thin line between hermetic self-obsession and porous openness to novelty —a balance that requires courage, since the purity worth protecting is something the world desperately needs yet rarely knows how to value.
Robot learning from generated video. Paper discussion: From Generated Video to Physically Plausible Robot Trajectories. Solve data scale problem for robotics by generating videos at scale, use RL to train robots in sim and then sim to real. First, generate the video with human moving. Second, adjust generated videos to realistically reflect humans. Third, close the morphology gap (human to robot) by adjusting the sizes to the robot. Using key points (joint points, etc.) to match the morphology. Fourth. Train in sim using RL. Fifth. Sim to real. Limitation. No reasoning applied. Future work requires reasoning to adapt to unseen and untrained problems.
Book impressions ‘Stoner’ by John Williams. Part 2. A student entering university is like an explorer discovering territory, not a random sequence of events. Believing in a better future through eduction might be quixotic (as Masters says), but better then alternative. What else is there? Pascale’s wager. You’re better off believing in something, even if you’re wrong.
Book impressions ‘Stoner’ by John Williams. Part 1. The university is there for them, not for the students. Eludes to a misallocation of resources in academia. Mission creep. Academics serve their own interest, not the students. In my opinion, the purpose of university is to improve the future of its students. And the purpose of science is to lower cost of energy, healthcare, housing, defense and eduction, the most important items for regular people. As a society we made a Faustian bargain with science. We are betting on progress at the expense of ending up like the Sorcerer’s apprentice, destroying ourselves. In return we must push for better lives for most, not just academics.
Learning from video solves the data scale problem for robotics. Jitendra Malik presentation. Pose estimation. 3D representation and transformation into 4D. Translate imitation learning from video into robot actuator information. End to end or hierarchical. Engineering robot learning data. Next step is generate videos and then develop architecture where robot decides what data it needs to learn tasks. Crate your own distribution of data learning and solve out of distribution problem.
❤️ The future of AI and the future or reasoning. What data do I need to get better? How many iteration around which dataset do I need to get better? Learn how to generate the data necessary to learn skills when pursuing goals. Train yourself in runtime. Find ways to better learn how to best achieve a goal.
Book discussion ‘Never Let Me Go’ by Kazuo Ishiguro. The purpose of a novel is to describe the human condition. This novel does a great job in doing that. Nietzsche, Dostoyevsky. What is the core of being human? Belief. What is love? Souls combine. Memory. Human existence requires more than love and passion, fear and envy. It’s about controlling your own body, own destiny. Pascale’s wager. Better believe in something than just list facts. Never do to humans what you would not want to be done to you. But keep it in perspective. Why aren’t they rebelling?
Discussing presentation by Sergey Levine about Robotic Foundation Models. Action chunks enable better application for test time compute in real world AI. ‘One of the more interesting innovations in language models has been the introduction of test time compute, which is the ability to use compute at test time to solve problems, i.e. reasoning.’ VLA, vision language action, models use the ability to reason (from LLMs) to solve problems in real world. Autonomous learning. How to make the system conducive to learning? By generating the ‘right’ data for learning. Set goal and let robot figure out how to improve by generating the ‘right’ environments and data sources to learn from. Action chunks are compressed actions sequences. Out of distribution inference. What is reasoning? Game theoretic (recursive) sequence of events. Reasoning is what Sutton (Bitter Lesson) calls search. We need all five layers of the cake to move AI forward. Design a system that is conducive to learning.
Book impressions ‘Never let me go’ by Kazuo Ishiguro. Part 2. Why did Madame take their art? In order to determine whether they are truly in love. This is Huxleyian, but inverse, instead of love not permitted, love is the highest achievement. But it must be proven. How to prove you love? With art. Love even exists without reproduction. Separate love from sex and reproduction. Love is a privilege.
Book impressions ‘Never let me go’ by Kazuo Ishiguro. Part 1. Scene where Cat finds the tape. Fine grained emotional strand weaves through novel. They are different but they are human. What does it mean to be human? The purpose of literature is to describe the human condition. And Ishiguro does a great job at that. Cat and Tom are deprived of basic human functions but they still show human like emotions and reaction. By dehumanizing humans and still depicting very human like conditions, Ishiguro shows great artistic presence.
The purpose of the Fed is to finance government spending. Anybody who says different is ignorant and/or a charlatan. Particularly detrimental is when professional economists use jargon to divert from the real question, which is how we stop the Fed from monetizing government debt. The US government has over 40 trillion USD in debt. That’s the Fed legacy. This must stop. Otherwise our society will drift into Socialism. In this interview Douglas Dimond says the right things, but he stays away from calling for a stop of Fed driven government financing.
