Alloy of Our Intelligence

Richa Sharma, September 13, 2026

Tags:move 37navier stokesswarms

Lee Sedol stunned after seeing an unprecedented Move 37.

CONTENTS

Move 37 against the Local Maximum

Go was conceived more than two millennia ago, yet its rules are simple enough to pick up as a hobby in a single afternoon. It hands you a simple wooden grid with stones that can be arranged in over 1017010^{170} legal positions, dwarfing the number of atoms in our observable universe. And yet that combinatorial abyss disappears behind our curiosity and intuition that picks a handful of moves worth pursuing, compressing these endless possibilities into shapes, threats, territory, and patterns learned, and passed on to other human players.

Many scientists, attempted to beat a human master at the Game of Go, but it possessed an almost paralyzing breadth. Algorithms like Deep Blue which defeated Garry Kasparov, proving machines could brute-force the depth of a game like chess, still failed to win at Go.

Then a machine played Move 37, against Lee Sedol.

AlphaGo placed a black stone on the fifth line in Game 2. It was so contrary to human priors that its own network which was trained on over 30 million positions from human games, estimated only a 1-in-10,000 chance of a human ever playing it. Yet in retrospect, it was the perfect move to decide this game 100 moves later. As if the machine made it presciently, ahead of time.

Thousands of years of play had compressed human intuition into strategies that turned out to be a local maximum. Intuition is compressed judgment: efficient, but capable of trapping us in a conceptual dead end.

Now the board has shifted, from games played in leisure to problems that have plagued us since we first tried to model the universe. Navier–Stokesρ(∂v∂t+(v ⁣⋅ ⁣∇)v)=−∇p+ρg+μ∇2v∇ ⁣⋅ ⁣v=0\begin{gathered}\rho\left(\frac{\partial\mathbf{v}}{\partial t}+\htmlClass{navier-acceleration}{(\mathbf{v}\!\cdot\!\nabla)\mathbf{v}}\right)=-\nabla p+\rho\mathbf{g}+\mu\nabla^2\mathbf{v}\\[5pt]\nabla\!\cdot\!\mathbf{v}=0\end{gathered}In the highlighted term lies the seed of turbulence: ∂v∂t+(v ⁣⋅ ⁣∇)v\frac{\partial\mathbf{v}}{\partial t}+\htmlClass{navier-acceleration}{(\mathbf{v}\!\cdot\!\nabla)\mathbf{v}}. The flow changes its own motion, creating swirls within swirls. This feedback makes turbulent motion so hard to predict, with no general exact solution to describe it.The second equation, ∇ ⁣⋅ ⁣v=0\nabla\!\cdot\!\mathbf{v}=0, expresses conservation of volume: stretching a vortex makes it narrower, allowing its spin to intensify.radius 1edge speed 1×spin rate 1×Narrower vortex, faster spin akin to a skater pulling their arms in (ideally, spin ∝ 1/r²). is one such deceptively simple problem. It consists of equations that govern anything that flows: the cream curling into your morning coffee, the blood your heart pushes out with its every beat, and the air leaving your lungs as you breathe.

If mathematics is humanity’s attempt to consolidate nature’s infinities into a language, classical physics is our attempt to map its continuity… and to find out exactly where that continuity fractures.

One such fracture hides within Navier-Stokes, planted like a seed in the mechanics of turbulence. The equations began as an attempt to apply Newton’s second law of motion to a fluid. But unlike a solid, which is an unchanging object, a fluid's motion is self-referential. Its overall acceleration depends not just on time, but on convective acceleration: the water literally carrying itself along from one point to another. Because the fluid is incompressible, its volume must remain constant. When a spinning vortex is forced through a narrower junction, it stretches, thinning out and violently speeding up its spin, akin to a skater pulling in their arms. This non-linear feedback loop introduces turbulence, and it threatens to spiral out of our mathematical control.

So the natural question to ask is: what happens if the vortex keeps stretching? Could it spin so fast that its speed becomes infinite?

Conceptually, human beings struggle with mathematical singularities. Our biological intuition guides us away from what cannot be cleanly visualized; after all, our physical intuition is an artifact of a primarily visual cognitive evolution, shaped to navigate physical space, forage for food, and hunt. Even minds like Einstein initially rejected singularities within their own equations, dismissing them as mathematical artifacts without physical significance—artifacts we now famously image as black holes.

And since we have never seen a river or an artery tear the fabric of reality with infinite speeds, humans were naturally drawn to the conclusion that the math must eventually smooth itself out.

Trapped in a conceptual local maximum akin to orthodox Go openings, mathematicians spent decades trying to smooth out equations that they intuitively believed had to remain continuous.

