"Not Even Wrong" Podcast
Investing in fundamentally new concepts and engineering practices with large impact.
Video to robot action models. Paper discussion. Pre-training a model on desktop video-action pairs to learn general sensorimotor primitives.Using Generalist-IDM (inverse dynamic model, i.e. learn actions from observations) to scale the dataset with pseudo-labeled internet videos. Abstract desktop actions to robotic control commands. (Mouse forward is abstracted away to forward and then the robot figures out how to go forward). Fine-tuning on robot-specific data to adapt to physical environments.Eventually, robotics will be trained on generative AI. Same concept like video to action, but video scenarios are generated by AI.
Bitter Lesson Everything. Discussing essay. What works for AI works for everybody else. Develop general methods that leverage compute. Whatever you do, think about how compute will change your world. Whether it’s in investing or in your own career. Example. Simulation driven by generative AI. In virtue product design, testing and development. Optimus leverages compute for labor. Omniverse is bitter lesson pilled endeavor by Nvidia to flip manufacturing on its head.
Bitter Lesson for careers. What works for AI will work for companies and careers. Build a general skill set that leverages compute. Example. Physics simulation. Scales with compute, gets better and eventually replaces real world prototyping, tinkering and traditional engineering. The shop goes virtual. Manufacturing will be like chips, design in virtu, tape out and production final step. What will not change when compute scales? Model architecture design, benchmark design, evaluation, data generation, data management and processing. In virtu manufacturing is turning the manufacturing process on its head. First generate data that trains AI to design the product. Then tape out and manufacturing.
❤️ The Bitter Lesson for Investing. Discussing our post on X. ‘Companies that focus on general methods that leverage computation are ultimately the most valuable.’ Wealth creation is predicated on utilizing computation effectively. Wealth creation is predicated on building processes that enable products and services that get better and create more differentiation the more computation is applied. Examples: FSD leverages compute for transportation. Optimus leverages compute for labor and colonizing Mars. Video generation leverages compute to accelerate robotics simulation. Inspired by essay. ‘Bitter lessons building AI products’.
Market commentary. Trump threatens trade war with China. Stupid. Just diverting attention from his home made mess. The problem with US is not China. It’s ourselves. Corruption, large deficits and activist judges undermine the fabric of our society. That’s what Trump should be tackling instead of picking fights with Xi. China worships money and math and disregards knowledge. US worships knowledge through free thought and undermines math and money. Fix that and leave Xi to his own devices. Cut access for Chinese students to US universities and they will do whatever it takes to get it back.
Eindrücke “Der Mann ohne Eigenschaften” von Robert Musil. Warum machen wir nicht aktiv Geschichte? Wir haben es in der eigenen Hand. Doch die meisten wählen Passivität. Aktive Passivität ist der Zustand des ‘Manns ohne Eigenschaften’ und auch der Moderne. Die Frage ist nicht wer du bist, sonder wo bist du. Multiverse. Wo bin ich? Nietzsche Jahr. Ulrich Jahr. Ich Jahr. Jedermann könnte grosses leisten, oder als ‘grosses geleistet’ betrachtet werden. Kubismus. Einstein. Raum-Zeit. Quanten. Alles was möglich ist, wird passieren. Emerson. Self Reliance. Nimm das Leben in die eigene Hand. Sonst ist es nicht Leben.
Eindrücke “Der Mann ohne Eigenschaften” von Robert Musil. Arnheim schreibt. Er will verstehen. Lehnt sich aber gegen alles was seinem Verstand fern bleibt auf. Er ist wie ein religiöser Seelenfeilscher, aber ohne Ideologie. ‘Wie das weibliche Auflehnen gegen Geld, Wissen und Rechnen’. Nicht weil man es nicht akzeptiert, sondern weil man es nicht hat/kann. Opposition gegen das nicht-Haben. China-USA Unterschied. USA setzt auf Wissen und bekämpfen Geld und Rechnen. China setzt auf Geld und Rechnen und beschneidet das Wissen, weil ohne freie Meinungsfreiheit kein Wissen zu haben ist. Arheim’s Idee des Kaufmanns manifestiert sich fünfzig Jahre später im Silicon Valley. Der spirituell durchwachsene Unternehmer, der reich wird und sich dabei seelisch bestätig und geistig wohl fühlt. Moderne, die auch heute noch sehr stark zu spüren ist.
Unifying FSD with Optimus is a big deal. Robots learn to pursue goals which are predicated on solving navigation problems. FSD is first solved robot problem at scale with real customers. Bootstrap general robotics problems. In his PhD thesis Jim Fan introduces Rubiksnet, a video learning architecture which leverages video action recognition. Unification of FSD and Optimus a step in that direction. ‘making RGB images to action sequences’. This is what FSD does. Skills can be learned implicitly while learning to pursue goals. Instead of ‘doing the dishes’ or ‘sweeping the floor’, learn to keep a room clean. Compute-time is by far the most important driving factor in developing intelligence. Bitter Lesson.
Is there an organic growth rate limit for wealth creation? In nature yes. But not with human at center. Explanatory knowledge, risk taking and innovation drive much higher growth rates. Bring down cost by 100x (like launch cost for space through Starship) or cost per mile of driving from 2$ to 0.2$ you drive growth. Compound growth rates much higher than natural rates. Risk is that the 80% of people who don’t participate directly ask for bigger slice of the pie as reflected in gold price.
Eindrücke “Der Mann ohne Eingeschifften” von Robert Musil. Es gibt kein Dazwischen im Recht. Vielleicht im Gesetz, doch nicht im Recht. Gerechtigkeit ist, den Menschen vor Gewalt zu schützen. Irre Täter sind Täter, wenn sie Schaden anrichten. Es gibt kein Dazwischen. Ulrich’s Vater ist besorgt, dass die moderne Gesellschaft sich mit der Lockerung der Justiz in eine gefährliche Richtung bewegt. Die moderne Gesellschaft hat den Absolutismus mit Justiz ersetzt. Sie muss sich daher sehr genau überlegen, wie sie in Zukunft eine gerechte Gesellschaft schaffen will. Ein Mensch hat das natürliche Recht zur Verteidigung des eigenen Körpers. Wenn es der Staat nicht macht, dann muss ich es machen. Ohne Schutz gibt es keinen Staat.