Thesis discussion Baifeng Shi PhD defense. Scaling vision models or not? Instead of larger models (more parameter count) train models with different resolutions. Task oriented gaze. RL for learning how to mask model so focus on gaze. For example, only focus on what moves. Good example of how robotics can be solved with AI instead of sensor problem (event driven sensors).
Book discussion ‘Foster’ by Claire Keegan. Beautiful poetic exploration of what matters in life. Love, belonging, hope, care and kindness but also shame, neglect and fear. Keegan uses sparse language with a subtle emotional undertone. This is a poem written in prose. Finding your place in life starts with people. The characters blur with the landscape and form a stage in which the narrator evolves into a person.
Book discussion ‘A Mind at Play. How Claude Shannon invented the Information Age.’ Great story explaining how Shannon created new theory of information. Communication is fundamentally about prediction. Storage is transmission through time, communication is transmission through space. Two key contributions of Shannon are 1/2 Combining Boolean logic with circuits. Adding physical layer to the concept of binary logic. Enabling information processing based on Bits. 2/2 Information is about reducing entropy. Adding redundancy to make messages predictable. Communication is problem that can be solved with engineering. Two more takeaways. One is importance of physical substrate of information processing. If we have more than Bits, like in a quantum computer, we can do different math and logic. Also, interesting is his take on investing. Wealth creation through investing is taking stakes in technologies that create real value and hold on.
Book impressions ‘A Mind at Play. How Claude Shannon invented the Information Age.’ Part 3. Shannon develops theories about how to play stock prices. But eventually, he decides to drop the endeavor since its futile. He makes his money with technology investments, in things he knows something about. That's wealth creation with capital. There is money to be made with modeling stock prices. But that has nothing to do with investing. Algo trading is a fancy word for taking liquidity risk or even worse, speculation with a third party holding the bag when things go wrong. Wealth creation is about holding on to technologies that deliver real value.
❤️AI anxiety and what to do about it. Focus on what is not going to change.Lower cost, more energy at lower cost. More convenience, more safety. This applies to businesses as well as people. When deciding what to learn and how to deploy your human capital focus on those factors. For people, build a pipeline of habits that make you indispensable in an economy where those things matter. For businesses build scale and develop processes for rapid iteration to compound knowledge.
Copernican revolution of human relationship with work. Instead of defining yourself through work, put yourself in the center. Humans optimize for two things in life; admiration external and internal. Neither wealth, nor career nor the beach house, Porsche or trophy wife matter. What matters is admiration. We think we want it from others, but what we really care about is to be happy with ourselves. With AI coming, flip the problem. Focus on what makes you admire yourself. Do good deeds, care about loved ones. Be kind and think about what you want to tell your descendants in hundred years. 2 types of jobs remain for humans. 1/2 Feeding the machine. 2/2 Human to human. If you choose the former, focus on defining the problems and data necessary to feed the machine so it can help you find solutions.
Tesla delaying scale rollout of Cybercab but better FSD supervised adoption. Elon is embarrassed on the call because he talks a lot about the future and not enough about current business drivers such as Cybercab and FSD unsupervised. Problem seems to be the AI. Longer time frame testing investor patience. Good news is that FSD supervised is finally getting material traction with customers and driving better demand. Five vectors for higher stock this year. 1/5 FSD supervised rolls out globally and drives sales. 2/5 FSD unsupervised rolls out in selected geographies and drives sales. 20% volume growth. 3/5 Semi starts scale production 4/5 Energy grows at least 20% in sales 5/5 Margin expansion.
Book impressions ‘A Mind at Play. How Claude Shannon invented the Information Age.’ Part 2. Information is reducing entropy in message. Redundancy increases predictability. Communication is predicting. The more redundancy, the better error correction. Shannon introduces this concept and creates communication theory.
Book impressions ‘A Mind at Play. How Claude Shannon invented the Information Age.’ Part 1. Shannon combined Boolean logic with circuit design. Great innovation and birth of mechanized information processing. Lays foundation for a whole industry. Great example of mathematics. The purpose of mathematics is to describe nature. Shannon took the description of logic (Boole) and applied it to circuits. Bridge between concept and the real world. Math and physics should be like that. The purpose of a University is to design the best possible future for its students.
