It has been 25 years since a report of original research was last submitted to our editors for publication, making this an appropriate time to revisit the question that was so widely debated then: what is the role of human scientists in an age when the frontiers of scientific inquiry have moved beyond the comprehensibility of humans?
This is how Ted Chiang opened his visionary short story “Catching Crumbs from the Table”, published in Nature back in June 20001 (later anthologized as “The Evolution of Human Science” in Stories of Your Life and Others).
Exactly twenty-five years later, in November 2025, James Evans and Eamon Duede used that same quote to headline their article2 in Science titled “After Science.” They warned that we are witnessing a curious inversion: our AI systems can instrumentally control and predict nature far faster than humans can understand it. In some instances, it appears possible without human understanding at all. AI is changing the balance between understanding and control in science.
The debate over humans being surpassed by artificial intelligence has been raging for over two centuries. Frankenstein’s monster was able to proccess information much better than his creator. Samuel Butler anticipated most of the 150 years of debate that followed regarding the possibility of being surpassed by machines, and even a brilliant Spaniard was able to ironize on the question.
On this blog, we have frequently tracked the evolving social fabric of discovery; exploring “The Question Concerning Scientific Research;” pointing out that science advances one funeral at a time (to borrow from Max Planck); asking to what extent we can ever predict future scientific discoveries; analyzing the data-heavy imagination of the future of science, where large language models were always bound to rewrite the rules; and anticipating that algorithm-led discovery opened up the possibility of discoveries that no human being can ever understand.
That moment has already arrived. It feels remarkably similar to the moment nearly ten years ago when AlphaGo defeated the World Go Champion Ke Jie.
This week, OpenAI announced that an internal, general-purpose reasoning model successfully disproved a longstanding conjecture in discrete geometry known as the planar unit distance problem. First proposed by Paul Erdős in 1946, the problem asks for the maximum number of pairs of points that can be exactly one unit apart when $n$ points are placed in a plane. For nearly 80 years, human mathematicians believed that simple square grid layouts were essentially optimal. The AI system proved otherwise, constructing an entirely novel, infinite family of point configurations that achieved a polynomial improvement.
What makes this an epochal milestone is that the proof didn’t come from a niche tool specialized for mathematics, but from a general reasoning model. It marks the first time a prominent open problem in a core subfield of mathematics has been solved completely autonomously by AI. As Fields Medalist Tim Gowers noted in the companion paper, this is a profound milestone in AI mathematics. The system bypassed human intuition entirely by forging unexpected, sophisticated links between elementary geometry and algebraic number theory—realms humans hadn’t successfully bridged for this problem.
For the time being, it seems mathematicians still understand the result and are able to explain it. However: mathematicians, beware!
Actually, as I watched all this unfolding, the single best explanation of this mathematical leap was actually provided to me directly by ChatGPT.
My question: So… what exactly is it that has been DISPROVED here?
ChatGPT answer:
The conjecture that was believed to be true
The dominant intuition was:
“The square lattice is already essentially close to optimal.”
Formally, Erdős conjectured that:
Here:
- means “something that tends to 0.”
- In other words, the exponent was expected to approach 1 indefinitely.
In less technical language:
The maximal growth was supposed to be “almost linear.”
That is, there could not exist a construction with a fixed exponent strictly greater than 1.
What did the model find?
The model found an infinite family of configurations satisfying:
And this was later refined to something like:
What has not been solved
This is crucial.
The problem has still not been solved completely.
We still do not know the exact asymptotic growth of .
What we now have is:
- a new lower bound:
- and the classical upper bound:
In 1962, Thomas Kuhn published The Structure of Scientific Revolutions. He argued for an episodic model in which periods of conceptual continuity and cumulative progress, referred to as periods of “normal science”, were interrupted by periods of revolutionary science—paradigm shifts. By 1996, John Horgan was asking if we were Facing the Limits of Knowledge in the Twilight of the Scientific Age, a thesis recently updated by large-scale studies confirming that papers and patents are becoming fundamentally less disruptive.
Are we actually running out of things to discover, or are we just hitting the ceiling of what the biological human brain can process and share? Is it AI here to help us build Ted Chiang’s “metahuman” science?