Eindrücke “Der Mann ohne Eingeschifften” von Robert Musil. Ein vollkommenes Leben bedarf weder der Kunst noch der Wissenschaft. Es gibt keine Vollkommenheit, weil sonst die Wissenschaft und Kunst ein Ende hätte. ‘The Beginning of Infinity’. Unser Geist ist die Quelle des Lebens. Das Ich bestimmt was ich lebe. Vollkommenheit ist anti-Kunst.
Eindrücke “Der Mann ohne Eigenschaften” von Robert Musil. Ulrich ist ein gefährlicher Mann. Äußerlich frei, innerlich steif. Gegenteil von anderen Menschen, die innerlich wild sind und äusserlich steif. Menschen sind Bestien, die Verstand und Geist wie einen Schleier mit sich herumtragen. Philosophie ist die Rechtfertigung, anderen zu schaden. Geist ist ein gefährliches Wort. Innerlicher Geist. gefährlich. Äußerlicher Geist, oder Geist der Beobachtung ist gut. Ulrich ist Prince Mishkin, Elon Musk. Widersetzt sich dem Drang der menschlich hierarchischen Ordnung.
Eindrücke “Der Mann ohne Eigenschaften” von Robert Musil. Worum geht es beim Wissen? Ist der Drang nach Wahrheit eine Sucht? Seit Galileo Galilei hat sich die Menschheit auf das Beobachten konzentriert und dabei auch viel Erfolg gehabt. Faust. Prometheus. Zwiespalt zwischen dem verfeinerten (Schein)Wissen der Menschheit und der Genauigkeit der Physik. Popper. Probleme lösen. Wer profitieret denn von Technik? Krieger, Jäger und Kaufleute. Was ist denn schon ein Problem? Wir schaffen Probleme, damit wir Wissenschaft betreiben können. Fortschritt und Freiheit sind das Gleiche. Fortschritt bedeutet neue Oberklassen. Freiheit ist sehr eng mit der Frage des Fortschritts verbunden.
Thin Air in AMD Open AI land. Financing structure, is absurd. But more importantly, AI is about scaling. Build the machine that trains the machine. Determining skill. AI is like the chip business. Large capex, high asset turnover and increasing capital commitments. It’s not the AI, it’s the scaling - Stupid. We are concerned that neither Open AI nor AMD are talking about challenges of scaling. Nvidia has recognized the problem Learned from Colossus 1. Musk understands scaling. It’s like in the EV FSD business. Power, voltage, interconnect, model parallelism, Jensen’s law. All these factor determine performance Scaling cannot be outsourced.
Eindrücke “Der Mann ohne Eigenschaften” von Robert Musil. Präzisionslust des Menschen ist verlockend. Eine Sucht. Doch nur in der Physik und Technik.‘Wir haben dem Ingenieur den Drang zum Genauen hinterlassen und widmen uns wieder dem Ungenauen.’ Kaum Fortschritte anderswo. Die moderne Gesellschaft zerfrisst sich an dieser Zwiespalt. Das Problem ist die menschliche Vernunft in das biegsame, ungenaue Ganze hineinzupressen. Jurisprudenz Problem. ‘Wenn wir nur dann handeln würden, wenn uns das Gemüt explosiv dazu zwingt, würden wir viel mehr ‘Besseres’ erreichen, doch ein Grossteil der Bevölkerung wäre dann Ameisenhaufen. Wirkliche Berufe sind nur für Wenige.
Pre-training under infinite compute. Paper discussion. Regularization and ensembling improves model performance on fixed data. Regularization reduces impact of individual weights that can dominate a model. Ensembling, same data, train on same model with different hyper parameters. Then average prediction. Is there a world where compute is infinite and data constrained? No. Compute creates its own data. We will never be compute, nor data constrained. We are goal constrained. Malthusian view on data.
Eindrücke “Der Mann ohne Eigenschaften” von Robert Musil. Spanning zwischen Industry und Adel. Privates Eigentum. Freie Marktwirtschaft oder Säbel-Bürokratie. Vernunft, Ratio im Gegensatz zu Stand, Gefühl und Tradition. Nationalismus. Auflehnung gegen die Vernunft mit Gefühlen aus tief menschlichem Sumpf. Es gibt keine ganze Bildung mehr, keinen Leibniz. Wissen ist ein Fluss ohne Anfang und Ende. Was kommt nach der Vernunft? Emotion, Nationalismus, Ameisenstaat? Der Drang nach Wissen ist keine Wahl, es ist eine Sucht. Sie schleicht dem Menschen nach. Menschen ändern sich nicht, trotz allen Wissens. Faust. Klassik. ‘Unsere Zeit hat das Wichtige verloren. Wenn wir einen Homer heute hätten, würden wir ihm zuhören?’ Eine Gesellschaft ohne ‘Wichtiges’ hat weder Bedarf an Helden noch an Ehre. Ist der Unternehmer besser als der Adelige? Spielt es eine Rolle, wer den Menschen schadet? ‘Nur Verbrecher schaden anderen ohne Philosophie’. Gesellschaft modert in sich hinein, wie eine Ehe, bei der man ein ‘sich Zusammensein’ vortäuscht. Demokratie ist nicht so. Österreich war so. Industrie ist heute weit besser als zu Musil’s Zeiten. Mehr Wertschöpfung, als Schaden.
Thoughts on Tesla and Nvidia after Q3. The God and the Bad. Bad; China (competition. US (lackluster adoption of EVs), Europe (market share low). What will not change? Lower cost and improved user experience drive demand. Tesla winning on both. God; AI. FSD. Robotaxi. It is one year after Cybercab unveil. Where do we stand today? FSD works. Robotaxi launched. Unsupervised FSD by year end. Deployment? Can we deploy cybercab at scale? 100,000 robotaxis by 2026, 5,600,000 by 2030 (cumulative). Deployment risk is safety, technology, compute, network effect, network optimization and demand. Moat. Nvidia. Token factories create applications which drive demand. Different than internet. Infrastructure not only supporting AI, but driving AI.