Book discussion ‘The Art of Doing Science and Engineering’ by Richard Hamming. You can train yourself to do great things. Luck favors the prepared. When rigor enters, meaning departs. What you learn from others, you can use to follow. What you learn for yourself, you can use to lead. Doing great things and self confidence are trainable muscles. In the short run most endeavors are nothing more than a Rorschach test. In the long run, your career trajectory is a good proxy for intelligence. Longterm, it’s the habits that determine success. You get what you measure. Be careful what you measure. The story about the fishermen who use the net to determine the size of the fish. Tool determines theory. Dangerous.’ This book is written for young engineers. Vision mattes. Pascale’s wager works for every large decision in life. ‘Drunken sailor with pretty girl.’ ‘Universities must be obsessed with the future of their students, not with the past of their professors.’ The purpose of mathematics is to describe physics. Learn how to learn. Overcome arrogance, inject humility and avoid overconfidence. As a culture we are at crossroads. Where do we want to be? In an Richard Hamming world or in a Woke infested relativistic de-growth world.
Wayve has good tech but not so good business. They have the right approach towards robotics. End to end, algorithmic solution. Problem is data and business model. Selling to OEMs is wrong approach. Wayve needs to fully integrate to compete with Tesla and Chinese players like Xiaomi. Lack of vision. Best case under current model is acquihire. But they could become a real player if they integrate into hardware.
Book impressions ‘The Art of Doing Science and Engineering’ by Richard Hamming. Part 2. What is AI? Wrong question. What purpose does it serve? What goals is it pursuing? AI is Computational part of pursuing goals. Human-AI interaction is not about work. It’s about value. Humans are discriminators. They decide what is valuable. AI must must lower cost and increase value. Role of humans is to run their own agentic companies. The CEO economy. There should be no human on the battlefield nor in the workplace.
Book impressions ‘The Art of Doing Science and Engineering’ by Richard Hamming. Part 1. Vision matters. Accuracy not as important. But belief is direction to greatness. ‘If you know what you’re doing in science, you shouldn’t be doing science. If you don’t know what you’re doing in engineering, you shouldn’t be doing engineering’. Mindset matters. Match your frequency and wavelength with reality. Must be yours. History matters. You might not be able to extrapolate. But it’s best we have.
Book discussion Crux, a novel by Gabriel Tallent. Purpose of life story. ‘Catcher in the Rye’ meets ‘Into the Wild’. Social activism trickles through story. Unfortunate. Tama character unlikalbe first, then ends up likable. Pivot of character development could be more smooth. Stories are built on characters. Eventually, love is the base of all endeavor. Tama finds purpose in love and care. Dan doesn’t. Adults fail their kids. Ultimate purpose of life is to enable a better future for next generation. It’s not podiums and posters. Even inspiration doesn’t count. It’s when you do things for others that don’t nurture your narcissism that actually make you stronger. Kindness. Fathers don’t matter. Why? Climbing as a metaphor for life is good choice. Social activism trickling through the climbing world, not good.
Book impressions Crux, a novel by Gabriel Tallent. Part 2. Why go to college? What does education actually do? It’s not about getting a job nor about securing a regular, normal life. That’s precisely not what a college eduction is supposed to deliver. Learning how YOU can learn and think. It’s the first step in defining YOUR path. Developing into a real person with a unique path. Climbing is a good metaphor for life and dreams. But it’s not about chasing numbers or podiums. It’s about creating your own path. Start with a college education. Take it seriously. Go in depth and then follow path. Dan should never, ever give up college for climbing. That’s stupid.
Discussing presentation by Mac Schwager. Are foundation robotics models general AI or are they just memorizing? Mostly memorizing. But what does it mean to be general vs. memorizing? Schwager studies tokens and how they appear. General means, they appear in many different, somewhat related scenarios. Memorizing means, they only appear in specific scenarios. How to remedy? Lots of diverse data and general models. Ideally, the robot learns from its own experience. Generality is the ability to solve problems in previously unseen scenarios. Bitter Lesson.
Book impressions 'Crux', a novel by Gabriel Tallent. Novels with political undertone are redundant. Using climbing as a metaphor for life. Good idea. Step by step solutions to problems. We’re not happy about using climbing as a veneer for social activism. Athletes must not engage in political activism. Focus on performance and keep sport out of politics. Character of Tama is unlikable. Risk of stereotyping lesbians as sociopaths like Tama.
Discussing presentation by Daniela Rus about energy efficiency inference models. Liquid networks. A network of neurons each defined by a differential equation. More information per neuron. Less I/O and math per matrix. More inference per Watt. Strong models. Sequential. Good for robotics or real world AI. Drawback; context. Is this just repacked ODEs? No. Neureal nets defined by ODEs with learned coefficients. What about continuous learning?
