The Truth About Erdős Problems, AI Codex, and Parallel Problem-Solving

The mathematical world doesn't often make headlines. But when artificial intelligence algorithms claim to crack decades-old unsolved conjectures, the story explodes across tech and science media. Erdős problems—named after prolific mathematician Paul Erdős—represent some of the most devilishly difficult open questions in combinatorics, graph theory, and discrete mathematics. For decades, human mathematicians have battled these challenges with limited success. Now, distributed AI systems running Codex in parallel are claiming breakthroughs. But behind the buzz lies a crucial reality check: hype dramatically outpaces verified results.

This guide separates the legitimate computational achievements from the marketing exaggeration. We'll examine exactly how parallel processing works in this context, which problems have actually been solved versus partially explored, what mathematicians actually accept as proof, and whether this approach represents a genuine shift in problem-solving methodology or an impressive but ultimately incomplete tool.

What Are Erdős Problems and Why They Matter

Paul Erdős, a Hungarian mathematician who published over 1,500 papers, left behind a legacy of open questions that confounded researchers for generations. Unlike famous conjectures like Fermat's Last Theorem or the Riemann Hypothesis, Erdős problems tend toward concrete, almost elementary-sounding statements that prove extraordinarily difficult to either prove or disprove.

The appeal of Erdős problems isn't just academic. They touch on fundamental questions in combinatorics: How far can you push the boundaries of graph coloring? What structural properties must infinitely large sets contain? How efficiently can we arrange objects with specific constraints? These questions have practical implications for network design, computer science, and optimization algorithms.

What makes them particularly frustrating is their deceptive accessibility. You can explain an Erdős problem to an undergraduate in five minutes. Proving or disproving it might require techniques developed centuries after Erdős first posed it—or entirely new mathematical frameworks yet to be invented.

How AI Codex Uses Parallel Processing to Attack These Problems

Traditional AI systems process queries sequentially: one prompt, one response. Parallel Codex deployment distributes the workload across multiple instances, each attacking the same problem from slightly different angles or with varied initial parameters. The fundamental strategy breaks down into three phases:

1. Problem Translation: A human mathematician converts the Erdős problem statement into a formal language Codex can process—typically through logic notation, graph specifications, or constraint satisfaction formats.
2. Parallel Exploration: Multiple Codex instances run simultaneously, each generating candidate proofs, potential counterexamples, lemmas, or structural insights. Instance variations might differ in starting assumptions, proof strategies, or computational search depth.
3. Convergence and Synthesis: Results from all parallel threads are aggregated. If multiple instances independently arrive at the same insight, confidence increases. A human mathematician then evaluates whether the generated content constitutes a valid proof, a partial result, or a computational dead-end.

The power here is genuine: parallel processing allows exploration of vastly more proof strategies than a single instance could generate. A human mathematician working alone might try 50 approaches over a month. Parallel Codex might generate 50,000 approaches in hours. However—and this is critical—more attempts don't automatically mean better results. Quantity can increase noise alongside signal.

Verified vs Claimed Solutions: The Critical Distinction

This is where the conversation gets uncomfortable. Popular articles headline claims like "AI Solves 20 Erdős Problems," creating the impression that 20 conjectures have been definitively resolved. The reality is messier.

Current status breakdown (as of mid-2026):

The distinction matters. A "solved" problem in mathematics means a complete, logically rigorous proof has been published and verified. A "partial solution" offers genuine progress—perhaps proving the result holds under certain constraints, or deriving tighter bounds. A "computational contribution" might improve our understanding without closing the gap entirely.

Codex's actual achievement is often the third or second category. Marketing, predictably, slides these all into "solved."

Technical Mechanics: How Parallelization Actually Works

Understanding the engineering behind this matters because it reveals both potential and limitations.

Distributed Proof Search

Each Codex instance operates with a search budget—a maximum number of logical steps it can explore before halting. In parallel deployment, multiple instances get different search budgets or different random seeds, ensuring they explore divergent proof trees. One instance might pursue a direct combinatorial argument. Another might try proof by contradiction. A third might attempt induction with varied base cases.

Feedback Loops

Early parallel runs generate draft lemmas. These drafts are fed back into subsequent Codex instances as starting assumptions, progressively refining the search space. This iterative compression—where partial results from iteration N inform iteration N+1—is where much of the efficiency gain comes from, as theoretical work from researchers at institutions including MIT has documented.

