"Not Even Wrong" Podcast
Investing in fundamentally new concepts and engineering practices with large impact.
Paper discussion ‘Deep Networks for self supervised RL’. Turn reinforcement learning into a classification problem. During training look at what successful end states agents reach (exit maze, win game). Those are goals. Then look at how intermediate steps relate to end states (is this state in the same class as the end state, does it correlate, does it compare?). Understanding of intelligence. For example, if you want to become a successful hedge fund manger, what should you look at? Next year’s returns, the size of your office or longevity and what managers do who have survived 30 years in the business? I bet on the latter. Similarly, this paper argues that comparing current states with end states is a powerful way to deal with RL. This isn’t even real RL because there is no reward function. It’s self supervised deep RL without the need for had crafted reward functions.
Reaction to Jensen Huang’s comments on autonomous driving and Alpamayo. Give OEMs the illusion of self driving. Android business model. That’s ok. But Jensen should have been clear that Nvidia is doing this because nobody other than Tesla has done it right. Also, he should have said that OEMs need to up their game to succeed in autonomous driving. This is not a side project. OEMs must become AI first companies. Jensen is pushing against a rope and falsely bragging bout being the ‘first ever autonomous driving reasoning model’. If you solve for a task, you will lose. If you solve for a goal you will win. Nvidia is selling autonomous driving. Tesla is solving for low cost, high safety and high comfort transport (Cybercab, FSD, Grok). Honesty and humility pay in the long run.
Eindrücke “Der Zauberberg” von Thomas Mann. Teil 2. Das komplizierte Verhältnis von Mensch und Zeit. Inspiriert von Einstein’s Relativität. Künstler interpretieren neue Weltsicht. Was ist heute neu? 1/3 Künstlichen Intelligenz - ist Intelligenz ein Anreihen von Matrix Multiplikationen? 2/3 Higgs Boson. Ein Teilchen, das durch die Interaktion mit dem Higgs Feld anderen Teilchen Masse gibt. 3/3 Ausdehnung des Universums.
Eindrücke “Der Zauberberg” von Thomas Mann. Teil 1. Der Gast, der nie richtig ankommt. Und doch ist der da. Warum? Was heisst es, krank zu sein. Eine Ehre? Gegenteil? Geist und Körper sind Eins. Wenn ein Teil davon nicht funktioniert, gehts nicht mehr. Nietzsche. Das Sanatrium spiegelt die starre Weltordnung. Settembrini ist ein frischer Wind, der die starre Weltordnung gewandt in Frage stellt. Doch viele seiner Ideen haben fragwürdige Konsequenzen. (Sozialismus, Freiheit, Fairness). Es ist eben schwer, eine humanistische Gesellschaft aufzubauen. Die Gesellschaft befindet sich in einem unangenehmen Schneidepunkt. Man hat Gott aufgegeben. Man befindet sich 'Jenseits von Gut und Böse'. Wie soll man sich den benehmen? Hans Castrop will nicht dazu gehören. Doch er tut es trotzdem.
Paper discussion FoundationMotion. Learning from video. Automate video segmentation and semantic reasoning. Help robots learn from video by understanding spacial positioning and reasons for why and how moving in space. Spacial positioning. Segment. Spacial positioning. LLM and VLM semantics. Low cost. High throughput auto labeling of video. Internet scale robot learning from video. Milestone for video generation for robot learning.
Book discussion “A Man Called Ove” by Frederik Backman. Satire. Tragic and funny. True comedy. 1/3 A man needs something to fight for and somebody to love. Like Hemingway ‘For Whom the Bell Tolls’. 2/3 Absurdity reveals humanity. Beckett. 3/4 It’s what you do, not what you say. Existentialism. You can be kind and not polite. 4/4 Woke undertone of the book spoils the fun. Like a Netflix show.
Wealth creation is constrained by the ability to pursue interesting goals. Agency, not intelligence. Intelligence is predicated on access to energy. But growth is predicated on pursuing and executing against interesting goals. Inspired by Richard Sutton. Agents evolve by calibrating themselves against nature and succeed pursuing interesting goals. What is an interesting goal? Projects that leverage compute to generate value for society and allow the company to capture value. Intelligence is important but not sufficient to make money. It’s agency. Same applies to individuals. Example: Space X and Tesla build AI datacenter is space. Akin to Tai Pans conquering Asia. Profit motif drives progress.