Buchdiskussion ‘Unterleuten’ von Juli Zeh. Postmodernes Theaterstück auf der Bühne eines Romans. Wie im Theater, so muss der Leser auch hier im Spiel der übertriebenen Simulation eines Soziotops mitmachen. Don Quixote trifft Godot, trifft den Besuch der Alten Dame. 1/3 Realität vs. Illusion. Es gibt keine Wahrheit, nur Geschichten. Kein Sinn, nur spielen, Gewinnen. Entschleunigung der Gesellschaft. Emotionen passen nicht mehr. Menschen ohne Zweck und Berufung. Wo bleibt die Tugend? Wo ist die Frau und wo der Mann? 2/3 Godot. Kein Sinn. Sogar die Rossfrau hat nur Gewinnen im Kopf. Es fehlt an Humanität und Liebe. Vor-apokalyptische Gesellschaft, die im Sumpf der Postmoderne versunken ist und sich am Leichengift moralischer Überlegenheit vergriffen hat. 3/3 Besuch der alten Dame. Wenn plötzlich Geld fliesst, spielen die Dorfbewohner verrückt. Verlockung durch Reichtum verändert Gemüter und beeinflusst das Verhalten. Deutschlandbild einer fahlen Gesellschaft, die in sich selbst zerfällt. Spiegelt nicht nur Deutschland, sondern Westen.
To generate wealth you must concentrate. Invest in maximum epistemic uncertainty with large upside. Only constraint; survive. Christopher Columbus style of investing. Legacy investment theory has it all wrong. It’s not about diversification. It’s about maximizing concentration on positions where epistemic uncertainty and potential returns are high. Legacy investment managers measure uncertainty as standard deviation of stock prices. Useless measure of risk and wealth creation. Wealth is predicated on solving difficult technological problems. Such problems contain high epistemic uncertainty, that is knowledge that is not available right now but can be obtained through iteration and error correction. Engineering is another word for reducing epistemic uncertainty. Concentrate on deep tech with high epistemic uncertainty and entrepreneurs tackle such problems at scale. Constraint is; survive. Darwin. Niche. Develop knowledge to occupy new niche. Large, profitable and scales with customer adoption.
Productivity Technology usually benefits the few and leads to increased inequality. But technology as a whole benefits all. That’s why we tolerate engineering. Only about 5% of top developers are getting better with AI. The rest falls prey to mediocrity. The same dynamics played out when calculus was invented in the 18th century or software in the 20th. Technology is a divider when it comes to productivity because some people figure out how to out perform others. The gap exponentially wides. But technology in general is an equalizer. The internet, self driving electric cars, penicillin, low cost transport and most other engineering marvels benefit all. But productivity tools don’t. Society has to deal with this to stifle the rise.of Socialism driven by feelings of jealousy and envy.
Eindrücke ‘Unterleuten’, ein Roman von Juli Zeh. Teil 3. Charakter von Linda Franzen ist unerträglich. Egoistisch, obsessive Narzisstin. Rossfrauen Dasein ist bloss ein Vorwand, um ihren narzisstischen Trieb voll auszuleben. Doch sie erinnert mich an Elon Musk und andere Unternehmer, die ich mag. Wie kommt das? Der Unterschied ist, dass Musk und andere (Jeff Bezos, Larry Page, Palmer Lucky u.a) eben Unternehmer sind, deren Drang nach Schaffen zu enormen gesellschaftlichen Fortschritten führt (Elektroautos, Internet, Energie, Verteidigung). Allen gemeinsam ist, dass sie durch Technologie Kosten senken und so den Zugang für viel mehr Menschen ermöglichen. Linda’s Drang dient nur der Ablenkungen ihrer tiefgründigen Unsicherheit gegenüber dem Leben. Nutzt niemandem, nicht einmal ihr selbst.
Tesla investors waiting for Goal or Godot? Holding on to Tesla has not been rewarding in the past five years. Patience or stupidity? Today’s delivery numbers are disappointing. Why are people not buying more Teslas when there is FSD and best in class product? Not knowing why things aren’t working is unnerving. Why are we holding on to Tesla? Fully integrated robotics company. Transport as a service. Labor as a service. Cybercab is transitioning from technology risk to business risk. Solving tail end with data driven learning takes time. Scalable. Problems Tesla is solving are conducive to data driven learning. Why is Tesla storage business not growing when energy demand is accelerating? What to do now? Stick with fundamentals. For example, we’re going to CVPR this year because vision is key to robotics.
Tesla should buy Physical intelligence. Sergey Levine interview and his take on Moravec’s paradox. 1/7 Moravec paradox applies to all AI. It’s not the difficulty which matters but how conducive a job is to learning from data. 2/7 Robotics an AI problem, not sensors. For example, haptics can be solved with vision. Robots have different constraints than humans. Don’t anthropomorphize. 3/7 Neurons fire based on tip of tool, not tip of hand. 4/7 Generality is best achieved if a system improves while doing it’s job and pursuing its goal. Tesla is solving tail end of FSD through AI, not explicit interventions. Scalable. 5/7 What’s the kindest thing somebody has ever done to you? 6/7 Create systems that get better at the thing the more they do the thing. 7/7 Physical Intelligence could be for Tesla what Deep Mind is for Google.