Inference and training compute on the edge. What happened to the cloud will happen to the edge (see Nvdia vs. Intel) Large AI deployment, with compute and memory infrastructure buildout for cars and robots. Tesla has unique position as vertically integrated real world AI company to dominate inference at the edge.
If you want to train an AI to write poems, don’t train it on Shakespeare. Let it write billions of poems and test their effect on real people. Another take on Richard Sutton’s idea of learning from experience and why LLMs are not bitter lesson pilled. Sutton is not questioning the paradigm of learning from data. In fact, he is even more extremely focused on learning ‘just from data’. Data in this context means the agent is influencing the environment and recursively adjusting its behavior on what it experiences. Learn from behavior - not priors.
Eindrücke von ‘Der Mann ohne Eigenschaften’ von Robert Musil. Unterbewusstsein. Momente der Annäherung an die Wirklichkeit. ‘Das sind die Momente, für die Gott die Welt geschafften hat’. Inverse Relation von Wissen und Stand zur Wirklichkeit. Je mehr man hat, desto eher neigt man, nicht zu denken, weil man es sich leisten kann. Ulrich nimmt sich frei von Ratio und Stand. Weg von Wirklichkeit und hin zum Möglichkeitsmenschen . Doch ‘der Mann ohne Eigenschaften ist Eigenschaften ohne Mann’. Stand, Reichtum hat mechanistische Wirkung einer Scheinwirklichkeit, die zur Person wird, indem sie das Ich umschleiert. Wie kann man sich dem Fluss des Lebens entziehen und doch das eigene Schicksal in der Hand behalten? Wirklichkeit ist eine ‘Equalizer’. Niemand kann sich ihr entziehen.
If you want to study at university - study physics. Physics has most reach. in the sense that it captures most complexities of nature with the most conscise explanation. It’s poetic. The curriculum not only teaches you valuable knowledge, but more importantly, trains you to calibrate knowledge against nature. Nature is the only ground truth that matters. Musk. Sutton. Reduce the risk of “cargo cult science”.
Erste Eindrücke von “Der Mann ohne Eigenschaften” von Robert Musil. Wirklichkeitsmensch. Möglichkeitsmensch. Es könnte auch anders sein. Ich bin Potential. Bildung ist Brücke des Spirituellen zur Wirklichkeit. Möglichkeit ist Potential des Menschen. Ulrich will Bedeutung. Zuerst wird er Soldat, dann Ingenieur und schließlich Mathematiker. Mathematik denkt anders als der Mensch. Beschreibt sie die Natur trotzdem bestens. Grundidee der Moderne, der westlichen Gesellschaft; Zukunft ist nicht vorbestimmt. Sie kann von uns geprägt werden. Weder Hellenisches Schicksal, noch relativistischer Unsinn. Zwischen Schranken der Wirklichkeit dehnt sich das Potential aus.
Periodic Labs AI science. Physics as ground truth. Automate experimentation and simulation. Scale trial and error. Create new knowledge by trying more at scale. Four different eras of science: Galileo: Model the world based on observation. Do experiments to test hypothesis. Science AI 1.0. Do the same but collect more data and find new patters. Science AI 2.0. Use search and generative AI to find even better answers. Science AI 3.0. Data first. Design experiments with data generation in mind. Produce data so AI can learn faster and better. Hiring expertise in AI, experimentation and simulation.
Reaction to Richard Sutton Interview on Dwarkesh Patel. Sutton is the most important intellectual of our time. His approach toward building intelligence is key to the future of humanity. ‘Intelligence is the computational part of an agent pursing goals’. Solving problems by learning from experience. Paradigm of training is wrong fork in the trajectory of intelligence. Learning from experience is better. Agent reacts to environment and changes behavior based on reward feedback. Intelligence is goal bound. Transformer not good enough. OAK architecture better direction. There are only a few successful real world AI breakthroughs. Alpha Zero, FSD, Drone racing. Agents that can set their own goals. Robinson Crusoe.
Discussing new essay : “Is demand for intelligence always growing? Yes. Intelligence is goal bound, not constrained by compute. More goals create ever more goals and drive intelligence and wealth. Compute is necessary but not sufficient. Technology is not enough (CCP’s China). The key to intelligence is to build a moral, ethical and social foundation to create more ‘Möglichkeitsmenschen’ as described by Robert Musil in ‘Der Mann ohne Eigenschaften’. Society of openness, unbound by fixed traits and rigid convictions. Karl Popper’s ‘Open Society’ is a blueprint for an intelligence seeking culture. The purpose of freedom is to pursue goals and foster intelligence.
Adaptive agents are more deterministic than deterministic models. Intelligence is about goal setting. In order to build a robust agentic platform we must view nature and experience as ground truth. Ironically, deterministic models are more brittle and prone to mistakes than evolutionary, adaptable agents that ere not deterministically architected but use nature as ground truth. The key challenge is to set goals. Ideally, agents will eventually learn how to set their own goals.
Book discussion “Tell me lies” by Carola Lovering. This is not a love story. Evolution of the individual. Good and Evil. Lies are like a rock you’re dragging along. Why is Lucy Good? Because she internalizes nature and expresses herself as part of it. Ultimate goal is love. Lucy is her life. Her life is Lucy’s kingdom. Steven is the opposite. He thinks nature is something you take advantage of. He lives despite of himself. Worst thing about Steven is, he is as conventional as it gets. Evil and conventional is unacceptable. If this novel had been written in the 19th century Russia it would be called ‘Good and Evil’. There is Right and Wrong and there is Good and Evil. Fleetwood Mac songs. Emotional bond. Good choice. Depicts college life. Decadence. No learning. Drugs. Zombie intellectuals.
