Google DeepMind's AI Solves Nine Open Erdős Problems, Including Two Uncharted for 56 Years
Google DeepMind announced on 24 May 2026 that its AI agent had autonomously solved nine of the 353 open Erdős problems — a collection of mathematical conjectures named after the prolific Hungarian mathematician Paul Erdős, many of which have remained unsolved for decades. Among the nine was a pair of problems that had resisted proof for 56 years.
The announcement, posted to the Polymarket platform and amplified by the Cointelegraph wire on 25 May, marks what the research community is rapidly framing as a step-change in automated mathematical reasoning. DeepMind's system worked through the corpus independently, without the iterative human prompting typically required by large language models.
Nine Problems, Two Half-Century Holdouts
Paul Erdős, who died in 1996, was among the most prolific mathematicians in the twentieth century. He formalised the concept of the Erdős number — a tongue-in-cheek measure of collaborative distance from his own published work — and he posted悬赏 problems — prize problems — to mathematicians worldwide, funding awards from a personal fund for solved cases. The 353 problems that remained open at the time of his death represent some of the combinatorics, number theory, and graph theory discipline's most stubborn open questions.
DeepMind's agent solved roughly 2.5 percent of that outstanding backlog in a single automated run. Two of the nine problems had been formally open since the late 1960s, placing them among the most durable unsolved items in the collection. The identity of the specific problems was not detailed in the initial announcement, though DeepMind indicated full proof writeups would be published through standard mathematical venues.
The result lands alongside prior DeepMind milestones — including AlphaFold's disruption of protein structure prediction — in a pattern of AI systems making genuine contributions to established scientific disciplines rather than merely synthesising existing knowledge.
A Different Order of Reasoning
What distinguishes this from earlier language-model results is the autonomous and self-contained nature of the proof generation. Mathematical proof requires not merely pattern recognition but logical chaining across multiple variables, an ability to know when a proposed path has failed, and the formal discipline to construct arguments that satisfy peer review. Systems that pass through plausible-sounding text without that structure have long been dismissed as sophisticated parrots. DeepMind's announcement suggests its agent operated differently — producing outputs rigorous enough, the company claims, to withstand scrutiny from working mathematicians.
The announcement did not specify whether independent mathematicians had already verified the nine proofs as this article went to publication. That caveat matters: the history of automated reasoning includes high-profile claims that collapsed under expert review. The mathematical community's response — whether the proofs hold or whether gaps emerge — will define whether this is a landmark or an overclaim.
Oil markets registered the broader signal simultaneously. Brent crude fell nearly 5 percent on 24 May, reaching a two-week low as traders priced in growing optimism around a US-Iran nuclear framework agreement. That parallel movement reflects a market reading of DeepMind's announcement not narrowly — as a discrete AI result — but as part of a broader acceleration in AI capability that carries investment and industrial implications. Lower oil prices compress margins for energy-intensive AI training runs, a relationship traders are now beginning to price into longer-dated contracts.
Structural Weight: What the Pattern Implies
The DeepMind result sits within a structural shift in how AI systems are being deployed against high-value, formally structured problem domains. Mathematics, unlike language, offers built-in verification: a proof either holds or it does not. That makes it a unusually clean testbed for claims about autonomous reasoning — and unusually embarrassing when those claims fail.
For the AI industry, the commercial implications are significant. Proof-assistance software already exists in niche markets — formal verification tools for software, computer-assisted theorem provers used in academic settings. An AI agent that can generate novel proofs at scale across combinatorics and number theory would alter the economics of fields that have relied on specialist human reasoning. It would also intensify the ongoing debate about what constitutes a genuine intellectual contribution when a machine produces the argument and a human signs off on it.
For the Ethereum Foundation, announced separately on 24 May, the context is different but the structural dynamic is analogous. Foundation co-founder Vitalik Buterin said the organisation was moving toward a leaner role with reduced ETH sales and less direct control over ecosystem expansion. The parallel is not exact, but both stories reflect institutions reckoning with a shift in where authority and capability sit — whether in the case of mathematics, between human proof specialists and automated systems, or in the case of a cryptocurrency foundation, between a central coordinating body and a decentralised network that no longer requires the same degree of institutional management.
What Comes Next
The critical near-term question for DeepMind's result is peer review. Mathematical history is littered with announced proofs that failed at the detail level — problems where the overall strategy was sound but the execution contained gaps that took months or years to surface. If the nine proofs survive expert scrutiny, the result will accelerate a trend already underway: AI systems moving from synthesis roles into genuinely generative ones, producing outputs that human practitioners must engage with rather than merely verify.
That shift carries consequences for the research economy. Mathematics has historically been a field where human intuition and training created the barrier to entry. Automated reasoning, if it matures, compresses that barrier in ways the academic community has not yet developed norms to address. Credit, citation, career structure, and the meaning of a proof are all concepts that may need renegotiation.
For markets, the immediate signal is the oil-price move: a 5 percent fall in a two-week window, driven by a geopolitical development that is itself contingent on diplomatic outcomes, underscores how fragile consensus on energy pricing remains. The US-Iran framework, if it holds, would add significant barrels to a market that has been pricing geopolitical risk premia for years. That relief trade is now visibly competing with AI-acceleration anxiety — a tension that will define investment flows through the second half of 2026.
This piece was developed from Cointelegraph and Polymarket wire reports. Monexus will follow DeepMind's publication of formal proof writeups for independent verification.
Wire provenance
This editorial synthesis draws on the following public wire/social posts:
- https://t.me/cointelegraph/198987
- https://t.me/cointelegraph/198982
- https://t.me/cointelegraph/198965
- https://x.com/polymarket/status/1924472187697479789