Five most favorite books read in 2025.
1/5 “Angle of Repose” by Wallace Stegner. Breaching the trust of love is the ultimate Angle of Repose.
2/5 “For Whom the Bell Tolls” by Earnest Hemingway. Life is when you fight and love.
3/5 “The Noise of Time” by Julian Barnes. Integrity is like virginity. Once lost, you can’t recoup it.
4/5 “Der Mann ohne Eigenschaften” by Robert Musil. “We do not have too much science nor too little soul, but not enough science in matters of the soul”.
5/5 “Ralph Waldo Emerson. A mind on fire” by Robert D. Richardson. Intellectual declaration of independence.
❤️ What matters for next year. ’If you’re serious about software you build your own computer.’ Jensen Huang. Compatibility, Composability. Co-design. Flexibility. Speed. Robustness. All these KPIs cannot be optimized unless you control the stack. The interesting thing is that the definition of computer is changing. Datacenter Nvidia. Car, Datacenter, Robot, AI in Space, Tesla. Full stack iteration from chips, compute to end user. AI starts with the training compute and ends with the end user actually using the intelligence. No more distinction. Software + Hardware = AI.
Financifiaction of the economy and why the Federal Reserve is the big problem of our time. When the tail wags the dog. Finance is the tail and engineering and innovation are the dog. There are ten finance people per engineer. It should be the other way around. Money is a technology to allocate labor. The Fed is distorting the economy and is inherently unfair. Finance, consulting and politics are all collecting ‘Universal High Income’ as part of the Fed gravy train. Trump must rein in. ‘Drain the Swamp’ hasn’t happened yet. Stocks like Goldman Sachs and JP Morgan outperformed many deep tech companies. Graduates can make more money on Wall Street than working for deep tech. That’s a shame. High deficits and government debt are on Trump. We must rein in on the Fed gravy train. One solution is to de-dollarize and adopt Bitcoin.
Discussing Presentation by Sergey Levin on Vision Language Action models and dexterous Robotic foundation models. Take an LLM with semantic understaffing. Add vision encoder and enhance it with action encoder. Cross embodiment training. RT-X kicked off excitement for robotic foundation models. Post training is telling the model how to do the task. Pre-training gives it the knowledge of what is possible. Key is to train the model for tasks it hasn’t see before. That’s where pure imitation fails. RL can help a robot do a task by itself and learn from its own action. “The job of the foundation model is to learn from the RL based task specific actors and generalize their learnings more broadly.’In this paper, DSRL is diffusion steering RL, where the authors pre-retrain a small actor-critic to structure the noise in such a way that the VLA has better chance at choosing optimal action. Kind of like a fighter titling to the left expecting a blow from the opponent. Robots will scale with applications that don’t even exist today such as building AI data centers in space.
Discussing essay ‘Keep it Honest’. On why market hypes are normal and how to navigate them. Keeping it honest means tethering big dreams to reality: Begin with scientific proof-of-concept, advance to engineered products, then scale affordably using Wright's and Moore’s law. This framework helps separate generational winners from fleeting hype.
Book discussion “Tai Pan” by James Clavell. Part 4. 1/3 Exploration. Wealth is ultimately constrained by the ability to pursue interesting goals. Exploration has always been a catalyst for knowledge and wealth creation. Same today with space. 2/3 Clash between West and China. Epitomized in the relationship between the Tai Pan and his Chinese mistress MeiMei. She is like water, ‘slip through fingers and holds up a ship’. 3/3 The importance of a strong men. Dirk Struan is the Tai Pan. He drives society forward. Deep tech entrepreneurs are the Tai Pans of today. Elon Musk, Jeff Bezos, Jensen Huang, Steve Jobs.
