I’m joining @GoogleDeepMind to build world models and interactive media experiences. This move means I’ll be dropping out of @StanfordGSB and stepping down as its co-student body president. I’m incredibly excited for what’s next. I still believe deeply in the value of an MBA, and I’ll write separately about how I made the decision to leave Stanford. For now, I wanted to share some of the thinking behind what’s next:

When a camera pans away from a table and pans back a minute later, a pure pixel predictor video model usually forgets what was sitting on it. This spatial blind spot is a huge bottleneck holding back physical AI and robotics. An embodied agent operating in a home or warehouse has to reason about occlusion, metric distance, and continuous trajectories in real time. Solving this requires world models that maintain persistent spatial memory and respect physical constraints.

Most people assume training on millions of hours of video will automatically teach models spatial reasoning. And video models can definitely generate convincing sequences of pixels but they still can't really maintain good object permanence, consistent camera geometry, or basic physical dynamics for long session lengths.

Most VLMs treat images as a flat grid of 2D pixels that get mapped into the same token space as text. That’s great at semantic pattern matching and identifying objects, but they lack a good way to represent depth, volume, and geometry. Recognizing what an object looks like in 2D is very different from understanding how it occupies 3D space.

A frontier vision-language model can describe a picture of your living room but it can't tell you if your new couch will fit through the doorway. This is an important limitation in how today's multimodal models perceive the world.

A frontier vision-language model can describe a picture of your living room but it can't tell you if your new couch will fit through the doorway. This is an important limitation in how today's multimodal models perceive the world. Most VLMs treat images as a flat grid of 2D pixels that get mapped into the same token space as text. That’s great at semantic pattern matching and identifying objects, but they lack a good way to represent depth, volume, and geometry. Recognizing what an object looks like in 2D is very different from understanding how it occupies 3D space. Most people assume training on millions of hours of video will automatically teach models spatial reasoning. And video models can definitely generate convincing sequences of pixels, they still can't really maintain good object permanence, consistent camera geometry, or basic physical dynamics for long session lengths. When a camera pans away from a table and pans back a minute later, a pure pixel predictor video model usually forgets what was sitting on it. This spatial blind spot is a huge bottleneck holding back physical AI and robotics. An embodied agent operating in a home or warehouse has to reason about occlusion, metric distance, and continuous trajectories in real time. Solving this requires world models that maintain persistent spatial memory and respect physical constraints. So, unlocking general-purpose physical intelligence depends on building models that understand the geometry of the world.

When a model trains on unverified synthetic data, hallucinations compound, edge cases disappear, and the system gradually degenerates into bland mimicry. It's the digital equivalent of making a photocopy of a photocopy over and over again.

Training new AI models on data generated by older AI models is becoming a trap. As the public internet fills with machine-generated text, labs that don't verify their synthetic data run straight into model collapse. When a model trains on unverified synthetic data, hallucinations compound, edge cases disappear, and the system gradually degenerates into bland mimicry. It's the digital equivalent of making a photocopy of a photocopy over and over again. Many teams assume they can bypass data scarcity simply by prompting a frontier model to generate millions of new training examples. That approach works for basic things like syntax formatting, but it doesn't teach a model advanced reasoning or emergent capabilities. Unverified synthetic text simply recycles existing priors. High-value synthetic data relies on deterministic verification engines. Engineering teams are coupling generative models directly to code compilers, formal math solvers, and execution environments. The model proposes a solution, an external system validates it against ground truth, and only verified successes make it into the training curriculum. We're seeing this dynamic expand into spatial intelligence and robotics as well. Generating synthetic 3D environments allows models to interact with real causal physics instead of guessing from flat 2D images. When an agent knocks a virtual cup off a counter, gravity provides an uncheatable evaluation signal. In this next phase of AI development, competitive advantage comes from the rigor of your testing environments rather than the sheer volume of generated tokens.

I think the next wave of AI winners will be defined by post-training mastery. Base models provide the commodity substrate, but the teams that build the most disciplined post-training infrastructure will capture the enduring value.

Pre-training frontier models is quickly becoming a commodity and post-training is becoming the competitive differentiator. The moat is shifting to post-training.

The most effective agent workflows constrain the model inside a deterministic software harness. You give the system concrete verification steps, dry-run environments, and explicit fallback paths whenever a check fails

I think enterprise defensibility belongs to the teams that build verification layers, deterministic rails, and robust error-recovery infrastructure.

If each step in an autonomous agent’s workflow is 90% accurate, a six-step chain succeeds barely over half the time.

I think the AI ecosystem would benefit from fewer synthetic benchmarks and more public arenas that test autonomous systems end-to-end

We need to bring back large-scale engineering competitions that give away a ton of prize money. The modern autonomous vehicle industry traces back to the DARPA challenge, which pushed academic robotics teams to build working prototypes which led to tons of companies, like Waymo! Physical competitions have always forced a level of pragmatism that siloed software/AI leaderboards just can't match. Autonomous racing leagues and AI motorsport challenges in China (like in this video, where the cars are all self-driving) have accelerated their progress in real-time world models and spatial intelligence because competing on a track at insane speeds forces you to optimize low-latency inference under unforgiving physical constraints. I think the AI ecosystem would benefit from fewer synthetic benchmarks and more public arenas that test autonomous systems end-to-end. Bringing back high-stakes challenges is a fast way to turn research into actually deployable products.

The ultimate promise of RSI is turning the manual craft of building models into a scalable engineering discipline. So, I think the real significance will be the industrialization of AI research. But long-term defensibility belongs to the teams that build rigorous verification harnesses and automated training environments.

RSI usually sparks some anxiety about an AI rewriting its own code and ascending to uncontrolled superintelligence overnight. But RSI still runs into physical engineering constraints; models trained on their own ungrounded outputs suffer from synthetic collapse without good external data sources. Compute budgets, power constraints, and real-world edge cases will remain friction that software alone can't really erase.

The danger in AI-doomerism is that it distorts how we build and regulate technology. When policy conversations center on mythical rogue superintelligences, regulators waste time drafting rules for sci-fi scenarios. Meanwhile, we need to focus on bias, data provenance, synthetic data contamination, evaluation integrity, and security.

Large models don’t possess will, desires, or a drive to preserve themselves against human intervention. They’re mathematical functions that map inputs to outputs based on historical training data. The danger in AI-doomerism is that it distorts how we build and regulate technology. When policy conversations center on mythical rogue superintelligences, regulators waste time drafting rules for sci-fi scenarios. Meanwhile, we need to focus on bias, data provenance, synthetic data contamination, evaluation integrity, and security.