Building infrastructure or a House of Cards? Nvidia ‘invests’ in customers. Why is Jensen investing in Open AI?’ The Good, the Bad and the Ugly. The Good. Nvidia is building an ecosystem based on token factories. Only Elon Musk is really pushing forward. In order to build a global infrastructure based on Nvidia Jensen needs a horse to pull the wagon. Jensen knows where this wagon is going. Must build a bridge for others to see it through. The Bad. Open AI is not set up to drive an industry. Lack of leadership, lack of capital. Culture of deceit. Research, not large infrastructure buildout. The Ugly. Open AI is building a house of cards . Pulling in has-beens such as Oracle, Microsoft. Disturbed relationship with truth. How far can this go? Is intelligence the kind of service where demand is always growing? How much GW is too much? The risk to AI is to the upside. We don’t know where this can go. But we feel it’s big. Bigger than the internet. Sputnik, China competition.
Episode, September 18 2025 III
Richard Dawkins on poetry. A critique. Language has been around much longer than speech. Poetry is highest form of human intelligence. Occam’s razor. Poetry is speech. Math is nature.
Trump is the best thing that can happen to US academia. Cut corruption. Cut reliance of science on government funded research. Let private sector determine research. Purpose of academia is to teach. Research is for private companies. Much better for society. There is no such thing as science. There are scientists.
Book discussion “Ralph Waldo Emerson. A mind on fire” by Robert D. Richardson. 1/3 Great life, great story and great poetry. Ralph Waldo Emerson’s life seen through his writings. Poetry is the highest expression of human intelligence. Occam’s razor. Language has been there before speech. Poetry is speech. Math is nature. 2/3 History. Emerson is a child of the new America. Intellectual declaration of independence. Eloquence and debate are the fundamental drivers of a new, democratic society. Declaration of independence for US but also for the individual. Independence from coercion and imitation. ‘I must find someone who can make me do what I can do’. ‘Cherish intelligence, not knowledge’. 3/3 Importance of liberty. Emerson’s poetry is about how the individual liberates themselves from coercion. Founding fathers. Emerson, Melville. ‘Self Reliance’. Independence of the mind. Habeas Corpus of the Mind. ‘Each men’s house is his castle’. ‘Each men’s mind is his world.’ Choice is fundamental. America is the magnet. Our mind determines and shapes the universe.
From Data to Experience. Two vectors drive AI 3.0. 1/2 Data. In the age of experience, data comes from experience. Experience factories produce tokens which train agents to pursue goals. 2/2 Compute time. More investment in compute than data. That will change. Companies like Surge.ai are AI 1.0 type business. AI 3.0 is the token factory which it creates from experiences (generated worlds). Many worlds interpretation of AI. RL. Reward is enough. Goals. Intelligence is Kolmogorov complex. No shortcuts to intelligence. Iterate through trajectories. Sympathy with intelligence, not knowledge. Knowledge fosters benchmark hacking, experience based agents hack the only benchmarks that matters - real world. Reward function design is hottest programming language.
We make order. We are actively influencing the cosmos. Human mind has the ability to shape the cosmos. From the small to the large. Deutsch. The Beginning of Infinity.
Knowledge subsists according to the nature of that which knows. Not of knowledge. It’s about the individual and his interpretation of observations which creates knowledge. It happens in the mind, human or AI. Truth is that which is always true.
Episode, September 16 2025 III
Discussing Ralph Waldo Emerson Biography. Part 1. ‘A mind of fire’. American Intellectual declaration of independence. Evolution, not revolution. The value of poetry. Compact representation of intelligence. The art of compression. Highest form of human intelligence. Poetry is the least Kolmogorov complex way to express concepts. ‘All important truths must be self evident.’ ‘Nothing good ever happens without enthusiasm’. ‘All value lies in the sweat and blood invested by the poet.’ ‘There is only one regret. Not the things you’ve done but the things you’ve not done’. Emerson’s view on the individual. ‘Self Reliance’. ‘The teacher is the one which can assist a child in obeying his own mind. The value of Shakespeare comes to fruition in oneself. Everybody has their own Shakespeare in their soul. The highest form we can attain to is not knowledge. It’s sympathy with intelligence. Individual expression of nature is what counts. Zero to one. Beginning of Infinity. A single human mind is the starting point of infinite possibilities because we absorb nature and shape it with our thoughts, passion and enthusiasm. The artist is the one who touches as many souls as possible. Kolmogorov definition of art. Short sequence of strings (representation of art) touching most souls (Picasso, Emerson, Goethe). There might be a universal truth out there. But what counts is your individual interpretation of truth. All happy poets are happy alike. All unhappy poets are unhappy in their own way.
Human mind shapes reality. Definition of human is the entity which changes and creates order in the universe. Quantum even in classical dimensions. Observations are theory laden and human mind creates new reality( Beginning of Infinity).
❤️Data flywheel. Business model of AI 3.0 Market undervalues Tesla but is not wrong. Just doesn't see data flywheel? Feed data back into algorithm to improve products and services. AI 1.0 Transactions create data. Collect data and figure out how to use it (Banks, legacy media, Saas). AI 2.0 Real transactions create data which directly improves the user experience. (Google, Tesla FSD). The data creates the product. AI 3.0 Data factories produce tokens which train agents how to pursue goals. AI 3.0 businesses are compute and data constrained. The successful business of AI 3.0 takes human out of production and puts human in center of product experience.
Kolmogorov Complexity, Kurt Gödel and Richard Sutton. At some point systems become so complex, we can’t formally verify and predict. Operational form of Gödel’s Incompleteness. Intelligence is such a system. Sutton’s continuous learning combined with generative AI leads to more intelligence thourgh massively parallel iteration.
Debate doesn’t foster guns but guns enable debate. Celebrating Charlie Kirk’s murder is despicable and stupid. Debate doesn’t create violence. But without guns there is no debate. Only if people have the means to stand up to power, power will take them seriously. Guns enable debate. Without guns there is only make belief democracy.
