New Math, New Margins: What AI Math Means for Operations

On July 20, Harvard mathematician Levent Alpöge announced a counterexample to the Jacobian conjecture. The problem had been open since 1939. He found the counterexample with Claude Fable 5.

The solution is, by the standards of the field of mathematics, surprisingly simple. But you don't need to follow the algebra to understand the story:

87 yrs

The problem was open, from Keller's 1939 formulation to July 2026

3 lines

The length of the counterexample. It is explicit, and can be checked by hand

The answer is easy to check but was difficult to find. The space of possible counterexamples is enormous. The real work was to search it and recognize the right object. In other words, AI systems are getting better at finding a needle in a haystack.

Now, this counterexample will not change warehouse operations tomorrow. Or the next day. But the same technology that cracked the Jacobian Conjecture can design algorithms, build optimization models, and check constraints. Routing, scheduling, forecasting, and pricing all depend on that technology.

How AI mathematics developed

In 2020, language models struggled with basic arithmetic. They proved capable of memorizing simple operations but little more. How much difference can six years make? Well, models learned to use tools, to write and run code instead of calculating in prose, and to check their own work against mechanical software not prone to the sort of hallucinations that pester LLMs. By 2024, they were solving Math Olympiad problems, albeit with some handholding. A year later, they could do so from the written problems alone, in natural language, inside the time limit. In 2026, the results crossed from competition into research: formal proofs of open conjectures, novel counterexamples, and now the Jacobian result.

AI mathematics, from imitation to discovery

1. 2020GPT-3 copies patterns.The model reproduces known patterns and does basic calculations. It loses track in multi-step reasoning.
2. Nov 2022ChatGPT explains.The model gives mathematical help in conversation to everyone. Its proofs look correct but often are not.
3. Dec 2023FunSearch finds new results.DeepMind's LLM-driven search produces the first verified new results on open problems. This includes bin-packing methods that apply to data centers.
4. Jul 2024Olympiad silver.AlphaProof and AlphaGeometry 2 perform at silver-medal level on International Mathematical Olympiad problems.
5. Jul 2025Olympiad gold, in natural language.Gemini Deep Think performs at gold-medal level. It reads the written problems and writes its proofs in natural language, in the time limit.
6. Feb 2026Research proofs start.OpenAI sends research-level proofs for review, and withdraws one after expert criticism. DeepMind's Aletheia reads the literature and corrects its own proofs.
7. May 2026Open problems start to fall.Formal agents solve 9 of 353 attempted Erdős problems and 44 of 492 OEIS conjectures. An OpenAI model disproves a central conjecture in discrete geometry.
8. Jul 2026The Jacobian conjecture falls.Alpöge and Fable 5 produce an explicit three-line counterexample to the 87-year-old conjecture.

Three things drove this unprecedented acceleration. First, the models themselves improved: each generation reasons over longer chains with fewer dropped steps, and much of the progress in the timeline is simply that. Second, tools reduced the calculation errors: models learned to run code and call symbolic software rather than attempt arithmetic in prose. Third, proof assistants such as Lean allow for quick and accurate evaluations for the validity of mathematical proofs. This is very helpful. A system can propose 1,000 candidate arguments, and the proof assistant rejects the 999 incorrect ones. Search starts to do the work of scarce expertise.

And thus a revolution is born. Until now, bespoke mathematical work was out of scope for most businesses. This was not without good reason. Mathematical skill was a deeply limited resource: few people, long training, slow output. The rate of appearance for research-grade results is low, and most output is not important. AI overturns this equilibrium. Math is now a commodity.

Your business already uses mathematics

The Jacobian conjecture has no direct effect on businesses. But it is a clue to something larger. It shows that AI can search vast mathematical spaces. It shows that the output can be compact and easy to check. Discovery is becoming scalable with compute.

Businesses use applied mathematics constantly, usually without the name. Delivery routing, inventory policy, workforce schedules, manufacturing plans, cloud capacity, prices, financial risk, fraud detection, energy management, and network design are all mathematics problems. Companies see them as cost, service, capacity, and risk problems.

Examine how companies maintain this mathematics today. The solver in your planning software contains decisions from an operations-research team. That team made those decisions years ago, for a generic customer, with assumptions that your company has never seen. To adapt the solver, a company must hire specialists from a small pool, or pay for a consulting project. Most do neither. They adapt the business to the tool. The cost is permanent: extra capacity, extra stock, and manual corrections at each point where the template does not match the real operation.

How AI mathematics applies to operations

Custom optimization instead of generic workflows. Most companies use fixed business rules, standard solver packages, spreadsheets, and heuristics often arriving prepackaged. A system that does mathematics on demand changes the cost of adaptation. It can find constraints in policies and operational data. It can separate hard constraints from preferences. It can propose objective functions. It can design solution methods for your specific network. It can update the model when conditions change. The gain isn't merely faster calculation, it's optionality. The gain is custom models of your operation, at the price of templates.