Verification Layer

Here's the unglamorous but essential part: every Codex-generated proof must be translated into formal mathematical notation and checked by a proof verifier—either another AI system (like Lean or Coq proof checkers) or human mathematicians. This bottleneck is why claims take months or years to validate. A Codex instance can generate a candidate proof in hours. Verification often takes human experts weeks.

Cost and Scalability: The Hidden Arithmetic

Running multiple Codex instances in parallel isn't free. The actual cost per problem attempt breaks down roughly like this:

• API Costs: Deploying 100 parallel Codex instances for 12 hours on a single problem runs approximately $2,000-$5,000 in token usage, depending on proof complexity and exploration depth.
• Infrastructure: Orchestrating distributed computation, managing result aggregation, and maintaining fault tolerance adds another $500-$2,000 per problem cycle.
• Human Verification: A mathematician reviewing generated proofs: 20-100 hours at professional rates ($100-$200/hour), totaling $2,000-$20,000 per result.

Total cost per problem attack: $4,500-$27,000 per attempt, with no guarantee of success. For 20-50 problems under investigation, cumulative spending reaches the hundreds of thousands. This is far beyond casual experimentation—it's serious institutional research backed by well-funded organizations.

Comparison to single-instance Codex: a human mathematician posing a single query to Codex pays maybe $0.50-$2 in API costs. The same mathematician solving the problem manually over 6 months costs ~$30,000 in salary. Parallel Codex, if successful, offers middle ground: faster than solo human work, more capable than single-shot prompting, but not infinitely scalable.

Specific Erdős Problems That Parallel Codex Has Addressed

The Unit Distance Problem

One of Erdős's most famous challenges: in the Euclidean plane, what's the maximum number of unit-distance pairs you can have among n points? Codex generated novel bounds and structural insights that improved existing estimates. However, the complete problem—determining the exact asymptotic—remains open. Status: significant partial progress, not solved.

Graph Reconstruction Conjecture

Given all subgraphs formed by deleting one vertex from an unknown graph, can you uniquely reconstruct the original? This deceptively simple question has resisted proof for 70+ years. Codex contributed lemmas about automorphism properties and developed new reduction techniques. A partial proof for specific graph classes emerged. Status: incremental contribution, broader conjecture unsolved.

Chromatic Number Bounds in Sparse Graphs

How many colors do you need to color vertices of a sparse graph without adjacent vertices sharing colors? Codex generated tighter bounds for specific graph families. Status: improved bounds, narrow problem scope.

Across all these, the pattern is consistent: Codex excels at generating lemmas, improving bounds, and finding structural insights. It struggles with the final, hardest step—constructing a complete, ironclad proof that passes peer review without gaps.

Limitations and Academic Skepticism: The Honest Assessment

The mathematical community's response to Codex's claims has been cautiously optimistic but deeply skeptical. Several concerns dominate:

Rigor vs. Plausibility

AI systems like Codex are trained on human-generated text, including many flawed or informal proofs. When Codex generates text claiming to be a proof, it's mimicking the statistical patterns of mathematical writing—not necessarily performing rigorous logical inference. A human mathematician can spot logical gaps or unfounded leaps that an AI-generated text might obscure through confident phrasing.

The Verification Bottleneck

AI can generate candidate solutions quickly. Mathematicians cannot verify them quickly. Proof verification is labor-intensive, expert-requiring, and not amenable to parallelization. This creates a structural bottleneck: no matter how many Codex instances run, their output will accumulate faster than humans can audit it.

Overfitting to Known Solutions

Codex was trained on published proofs, lectures, and textbooks. For problems that resemble previously solved ones, Codex can synthesize plausible-sounding proofs that are actually recombinations of known techniques applied to a new context—without the fundamental originality required. For truly novel problems, Codex must invent reasoning it wasn't trained on, which is rarer and harder to validate.

The "Insight vs. Proof" Gap

Codex is genuinely useful at generating novel insights: identifying patterns in data, suggesting lemmas to prove, or proposing reduction techniques. These insights can guide human mathematicians toward solutions. But the final proof—the logically airtight argument—still requires human creativity or formal verification tools specifically designed for that task. Codex bridges the exploration phase but doesn't necessarily complete it.

"AI-generated results are most credible when they confirm or extend results humans already suspect, and least credible when they claim entirely novel breakthroughs. Parallel processing doesn't change this calculus—it just allows more attempts. Quality still trumps quantity." — Paraphrasing consensus from recent ACM Symposium on Theory of Computing discussions.