❤️ Investing as case study for ‘Reward is enough' . If wealth is the goal. What is wealth? David Deutsch: Optimize for physical transformations. Invest in companies that increase possible physical transformations. Two extremes. Tesla and Fusion energy. Tesla allows for intermediate rewards. Fusion doesn’t. Sparse reward system might work but is difficult to make money with. Tesla's intermediate rewards allow for more intelligence (weighted by Kolmogorov complexity) and less risk with better upside.
Many worlds interpretation of AI enables induction as driving force of intelligence. Iterate through many different worlds and learn from induction what action to predict. ‘Turkey won’t be fooled’. Models will make right predictions from an evolutionary perspective. But what about explanatory power? Can the turkey explain why being fed well is not a good idea? Universal representation of world models. Universal Unsupervised Environment Design (UUED).
AI scaling law 2X in X years? Elon Musk says that with a 10X increase in compute you double intelligence. So 10X compute = 2X intelligence. That means intelligence is a function of compute and hence power. ❤️Building power infrastructure at scale and in AI time (i.e. rapid deployment) is going to become the key competitive advantage of companies and countries. Tesla. Nvidia. Oracle? Can legacy companies deploy at scale and speed? Electricity used to be adjacent to products and services. Now electricity is the product.
Two important comments about AI and compute. 1/2 Musk says that a 10x increase in compute doubles intelligence. Statement implies you can quantify compute and intelligence. Former easier, latter harder. Marcus Hutter has formal definition of intelligence which is quantifiable. Develop agents that pursue goals. Reach more goals with less complex algorithms means more intelligence. 2/2 In this x post Xai employee says that robotics will be solved when video generation enables many worlds to be simulated. This is in line with our previous post about many worlds interpretation of AI and continuous learning (Sutton).
❤️ Towards truly open ended reinforcement learning. Jack Parker-Holder PhD thesis. “We believe a possible path towards generally capable agents is a system that begins with simple configurations but makes it possible for them to automatically become more complex over time.” 1/5 RL. 2/5 Auto RL, learn meta parameters to adaptively change behavior. 3/5 Diverse Environments thourgh unsupervised environment design (UED) 4/5 Learn world models from data. 5/5 Genie like many worlds video generation. Goal of Genie is to build a generated video model for real world agents to learn in simulation. Create many worlds in which agent optimizes behavior for many different scenarios. Embed many worlds in real word data. Learn from experience. Eventually combine open ended learning with continuous learning (Sutton).
Lightmatter. Interesting company. Photonic interconnects. Solves I/O problem of AI and Dennard scaling problem. Use fiber instead of copper to connect chips and connect memory with ALU. More data, higher compute utilization, less energy. Problems: Cost, Software. New architecture requires adaptation. Not just a faster horse. It’s a new vehicle. Dennard scaling. Power density of chips remains the same with shrinking. This scaling law has been broken. That’s partly due to the interconnects within chip and from chip to chip. More I/O raises power consumption per compute and kills Dennard scaling.
Energy. Intelligence is compute and thus energy constrained. One gigawatt datacenter will require electricity like 2.5 times San Francisco. Where are the electrons going to come from? 1 GW datacenter requires roughly 4 GWh storage. Tesla Megablock launch is perfect for this market. Move electricity to datacenter. Don’t rely on legacy providers.
Kolmogorov foundation of Statistics and AI. Formally prove inference conditional probabilities from data. Data is all the information you need. But how to extract it? Create filters and infer intelligence from data. Learn filters. Pre-training means create filters. RL find optimal paths in filter space. Inference is matching input with appropriate filter to do action. Optimize linear algebra computation. Fundamental paradigm of AI is more data, more general, more diverse and more compute leads to better intelligence. Intelligence is ultimatly energy constrained.
Many worlds of AI. Jack Parker-Holder PhD thesis discussion. Part 1 . Towards truly open ended reinforcement learning. Learn from experience. Unsupervised Environment Design. Agent progressively changes the environment and learns behavior. Optimize for diversity. Determinant of matrix as measure of diversity. Key Problem: Designing adaptive curricula to train robust RL agents that avoid premature convergence, handle diverse tasks, and achieve open-ended learning. Scaling. Bitter Lesson. Three optimization problems. 1/2 Agent behaving within environment. 2/2 Environment changing. UED. 3/3 Continuous learning (Sutton).
Emerson is the Beatnik of the early 19th century. Literally goes ‘on the road’, sells everything and also gives up his belief structure. ‘Free yourself from conventions to start living an authentic life.’ Intellectual declaration of independence. Prelude to Moby Dick. America lives by what it does and what it produces. Learn from experience. Science is in relationship to the human mind. In the end it’s about us, not about understanding an abstract deity. God is what we make him to be. Socrates lays the ground, Emerson carries the torch.
Emerson’s instructed Eye and David Deutsch. Both argue that observations are filtered through human mind. Emerson focus on ideas. Deutsch on theory. There is an objective truth but what matters to us is how we interpret it. What we observe is what we think. ‘The painted word’ becomes ‘The observed word’. German transcendentalists were focused on the ‘Ding an sich’. Emerson saw that within us.
Milton, Emerson, Faust. Impression of Emerson biography. Relation between Milton (Paradise Lost), Emerson’s self reliance and Faust. Free will. Steinbeck East of Eden (Thou mayeth). Be yourself. Live authentic. Imitation kills. Instagram and Linkedin destroy authenticity. Faust is about temptation, Emerson’s self reliance about choice.
Cellular automata and Von Neumann. Stephen Wolfram - computational irreducibility. The limit of compression and predicability. How much iteration do you need to reproduce a process? Solomonoff Induction and Kolmogorov Complexity. Solmonoff takes KC and assigns higher weight to lower KC models when predicting future sequences. Von Neumann developed the idea of cellular automata in 1940s about self-reproducing systems governed by simple rules. Wolfram thought about the emergence of complexity. Richard Sutton similar approach towards intelligence. Simple rule (‘reward is enough’) and exposure to nature yields complex behavior. The difference between human and digital form of AI is that AI has many more iterations to go through.