Eindrücke ‘Unterleuten’, ein Roman von Juli Zeh. Teil 2. Zitat von Gerhard: ‘Wir leben in einer Welt, in der Ärzte das Gesundheitssystem zerstören, Akademiker das Wissen zerstören und Politiker die Freiheit vernichten.’ Obschon ich politisch mit Gerhard nicht übereinstimme, bin ich mit diesem Zitat einverstanden. Zeh zeigt wie politische Meinung verwickelt ist und die vermeintlich klaren politischen Demarkationslinien falsch sind.
Robotaxi competition is about cost, comfort and scaling. Demand for transport grows exponentially. Three levels of competitive dynamics between Tesla and others. 1/3 Short term, Due to inferior tech, Google pumps billions into Waymo in hope to suffocate Tesla. How can Tesla respond? Much lower cost, lower wait times, more comfort, network effects and scaling anywhere globally. Every company is ultimately defined by their cost structure. 2/3 Mid term. Rapid iteration across Robotaxi value chain. Automate tail end. Waymo is self driving car on digital rails. Cannot scale due to human labor and explicit code growing exponentially with tail end. Risk; Waymo fights with regulatory capture. Incumbents haven’t been able to lower price. 3/3 Tesla is launching a service which is designed from the ground up to be fully automated for scale, low cost, high comfort and reach. Customers can turn their robots into the network. Cybercab. Optimus (maintenance, logistics). Grok, network management. Digital Optimus, agentic logistics.
Energy, Defence and Transport. Invest your money and career in those areas. Pascal’s wager. Believe in God even if you don’t. It’s a better bet. Same applies to career and money. Define areas of belief and and train yourself to become passionate about them. Then focus, learn and iterate. Deliver lower cost, safer, cleaner and scalable energy, transport and/or defense. Purpose. Through history humans mostly cared about energy, transport and defense. It’s where real wealth is created. No Ponzi needed.
Eindrücke ‘Unterleuten’, ein Roman von Juli Zeh. Teil 1. Beschreibt Charaktere mit spitzer Genauigkeit ohne der Pedanterie zu verfallen. Satirisch übertrieben und unterhaltsam. Fortsetzung von Middle England - die Deutsche Version. Zeh zeigt die Wurzeln der gegenwärtigertigen Frustration in Deutschland von beiden Seiten, Links und Rechts. Fehlender Bezug zum Leben, ohne Zweck und mit viel Lärm und teurer Unterwerfung in der Stadt und gleichzeitig fehlende Relevanz auf dem Land. EU als Bürokratiekorsett mit dem Ziel, den Fortschritt zu stoppen. Die Progressiven von heute sind an Erhaltung des Bestehenden interessiert und sehen alles Neue als Gefahr. Aufschwung von Rechts, weil sich die Rechte zumindest verbal um die Leute kümmert, während die Mitte und Links in einem zwecklosen Idealismus von Europa, Ökologie und falscher Toleranz versinkt.
Diffusion models are the future of AI because most bit streams will be video. Stefano Ermon introduces diffusion for text. But that’s just a side show. Elon’s X post ‘The vast majority of AI workload will be video understanding and generation, so good chance diffusion is the biggest winner overall.' And Stefano’s response: ‘For video and world models, diffusion is simply the right tool: more parallelism and better scaling.’Diffusion finds signal by symmetry breaking. Much better for video. Robotics based on vision is driven by diffusion. Tesla FSD using diffusion for video generation in training. Diffusion also utilizes compute more efficiently thanks to parallelism. Elon’s comment: ‘ratio of compute to memory bandwidth will increase.’
Negar Mehr. Multi-agent collaboration. Robots and digital agents. What is an entity? One robot, a system of robots? Game theory. Multi agent collaboration through control theory. Or RL? Using LLM for high level semantic decision making. LLM guides step by step. Good at predicting next step. Like an operating system for multi agents. (Optimus managing Cybercab network). Collaboration is a big deal. Even human intelligence is not fully understood when it comes to collaboration. Coase theorem. Why do people work together. Collaboration? LLM, maximum likelihood method. Modeling reward functions. Drawback in robotics is LLMs lack spacial understanding.