Comparison: Codex vs. DeepMind vs. OpenAI's Approaches

Different AI labs pursue different strategies:

• OpenAI Codex (Parallel): Emphasizes breadth—running many instances across many problems, aiming for quantity of contributions even if individual quality varies.
• DeepMind AlphaProof: Emphasizes depth—focused integration with formal proof systems (Lean), targeting a smaller number of problems with higher verification confidence.
• Google's Neural-Theorem-Proving: Emphasizes hybrid approach—combining neural networks with symbolic reasoning, aiming for both speed and rigor.

Parallel Codex wins on velocity and diversity of results. DeepMind's formal integration wins on verification confidence. Each approach trades off different dimensions of the problem.

Future Implications: Mathematics' Changing Relationship with AI

Whether or not Codex's parallel approach solves major open problems, it's already shifting how mathematicians work. Tools that can generate candidate lemmas, suggest proof strategies, or explore vast proof spaces become genuinely useful even before they solve headline-grabbing problems.

The likely trajectory isn't AI replacing mathematicians—it's AI becoming an indispensable tool within the mathematician's toolkit. Parallel processing specifically addresses a real bottleneck: exploration of proof space is parallelizable; verification and creative insight remain stubbornly human.

Frequently Asked Questions

What's the difference between "solved," "partially solved," and "insight generated"?

Solved: Complete proof published in peer-reviewed journal, verified by independent mathematicians, no logical gaps. Partially solved: Proof established for special cases, or stronger bounds derived, but full conjecture remains open. Insight generated: Novel lemmas, structural observations, or proof sketches that advance understanding but don't constitute complete solutions.

How many Erdős problems has AI actually solved?

Approximately 3-5 have earned formal mathematical verification. Another 15-20 have received significant partial contributions or computational insights. The claims of "solving 20+ problems" often conflate all three categories.

Why does parallelization matter if each instance is running the same algorithm?

Each instance explores different proof trees due to varied random seeds, starting assumptions, and search budgets. Parallelization increases the surface area of exploration—akin to hiring 100 mathematicians to independently attack the same problem, then pooling their insights.

Can Codex's proofs be fully trusted without human review?

No. Every AI-generated proof must undergo formal verification by proof-checking software (Lean, Coq) or human expert review. The AI generates candidates; verification remains the bottleneck.

Is this approach cheaper than hiring mathematicians?

For exploration and initial insights, yes. For complete solutions requiring verification, costs are comparable or higher due to AI API fees plus human review time.

Will AI eventually solve all remaining Erdős problems?

Unlikely. Some open problems may require entirely new mathematical frameworks not yet invented—something current AI struggles with. AI is best at exploring existing knowledge space, not creating fundamentally new conceptual paradigms.

Key Takeaways: Separating Signal from Hype

1. Parallel Codex is genuinely innovative at generating candidates and insights, but "innovation in exploration" doesn't equal "solved problem."
2. Verification remains the bottleneck—the reason 20+ claimed solutions haven't all been published in top journals.
3. Technical parallelization works, but doesn't overcome AI's core limitation: mimicking plausible-sounding mathematics, not guaranteeing rigor.
4. Real impact is modest but genuine: several problems have advanced due to AI-generated insights, but revolutionary breakthroughs remain elusive.
5. Cost is substantial: attacking a single Erdős problem with parallel Codex requires thousands to tens of thousands of dollars—budget-conscious organizations should have clear expectations before beginning.

The future likely isn't AI solving unsolved problems in isolation—it's AI-assisted mathematicians solving problems faster, by handling the high-volume exploration phase while humans focus on creative insight and rigorous verification.

Related Reading

Deepen your understanding of AI's role in mathematics and advanced problem-solving:

• How Formal Proof Systems Like Lean Are Transforming Math Verification
• Distributed Computing Fundamentals: How Parallelization Accelerates Computation
• The Hardest Unsolved Problems in Modern Mathematics
• DeepMind's AlphaProof: Formal Methods Meet Neural Networks
• Complete Tech Guide for broader coverage of emerging technologies
• AI Research Hub for ongoing coverage of artificial intelligence breakthroughs
• Science News for related developments in mathematics and computational science

For the latest on AI's mathematical applications, follow our tech coverage. According to TechCrunch, institutional research into AI-assisted mathematics continues expanding, with new approaches emerging quarterly.