❤️ The arch of AI. Von Neumann, Hopfield, Sutton. From probabilistic logics to neural nets to open ended learning. Von Neumann introduces the idea of automata. In Hopfield’s 1982 paper he shows how ‘simple binary neurons, following local rules, can lead to emergent global computational properties'. Richard Sutton’s latest work on OAK architecture introduces a model of continuous learning and open ended adaptation. 'Reward is enough’ simple rule can lead to complex behavior if exposed to nature. Von Neumann’s game theory lays the foundation for RL.Mathematical framework for interacting agents. Dont' underestimate the math!
First impressions on Ralph Waldo Emerson biography. Declaration of intellectual independence from Europe. Religion is spiritual cement. Inspiration is what counts. Learn yourself by reading history. You are the master of your own destiny. Always do what you’re afraid to do. Emerson is Captain Ahab’s spiritual inspiration. Emerson’s American Transcendentalism makes the
difference between the US and all countries. Over-soul. Every individual has it. Nobody can be pushed aside because we all are part of the over-soul. Self Reliance. Strong belief in self and constant error correction forges character and approximates individual soul to over-soul. Plato. Constant change. Flux.
Book discussion “The Man from the Future. John Von Neumann” by Ananyo Bhattacharya. The right mind at the right time. WWII catalyzed development of computer. Game theory is crucial today as it’s the base for reinforcement learning. For example, min-max algorithm is used in AlphaGo. Anticipated Nash equilibria. Agent interacts with environment. Simple rules (goals, scalar reward is enough) lead to complex behavior. Von Neumann-Morgenstern utility theory underpins value concept in RL. Von Neumann - John Hopfield - Richard Sutton. Knowledge is good, inspiration is better. WWII led to evolution of computing and the birth of AI. America is coming back to 1950s. Optimism. Can do society. Build and take risks. China helps us on this journey. Sputnik effect. Applied math, EE and physics back en vogue. China is forcing function for US to get back real jobs and not waste talent on paper pushing careers. Author gets caught in pretentious technical pedantry. Not good. This Story is too good to be screwed up.
An agent doesn’t have to understand itself to become more intelligent than humans. Simple computational rules exposed to reality will over time yield complex behavior that is more intelligent than humans. Achieving goals.
Reduce entropy as goal for RL agent. Nature is caught in a cosmic prisoners’ dilemma, where each entity tries to reduce entropy while increasing entropy overall. Ends in heat death.
Fighting is a good proxy for RL. Agent interacts with environment. Fighters’ goal is to minimize entropy. Learns what is necessary to achieve goal. Reduce entropy (submit opponent). Bruce Lee, ‘be like water’.
❤️ Be like water. Bruce Lee and Richard Suttons' continuous learning vision for AI. Flexible, self learning AI agents are like water. They adapt to the environment. They learn the rules of the environment (MuZero). They learn what matters to be learned. No form is form. Adapt to the environment by learning what matters in order to achieve goals. The goal of fighting is to minimize entropy.
Optimal Data for training AI models. Interview with Ari Morcos. Paper: How to keep marginal utility of data for AI training from declining. Improve performance of AI models, (solution per dollar), with better data curation. Automate data pruning. Find structure in data so that additional training runs yield better results. Remove redundancy. Data curation underrated in AI culture. Bitter lesson. Optimal data is often a function of relationship with dataset. Unsupervised learning is big breakthrough in AI and enables scaling laws.
If you want to understand Nvidia take a proper linear algebra class. Nvidia has been misunderstood for years. Learn linear algebra to get Nvidia. The underlying force behind modern AI systems is math and in particular linear algebra. Nvidia has engineered a massive compute infrastructure to parallelize math and thus enable AI systems to scale massively. Jensen's law outpaces Moore’s law. 40% revenue CAGR for the coming 5 years. Stock 1000 by 2030. Real world AI learning from experience and generating its own trajectories. 4 tril. USD AI capex by 2030. Inference and trading will merge. Sutton. Continuous learning AI is limited by compute.
Can we build an intelligence that is more intelligent than us? Can a system build something from within that’s inherently different? Yes. Because digital systems can run more rollouts than humans. Try and fail much faster at much bigger scale. Self referential, i.e. using its own weights to adapt the weights with nature as ground truth. Learn from experience. Richard Sutton.
Math is at the core of AI. AI is math. First impressions of Von Neumann biography. It was mathematicians who created the modern computer. Math riddles such as Turing’s halting problem which arose from Hilbert’s Entscheidungsproblem, Goedel’s inconsistency and Von Neumann’s idea of general computer architecture. Chess is just math. Game theory is more, it reflects real life with messy outcomes. AI is driven by math. Riddles are often avenues for new discoveries. Today we’re asking the question : Can we create a better intelligence than us? Yes, but how. Richard Sutton on AI that learns continuously.
❤️ Generative AI can create many worlds and lead towards AGI overcoming high Kolmogorov complexity. Reality is non stationary. That’s why iteration is the best and only way to learn how to adapt. Intelligence is adapting to a highly complex environment. Learn in run time through experience. But it take more than one world for an agent to become superhuman. Generative worlds are the path towards superhuman AGI because the agent can iterate through many worlds in parallel.
Richard Sutton presentation on his quest for a general continuous learning AI. ‘My vision: build an intelligence that continuously learns based on experience with nature.’ The progress of intelligence is only constrained by compute. Simple goals tied to rewards yield to complex behavior. Design space. Runtime space. Agent learns only from interaction with nature. No design. No priors. Just experience and methods of learning embedded in algorithm. Goals are represented by reward. Reward is enough. Scalar. Small reward in small steps yield complex behavior. Nature is non-stationary. High Kolmogorov complexity. . Computationally irreducible. You can’t predict. You must iterate and learn from experience. Key to Sutton’s OAK architecture is self discovery of features (like crawl, cry, swing..) and options, which is the solution to feature (learn how to crawl). Options expand the state vector. Caveat: Agents only learns from its own interactions. How about many worlds, generative worlds, so the agent can iterate over many scenarios?
