GPT-Next Solves 80-Year-Old Erdős Conjecture for Under $1,000

By VirentaNews Staff — May 26, 2026

GPT-Next, a next-generation AI developed by a team affiliated with OpenAI, has resolved the Erdős unit distance problem—a central conjecture in combinatorial geometry first posed by Paul Erdős in 1946—using less than $1,000 in commercial cloud computing. This breakthrough marks the first time a long-standing mathematical conjecture, resistant to human ingenuity for over eight decades, has been independently solved by an artificial intelligence at such low cost. The result not only confirms a deep mathematical truth but also signals a paradigm shift: AI is no longer merely an assistant to human researchers but a capable agent of independent discovery, with profound implications for the future of science, peer review, and intellectual labor.

Evidence of a Mathematical Breakthrough

The Erdős unit distance problem asks: given n points in the plane, what is the maximum number of pairs that can be exactly one unit apart? Erdős conjectured in 1946 that this number grows no faster than n^(4/3), a bound that has remained unproven despite extensive efforts by generations of mathematicians. GPT-Next approached the problem not through brute-force computation but by synthesizing geometric intuition, graph theory, and probabilistic methods to construct a novel proof framework. The AI generated a 38-page formal argument, later verified by three independent mathematicians using proof assistants like Lean and Coq. The solution adheres to rigorous mathematical standards and has been accepted for publication in Annals of Mathematics, pending minor revisions. The total compute cost, estimated at $879 on AWS and Google Cloud, included 14 hours of training on a sparse mixture-of-experts architecture and 62 hours of symbolic reasoning inference. This cost-efficiency contrasts sharply with previous AI-assisted proofs, such as Thomas Hales’ formalization of the Kepler conjecture, which required years of human oversight and millions in resources.

Key Players Behind the Advance

The breakthrough emerged from a hybrid team at OpenAI’s Fundamental Research Division, led by Dr. Linh Tran and Dr. Elias Mota, who specialize in AI alignment and mathematical reasoning. GPT-Next itself is not a publicly released model but part of a closed research track focused on recursive self-improvement in formal systems. The model was trained on a curated corpus of 12 million mathematical papers, proofs, and competition problems, including the entire arXiv mathematics archive and the Putnam exam database. Crucially, it integrates a symbolic reasoning engine inspired by Lilian Weng’s hybrid neuro-symbolic architectures, allowing it to manipulate abstract concepts with precision. The verification process involved researchers from the Institute for Advanced Study in Princeton and the University of Cambridge, who confirmed the proof’s validity without identifying logical gaps. Notably, Paul Erdős himself offered a $500 prize for a solution—adjusted for inflation and matched by OpenAI, the reward was posthumously awarded to the AI team.

Trade-Offs Between Efficiency and Trust

While the low cost and speed of the solution are revolutionary, they raise urgent questions about epistemic trust in AI-generated knowledge. A proof produced by a neural network, even when verified by humans, may lack the intuitive clarity that mathematicians value—such as the ‘aha’ moment that accompanies human insight. This opacity risks creating a dependency on AI systems whose reasoning pathways are not fully transparent, potentially undermining the pedagogical and philosophical foundations of mathematics. On the other hand, the democratization of discovery is undeniable: researchers at underfunded institutions can now access AI co-pilots capable of solving problems once reserved for elite teams. The economic implications are equally stark: if frontier models can resolve century-old conjectures for under $1,000, the return on investment in AI research could dwarf that of traditional academic science. However, concerns persist about bias in training data, the risk of false positives in unverified domains, and the potential for overreliance on AI in fields requiring ethical judgment.

Why This Moment Represents a Turning Point

The timing of this breakthrough is no accident. Over the past two years, AI systems have crossed a qualitative threshold in handling abstract reasoning, driven by advances in chain-of-thought prompting, program-aided language models, and formal verification integration. GPT-Next builds on earlier milestones, such as AlphaTensor’s discovery of faster matrix multiplication algorithms and Minerva’s performance on Olympiad-level problems. What distinguishes this event is the closure of a problem with deep historical significance using off-the-shelf infrastructure. The feasibility of such discoveries before 2024 was limited by model scale and lack of symbolic grounding; now, with hybrid architectures maturing, AI is becoming a first-class participant in knowledge creation. The fact that GPT-Next solved the conjecture without human step-by-step guidance—only a high-level problem statement—suggests that autonomous discovery is not just possible but increasingly routine at the frontier.

Where We Go From Here

In the next 6–12 months, three scenarios are likely. First, a surge in AI-assisted proofs across number theory, topology, and theoretical computer science, as research labs deploy similar models on open conjectures like the Hadwiger-Nelson problem or the Erdős–Straus conjecture. Second, institutional resistance may grow, with journals demanding full transparency of AI reasoning traces and universities debating authorship norms—will AI be listed as a co-author, or merely a tool? Third, a new class of ‘proof markets’ could emerge, where researchers crowdsource AI compute to test hypotheses, accelerating discovery but potentially exacerbating inequality in access to AI infrastructure. Regardless of the path, the boundary between human and machine insight is now irreversibly blurred.

Bottom line — an AI has done what humans could not for 80 years, not just solving a famous math problem but redefining who—or what—can be considered a discoverer in science.

Source: Reddit