!["Another explanation is that the solution required ideas from fields that most researchers working on this [Erdos unit distance] problem were unfamiliar with... these explanations should make us uncomfortable. They suggest that the incentives toward specialization and siloing, however understandable they may be, have deprived us of high-quality scientific work." - Daniel Litt - Mathematician](https://i0.wp.com/globaladvisors.biz/wp-content/uploads/2026/05/20260521_05h01_GlobalAdvisors_Marketing_Quote_DanielLitt_MW.png?fit=1200%2C1200&ssl=1 "20260521_05h01_GlobalAdvisors_Marketing_Quote_DanielLitt_MW")
“Another explanation is that the solution required ideas from fields that most researchers working on this [Erdos unit distance] problem were unfamiliar with… these explanations should make us uncomfortable. They suggest that the incentives toward specialisation and siloing, however understandable they may be, have deprived us of high-quality scientific work.” – Daniel Litt – Mathematician
The disproof of a famous geometric conjecture using algebraic number theory and a large language model is not just a curiosity of modern mathematics; it is a case study in how institutional incentives can steer entire communities away from fertile ideas 1. For decades, an apparently combinatorial question about distances between points in the plane resisted traditional attacks. When a solution finally emerged, the crucial tools came from a direction that many of the field’s specialists had not seriously invested in. The episode crystallises a broader problem: structures that reward narrow expertise, short feedback cycles, and well-mapped research programmes can quietly suppress the kind of boundary-crossing work that transformative discoveries increasingly demand.
At issue is not merely that one problem took longer to solve than it might have. It is that an entire research ecosystem can become locally optimised yet globally suboptimal. Individual researchers make rational decisions: specialise deeply, follow the literature of your own area, build on the techniques that your peers recognise and review committees understand. Collectively, however, these decisions can create blind spots where powerful methods remain underused, or entire theoretical bridges are never built. When a breakthrough comes from an unexpected junction of fields, facilitated by a tool few insiders fully understand, it becomes a mirror held up to these collective choices.
The Erd?s unit distance problem provides a particularly vivid example because its formulation seems so elementary. Place n points in the Euclidean plane and ask: how many pairs of points can be exactly one unit apart? Erd?s conjectured that the answer grows at most on the order of n^{1+o(1)} as n increases 2. A simple lattice construction already shows that one can achieve roughly n^{1+1/2} unit distances, i.e. on the order of n^{3/2} 3. The challenge was to prove that no configuration can do substantially better than almost linear in n. Despite intense attention, the best known upper bound stayed stuck at something like O(n^{4/3}) for decades, a substantial gap from the conjectured behaviour 2,3.
That such a long-standing problem would eventually be resolved is not surprising. What is striking is how it was resolved. The solution combined ideas from algebraic number theory, sophisticated combinatorics, and automated reasoning using a powerful AI system 1. For many experts in discrete geometry, deep algebraic number theory sits outside the core toolbox they use every day. Conversely, number theorists might be only passingly familiar with the fine-grained incidence geometry and extremal combinatorics that frame unit distance questions. The fact that the argument could be assembled at all depended on a small group of mathematicians who spanned several of these areas, and on an AI model capable of exploring lemmas and calculations at scale.
This is where the question of incentives becomes inescapable. The mathematical community had no shortage of talent over the decades in which the unit distance problem remained open. What it arguably lacked was a dense mesh of people at the intersections: researchers who were not only technically capable in at least two distinct subfields, but also rewarded for spending years developing that mixed expertise. In most academic systems, such profiles are harder to cultivate. Hiring committees are organised by field; journals are structured by area; funding panels ask for clear, disciplinary narratives. A young mathematician who spends substantial time learning algebraic number theory while nominally working in discrete geometry may appear unfocused or risky on paper, even if exactly that breadth is what some of the hardest problems now require.
The geometry-number theory interface and hidden toolsets
When a problem about distances between points in the plane calls on algebraic number theory, it is natural to ask whether this is a quirky accident or a sign of deeper structure. In modern mathematics, geometric configurations often encode arithmetic information. Classic examples include the way rational points on curves correspond to solutions of Diophantine equations, or how algebraic geometry packages systems of polynomial equations into varieties whose properties can be studied using cohomology and sheaf theory. In that sense, it is not inherently surprising that counting unit distances might eventually touch fields concerned with algebraic relations among lengths, coordinates, and symmetries.