Yuke Zhu, GEAR. On general robot models. From his PhD thesis. Vision action loop. Vision generates actions, generates vision. Repeat. Solution: Dissect vision-action in three levels: 1/3 Low level. See and react. 2/3 Second level. Plan. ‘Find this apple’. 3/3 Third level. Longterm planning. ‘Keep the solar array functioning’. Learning from video data: 1/2 See video and dissect the video into semantic subgoals ‘pick up cup’, ‘avoid obstacle’. 2/2 Take those subgoals and execute visual motor skills. Digital Cousin. Start with a setup of real robot data. Then use gen AI to build scenarios in virtu. Key idea is to mix real data with synthetic and make sure the real data has enough representation. Learning from World models: DreamGen Paper. In this talk Zhu says that eventually the data pyramid should be flipped (base is human video, middle is synthetic and top is real world robot data). Flip the data pyramid. Real world data will be largest amount of data. Hinges on how can we solve the bootstrap problem by creating an RL data flywheel. Dexterity. In this paper, the author decomposes dexterity into a multiple of waypoint contact problems. Then uses RL to train the robot to manipulate an object while observing waypoints and object symmetry. Space AI is where robotics meets purpose. Will drive innovation because it’s a clearly defined goal with profitable outcomes.
Book discussion Tai Pan by James Clavell, Part 3. Exploration is deeply engrained in nature. China, Hong Kong. Symbols of exploration. Space X. AI in space is new kind of exploration. Humans need exploration and the lure of financial profits to advance. Space offers this opportunity. Tai Pan is risk taking entrepreneur. Deep tech entrepreneur is the modern Tai Pan. Progress comes only with exploration because new problems must be solved. In Tai Pan they have to learn Chinese, sailing (steamship invention), learn how to deal with Malaria, plant tea outside China. In today’s space exploration we are iterating around new forms of transport (Starship), new forms of labor (Optimus), new forms of AI (space based inference). Ultimately a society is constrained by its ability to pursue new, interesting goals.
It’s the money. Stupid. We have a commercial use case for space and Optimus. It’s AI in space. Inspired by Joe Bhakdi’s comments on the economics of Space AI. Humans explored earth not because they’re curios. It’s because there was money to be made. Same applies to space. AI data centers for training and inference are a great commercial use of Space X transportation, Tesla chips and Optimus. Robots will build and maintain gigawatt scale damtcenters. Tie dollar to Kilowatts.
Neurips 2025 Paper discussion. Model based RL that natively seeks uncertainty. Pieter Abbeel. How can model based RL be proactively explorative? Optimize for trajectories where epistemic uncertainty is high (more data might help improve predictions). Good for sparse reward problems (Robot must plug in to charge itself).
Book discussion Tai Pan by James Clavell, Part 2. Scene when the Tai Pan intercepts ship to fetch mail and has information about markets before everybody else . Value of information. We are modern Tai Pans. We trade on information. But instead of instant market news, we focus on trajectory about technology. By analyzing current trends in science and technology we aim to predict where wealth will be created in the near future.
The curse of performance and the importance of measurement. How to measure a hedge fund. Performance? Yes, but performance can mislead. Key aspect to consider is time. Match time with technology trajectory of investments. Tesla has underperformed less innovative stocks such as JP Morgan in the past five years. Does that mean it’s a bad investment? No. The arch of Tesla’s technology projects projects much higher returns.
Neurips 2025 Paper discussion. Model based RL that natively seeks uncertainty. Seek trajectories with high epistemic uncertainty. That is, if I had more data, I’d be more certain. Optimize for choosing those trajectories. Good for sparse reward situations (robot must plug itself in in a room it has never seen before). Greedy strategies are suboptimal in sparse reward situations. Look for counter factuals that improve the model.
Neurips 2025 paper discussion. Whole-Body Conditioned Egocentric Video Prediction. Jitendra Malik. Using a transformer based, toekinezed model to predict video trajectories based on what an ego centric agents sees and how it’s limbs relate to each other. Instead of words the grammar is geometry and angels. Ego centric video prediction. Inverse; can you predict actions based on video sequence? Vision serves as a natural signal for long-term planning . In order to understand the world we need to move in it. Vision is not passive. It involves prediction. To address these challenges, we develop a novel approach PEVA that combines several key innovations. First, structured action representation that preserves both global body dynamics and local joint movements. Second, novel architecture based on conditional diffusion transformers that can effectively model the complex, nonlinear relationship between body movements and visual outcomes. Third, large-scale dataset of synchronized egocentric video and motion capture data. PEVA incorporates vision, action and prediction of trajectory. (Von Helmholtz)
Neurips 2025. Sergey Levine. scalability of offline reinforcement learning (RL). Horizon Reduction. Three contributions. First, we empirically demonstrate that many standard offline RL algorithms scale poorly on complex, long-horizon tasks. Bias. Estimation error. Second, we identify the horizon as a main obstacle to RL scaling, and empirically show that horizon reduction techniques can effectively address this challenge. Third, we propose a simple method, SHARSA. Problems that require long horizons are not good for offline RL. Others, where short horizons work are good. Inverse. Mitigate with chain of thought. Break down problem, i.e horizon into shorter periods. ‘This is akin to how chain-of-thought reasoning improves LLMs, which shows that decomposing a problem into multiple simpler subtasks is more effective than producing an answer directly.’
