"Not Even Wrong" Podcast
Investing in fundamentally new concepts and engineering practices with large impact.
Discussing Luca Carlone’s presentation on robotics and memory. Most of what robot models do is memorizing (Mac Schwager). Better solve the energy problem and memorize as much as you can. Or find more efficient ways to learn and combine memory with task specific problems? Carlone. 1/3 Carlone comes from old school engineering, optimizing and Frankenstein robotics. Wants to pivot to data driven, end do end and learning based. 2/3 End to end versus engineered solution. Spacial intelligence augmented with foundation models. Add semantics to robot models. 3/3 Bitter lesson. Experience. Let model learn what matters, avoid inductive bias. Tesla FSD has spacial intelligence without priors. Goal is reduce time to market for real world AI. Robot learning pipeline; pre-train from internet scale video. Mid-train for skill acquisition and fine tune against benchmarks. Intelligence and scale manufacturing is key. New compute and chip architecture. Davison (Graph based design).
Book discussion ‘Flesh’ by David Szalay. Coming of age story of a man in Hungary pre and post Communist era. 1/3 Human condition. Deep and subtle described in raw, direct form. 2/3 Prince Myshkin meets the Brothers Karamazov. Morals, justice and what it means to do the right thing. 3/3 Life happens to István. Eventually ends up in cul-de-sac. You must take agency and responsibility. Not taking action when wrong things happen is the same as doing wrong.
Tesla Cybercab is neither Cybertruck nor ChatGPT. It solves a real problem at low cost. Waiting to monetize FSD through car sales, software and transport as a service. Cybercab is inverse of Cybertruck. Cybertruck is high cost for not enough value due to lack to battery technology. Cybercab is designed to deliver transport as a service at low cost. FSD unsupervised solves real problems such as parking and autonomous transport. Zero to one technology cannot be reasoned by analogy. Compulsion over planning. Solving new problems creates more interesting problems (innovation stack). Time to solve problem is not crucial, but being on the right path is.
Andrew Davison at ICRA 2026. Spatial AI. Representation. Similar to Ranjay Krishna, Chain of thought for spacial understanding through efficient representation. Is it effcient to go from pixel to torque? Every model depends on priors. The most important prior is the choice of training data and architecture. Reasoning in space. Tesla FSD exhibits spacial intelligence. Because it’s trained on native driving trajectories. Bitter lesson. Hardware for embodied AI . Graph network like. Distributed compute and memory. Our approach to robot learning pipeline. 1/4 Pre-training on large scale video data 2/4 Mid training for skill acquisition through imitation, behavioral cloning, teleoperation or corrective RL (Kumar). 3/4 Robot in loop. Robot as data collection agents with counterfactual (shadow mode). 4/4 Fine tune against benchmark.
Sergey Levine on robotic foundation models. 1/3 Pre train on large data, video, simulation, imitation, cloning. Goal is to learn about performing tasks in real world. 2/3 Skill acquisition through automated RL. Robots must learn skills by interacting with real world. 3/3 Robot in the loop. Treat robot for real world data acquisition. Feed specialist knowledge back into general model. Focus on mid training. Sample efficient RL. Fast learning from real world interaction. Correcting, not showing. Let robot find its own trajectory. Future vision (our take); video generation to generate trajectories and select paths. On policy trajectory in real time. Combine with automated RL. In order to learn, robot must have counterfactuals to select from.
Embodied intelligence must happen through space. Commenting on Vladlen Koltun’s presentation at ETH Zurich. Artificial General Intelligence is a bad problem to solve. Intelligence must be useful to humans. Embodied intelligence can be accelerated with cognitive intelligence such as Grok or GPT. But spacial reasoning requires spacial learning. Counterfactuals in space (I move here but should have moved there..). Tesla FSD has emergent spacial reasoning from end to end training. It can be accelerated through cognitive and simulation tools. Robot learning pipeline; 1/3 Pre training from video. 2/3 Mid training through direct skill development in simulation. 3/3 Robot in loop. Robot is data collection engine for real space interaction (dataset combining video feed and corresponding actuator data). Feed data back into simulation, improve model through counterfactuals (Robot did this..but model predicted that). Eventually, learning from experience. No rewards, just behavior.
