I’m Debajyoti Ray.

What I’m working toward

AI will matter most in two places our society has long struggled: keeping itself grounded in physical reality, and allocating its resources well. Almost all of today’s attention is on the screen: chatbots, copilots, content. I’m more interested in what happens when AI moves from screens to atoms, and from advice to allocation. Three convictions follow from that.

1. 1
   Grounding

   Intelligence has to touch the world

   An AI system that never makes contact with physical reality has nothing to correct it. It learns from its own outputs and the outputs of other models, and without an external signal the loop folds back on itself, a hall of mirrors that hardens into a kind of self-reinforcing delusion. At the scale we’re building, that’s not a quirk; it’s an attractor a whole society can fall into. I came to that conviction firsthand, having built the foundational technology and data platforms behind VideoAmp, Rivet AI, and Inference Cloud: our society’s future depends on handling these closed-loop attractors far better than we do today. Physical embodiment is the corrective. Grounding is what keeps intelligence answerable to something outside itself. More on this here.

2. 2
   Screens to atoms

   The physical economy dwarfs the digital one

   Almost everything we’ve built AI for so far lives in the digital world. That is about to change. The physical-AI economy, spanning robotics, manufacturing, energy, the oceans, and the space economy, will eventually be two to three orders of magnitude larger than the digital one, because that’s the actual size of the world. We are at the very beginning. As launch costs collapse, the cost of mapping the Earth collapses with them: more than eighty percent of the ocean floor has still never been mapped in detail, and most of the planet’s resources remain uncharted. That all changes downstream of cheap access to orbit, and the SpaceX IPO is a testament to how real this shift already is. The same forces are turning science itself into something AI can accelerate, which is why I’m behind the mission of Coactive Science. Here’s the longer argument.

3. 3
   Allocation

   AI is an allocator, and that’s where government comes in

   Most of the consequential decisions a society makes are allocation decisions: who gets capital, where it goes, what gets built. Humans make these decisions with a handful of variables in working memory. AI can weigh millions of variables and competing priorities at once, toward outcomes far closer to a genuine social optimum. Allocation has a mirror image: waste. Fraud, waste, and abuse drain resources away from the people who deserve them, which is morally wrong, and AI should be used to identify and eliminate it. That is why AI for government will be the most impactful frontier of the next five to ten years, and why I’ve chosen to work on it in housing, the single largest component of the economy and one of the places where bad allocation is felt most directly. I unpack this here.

How I got here

I’ve been drawn to mathematics for as long as I can remember, and not long after, to writing code. I first learned databases and used them to help small businesses, which paid for much of my education, and later built AI and machine-learning models for hedge funds. I competed in the Olympiad, and that passion took me to the University of Toronto to study pure mathematics.

While preparing for the Putnam competitions, I realized I cared less about proving the theorem than about building the system that could. Around the same time I was taking Geoffrey Hinton’s machine-learning course, and during it I asked him if I could write code and prove theorems for him. I ended up working alongside members of his group and contributing to four papers.

From there I went to Caltech for a Ph.D. in Computation and Neural Systems, the program John Hopfield helped found. Hopfield and Hinton would later share the 2024 Nobel Prize in Physics for the ideas that became modern AI, a lineage I still find a little surreal to sit downstream of. The question that animated my doctorate was a simple one to state and hard to answer: how do you make a neural network honest about what it doesn’t know? I worked on modeling uncertainty in deep networks, which led me to active learning. I’ve come to believe that uncertainty, knowing what you don’t know and choosing what to learn next, is the key that unlocks both physically embodied AI and the diminishing-returns wall that today’s large language models are starting to hit. I wrote about why that is here.

Writing

I write to think out loud about the ideas above: uncertainty, grounding, the physical economy, and AI in the public sector. A few starting points:

• Why active learning is the key to physical AI

  Knowing what you don’t know is the bottleneck for both robots and LLMs.

• Grounding, and the hall of mirrors

  What happens to intelligence that never touches the world.

See all writing →