Book discussion Tai Pan by James Clavell. Scene where Dirk Struen challenges his son after he crossed him on the land sale issue. ‘It’s who you cross which defines the respect you get’. Clavell’s novel describes China very well. In order to achieve great things you have to stand up to great men. Investing in people. Musk. Jensen. China is a circle, not a cross. No clearly defined lines. Face means doing the right thing in a complex web of trajectories.
Neurips 2025 Paper discussion. DataRater: Meta-Learned Dataset Curation. Solving the problem of which data should we use for training? Better data - better models. David Silver et. al. Evaluate data used to train and classify what is useful. DataRater turns the vague question → “Which data is good for my goal?” into an exact optimization problem that it solves with calculus (meta-gradients). Key problem is the cutoff. How do you know when to cut off the meta learning process? Keep all data, but weight it lower.
Neurips 2025. World Models for robot learning. Embodied AI. World models. Science. Technology. Entrepreneurship. What is a world model? From Nvidia’s World Model paper: “Our goal is to train a World Foundation Model (WFM) that can serve as a general-purpose physical world simulator and a versatile visual world encoder, capable of modeling the physical dynamics of diverse scenes and objects.”Counterfactuals for learning. Controversy. Video data curation. Choose interesting clips, where counterfactuals have consequences. Emulator saved Nvidia. Emulator is predecessor of World Model. Emulate real world. Reality is best ground truth we have to learn how to adapt. Intelligence is the computational part of learning how to adapt to reality. If we can emulate reality in world models, we can learn faster, lower cost. Not tasks but goals.
Neurips 2025 Our focus is on world models, RL for real world AI. Workshops. 1/3 Embodied World Models for Decision Making. Topics: Model-based RL and long-horizon planning. Aligning simulation and physics for robot learning. Interactive scene generation and downstream tasks. Video-language-action (VLA) models and leveraging the world knowledge encoded in LLMs. 2/3 Unifying Representations in Neural Models. Findings in neuroscience and artificial intelligence reveal a shared pattern: whether in biological brains or artificial models, different learning systems tend to create similar representations when subject to similar stimuli. When are two models functionally identical despite different weights. Platonic representation of intelligence. 3/3 Reinforcement Learning Experiment and Theory. Schedule
Neurips 2025 Paper discussion. Enact, Benchmarking world models. World models for embodied agents must rely on embodied intelligence. Words alone will not do. How to train world models. Forward: generate videos from scenes and actions and Inverse: predict actions based on video sequence. Q&A to tesl whether world model can do forward and inverse well. Agents must learn from counterfactuals. Sanctioning mechanism deciding which counterfactual is better.
Wealth creation is not constrained by energy nor compute. It’s by setting good goals. AI works best for problems that can be optimized with clear cut evaluation. A good goal is when an objective can be reached through rapid iteration and clear cut evaluation, so that rewards can be attributed to progress. Good goals are also subject to human evergreens such as lower cost (monetary, social) , more selection, more comfort. That’s what entrepreneurs do. We also need sanctioning of bad goals. That’s what capital markets do. Finding such goals is the limiting factor of economic progress. Wealth creation is increasingly determined by leveraging compute. Bitter Lesson Everything. Wealth is defined by the amount of possible digital and physical transformations. Deep Tech is solving difficult problems that exponentially increase wealth. Bitcoin, LLMs, Starship, Self Driving cars. Example, Musk’s recent challenge to Grok team to play League Of Legends from video. This is a good problem to solve because it has clear evaluation and serves a much broader purpose (robot learning form video, decision making under uncertainty).
