When Bits Hit Atoms

The Final Frontier of AI is Physical

There is a simple but universal reality that more are beginning to appreciate: the physical world is rate limiting.

Whether you believe that AGI is imminent, that digital tool-using agents will completely disrupt knowledge work, that abstract scientific and creative work will be automated, or even that AI is fundamentally not a normal technology…

…an unavoidable fact is that real-world wall time cannot be hacked or gamed. Spinning up more GPU clusters will not accelerate the laws of physics. You cannot CoT or search your way through latency-critical decisions. There is no `env.reset()` for physical systems operating in real-world scenarios. There is no escaping the conundrum that the two elastic ingredients which power the Bitter Lesson, compute and search, must confront the inelastic limiting reactant that is reality.

The problem itself, of course, gets combinatorially harder the more dimensions you add: extending the time horizon of physical problems, adding precision and accuracy requirements, increasing complexity and performance expectations, evolving the amount of real-world contact needed (from freeform flight to wheeled navigation to legged locomotion to industrial manipulation to multi-fingered dexterous hands). Adding any one of these dimensions, let alone several, increases the problem difficulty by at least an order of magnitude.

Moravec’s Paradox speaks to the comparative cognitive difficulty of tasks which come naturally to humans, but the structural and economic difficulties of real-world messiness are just as important. Just as there is a hardware lottery which has resulted in a self-selecting cycle of algorithms developing based on chip properties, there has also been an application lottery which has driven immense AI progress on applications where research and scaling are limited primarily by smart algorithms or diligent engineering. If you feel like a problem can be solved with just a dash of ingenuity and code, then you can bet that many other smart researchers feel the same, and more likely than not, that problem will indeed be accelerated and solved soon.

But the real world is not that. Everywhere you look, the compounding effects of physical rate limitation rear their heads, from real-world hardware iteration cycle times, infrastructure, and supply chains to large-scale data collection and evaluation, massive economic capitalization requirements, and long and uncertain timelines for value capture. The real world is hard. We are starting to see various realizations of this concept (albeit sometimes conflated with geopolitics) with commendable pushes towards re-industrialization, hard tech, general purpose physical AGI, and humanoids.

And I am glad for that. I can think of no problem which is more important, interesting, or novel. Because for any problem that touches reality, no matter how lightly, that touchpoint will always become the bottleneck. In a world of so many uncertainties, of this at least I am sure of: physical AGI will be the final frontier of AI as a whole. We are still so, so early. But the future is bright, and it will be a future built, tested, and ultimately measured against the physical world.