Essay
On Math, AI, and The Biological Rate Limiter We Still Need
Brains that learned the hard way, who lifted the mental weights, will become an antiquity.
“You’re one of the last of a dying breed,” I found myself telling my father yesterday, in the midst of our discussion on the latest set of AI-related, controversial math breakthroughs. His brain - that of a retired math professor, precociously self-taught in his early teens and later distinguished for decades in his field of applied math (fluid dynamics) - was built “the hard way”. While he’s often used computers for computations and modeling in his research, more often, his progress involves scrawling Greek letters on notepads and producing strange geometric doodles, as he considered Jupiter’s atmosphere or modeled the terrestrial jetstream.
As I scanned the heated discussions on X about Navier-Stokes, OpenAI and the elite math community, I wanted my dad’s perspective on the news as someone who’s been in the field since the late 70s and knows many of the cited paper-authors personally. Math didn’t used to be sexy, but in some strange ways, it is now, at least in a certain nerd-world my algorithm has me inhabiting. As a field with hard rules and written proofs, it appears uniquely well-suited to AI-driven progress. So much so that some leading professors have wondered whether it is prudent to send students down the traditional math track, or to push them into AI research instead, if their highest goal is mathematical progress.
My dad’s reaction to the OpenAI-authored paper was one of genuine excitement, albeit with some frustration and a whole lot of concern. My call interrupted him mid-read of the publication, which I assume perhaps 0.0001% of the population can follow honestly, he being among the lucky few. He was initially struck by the length of it, and by the sense that it didn’t feel deeply crafted or human in authorship, due possibly to the perceived nature of the proof as more “brute-force” than “elegant-conceptual”.
We began speculating about how OpenAI authored this paper. When 10,000 AI agents achieve a seemingly earth-shattering result, people have to check that work, deeply question it, prepare it for publication, and consider the host of implications it will have by virtue of the ends and the tech-enabled means.
Any regular AI user feels the microcosm of this in their comparatively simple day-to-day work, where AI’s apparent accelerative powers need to be throttled by honest, imperfect human ingestion and reasoning. The seduction of AI is to build the user’s trust so much that eventually, we wane in our initially studious and righteous attempts at following-along, challenging, and understanding what comes out. When we’re the biological rate limiter for productivity, we may question whether it’s better to remove the friction of human understanding altogether. It’s a fair concern in today’s modern workforce culture driven by productivity. How much does one really need to understand a given subject to make a decision? How informed have we truly ever been in the pre-AI era? What level of internalized mastery is appropriate for advancing a project or a piece of work? It’s a slippery slope of justifications for what can amount to laziness, and risky loss of intellectual control.
Back to my dad’s brain. Together we wondered about the generational shift that is likely to take place in this field, and all others. No human can match the raw compute of 10k agents alone, we’re bounded by time and biology and our own “pre-training”. When AI can be marshalled to progress without the slow-moving shackles of human computation and hard-earned learning, what’s the baseline level of human understanding to make something meaningful of the end result? Without the circuitry being forged over years through repeated synaptic firing, we lose our authority to arbitrate. We then must look to AI as arbiter. To understand AI outputs, we must use AI, and the hall of mirrors becomes one where we’re standing outside, looking through a pinhole at endless refractions of thought. We already see this pattern emerging in other fields: HuggingFace needing Kimi to counter rogue OpenAI agents, where automated defense is the only response to automated intrusion. As AI grows in complexity, using AI to govern and decipher AI becomes a default mode, leaving the human stranded on the periphery.
And brains that learned the hard way, who lifted the mental weights, will become an antiquity.
Humans trade on social currency. That means reputations and accountability matter. Someone has to reckon enough with any output to judge it, know it matters, review the reasoning, and then package it for consumption and action by others. We haven’t fully ceded control for agents to make all decisions for our society. But the way things are moving, where governments and corporations have no incentive but to make better, faster models, bigger and more sophisticated swarms, there is a tsunami that is sweeping away our old time-tested modes of grappling with truths, validating them together, questioning each other at the speed of our brains. When life and science moves at the speed of data-center-driven compute, we’ve engineered ourselves out of the driver’s seat.
The beautifully fine-tuned biology of my dad’s expert brain may never again be reached in quite the same way by a new generation of mathematicians, but that doesn’t mean we wouldn’t benefit from it.