Algorithm discovery gives repeated gains. Small algorithm improvements have large effects at scale. DeepMind's AlphaEvolve is a Gemini-based agent that improves code against an evaluator. Its data-center scheduling method is in production at Google:

0.7%

Of Google's worldwide compute, recovered continuously by an AlphaEvolve scheduling method

23%

Speed increase on a Gemini training kernel. This cut total training time by approximately 1%

100M

Miles of driving each year that UPS expected its ORION routing system to remove

A gain below one percent, applied to every machine every day, is not a one-time saving. It returns value continuously. IBM reports the same pattern in heavy industry. Its real-time site optimization at Suncor has an estimated value of tens of millions of dollars each year. The gain for each decision is small. The gain for each year is large.

Forecasting and simulation become adaptive. The system decides which variables are still important. It estimates uncertainty honestly. It detects when a historical relation stops. It uses the simulation budget on the scenarios that matter most. Planners know the current failure mode. A demand model learns from years of stable history. Then a structural break occurs. The model continues to predict with confidence, because nothing in the model knows that the world changed. The company finds the error at the end of the quarter. An adaptive system fails differently. It reports the break, increases its uncertainty, and moves simulation runs to the new scenarios. The end state is not a quarterly forecast on a calendar. It is an operational model that is only as old as its data.

Verification adds guardrails. The same logic tools that check proofs can check decisions. You do not only ask: "Which action is best?" You also ask: "Show that this action obeys each mandatory constraint." The first applications are the areas of the business with many written rules: authorization policies, financial controls, safety limits, contracts, cybersecurity rules, and regulatory workflows. In these areas the rules are already written, and that is what formal checks need. A violation also costs much more than a check. AWS sells a version of this today. Its Automated Reasoning checks compare AI outputs with formal policy rules. Its PitCrew case study reports that compliance cross-checks went from two weeks to approximately 30 minutes. Now you should always take vendor numbers with a grain of salt, but the direction is correct. We've previously written about why audit trails are necessary for agents in real workflows.

An example company

“Bespoke algorithmic solutions, tailored to their exact needs.”

Consider a regional manufacturer and distributor. It has several factories, multiple warehouses, variable demand, limited transport capacity, energy-intensive production, and service-level agreements. Today, six teams manage inventory, production, routing, staffing, purchasing, and energy separately. Each team has its own spreadsheet, software, assumptions, and KPIs. The six plans first meet in the real operation, and there they conflict.

Now give that company one shared operational model, built and maintained with AI mathematical tools. Business rules become explicit constraints. The system optimizes production, inventory, routing, and energy together, not in sequence. It searches for algorithms that match this specific network: bespoke algorithmic solutions, tailored to their exact needs. It simulates disruptions before they occur. It verifies that plans obey safety limits and contract limits. It explains the important trade-offs in language that plant managers can easily digest.

A thought experiment: 100 times more mathematics

Suppose AI systems expand the stock of correct mathematics to 100 times its current size. The number of theorems is not the important measure. The important question is different: can machines search, rank, compress, and apply them? A library that nobody can search is storage, not capability.

The first-order effects are practical: better optimization methods, faster numerical algorithms, lower simulation costs, stronger error-correcting codes, sharper cryptographic tools for attack and defense, and more reliable software verification. The second-order effects are larger, because everything else is built on mathematics. AI training improves. Scientific models become more powerful. Engineering design cycles become shorter. Good simulations replace costly physical experiments. The loop closes: better AI produces new mathematics, new mathematics improves AI, engineering, and science, and those improvements make better tools for more discovery.

A larger corpus does not create economic growth automatically. Implementation depends on data quality, physical capital, regulation, organizational readiness, and the people who deploy the systems. The bottleneck moves from the discovery of the mathematics to the absorption of it.

The failure modes that matter

The obvious risk is faulty algorithms. Formal verification removes most of that risk. But there are risks that can survive verification.

Specification errors. The system optimizes or proves exactly the wrong thing, without an error in the logic. It minimizes transport cost, and delivery reliability decreases. It optimizes labor use, and the schedules become unsafe. It proves compliance against an incomplete reading of a regulation. For businesses, this is probably the larger risk class. The result comes with a certificate, and the certificate is about the wrong question.

Correlated errors. Five agents that agree are not five independent opinions. The agents can have the same training data, the same assumptions, the same software libraries, or the same incorrect source. That is the same as asking one analyst six times. The consequence for review processes is practical. If you want independent checks on a machine-made result, you must engineer the independence: different methods, different data, or a person with different assumptions. AI systems simply agreeing is not verification.