It took stepping away from reality to break the spell. In 2016, Terence Tao engineered conceptual singularities in a modified version of the equations. This conceptual crack was later forced wider by mathematicians Diego Cordoba and Luis Martinez Zoroa, whom Charles Fefferman called the heroes of the story, who laid the rigorous groundwork proving that fluid equations could indeed tear themselves apart.

But the formalized Millennium Prize problemFefferman’s formulation asks whether smooth 3D flows stay smooth or can develop singularities. A simulation may suggest a breakdown, but the prize requires a rigorous proof.Clay prize page · Official problem statement concerning the true Navier-Stokes equations still needed a monumental, rigorous push. That push arrived in September 2026, when OpenAI unleashed over 10,000 agents with virtually no compute constraints to exhaust the sprawling combinatorial possibilities of what was left. In roughly 88 hours, the swarm produced a staggering 166-page proof. In doing so, the AI world established a new capability ceiling for artificial swarms: a relentless, unintuitive rigor that will help us reconcile the loose strands we leave behind in our human questions about Nature.

Yet, as it stands, Navier-Stokes fundamentally remains a warning about our failure to map the universe with equations. We can use equations as our lighthouse to approach the truth, but they are not the truth itself. There will always be mysteries and singularities that defy our intuition over and over and over again.

So now, even though these equations guide us to simulate fluids with uncanny accuracy, they still permit what no physical fluid has ever been observed to do. The machines found cracks in our mathematical description of an omnipresent observable.

But what happens when we try to go after mysteries we can barely see?

Mysteries We Can Barely See

Milestones of modern AI master a very specific game: selecting domains with massive state spaces paired with cheap, deterministic verification. AlphaGo operated inside binary win states. AlphaFold navigated the dense, empirically grounded truth of the Protein Data Bank. In pure mathematics and theoretical physics, problems with this structure already exist. Conjectures with clearly defined rules are the machine’s new board of Go. Even navigating the 1050010^{500} string theory vacua across Calabi–Yau manifolds is, at its core, an algebraic and topological search problem. It is vast beyond imagination, yet fundamentally tractable because the verification harness is well-posed.

The wall appears when verification ceases to be cheap.

Consider a thought experiment: what if we restricted a frontier large language model’s training data strictly to 1905 and ran it at high temperature? Hidden among thousands of hallucinated, mathematically sound dead ends, the model might very well generate the exact sentence: "The speed of light is absolute, but time is relative." In 1905, without the hindsight of a century of modern physics, any reasonable scientist would have discarded it as ungrounded slop. Move 37 became immortalized because the win state arrived minutes later on the board. But Einstein had no immediate win state. Relativity had to wait fourteen years for Arthur Eddington to travel to the island of Príncipe and photograph a solar eclipse—extracting the faint, sparse outlier data required to prove that mass genuinely bends spacetime.

When data is sparse, algorithmic search without grounding sometimes yields overfitting instead of insight. We saw this even with human intellect: when 19th-century astronomers noticed the anomalous precession of Mercury’s perihelion, they did not invent curved spacetime. To explain the anomaly, they invented Vulcan, an imaginary planet hidden behind the Sun. The notion of curved spacetime was leaps and bounds beyond the common sense of the time.

No matter how brilliant human minds were, we often ran into our limits at the edge of understanding, where a fundamentally new ontology—a new category—needed to be created. At the conclusion of The Principles of Quantum Mechanics, Paul Dirac acknowledged the breakdown of electrodynamics with a rare, quiet humility:

"It seems that some essentially new physical ideas are here needed."

When a young Richard Feynman read those words, he didn't see an optimization problem. He saw an open frontier: "And so there I was! Some new ideas were needed. OK—so I started to think of some new ideas."

A machine optimizing within known parameters cannot generate that Dirac moment. It will not spontaneously architect the Transformer when searching for variations of the LSTM, nor will it invent Riemannian geometry to explain gravity when it can simply add another epicycle or invent another Vulcan. Brute force across an existing loss function merely scales our current assumptions.

Intuition, Unbounded

This technology is not an oracle designed to render human wonder obsolete; it is the alloy of our intelligence that gives our human intuition teeth. Our biological intuition evolved on the African plains to hunt and forage; when confronted with the crushing mathematics of singularities and sparse outlier data, it naturally breaks down, waiting for a once-in-a-century mind like Einstein, Feynman, or Witten to push us forward.