Book discussion ‘Son of Nobody’ by Yann Martel. Where does Western culture come from? Bridging Homer’s Iliad with modern life. Idealized philosophy clashes with real life stories, both in ancient times as well as today. Glory and heroism meet suffering, quarrel and sacrifice. 1/3 History is theory laden. Winners write history. It’s good to get a different perspective. 2/3 Western culture is built on the idea of sacrifice and kings must keep that in mind. No unlined power. Life of common man matters as much as glory of kings. 3/3 War is bad. Troy is perpetual war, like today with Middle East. Absurd, Iliadesque absurdity of fighting with no purpose. Avoid war through deterrence because fighting war is horrible.
Keynesian economics is charlatanism because it can’t predict. Austrian economics can. Just look at US government debt. Any science discipline must develop models and tools to predict. Austrian economics predicts that a constellation with Fed, no gold standard, dual mandate and US congress dynamics will produce high government debt and high inflation. That’s exactly what has been happening in the US. Keynesian economics is a tool to justify government intervention and deficit spending. MMT (Modern Monetary Theory) is taking Keynesianism to the next level by instating perpetual deficit spending. Link to podcast that inspired this episode.
Robot learning pipeline. Mid training through correction, not showing. Aviral Kumar presentation at ETH. When mid-training and robot goes off into bad trajectory, create counter factual by brining robot back initial state and let it figure out a better trajectory. Pre-training is a collection of possible robot actions. Mid-training is teaching the robot how to execute task. But instead of imitation learning, correction learning is only brining robot back to initial state and let it learn how to do better. Improve reasoning. Paper.
Discussing Shuran Song’s presentation on UMI (Universal Manipulation Interface). Crowd sourcing robot data. Handheld device with camera. Easy to use, low cost and relevant for robot hand manipulation. Stanford culture is crowdsourcing, learning from scale data. UMI problem with implementation. Why should people use this? Robot learning pipeline. Pre train from video. Skill based manipulation (UMI and others) for mid training. Fine tune against benchmarks. Open loop. Robot in loop. To learn something you need counterfactuals.
AI must be safe and accessible. Deterministic explainability is not the overarching goal. Reaction to Gradient Decent podcast interview. The goal of Tesla’s self driving technology is to make superhuman safety accessible at low cost. Explainability is important, but not overarching goal. End to end training is about giving up representation but gaining in exploration and learning. Why stick to human level understanding? Iterate at pixel level and develop newer, better solutions for transport. Same applies to other forms of AI. The goal of AI must be to deliver more convenience, more safety, more access at lower cost.
Book discussion ‘Normal People’ by Sally Rooney. Psychological discovery of human existence. Subtle, everyday events lead to deep insights. 1/3 Socio economic background of Ireland in early 2010s. Zeitgeist pre Brexit and Trump 2016. Lack of purpose. Men must have something to fight for and it cannot be an abstract idea. It must be real. ‘Literature is a fetich for cultured people to take them on false emotional journey to separate themselves from the common people about whose lives they like to read'. They find meaning in abstract goals such as Justice because that is much cheaper and easier than engaging with real, solvable problems. 2/3 Identity. What does it mean to be ‘me’. Cannot define by others. Dangerous. Happens within. 3/3 Power of love. Love is power. In order to love you must submit to the other person which makes you extremely powerful and vulnerable.
Solve the data problem in robotics. CVPR 2026 Part 5. Jitendra Malik presentation. 3dfy everything paper. Take images from video, represent the objects in 3D and then learn in simulation to manipulate. RL and in the loop fine tuning, which is you use robot data and counterfactuals (shadow mode) to learn from robots about the physical world then feed back into model. Tesla FSD is solving for video reasoning. Jitendra’s approach solves manipulation. Trajectory for robotics learning at scale. Pre-train from video. Mid-train with skill specific manipulation, imitation, RL and simulation. Fine tune against benchmarks.
Not selling Tesla to buy Space X. Tesla catalysts near term. Space X longterm. Hot air, removed from fundamentals. Not Tesla. Driven by self driving technology, which is maturing and ready for scale. Space X driven by Starship, which is engineering and even science problem. Space AI requires successful reusability of Starship. Lots of iterations required. Similar to FSD five years ago. Once Starship solved, unit economics of space AI feasible. Tesla and Space X merger creates full stack enterprise AI for digital and physical work and transport. Merger at least 50/50.