Overload. A large increase in mathematics produces more valid results than people can read. Sorting the important results from the unimportant ones becomes its own discipline. The infrastructure for this is new: machine-readable conjecture registries such as the Formal Conjectures collection with its 2,615 formal statements, provenance records, and independent audits. Businesses get the same problem at a smaller scale. A system that produces 50 valid schedule options each hour does not help a planner who can examine three each day. Ranking results by importance, not only correctness, becomes essential in this new paradigm.

Being too hasty. One benchmark result gives balance. FrontierOR is a May 2026 benchmark of large-scale optimization problems. The best one-shot LLM system beat the industrial solver Gurobi, in efficiency and in solution quality, in only 31% of the tested cases. The progress is real, and so is the gap. AI is able to perform better than traditional tools in some circumstances and this performance will only improve in the coming months. However, it is not superior in all situations.

Where value comes first

The near-term areas have a common profile: decisions that repeat, objectives that you can measure, constraints that you can write, history that you can test against, and pilots that you can run safely. Each property is necessary. Repetition turns a small gain into a repeated gain. Measurable objectives make honest evaluation possible; without them, you grade the system on how good it sounds. Written constraints are the material for optimization and for verification. A low-cost pilot lets you learn before the failure modes become costly.

That profile points to supply chain and logistics, manufacturing schedules, cloud and data-center operations, pricing and revenue management, energy, financial risk and fraud, and software with heavy compliance loads.

The long-term gains are larger and slower: drug discovery, advanced materials, semiconductor architecture, and new algorithmic foundations. The organizations that build the near-term skills first will capture most of them.

What to do now

Six steps. None of them needs a research laboratory.

Find your mathematically dense operations. Look for repeated decisions with limited resources, competing objectives, and hard constraints: schedules, routes, prices, forecasts, and large compute costs.

Correct the representation before the optimization. Write down what you optimize, which constraints are mandatory, which trade-offs are acceptable, and where you record decisions and outcomes. In most companies, this is the bottleneck. The solver is not.

Build evaluation environments. Make backtests, simulations, shadow deployments, baselines, and failure tests. Do not evaluate an operational AI system on how convincing its recommendation sounds.

Start with limited, reversible projects. Select projects where you can check the outputs, deploy gradually, and review the decisions, and where failures have a low cost.

Combine AI with operations-research experts and domain experts. The near-term model is augmentation. AI searches. Solvers calculate. Experts validate. Operators implement. Governance systems monitor.

Keep provenance. Record data sources, assumptions, objectives, constraints, model versions, human overrides, and outcomes. This improves current operations. It also prepares the company for more capable systems.

Conclusion

Discoveries by AI are accelerating and expanding in scope. The same technology that finds counterexamples can find novel schedules, routes, prices, and compliance proofs. Mathematics becomes less costly to produce and easier to deploy. The advantage goes to the companies with clean data, explicit objectives, and disciplined evaluation.

Sources

1. 01 · Announcement of the Jacobian conjecture counterexample — Levent Alpöge on X, Jul 20, 2026
2. 02 · Harvard mathematician uses AI to crack 87-year-old riddle — The Times, Jul 21, 2026
3. 03 · Jacobian Conjecture — Wolfram MathWorld
4. 04 · The new counterexample to the Jacobian conjecture — Secret Blogging Seminar, Jul 20, 2026
5. 05 · FunSearch: Making new discoveries in mathematical sciences using LLMs — Google DeepMind, Dec 2023
6. 06 · AI achieves silver-medal standard on IMO problems — Google DeepMind, Jul 2024
7. 07 · Gemini Deep Think achieves gold-medal standard at the IMO — Google DeepMind, Jul 2025
8. 08 · Our first proof submissions — OpenAI, Feb 2026
9. 09 · Towards Autonomous Mathematics Research (Aletheia) — arXiv, Feb 2026
10. 10 · Advancing mathematics research with AI-driven formal proof search — arXiv, May 2026
11. 11 · An OpenAI model has disproved a central conjecture in discrete geometry — OpenAI, May 2026
12. 12 · Formal Conjectures: an open benchmark for verified discovery — arXiv, May 2026
13. 13 · AlphaEvolve: a Gemini-powered coding agent for designing advanced algorithms — Google DeepMind, May 2025
14. 14 · Large Language Models for Operations Research: a comprehensive survey — arXiv, May 2026
15. 15 · FrontierOR: benchmarking LLMs' capacity for algorithm design in large-scale optimization — arXiv, May 2026
16. 16 · AI-based real-time site-wide optimization for process manufacturing — IBM Research / INFORMS, 2022
17. 17 · 2014 Corporate Sustainability Report (ORION) — UPS, 2014
18. 18 · How Automated Reasoning checks in Amazon Bedrock transform generative AI compliance — AWS, Apr 2026
19. 19 · PitCrew verifies AI agent decisions with automated reasoning — AWS
20. 20 · Why language models hallucinate — OpenAI, 2025