Trump let us down. He started an unwindable war. What to do as an investor? Trump’s and Israel’s war against Iran is unwindable because there is no purpose, no clear definition of what a win means and no plan of what to do after. Like Vietnam, Iraq 2 and Soviet vs. Afghanistan. What to do as an investor? Make sure your daily agenda is aligned with your mission. In our case, we attempt to generate wealth by investing in fundamentally new concepts and engineering practices with large impact. Focus on deep tech such as robotics, energy, new concepts for AI infrastructure, agentic real world systems (such as Cybercab network). Deep tech is best achieved by fully integrated goal oriented companies optimized for iteration. Tesla. Nvidia. Space X. Anduril. Monitor leverage and cost structure of fund.
Book impression ‘Middle England’ by Jonathan Coe. Part 7. Immense antagonism of pro EU people after Brexit vote. Elites were relaxed about the debate because they thought they will win. After loss, bitter anger. Seeds sown for Covid and Wokeism. Similar developments in US after first Trump election. Urban aristocrats suddenly woke up to a world where 'real people' fight back. That’s when they chose to impose Covid and the radical censorship on us. They masked us and looked us up. Trump 2 gave the US a breather. Now he’s making a huge Iran mistake. Iran could become his Covid 2.0 with dire consequences for future elections. Europe still strangled by the Woke crowd. US cities lost to Woke crowd.
Book discussion ‘Middle England’ by Jonathan Coe. Great portrait of social and political climate in Britain prior to Brexit vote. Described through well built characters. 1/3 Modern Madame. Bovary. Decline of contemporary British aristocracy. Sophie and Benjamin are modern aristocrats. Privileged, multicultural grifters. They want advantage through convoluted culture of tolerance and acceptance whilst refusing their own. Sophie is a fruitcake. Good intentions. despicable actions. 2/3 Bifurcation of society in people with purpose and without. All pro EU characters in this book have shallow jobs, careers and lives. Devoid of purpose and love. The pro Brexit crowd has real jobs, with real purpose, but not fancy (forklifts, driving school). 3/3 Everybody is attached to an idea. People live in their heads. Idealism is ok as long as you have a real life. Otherwise it becomes a dangerous drug. What is the path forward. It’s purpose or not. It’s love or not. Deep-tech. Do something real that you understand. Avoid EU and US East Coast phantasy life devoid of purpose and wealth creation. Learn how to love your work and the people around you. Coe keeps characters raw
Book impression ‘Middle England’ by Jonathan Coe. Part 6. Sophie is the epitome of what’s wrong with urban, multicultural society. Suicidal empathy. No bad intentions, But her actions are despicable. Tolerance with everybody but her own. She excuses people who almost destroyed her and despises the ones who love her. Sophie lives in her head and acts like s fruitcake and closet racist. Coriander doesn’t just happen. You can stop this. Inaction is destructive. Coe does a great job is depicting Sophie’s character.
X is falling behind Anthropic. Is this a problem for Tesla? No, because Tesla is working on a different problem. Yes, because Woke AI must not garner resources. Tesla and X are pursuing real world AI. Problem solving in the real world with digital and physical robots. Tesla near term catalysts are energy, Semi, Robotaxi and unsupervised FSD. Longterm Terafab, energy and compute from space, digital and physical Optimus. Tesla is building a fully integrated supply chain. Important to iterate and solve problems for real customers in real world with robots. Nvidia risk is to the upside. Competitive risk weighing on stock.
Book impression ‘Middle England’ by Jonathan Coe. Part 5. Fruitcakes, Loners and closet racists. The trifecta of Woke mind virus. Fruitcakes. Idealists and utopists who think you can engineer a fair society. By doing so they create an even more unfair society. Loners, isolated hermits. Lack of feeling and empathy. Closet racists, social justice warriors who tolerate everything but Western society. They are racist, sexist and nationalist. Exhibiting suicidal empathy. Multicultural closet racists are often self depreciating. The hypocrite version of closet racists is those who want Wokeism for all but their own.
Book impression ‘Middle England’ by Jonathan Coe. Part 4. Sequence when Sophie gets harassed by social justice warriors shows in subtle way how bad Wokesim is spread in academia. Very Orwellian. It just happens and nobody is to blame. Well, no. There is a culprit. It’s Coriander and she must be reigned in. The problem is not her aggression, the problem is the passive aggressive inability of everybody else to fight back and put her in her place. Coe writes very skillfully. His character build up and social context is subtle, but clear. Trump freed us from this tyrannic the US.
Book impression ‘Middle England’ by Jonathan Coe. Part 3. The irony is that when there was less diversity of writers in English literature, when they were mostly men, they wrote about lots of progressive issues. Today we have more diversity but pretty much one option. The message which is convoluted, Woke infested. Soviet style, Orwellian. Art cannot flourish under the tyranny of political correctness and Wokeism. Art doesn’t just require freedom, it is freedom. Today’s Western art culture is not free. Just look at Hollywood and American literature. Coe is refreshingly Habermasian.
