Black Box at the Blackboard
Written by Sihoon Yang
Edited by Lorenzo Rave
With the emergence of the fundamental equations governing fluid mechanics, discovered by the French engineer Claude-Louis Navier and the British mathematician George Gabriel Stokes during the nineteenth century, humanity was presented with one of the greatest achievements of mathematical physics. Their equations presented a way of describing phenomena ranging from ocean waves and meteorology to blood circulation through the arteries and lift on airplane wings through a unified approach. For almost two hundred years, however, there was an unanswered question that stood at the very heart of the field: is it possible that fluid dynamics with smooth solutions may develop an infinite velocity, or mathematical singularity, in finite time? The Navier–Stokes existence and smoothness problem was selected as one of the seven Millennium Prize problems by the Clay Mathematics Institute in 2000, standing as a benchmark of human intelligence and endurance for decades (“Navier-Stokes”).
That brings us to this month, September 2026. OpenAI claimed that using an internal multi-agent architecture, a group of approximately 10,000 synthetic AI agents working in constant mode for 88 hours managed to create a complete formal proof of the existence of a blow-up of fluid velocity under certain boundary conditions (“On the Navier-Stokes”). Instead of one lone genius drawing equations on a blackboard, a massive computational engine took 130 billion output tokens to solve a mystery that had puzzled physicists for 90 years (Thjømøe). In an age where artificial intelligence is starting to unlock the most challenging mathematical problems, are we pushing the boundaries of human knowledge forward, or are we surrendering our human understanding?
Indeed, advocates of AI-enabled scientific discovery are right to note that formal theorem provers such as Lean do not suffer from the human shortcomings that often undermine the process of peer reviews, such as overlooking a boundary condition, making a subtle mistake in reasoning or even the phenomenon of “structural hallucination”. In pure mathematics, it takes only one unproven assumption or one overlooked case to invalidate decades of work. By combining multi-agent systems with machine-verifiable logic, researchers can examine millions of possible proof routes at once without being vulnerable to human cognitive biases and fatigue or dogmatic ideologies. What is more, automated proof generation frees human minds from the tedium of verifying each step to allow researchers to concentrate on the higher-level synthesis of ideas and application to the physical world. With the help of synthetic agents, complex problems in fluid dynamics can be solved much quicker than any human mathematician can check all the steps by hand, paving the way for practical applications in climate modelling, aerodynamics on supercomputers, and cardiovascular medicine.
However, we need to put our tech-utopian glasses aside and consider what is being lost in this transition. Mathematics has never been just a collection of verified tautologies or the record of logical steps. It is rather a human conversation driven by intuition and creativity. Historically, a breakthrough in mathematics was not just the proof yielding a definite “yes” or "no". It introduced a whole new set of mathematical instruments and intuitions, inspiring scientists of future generations to draw connections between different scientific disciplines. The elegance of Euler’s Identity or Fermat’s Last Theorem comes from the imaginative leaps required to discover concepts that illuminate the inner architecture of our universe in a way comprehensible to a human mind.
If a black box system creates a proof involving millions of lines of computer code, it gives the answer without giving the understanding. If there is no intuitive human comprehension of how a singularity forms, we have traded understanding for verification. We risk turning from architects of scientific discoveries into mere consumers of facts. If a theorem was proven in a digital labyrinth and no human mind comprehended its inner intuition, was mathematics actually advanced, or did we simply build a faster engine to churn out facts we cannot comprehend?
Just as alarming is the growing consolidation and privatization of fundamental science. Millennium Prize problems were created with a definite humanitarian vision: to push the boundaries of “common heritage of humanity” through publicly accessible academic research ("Navier–Stokes"). Today, tools that can take this work further require multimillion-dollar compute farms and proprietary training datasets owned exclusively by private tech monopolies. When science becomes restricted by private API paywalls or is only used to promote venture capital as a benchmark, the principle of the scientific ethos of openness begins to collapse (Kim).
Last but not least, this transition can undermine the intellectual pipeline of scientific research. If mathematicians are trained in an environment where complex and creative breakthroughs in mathematics are routinely outsourced to AI agent swarms, we will lose the very field in which human intuitions are fostered. The intellectual rigor achieved through the challenge of an impossible problem is what helps build the critical thinking needed to question the existing paradigms.
Of course, outsourcing computation to 10,000 AI agents is definitely a triumph of modern technology. However, an answer without human insight is only half of the triumph. Artificial intelligence should act as a powerful telescope that expands our perspective, not a substitute for our thinking. As we enter the age of automation, our top priority should be preserving the human spark of curiosity, making sure that the search for truth is a human endeavor rather than a corporate calculation.
Works Cited
Kim, Aerin. "Did OpenAI Really Solve Navier-Stokes? Verifying the Disputed Proof." Miraflow, 12 Sept. 2026, miraflow.ai/blog/navier-stokes-ai-proof-controversy-openai-astra-explained-2026.
"Navier–Stokes Existence and Smoothness." Clay Mathematics Institute, 2000, www.claymath.org/millennium-problems/navier-stokes-existence-and-smoothness.
"On the Navier–Stokes Millennium Prize Problem." OpenAI, 8 Sept. 2026,
Thjømøe, Hans Christian. "10,000 Agents in 88 Hours Prove a 90-Year-Old Navier-Stokes Problem." Neoteric, 9 Sept. 2026, www.neoteric.no/blog/openai-navier-stokes-10000-agents-88h.
Comments