Book discussion “Anna Karenina” by Leo Tolstoi. "Happy families are all alike; every unhappy family is unhappy in its own way." Happiness comes from experience not reasoning. Culture, habits, knowledge are good but can be hindrances. People are unique in their unhappiness because their minds drive them towards a state of unhappiness. Corollary to Richard Sutton’s “Era of Experience”. Runtime intelligence, don’t rely on design time. But how much design is necessary? What priors are needed? For example, is responding to rewards a type of knowledge we require for intelligence? Where does that come from? Is adding numbers knowledge required for intelligence? Platonic representation of intelligence.
Reaction to podcast interview “Redefining Energy. How big things get done.” Professor Bent Flyvbjerg answers question: Why do big projects succeed or fail. Key ingredient of success is decision making; think slow, act fast. Think through pros and cons and carefully lay out a strategy. Then act fast. Valley of doom, the time you are vulnerable. Modularize project. Make it small, lego blocks. Self sustaining. Like Elon Musk. Set big goals, modularize into self sustaining projects. Negotiate in advance against law suits.
Learning from Video. From Seeing to Experiencing. Paper discussion. Navigation. Grounded perception based on waypoints. RL to train agent how to navigate from video. Use human intervention as negative reward.
Why AI models loose plasticity. Richard Sutton. Plasticity is the ability of a model to improve learning from new data . ‘With current deep learning methods, it is usually not effective to simply continue training on new data. The effect of the new data is either too large or too small. How I really feel talk. Our take on this talk. Why not train and keep learning? Gradient decent is designed for transient learning. Solution to plasticity: L2 normalization and continual back propagation. Identify weights that have little utility to the network. Reinitialize them. The goal of a high plasticity model is to activate as many units (neuron, computational unit) as possible with new data. The technique is to update the outgoing weights of a neuron but keep the incoming weights. So it doesn’t forget.
Key concepts: Intelligence is a model that determines its own input data and rewards. Humans define goals, rest is done by AI. Continual Back propagation and L2 Normalization.
Tolstoy’s Anna Karenina and the Era of Experience. InPart 8, Chapter 9 Levin talks about reading philosophy when looking for answers about reality. “As long as I follow the definitions of words, I understand. But it doesn’t get me closer to reality. Exchange the word ‘Will’ in Schopenhauer with ‘Love’ and you have a totally different understanding”. Human constrained intelligence is precisely that, ‘human constrained’. Let the AI find its own way through nature, without human judgement. Boostrap AI with human judgement but then let it find its own way. Is love a good goal and/or constant?
Waymo CEO Dimitri Dolgov interview reaction. Technology debt holds companies back. Dolgov started in mid 2000s before Image Net and is still caught in the sensor fusion Frankenstein AI paradigm. Focusing on safety not enough. Optimize safety, cost and customer experience. Bitter lesson. Era of experience. Platonic representation of intelligence.Tesla is bitter lesson pilled. Waymo not.
❤️ Soft money fosters honest art. Hard money fosters human art. “In not lying they see poetry.” Quote by Levin in Anna Karenina. Prelude to Marcel Duchamp . Fiat money creates froth in society and that is reflected in art. Duchamp fights against perversion of art by trying to be honest beyond human instincts. “What is real?” Desperate early 20th century attempt to achieve honesty in an increasingly delusive society corrupted by central banking and fiat money. Conceptual Art is wrong interpretation of Duchamp. He wants to be real, not merely conceptual. Show reality as it is, not pretty, not pleasant. Just is. But he fails. Painted word turns Modern Art and Pop Art into farce. Creeping Socialism. Bitcoin stops that. Honest money fosters human art. It might not be honest in the sense that it reflects reality. It’s more organic, pleases human desires and senses. While current art scene is Woke infested, eventually art will become more human, more organic, celebrate beauty.
Liu Qiao, Fireworks CEO, Founder. Interview on TWIML podcast. Most data lives in the application. Not the internet. Inference first. Work back from inference to training to data. Model quality is data. Differentiation in AI is data and how to build data flywheel. Start with inference, backwards. Frontier models are converging. Moat is data flywheel. Two phases. 1. Experimentation 2. Scaling.. Vectors of optimization: Quality, Latency and Cost.
It’s solutions, stupid! Not Artificial Intelligence. High valuations for loss making companies like Windsurf, OpenAI or Perplexity are a sign of AI obsession. What matters is not intelligence, its’ solutions to real problems. For example, we don’t care about writing code, we care about building apps that solve real problems. We don’t care about pixels and videos, we care about stories. Companies that loose money every time they make a sale are not real companies. What matters is solutions at scale and low cost. Tesla. Nvidia. The key question for investors is: How much value are you creating for society and how much of that value are you capturing? The former should be at least ten times higher than the latter. But there must be value capture, otherwise it’s not a business.
Discussing Essay “Reinventing Physics with Self-Learning AI”. Understanding nature through self exploration. Human knowledge as a judgment is good but not good enough because 90% of unknowns are unknown. AI can explore by iterating on its own reward functions and ask for its own data. Simulation. Beyond human knowledge. Science as a game. Reasoning based on real consequences, not human judgment. In order to change science we must change the methods. Inspired by position paper “The Era of Experience”.
Discussing Essay “Beyond Human Minds”. A reaction to position paper by Richard Sutton and David Silver titled “The Era of Experience”. Current AI models are bounded by human knowledge. To go beyond we must design systems that evolve on their own terms by finding pathways that maximize a reward function and by doing so create new knowledge. Today human knowledge acts as a gateway without assessing the consequences of their actions. What is the goal? How do design reward functions. Kant. Actions must be reconcilable with common human goals. Reward functions can be shaped by the AI itself. Human intervention is not an option. Inverse Moral Hazard. If the AI knows humans will intervene, it will act as if it was constrained by humans even if humans decided not to intervene for now. We don’t know what we don’t know.