Book impression ‘Middle England’ by Jonathan Coe. Part 2. The opening of the 2012 London Olympics shows the beginnings of Woke virus. Britain is not a multicultural society. It’s a culture of tolerance at risk of being undermined by those who were shown tolerance. Ironic and sad. One of the characters writes a PhD on how painters show black ancestry in European writers. That’s something Nazis would do. Disgraceful. Quote ‘for the privileged equality is a step down’. Wrong! It was the privileged who enabled the idea of equality. All those asking for equality today either came from places that never had it or feel bad about who they are. Only wealth and industry enables opportunities for all. Complacency, riots and anger don’t.
Book impression ‘Middle England’ by Jonathan Coe. Part 1. 1/2 Political. Approachable. Beginning of end of Britain as we know. Common sense of injustice amongst British society. Always been the case, but until 2008, people thought they have chance. (Difference between Communism and Capitalism is that latter gives you a chance). Not anymore. 2/2 One party system. Political parties are for show. Politics used to be about catering to your constituency. Today it’s about catering to own privileges. London is center of privileged society, detached from rest. London, New York, more in common than London and Birmingham. Problem of Western society. Radical indecision. Todays politics in nutshell.
Book discussion ‘The God of the Woods’ a novel by Liz Moore. 1/3 Disappointing. Thriller made to thrill. Bad execution. Suspense driven to the absurd. Insults readers intelligence. Like B-TV show. 2/3 Woke, feminist message. Boring, obsolete. Rich, white men are evil, everybody else cannot do wrong even if they do. 3/3 Good character build up. But strong characters are not enough for a good story. You’d expect literature to be immune to Woke group think. It’s not.
More energy enables more power and more intelligence. Space AI is the quest for orders of magnitude more intelligence. Energy is the ability to do stuff. Dot product is the mechanics of intelligence. The more energy, the more dot products, the more intelligence. Space AI removes constraints of space and power.
Space X is trailing in the AI race. What is the race? 1/3 Performance and quality of foundation models 2/3 Data centers 3/3 Market share. Space X trails in market share, not in performance and quality of models nor in data centers. By going to space, Space X removes two constraints, space and energy. New compute architecture with abundance of space and energy. Intelligence is compute and demand is infinite. Space X is building a truth seeking AI, which will prevail over Woke AI. Also, iterating along real world AI with Cybercab and Optimus are scalable use cases. Iteration along real customers is best way to develop AI. The more usage, the better it gets.
Highest value is when intelligence solves human problems.
Tesla’s FSD is being stalled by regulators on behalf of the legacy auto industry. Movie Longitude shows how watchmaker in the 18th century solved the longitude problem for British Navy. It took decades for the technology to be adopted and the movie shows why. The establishment was stalling by continuously asking for detailed explanations and overwhelming the watchmaker with scientific pedantries. Similar situation with Tesla FSD. Regulators are stalling to gain time for their constituency, legacy OEMs, not the public. FSD would save lives and give disabled, older and younger people more mobility if the regulators were more honest about the technology.
Knowledge and Truth and the job of the future. Inspired by Jessica Moss talk about knowing what is what. Discrimination as driver of knowledge, the ability to distinguish this from that. But how do we avoid a Wittgensteinian word salad arms race? Explanation. Epistemic uncertainty and aleatoric uncertainty. Concepts converge. By reducing aleatoric uncertainty (increase knowledge domain) we increase epistemic uncertainty (there are many more problems to be solved, and solvable with more data and compute. The job of the future is to reduce epistemic uncertainty in specific domains (aerospace engineering will work on AI models to increase knowledge about how to solve problems). Foundation models must follow truth seeking explanations. Most value. Demand for intelligence is infinite.
Quantum for AI and AI for quantum. Two presentations that both tackle quantum physics but from different angles. Researcher 1: Jiaqi Leng uses quantum computation to accelerate finding optimal point with lots of local optima. Interference, superposition and hamiltonian evolution offer a faster look at the landscape and accelerate optimal solutions. His PhD thesis is about setting the foundation for quantum computation to solve highly non-convex problems with lots of local optima. Researcher 2: Liang Fu. Use a neural net and transformer architecture to guess and approximate many electron interaction. A “first-principles AI” framework is introduced in which neural networks as universal and systematically improvable variational wavefunctions (guessing Hamiltonians of fermions when interacting) for many-electron quantum states. Crucially, these neural wavefunctions are optimized entirely by energy minimization, without training data or input physics knowledge. Input Hamiltonian and optimize with physics constraints.