Yet, for many years, the dominant line of attack on the unit distance and distinct distance problems went through combinatorial and analytic techniques such as incidence geometry, crossing number inequalities, and polynomial partitioning 2,3. A canonical step in this tradition is to bound the number of incidences between points and curves. A generic statement might be of the form I(P,C) \leq A\,|P|^{\alpha}\,|C|^{\b\eta} + B(|P|+|C|), where I(P,C) counts incidences between a set of points P and curves C, and A,B,\alpha,\b\eta are constants depending on the family of curves under consideration. Such inequalities are geometric-combinatorial in flavour; they say little directly about algebraic dependencies among distances expressed in number-theoretic terms.
The new work essentially overlays another layer of structure. Distances between points with algebraic coordinates can live in number fields, and patterns of equal or repeated distances can reflect algebraic relations among those coordinates. One can imagine assigning to a configuration a field K generated by the coordinates of all points, and then studying the action of the Galois group of K on the configuration. In favourable circumstances, this action constrains which patterns of equal distances are possible, limiting the total number of unit distances. This kind of reasoning is familiar to algebraic number theorists but far less so to many geometers trained in the combinatorial tradition.
If most researchers tackling the unit distance problem are steeped in incidence bounds but not in Galois theory, they may fail to see that a problem formulated with Euclidean metrics can be reinterpreted in the language of fields and automorphisms. The conceptual bridge is short in hindsight but long in foresight. It requires a mental habit of asking: can these geometric invariants be recast as algebraic invariants, and if so, do the known theorems of algebraic number theory have something to say here? That style of thought is not simply a matter of intelligence; it is cultivated by training, environment, and the tacit messages about which connections are worth exploring.
AI as a catalyst and a mirror
The role of AI in this story complicates the picture further. A large language model, trained on massive corpora of mathematical text and code, does not experience specialisation in the way human researchers do. While its internal representations are not literally a library of cleanly separated domains, the model effectively has access to patterns spanning algebraic number theory, combinatorics, geometry, and beyond. When prompted appropriately, it can propose lemmas, outline proof strategies, or search for counterexamples that weave together tools no single specialist has fully mastered. In the unit distance case, an internal OpenAI model contributed crucial steps that human mathematicians then verified and refined 1.
This dynamic makes AI both a catalyst for cross-field reasoning and a mirror reflecting the human community’s blind spots. The model does not know which subfield a problem is officially assigned to; it simply sees formal structures and analogies. If there is a surprising number-theoretic route to a combinatorial geometry question, the model may stumble across it more readily than a human trained inside disciplinary boundaries. But the model is not an independent agent of scientific progress. It operates within human-curated workflows: mathematicians decide which prompts to issue, which candidate arguments to examine, and which lines of reasoning to trust. Where humans have no conceptual foothold or no incentive to explore a particular direction, AI’s potential contributions may never be solicited.
In that sense, the successful collaboration between an AI system and human mathematicians in cracking the unit distance conjecture does not fully exonerate the existing incentive structures. One might argue that, had the incentives been better aligned toward cross-pollination, a similar synthesis of algebraic and geometric ideas could have been attempted earlier, with or without AI. Instead, AI arrived at a moment when the necessary mathematical tools already existed but were not yet assembled in the right configuration. It helped search and assemble them, but the fact that they were lying dormant highlights the cost of siloed research habits.
Specialisation as both engine and constraint
To understand why this situation is not easily fixed, it helps to acknowledge the genuine power of specialisation. Modern mathematics is vast, and deep progress often requires years of focused immersion in a narrow area. Specialised communities develop technical languages that allow rapid communication of complex ideas. They build collective intuitions about what works and what does not, enabling researchers to avoid re-deriving known dead ends. In this sense, incentives toward specialisation are not arbitrary; they reflect the reality that no one can be an expert in everything.