Also this week, two major papers just dropped in Nature introducing Co-Scientist3 —a multi-agent AI system built on Gemini for structured scientific thinking and hypothesis generation— and Robin4 —the first multi-agent system capable of fully automating both hypothesis generation and data analysis for experimental biology. They crawl literature, simulate internal debates, generate entirely novel hypotheses, and analyze data—successfully discovering new therapeutic drug candidates for leukemia and blindness with a fraction of the human time. Humans are rapidly becoming the ‘lab-in-the-loop’—the physical hands executing the chores of an automated mind. Nature also warns5,6 that uncritical adoption risks jeopardizing the foundational uncertainty that science requires. We risk trading deep, paradigm-shifting comprehension for highly optimized, automated utility.
We have officially transcended the ‘tool’ era of AI. The frontier has shifted from AI assisting scientists to AI becoming the scientist.
In a recent essay for Noema Magazine titled “The Death of the Scientist,” astrobiologist Sara Imari Walker argues that the answer depends on something no one knows: What is science? She reminds us that science is a social, symbolic, and fundamentally intersubjective process of building shared explanations —consensus—, not just optimizing prediction.
And models? Oh my dear, I know this deeply upsets those who have declared war on science fiction, but models are fiction. Nancy Cartwright’s famous thesis (notably from her 1983 book How the Laws of Physics Lie3) is that the “laws” of physics are highly simplified, idealized models that don’t actually describe the messy reality of the universe perfectly—they just tell useful, localized stories:
A model is a work of fiction. Some properties ascribed to objects in the model will be genuine properties of the objects modelled, but others will be merely properties of convenience
Will AI transform science? Sara Imary thinks so. Will it replace scientists? Certainly not.
Will AI transform science? Certainly. Will it replace scientists? Certainly not. If we misunderstand what science is, mistaking automation of method for the human project of collectively constructing, debating and refining the symbolic representations through which we make sense of reality, AI may foretell the death of science: We will miss the true opportunity to integrate AI into the culture systems of science.
Science is not merely about prediction and automation; history tells us it is much more. It is about explanatory consensus, and an ongoing human negotiation of which descriptions of the world we will collectively adopt. That negotiation, the intersubjective binding of observations into shared meaning is irreducibly social and, for now, irreducibly human.
25 years ago Ted Chiang was optimistic. He concluded:
We need not be intimidated by the accomplishments of metahuman science. We should always remember that the technologies that made metahumans possible were originally invented by humans, and they were no smarter than we.
The ultimate irony might not be that AI models are too complex for the human mind to understand, or that we are facing an unprecedented paradigm shift. Quite the contrary, it might be the one thing that remains stubbornly unchanged: human nature. The state, the power, and the hidden ambition of autocrats to control everything.
Who Owns the Future of Science?
____________________
(1) Chiang, Ted. ‘Catching Crumbs from the Table’. Nature 405, no. 6786 (2000): 517–517. https://doi.org/10.1038/35014679.
(2) Evans, James, and Eamon Duede. ‘After Science’. Science 390, no. 6774 (2025): eaec7650. https://doi.org/10.1126/science.aec7650.
(3) Gottweis, Juraj, Wei-Hung Weng, Alexander Daryin, et al. ‘Accelerating Scientific Discovery with Co-Scientist’. Nature, 19 May 2026, 1–3. https://doi.org/10.1038/s41586-026-10644-y.
(4) Ghareeb, Ali Essam, Benjamin Chang, Ludovico Mitchener, et al. ‘A Multi-Agent System for Automating Scientific Discovery’. Nature, 19 May 2026, 1–3. https://doi.org/10.1038/s41586-026-10652-y.
(5) Mikula, Sarah. ‘AI Might Jeopardize the Uncertainty Required in Science’. Nature 653, no. 8115 (2026): 974–974. https://doi.org/10.1038/d41586-026-01605-6.
(6) Ahart, Jenna. ‘AI Can “Same-Ify” Human Expression — Can Some Brains Resist Its Pull?’ Nature, ahead of print, 11 March 2026. https://doi.org/10.1038/d41586-026-00781-9.
Feature Image: A 2026 reimagining by Gemini of the original artwork from Ted Chiang’s 2000 Nature publication. While the Y2K original envisioned a technological “Upgrade” to the human mind, this creative adaptation reflects today’s starker reality: a world of independent, autonomous AI science where the human observer is left on the periphery, watching a synthetic brain process high-dimensional truths while the background warnings whisper of data manipulation and state control. (This description of the image was written by Gemini, to whom I granted the right to speak here 😉
Mind the Post 