AI will reduce bigotry because returns of collaboration go up with higher productivity. Complex societies with advanced economies have high return of collaboration. People choose diversity because collaboration is profitable and the opportunity cost of bigotry higher (focusing on kinship, ethnicity) The big question is, what will happen if AI reduced returns on collaboration (Coasean Singularity) and everybody becomes the CEO of a one man company? AI will actually increase returns of collaboration because higher productivity drives returns of collaboration. The more productive people are (they make more with same input of time, energy), the more they benefit from collaboration.
Discussing our essay ‘CRISPR: Much Ado About Very Little’. Gene editing started the same year as AI. Today AI is a trillion dollar industry while Crispr is still not much more than a science project. Why? Because the government has turned life sciences into a trust fund. Science must serve the public and not the other way around. Two things we want to see to invest in gene editing again; 1/2 Image net style data for life sciences to capture complex relationships (similar to LLMs and language) and 2/2 Real entrepreneurs enter the field, not academics with side jobs.
Book discussion “Buckeye” by Patrick Ryan. Coming of age novel about US between 1940 and 1980. Showing critical social and political developments through stories of human characters. Powerful. 1/3 The best history books are novels. Transcend time and location with human narratives. 2/3 Role of women evolves. Some things don’t change. Responsibility, accountability, trust and love. Women post WWII went through transition from tradition to liberation. Challenging. 3/3 Betrayal and lies. Even if you are lied to, forgiveness is best way forward. Actions in the book are not as crucial. It’s the way characters feel about what’s happening. Interesting literary technique.
Humanoid supply chain waiting for Tesla to deliver form factor and intelligence. Forcing function for scalable technology requires large, credible player to commit to size. In the case of humanoids it’s new supply chain. Component supplies must wait for the right mix of components, chips, networking solution and intelligence. Intelligence dictates sensor suite and sensor suite determines intelligence. Tesla best positioned because they can iterate around use cases like factory, space AI data centers and Cybercab maintenance.
Waymo is competing with Tesla the same way Yahoo competed with Google. Driverless taxi on digital rails versus Tesla’s general intelligence. Short-term gain, longterm pain. Tesla scaling will dwarf competitors. Waymo and Uber are competing with price in markets where Tesla is present. Waymo advertising legroom and other perks. Predictable competitive reaction towards Tesla's lower cost and higher value offering. Tesla is disrupting transportation form both sides, lower cost and higher comfort. Incumbents like Waymo and Uber will suffer psychological breakdown since they can’t react. Only way to compete is to replicate supply chain.Chinese OEMs are real competitors because they are replicating Tesla supply chain. We expect 130 Billion Dollar revenue by 2030 for Tesla.
We live in a tech bubble and the money is being made elsewhere. That’s exhausting. Lots of wealth has been generated in the past nine months. And it’s not in tech. AI is driving everything but the money is being made elsewhere. Copper mining stocks have tripled in the past nine months. Memory makers, steel makers, turbine blades makers and many other basic industries are thriving while we are talking about Space AI and Cybercabs. It’s exhausting. What do we do? Stick with our mission. Use Elon’s way; stay on path and aggressively error correct.
The purpose of science is technological deterrence and survival of species. In war and in morals. We are the sorcerer and we must control the brooms. Give AI good goals so it can do science for us. Three questions. 1/3 What is the purpose of science? 2/3 How can we make sure AI follows human states purpose of science? 3/3 How do we actually implement that? What is the best way to organize society in light of this massive industrial revolution? Organic. Error correct. The purpose of science is to enable strong technological deterrence and morals. Survival of society. Protect against external and internal enemies. Moral infrastructure to handle technological power. Goal driven AI is the most promising architecture. ‘Era of experience’. Preserve fragile equilibrium of liberty and stability. Ownership of science important. Responsibility to pursue overarching goals. What is a good outcome? What is success in science? Today, science goals are not well defined. Leads to inefficiency and corruption. Good goals (technological deterrence). Bad goals (climate change). Goals must have measurable verification and accounatibclty. Open models vs. closed models. There is no such thing as independent, open model. All AI is biased. Pursue truth and error correct.
Space AI will lower cost of intelligence by 100x or more. Today AI is power constrained. Not just power, but also political. Nimby and geopolitics are making terrestrial energy production ever more expensive. Space AI will shift focus from token/Watt to token/kg. New industry will arise and lower cost of intelligence by more than 100x. Demand for intelligence is infinite and with massively lower prices it will grow exponentially. Space X is space AI company. Tesla is chip and terrestrial intelligence company. Engineering problems to be solved; Starship reusability and lower cost per kg to orbit. Lower weight chips and optimizing for compute/kg. Lower weight solar cells with higher efficiency. Low cost and low weight radiation. Optical networks for compute and communication.