Essay discussion: Nvidia has Risk. But it’s not what you think. The biggest threat is bad US policy restriction to sell into China. Demand will continue to explode because the industry is moving from 1D data processing to 4D (3D + time), which requires orders of magnitude more compute. Nvidia is betting the firm on optimizing this particular workload. The US has never been good at controlling and hiding. We are best in running faster than everybody else. Compete. Don’t hide.
Essay discussion. Grok 5 video game challenge is not just a gaming stunt. It paves the way for real world AI learning at scale. Musk challenges Grok 5 team to beat human gamers in League of Legends. Learn from vision only. Execute like a human by reasoning under uncertainty and with incomplete information at human speed. Transfers to Optimus and Tesla. Here is why we think Tesla should have a stake in x.ai. In order to succeed on robotics you must have 1/3 Dexterity 2/3 AI and 3/3 scale. X.ai brings the brain and Tesla the body.
Unser Aufsatz “Von Kakania nach Brüssel: Musil und die Gefahren einer elitären Verkrustung”. Kritiker sind gerne bereit, den aufkeimenden Rechtspopulismus, den Musil beschreibt, mit der heutigen Rechtstendenz in Europa zu vergleichen. Doch sie unterlassen den anderen Aspekt, den Musil kritisiert und das ist die Verkrustung der Regierungselite im spätkaiserlichen. Vakuum. Desillusionierte Jugend. Rechtspopulismus. “Fruchtbare politische Debatten entstehen erst, wenn beide Seiten zur Selbstreflexion bereit sind. Genau dazu lädt Musils Roman ein: Er bietet nicht nur eine treffende Analyse der Moderne, sondern auch ein zeitloses Bild einer offenen Gesellschaft, deren größte Bedrohung darin liegt, den Dialog mit der Jugend zu verlieren.”
Book discussion “Der Mann ohne Eigenschaften” by Robert Musil. Where do we stand today? Who is ‘Der Mann ohne Eigenschaften’ 2.0? Fluid identities. No commitment. Alibi issues like Black Lives Matter or Save the Planet are similar to the Parallel Aktion, in so far as they feed themselves without real content. Moral high-ground is easy when there are no morals. Excessive rationalism is winning because it’s effective. Humanities must stop serving themselves and focus on humans in the age of AI. There is a soul, something beyond rational behavior. It must be discovered like everything else. “We do not have too much science nor too little soul, but not enough science in matters of the soul”. STEM eduction dominates academic discourse because the humanities have let us down. STEM cannot be hacked.
Buchdiskussion “Der Mann ohne Eigenschaften” von Robert Musil. 1. Auflösung von fixer Realität und Bedeutung. Der Mann ohne Eigenschaften ist hochbegabt im Nichtstun. Anwendung ist harsch und verfällt dem Sarkasmus. 2. Ratio kolonisiert die Seele. Doch die Seele schlägt zurück; "Wir werden vom Gewissenszwang des Verstandes zur Gewissenlosigkeit des Gemüts gezwungen.” 3. Möglichkeiten über Tatsachen. Potential braucht die Tatsache und umgekehrt. 4. Es gibt eine Alternative zum Ratio. “Die Menschheit weiss vom ungenauen Ganzen kaum mehr als vor 2000 Jahren.” 5. Gutes tun. “Das Meiste am menschlichem Tun ist nutzlos. Es wäre viel nützlicher, wenn wir nur dann etwas tun würden, wenn und das Gemüt und die Emotion dazu zwingt.” (Drang). 6. Kakanien. Erstarren des Gemeinwesens. “Sind wir auf dem Weg zu einem Ameisenstaat oder einer anderen, des Menschen unwürdigen Aufteilen von Leistung.” Alibihandlungen wie Black Lives Matter, sind wie …”Die Parallelaktion ist ein Schaffen ohne Inhalt, das in sich selbst wächst.”
Eindrücke “Der Mann ohne Eigenschaften” von Robert Musil. Richtungsbilder sind ewige Wahrheiten, die nicht ewig wahr sind. In der Moderne formt sich ein Zeitgeist jenseits von Gut und Böse. Alles geht, Frieden ist Krieg und umgekehrt. Das beste für die Rüstungsindustrie ist ein ewiger Frieden, für den man sich aufrüsten muss. Dieses Paradox charakterisiert die Moderne und leitet die Postmoderne ein, wo Relativismus alles auf den Kopf stellt. Richtungsbilder sind Wegweiser im moralischen Labyrinth der Postmoderne.