The problem arises when the marginal gains from further specialisation are overvalued relative to the gains from developing a second or third competency. Consider a hypothetical young researcher choosing between two paths. On one path, they invest most of their time in mastering finer details of existing incidence geometry techniques, aiming to push a known exponent in a long-studied bound from, say, 4/3 to 4/3 - 1/100. On the other path, they devote substantial energy to learning algebraic number theory and the theory of Galois representations, with a hazier sense of how these ideas might eventually feed back into problems like unit distances. The first path is legible to hiring committees and already established specialists; the second looks speculative and harder to evaluate.
Given the short-term incentives around papers, grants, and jobs, it is unsurprising that many choose the first path. Over time, this collective bias can produce an ecosystem with very high local competence but relatively few people with the breadth to even suspect that an algebraic perspective could be decisive in a geometric question. It is not that nobody could, in principle, learn both sets of tools; it is that those who might are nudged away from doing so. The eventual breakthrough then looks like a stroke of genius or a lucky accident, but it may also be the visible tip of a submerged structure of missed opportunities.
Objections: is discomfort really warranted?
Some mathematicians and philosophers of science might push back on the idea that such episodes should make us uneasy. One line of argument goes like this: knowledge is path-dependent; some problems simply require a certain critical mass of tools and concepts that naturally emerge only after decades of work. The fact that the unit distance conjecture required algebraic number theory and AI-informed exploration does not imply that the community did anything wrong. It may merely indicate that the problem was genuinely hard and that the confluence of ideas needed to solve it took time to arrive.
A related objection emphasises that specialisation is a rational response to complexity. If everyone tried to hedge their bets by training broadly, the community might lack the deep specialists needed to make incremental but essential progress on core theories. Moreover, the existence of AI systems that can trawl across disciplines could alleviate some of the cost of specialisation. Specialists remain in their lanes, and AI plays the role of a connective tissue, spotting analogies and suggesting cross-field borrowings when useful. On this view, the combination of specialised humans and broadly trained models could, if anything, be an optimal division of labour.
These objections have force, but they are incomplete. The issue raised by the unit distance episode is not that specialisation exists but that its current incentive structure may be skewed. If the community systematically under-rewards people who build bridges and over-rewards incremental refinements within siloed areas, then even AI’s connective role will be limited. After all, someone has to ask the AI the right questions, recognise which of its suggestions are promising, and craft a coherent narrative that other humans can understand and trust. That work is more easily done by researchers who themselves span multiple domains. If such people are rare or precariously positioned, the system as a whole may still underperform.
Rebalancing incentives in an AI-enhanced era
The arrival of general-purpose AI systems capable of non-trivial mathematical reasoning creates an opportunity to rethink research incentives rather than an excuse to ignore them. One concrete direction is to shift some evaluative weight from short-term publication counts toward demonstrable cross-field competence. Hiring and promotion committees could, for example, treat high-quality work that genuinely integrates two distinct areas as especially valuable, even if the publication record is modest compared with a pure specialist’s. This would require careful judgment, but it would signal that building intellectual bridges is a recognised scholarly contribution, not a risky side project.
Another lever lies in funding schemes designed explicitly to encourage interdisciplinary mathematical work that engages with AI as a collaborator rather than a black box. Grants could support teams that combine, say, a discrete geometer, an algebraic number theorist, and an AI specialist, with the explicit aim of attacking problems known to require diverse tools. Critical to such programmes is the expectation that AI’s role will be documented and scrutinised, leading not just to solved problems but to new conceptual frameworks that humans can internalise. Otherwise, there is a risk of offloading too much of the reasoning to opaque systems, replicating the very siloing problem at a higher level.
Graduate training is another site for change. Instead of curricula that channel students quickly into narrow tracks, departments could build structured opportunities for cross-training: seminars co-taught by experts from different fields, reading groups on problems known to have multiple conceptual incarnations, and supervised projects that explicitly mix methods. Exposure to AI tools should be folded into this, not as a replacement for learning hard theory but as a way to explore how diverse theories interact. The goal is to normalise the idea that a serious mathematician may develop fluency in seemingly disparate areas and in AI-assisted exploration, without being viewed as unfocused.