❤️ Discussing essay “The Bitter Lesson about Self Driving Cars”. We should not teach robots how we think we drive, we should let them learn. Compute and data scale. Human heuristics don’t. In fact, they are a trap because they only offer short-term gains. Evaluation and data are still bottlenecks. Simulation is the solution. Tesla follows bitter lesson. Waymo follows Frankenstein AI. Sutton versus Burgard.
Berkley Agentic AI Summit Part 3. Sergey Levine. Multi turn RL for LLM Agent. Ask questions. Clarify. Even suggest data to be trained on and use GEPA style eval to improve. Training with suboptimal data is better. In this paper Levine introduces a benchmark for off line RL testing. Ed Chi, Google Deepmind. Currently, we are interacting with a search engine. In the future we will be interacting with a mind. Turing test - Can I get annoyed at the agent? Agents are exposed to uncertain environment. Next Gen. AI is designed to interact with probabilistic environment. How do you define reliability of a system that acts in uncertain environment. Best reliable intelligence is Navy Seals. Teamwork.
Berkley Agentic AI Summit Part 2. Ion Stoica, Berkley. Reliability. Trust in outputs. Validation and Evaluation. Problem of AI is probabilistic nature of models and environment. How to prevent cascading failures? Testing tools. RLVR (Reinforcement learning with verifiable rewards). How do we know a particular version is more reliable? Reliability is Goedelian. Matei Zaharia, Berkley Reflective Optimization. Automate RL. RLVR works well, but requires lots of rollouts. Can those be automated. Yes. Use language to debug LLM. The paper on GEPA points out that the boundary between weight based and prompt based learning is not well understood. Feedback engineering. What feedback is best for the model to get better? Data engineering, what data is best to update weights to improve model?
Berkley Agentic AI Summit Part 1. Bill Dally, Nvidia.Acceleration of compute driven by representation, special operations, sparsity and Moore’s Law. Parallelism. Pareto optimality between Inference latency and number of customers you serve. On the edge you only deal with one customer. Power constraint. Nothing new. Concerning? Biggest value add of Nvidia is programmability and flexibility. Ramine Roane, AMD. Tech drives on three axes. 1/3 Total cost of ownership. 2/3 Security and dependability. 3/3 User experience. Jared Quincy Davis, Foundry. Open source compositional framework for agents. Similar to PyTorch and JAX, Inference scaling. Use same model for sequential validation. Improve reasoning. Facilitate communication between LLMs. PyTorch of networks. Decentralized intelligence. Blockchain like.
Tesla valuation model. Sum of parts per share. Cyber Cab 250, Energy 130, Car 130, Robot 25. Ironically most of the value comes from businesses media is not focussing on today. Car deliveries 1/3 of value, cyber cab and energy 2/3. Searchlight paradox of investing. Looking where the light is instead of looking where the value is. The bulk of value comes from long tail revenue Cyber cab and Energy. How can we invest in something where value comes from businesses that barely exist today? Look at pace and direction if iteration.
Bitter Lesson Material Science. AI can help material science discover new materials and find better ways to solve difficult engineering problems. For example, the Starship heat shield or EV batteries with higher energy densities. But science is caught in a pre-Image Net dilemma of inadequate representation, narrow models and underutilized compute und memory power. How can this be solved? Material science must be "bitter lessoned". Platonic representation of intelligence. Maybe videos of how materials behave (at scale) can teach us more about properties than tedious chemistry. The periodic table is 19th century technology. AI is 21st century technology. Bridge the gap.
Stock discussion. GEV and ETN. 1. Electrification of US. 2. AI Token factories. Growth in electricity infrastructure. Token factories. Tokens are value added electrons. More electrons means more infrastructure. US electricity base 1.2 tw. demand 4000 twh. 50% of capacity not used. Batteries. Full transition of fleet to EVs 1000 Twh, Token factories 500 Gwh. Assume 1000 Twh more demand by 2030, that is 25% growth. Plus industry 1%, that is 5% growth per annum. AI Data center electrification ads to content, UPS (uninterrupted power supply), Vernova, power, electrification. Power, gas and aeroderivative turbines. Flexible, fast deployment of power. Eaton electrification, aerospace. Tesla’s projection at battery day 2020 is coming to fruition. More investment in electric infrastructure. Driven by AI and EV, both 50% adding 1000 Twh demand by 2030. Legacy companies don’t fit our mission. If we chose to divest current positions, Bitcoin is better investment (higher return, i.e >20%).
Podcast reaction “Lanz und Precht discussion with Luisa Neubauer”. Climate activists are the antithesis of the sustainable energy movement. Like Lenin and socialism, pushing for the right cause for the wrong reasons. Climate is not a problem, it’s a pretext. De-carbonizing energy is a problem and technology, innovation and entrepreneurs can solve it. Climate is for self aggrandizing activists who want to hijack society with fear. Fortunately, society has stoped to engage. Solutions to sustainable energy are being developed in research labs, not on podiums and “climate conferences”.
Waiting for Godot? Tiring. Future potential while present is looking difficult. EV tax credit and ZEV credit reduction. Tesla transitioning from selling product to selling robot intelligence like transport or labor. Traffic is like a large civil works project that has been optimized from humans (thought inspired by Mariana Minerals). Transitioning to machine operated traffic solutions is promising but difficult. FSD unsupervised and Cybercab are drivers of future revenue. But when? Energy Gross Profit 4 Bio. USD and growing. Opportunity costs with other investments. Same as talent war, there is an investor war due to opportunity cost. Cybercab will have lower cost per mile because no driver and optimizing hardware for autonomous use case (less performance, more range per kg, more cycles) drives down cost per mile. Optimizing for Cybercab allows for less degrees of freedom and more optimization which lowers cost per mile. If Bitcoin is cost of capital. How do I justify owning Tesla? Tesla is solving edge intelligence at scale, i.e optimize for intelligence per gigabyte (memory and energy constrained).