Eindrücke “Der Mann ohne Eigenschaften” von Robert Musil. Die Grenzen der Gesellschaft. Wer gehört ins Irrenhaus? Warum sind die Irren nur aggressiv gegen Normale? Ist Geschwisterliebe erlaubt? Warum hat die Gesellschaft Schranken? Hat der Mann ohne Eigenschaften trotzdem Eigenschaften? Er haltet sich an Regeln, gesellschaftliche Regeln, die eine Art Eigenschaft darstellen. Schranken müssen vorhanden sein. Freiheit muss eingeschränkt sein, um frei zu bleiben. China? Moralischer Abgrund des Fortschritts willen.
Tesla stock has been pedestrian in the past five years. Now it’s time to shine. FSD and Robotaxi are catalysts. FSD is the first software as a service distributed through real world robots. Scaling is key for the company to produce cash flows and contribute to fundamental re-valuation of the stock. Investors have been patiently waiting for the technology to mature. It’s here. Now the company must commercialize through FSD adoption, car sales, Robotaxi and Cybercab. FSD is a generational technology. Competition. Customer experience, cost and reliability.
Nvidia earnings. Market jitters about AI infrastructure investments. Nvidia lowest cost per Watt, lowest cost per token. Versatility. Nvidia is building global token factory. China business dead. Singapore driving sales to China. Everybody wants latest racks. Scaling laws. What we need now is goals to purse which will drive inference. Dichotomy in market between business and market perception. Open AI sparked the rally and now it’s the cancer threatening it. It’s like a boat with inadequate leadership riding a wave it cannot ride. Market not buying self fulling prophecy. Nand flash memory makers Sandisk and Micron down today. Fears of oversupply. Most of AI that is currently being deployed is recommender systems which are being upgraded to generative AI. Data centric approach to AI will drive need for versatility which is Nvidia’s strength. AI investments and usage is broad. Global. Good.
LLMs live in Plato’s cave. Sergey Levine on LLMs. LLMs have learned how to mimic cognitive tasks from observing humans. They live in Plato’s cave and the internet is shining the light. LLMs compress cognitive abilities of humans via learning from internet. The internet is a compressed version of human cognitive knowledge. LLMs mimic human cognitive abilities. Problem. LLMs cannot figure out their own problems. Robinson Crusoe robot finds its own problems to solve.
World models for robotics learning. Overview. Four steps: 1. Superficial reproduction of world in video. 2. Navigation through video. 3. Physics incorporation and planning. Move through video in realistic way and have environment react. 4. Stochastic scenarios. Train on edge cases. If I can’t predict all edge cases, train on all of them. Self play through world model with goal oriented agent levering compute. Bitter lesson in robotics requires some sort of in virtu learning. Learning from video. Connect visuals with actuators.
Mach 7 Physics. From lab to real world engines. Laser diagnostics. Chemistry and physics of supersonic, hypersonic, and space re-entry systems, including scramjet combustors and ignition under extreme conditions. Laser diagnostics feeding ground truth data back into simulation. Data avalanche. Working towards fully digital in virtu design of complex propulsion technology. Lab focus on air breathing propulsion, Drone and electric propulsion. Advanced diagnostics and modeling.
Discussing essay “Less is Never More in China”. Shallow capital markets make China's AI efficient - not dominant. China’s AI strategy revolves around building more efficient models. This isn’t because Chinese founders are uniquely obsessed with efficiency, it’s because they don’t have the money. Tesla is a Chinese company (obsessed with efficiency) with global access to customers and capital.
Eindrücke von “Der Mann ohne Eigenschaften” von Robert Musil. Moral. Ein Gefühlszustand, der alle Gefühle vereinheitlicht. Gefühlsquelle, aus der man schöpft, doch die es eigentlich nicht gibt. ‘Eines Tages wacht der Mensch aus dem Rausch aus und stellt fest, dass es Sünde gibt.’ Warum soll man sich einschränken? Moral ist ein Widerspruch, zumal es die Vereinheitlichung aller Gefühle ist, die man nicht versteht. Nicht so wie ein Tier leben. Eigenen Taten überdenken und verbessern. Doch was heisst, ‘verbessern’?. Frustriert, weil Widerspruch. Spornt zur Verbesserung an. Antrieb von Wissenschaft, weil Drang zum Wissen ist Drang zur Erklärung der Moral.