Why this episode matters beyond mathematics
Although the unit distance conjecture is a mathematical story, its underlying tension applies widely in science and technology. In many domains, from biology to climate modelling to economics, major breakthroughs increasingly emerge at the intersections of fields: genomics and statistics, physics and machine learning, behavioural science and network theory. As AI systems grow more capable, they are poised to act as cross-disciplinary amplifiers. However, if human institutions reward siloed expertise while relying on AI to stitch everything together, we risk creating a fragile arrangement where few humans truly understand the integrated picture.
In areas with direct societal impact, such as AI safety, epidemiology, or financial stability, that fragility becomes a governance issue. Decisions may rely on chains of reasoning that no single community fully owns or scrutinises. If an AI system suggests a policy-relevant result using techniques drawn from several fields, but no one feels both qualified and incentivised to examine the full path, accountability suffers. The warning implicit in the unit distance story is that the same structural forces that delayed a mathematical breakthrough could, in other contexts, lead to misjudged risks or overlooked solutions.
The discomfort that arises from recognising these patterns is therefore not a call for nostalgia or a rejection of modern research architecture. It is an invitation to re-engineer that architecture in light of what recent events reveal. Mathematics did eventually produce the tools needed to crack the unit distance problem, and AI helped assemble them. But the delay and the surprise are signals that the current balance between depth and breadth, between specialisation and synthesis, and between human and machine reasoning is not yet optimal. Taking those signals seriously is part of the responsibility that comes with wielding powerful new tools in a world where the easiest path for individuals may not be the best path for knowledge.
References
1 D. Litt, social media commentary on the Erd?s unit distance conjecture and AI-assisted proof.
2 Erd?s distinct distances and unit distance problems, overview articles in discrete geometry.
3 E. Szemerédi, “Erd?s’s Unit Distance Problem,” in Open Problems in Mathematics, Springer.
References
1. https://x.com/littmath/status/2057180667104870874 – https://x.com/littmath/status/2057180667104870874
2. Daniel Litt – https://www.daniellitt.com
3. Erd?s distinct distances problem – Wikipedia – 2010-12-26 – https://en.wikipedia.org/wiki/Erd%C5%91s_distinct_distances_problem
4. (Fun)damental Uses of AI in Math: Research – 2026-02-05 – https://datascience.columbia.edu/event/fundamental-uses-of-ai-in-math-research/
5. AI math capabilities could be jagged for a long time – Daniel Litt – 2026-01-29 – https://epochai.substack.com/p/ai-math-capabilities-could-be-jagged
6. Amazing: Erd?s’ Unit Distance Problem was Disproved! It was … – 2026-05-20 – https://gilkalai.wordpress.com/2026/05/21/amazing-erdos-unit-distance-problem-was-disproved-it-was-achieved-by-ai/
7. Advancing science and math with GPT-5.2 | OpenAI – 2025-12-11 – https://openai.com/index/gpt-5-2-for-science-and-math/
8. Daniel Litt on AI and Math – Marginal REVOLUTION – 2026-02-23 – https://marginalrevolution.com/marginalrevolution/2026/02/daniel-litt-on-ai-and-math.html
9. Erd?s’s Unit Distance Problem – IDEAS/RePEc – 2016-02-02 – https://ideas.repec.org/h/spr/sprchp/978-3-319-32162-2_15.html
10. Improving mathematical reasoning with process supervision – OpenAI – 2023-05-31 – https://openai.com/index/improving-mathematical-reasoning-with-process-supervision/
11. Daniel Litt – Semantic Scholar – https://www.semanticscholar.org/author/Daniel-Litt/81119117
12. 90 | Erd?s Problems – https://www.erdosproblems.com/90
13. New ways to learn math and science in ChatGPT – OpenAI – 2026-03-10 – https://openai.com/index/new-ways-to-learn-math-and-science-in-chatgpt/
14. AI math capabilities could be jagged for a long time – Daniel Litt – 2026-01-29 – https://www.youtube.com/watch?v=jFJku8sxLWY
15. Solving a Conjecture of Erd?s – Rényi AI – 2023-12-10 – https://ai.renyi.hu/posts/solving-erdos-conjecture/
16. The Future of Math Research in the Age of AI – Silicon Reckoner – 2026-02-21 – https://siliconreckoner.substack.com/p/the-future-of-math-research-in-the