What if the escape from our scientific local maximum is finally within our reach? To grasp it, we have to leapfrog our fears of intellectual obsolescence. If I may be a romantic for a minute, I hope this paradigm shift forces the educational curriculums of our future to stop focusing merely on tests, and to center entirely on discovery instead. By offloading the tyranny of breadth to the machine, the ultimate end state of our intellect transforms from the simple replication of what we already know, to the discovery of a novelty that did not exist before.

The dread we feel right now is real, and I have felt it viscerally this past month. We cannot yet touch the human part of this technology. But humanity has always been driven by tangible awe: a man stepping onto the moon, the image of a black hole that once existed only in theory. Our hearts race when we hear those stories. The next one may begin with a kid somewhere, following their curiosity and rejecting a physical notion none of us ever thought to question, simply because the daunting breadth of scientific rigor finally rests gently in the palm of their hand.

Centuries from now, our descendants will find it quaint that intelligence was ever confined to organic matter, and that we spent decades terrified of its emergence in silicon, when it was simply destined to become the alloy of our mind.

References
  1. Mastering the game of Go with deep neural networks and tree searchDavid Silver et al. · Nature, 2016
  2. AlphaGo: Move 37 and the Lee Sedol matchGoogle DeepMind
  3. Deep BlueIBM · History of the 1997 Kasparov match
  4. Superhuman artificial intelligence can improve human decision-making by increasing noveltyMinwoo Shin, Jin Kim, Bas van Opheusden, and Thomas L. Griffiths · PNAS, 2023
  5. KataGUI: Study with KataGoGo visualization resource
  6. Moka v1Million · Go visualization resource
  7. Navier–Stokes Equation: Millennium Prize ProblemClay Mathematics Institute
  8. Existence and Smoothness of the Navier–Stokes EquationCharles L. Fefferman · Official Clay problem statement
  9. Vortex Lines, Vortex Tubes, and Vortex StretchingMIT · Marine Hydrodynamics, Lecture 8, 2003
  10. The 7 Million-Dollar Math Problems, Explained VisuallyNontrivial · Navier–Stokes video explanation
  11. Finite time blowup for an averaged three-dimensional Navier–Stokes equationTerence Tao · Journal of the American Mathematical Society, 2016; preprint, 2014
  12. Finite time singularities to the 3D incompressible Euler equationsDiego Córdoba, Luis Martínez-Zoroa, and Fan Zheng · 2023 preprint
  13. Finite time blow-up for the hypodissipative Navier–Stokes equationsDiego Córdoba, Luis Martínez-Zoroa, and Fan Zheng · 2024 preprint on forced fractional equations
  14. On the Navier–Stokes Millennium Prize ProblemOpenAI · September 2026 announcement and research context
  15. Finite Time Blowup for Navier–StokesOpenAI · 2026 proof manuscript
  16. NavierStokesAndEuler: Lean certificatesOpenAI · Formal verification repository
  17. On a Stationary System with Spherical Symmetry Consisting of Many Gravitating MassesAlbert Einstein · Annals of Mathematics, 1939
  18. Highly accurate protein structure prediction with AlphaFoldJohn Jumper et al. · Nature, 2021
  19. Distributions of flux vacuaFrederik Denef and Michael R. Douglas · 2004
  20. A determination of the deflection of light by the Sun’s gravitational fieldF. W. Dyson, A. S. Eddington, and C. Davidson · Observations of the 1919 eclipse, published 1920
  21. Explanation of the Perihelion Motion of Mercury from the General Theory of RelativityAlbert Einstein · 1915; CERN catalogue of the collected papers
  22. VulcanNASA Ames · Astrogram, February 1959, page 2
  23. The Principles of Quantum MechanicsP. A. M. Dirac · Oxford University Press; CERN catalogue
  24. Take the World from Another Point of View: Feynman on Dirac and new ideasRichard P. Feynman · Yorkshire Television interview, 1972; reproduced in The Quotable Feynman, page 7
  25. Attention Is All You NeedAshish Vaswani et al. · 2017
  26. Long Short-Term MemorySepp Hochreiter and Jürgen Schmidhuber · Neural Computation, 1997
  27. Apollo 11NASA · The first crewed Moon landing, 1969
  28. First M87 Event Horizon Telescope Results. I. The Shadow of the Supermassive Black HoleEvent Horizon Telescope Collaboration · 2019

Please cite this work as:

Richa Sharma. “Alloy of Our Intelligence.” September 13, 2026. https://richas.xyz/blog/alloy-of-our-intelligence

Or use the BibTeX citation 
@misc{sharma2026alloy,
  author = {Sharma, Richa},
  title = {Alloy of Our Intelligence},
  year = {2026},
  month = {09},
  url = {https://richas.xyz/blog/alloy-of-our-intelligence}
}