China Trip. Part 4. Dynamics between large model providers and smaller vertical players. BABA, Tencent, Bytedance, BIDU are selling tokens. Smaller players are using tokens to provide services to verticals such as healthcare, finance or transport. Unisound is such a company. Smart home, transport and healthcare. Smaller, proprietary model. AI native. Started as a voice technology company in the 2010s, pivoted to BERT and now using own multi modal model. Interesting dynamics. Same in US. Stack: Energy, chips, models, verticals. Money is made in energy and chips. Investment in all parts of the stack. Can smaller players attract talent? Can they keep up with model development? China discount at least 10X. Why? Chinese entrepreneurs not ready for big leaps. More myopic.
World Labs launches Marble. Generating real world simulation in virtu for robots to train on counterfactuals. World models are one of three essential ingredients for real world AI. 1. Planning algorithms such as Q-learning. 2. Evaluation function (what is a good action? What gets high values and what not?) 3. World Model. World Labs attempts to solve the World Model problem. Real world simulation in virtu. Two key factors. One is native 3D-ness of data representation. As FeiFei Li says in there essay: “ LLMs tokenize in 1D. World Lab uses native 3D representation of data. “ 3D representation allows for more affordances. Not just what can be seen, but what action is possible? Interview. World Labs can train on 2D data and create native 3D representation of scene due to Strong mathematical connection between 2D and 3D. Enable counterfactuals for robot to learn in virtual space.
Google vs. Nvidia chip strategy example of wealth creation in tech. Google developed TPU and transformer to take advantage of matrix multiply. Nvidia garnered most of the gains. Why? Because scale is what matters. Google knew what, but Nvidia knows how to do what’s necessary to scale. Scale engineering and manufacturing with global reach of software ecosystem are the blueprint for wealth creation in tech. Tesla. Native Chinese market presence with scale manufacturing and global software reach. This combo is unique and unfair advantage of Tesla. Like Nvidia.
China Trip. Part 3. Tesla’s position in the global autonomous EV market. The God, the Better and the Ugly. 1. The God: China’s avenues are driverless and noiseless. Future. Tesla is player in this highly competitive market. Supply chain, fighting for consumers. 2. The Better: Tesla is the only company with native Chinese market presence and global software and AI reach. Chinese market presence with global brand and general software suite is unique to Tesla. 3. The Ugly: Chinese overcapacity in EV’s and hard core focus on second mover copy prowess will be a problem and must be weathered.
China Trip. Part 2. Technology and modern lifestyle of China. How do young, modern tech people live in China? One big takeaway; the Chinese entrepreneurial class is still caught in myopic behavior. Make money fast. Lack of vision for longterm disruption. Low tolerance for disruptive innovation. Copy and cut cost. Clutter culture. Little patience for slick, Western influenced minimalist tech culture. Opportunity for Tesla. Opportunity for Chinese tech entrepreneurs. No desire to reinvent things from scratch and/or go back to first principles and reinvent a space. Every time I visit China I learn a lot about myself and the society I live in. China = copy hard. US = invent hard. China = 1 to n. US = zero to one. ‘Would you hire a young person that wants to unsettle what you built?’ If you are CTO of Baidu, would you hire a person that tells you to get rid of all sensors but cameras?
China Trip. Part 1. Focus of China trip was to learn about Chinese tech culture. visit companies, talk to people, researchers. See how Chinese tech scene works, lives and plays. Key take aways. 1 Manufacturing. Large capacity for cars. 2. AI for cars and robots, mostly old school. Some players like Xiomi, Xpeng copying Tesla FSD. Looks and feels similar. 3. China tech is advanced but not threatening US. Talent in China doesn’t want to work in China. Cultural centers like the West Bund in Shanghai are harbingers of things to come. An emancipated native Chinese tech culture. It’s hard to live and compete in China. 4. Car market is segmented into China clutter and slick Western design. Most local competitors compete on the clutter dimension, few, like Xiaomi, try the Western slick. We saw restaurants that are succeeding in Neo Chinese cuisine and look. Same could happen in tech. Chinese apps are full of clutter. U/I not well developed. China is a place of better future for kids, as opposed to Western societies which are caught in suicidal culture aberration.