- Is a pure math problem really worth your attention?
- What really happened: the race for mathematical reasoning
- What does "serious overhang" mean for your company?
- Why it's in the news now: the Italian AI market in numbers
- From the Erdős problem to business processes: the concrete link
- Three concrete scenarios: how does the situation change for an Italian company?
- Anthropic vs OpenAI vs Google: which platform should you choose?
- What concrete risks do you need to manage before adopting these tools?
- How to get started without wasting budget
- What mistakes do SMEs make when approaching advanced AI?
- How can you tell if AI adoption is delivering concrete results?
- Is investing in advanced AI right now really worth it for your company?
- How we see it at SHM
- Frequently Asked Questions
- Claude Mythos solved the 1946 Erdős conjecture in a weekend; Google DeepMind solved 9 in the same period with AlphaProof Nexus.
- The formal reasoning capabilities behind these results are the same ones that power logistics optimization, contract review and dynamic pricing.
- 76% of Italian SMEs have yet to invest in AI: the gap with large companies widens every quarter.
- The right strategy isn't rushing to adopt every new thing, but experimenting in a controlled way and scaling only when results are measurable.
Updated October 5, 2026
One weekend to solve a problem that had been open for 80 years. On May 26, 2026, Anthropic's Claude Mythos produced an elegant proof of the Erdős unit distance conjecture, and almost at the same time, Google DeepMind solved nine more with AlphaProof Nexus. For Italian SMEs, the question isn't "how impressive is this result?" but "what changes for me starting tomorrow?"
Is a pure math problem really worth your attention?
Because the capabilities that solved the Erdős conjecture are the same ones AI models use to optimize complex business processes. This isn't an academic exercise: it's a signal of what these tools can already do.
On May 26, 2026, an Anthropic engineer announced that Claude Mythos had solved the unit distance conjecture formulated by mathematician Paul Erdős in 1946. The solution came "over the weekend" and was described as an elegant proof, not a computational brute-force.
Almost simultaneously, Google DeepMind published the AlphaProof Nexus paper on arXiv, which autonomously solved 9 of the 353 open Erdős problems and proved 44 OEIS conjectures (the large online encyclopedia of integer sequences), using formal proofs verified in Lean 4, an automated verification language, at a cost of a few hundred dollars per problem.
Two labs, two different approaches, similar results on the same problems. For the technical details of the original announcement, the original source on The Decoder reconstructs the whole story.
| Your situation | Is it worth a closer look? | why |
|---|---|---|
| Tech SMEs with complex decision-making processes | Yes | Models with advanced reasoning can already automate analyses that today require consultants |
| Manufacturer with complex logistics | Yes | Routing and storage optimization is one of the most mature use cases |
| Professional firm (legal, accounting) | Yes | Contract review and compliance are direct applications of formal reasoning |
| SMEs without structured data or documented processes | Not yet | You need foundations first: clean data and defined workflows |
| Company that has already adopted basic AI | Yes | It's time to consider the leap to models with multi-step reasoning |
What really happened: the race for mathematical reasoning
Within a few months, three of the world's leading AI labs have solved math problems that had been open for decades. The progress is fast and well documented.
The Claude Mythos news doesn't come in isolation. It's part of a clear trajectory that Quanta Magazine and others have traced as the fastest acceleration ever seen in computational mathematics.
| Year | Model | Result |
|---|---|---|
| 2024 | AlphaProof (Google DeepMind) | Silver-medal equivalent at the IMO (International Mathematical Olympiad) |
| 2025 | Gemini Deep Think (Google) | Gold medal at the IMO, solved autonomously within the standard time |
| May 2026 | Claude Mythos (Anthropic) + OpenAI | Erdős unit distance conjecture solved independently |
| May 2026 | AlphaProof Nexus (Google DeepMind) | 9 open Erdős problems solved, 44 OEIS conjectures proved in Lean 4 |
| August 2026 | OpenAI (undisclosed model) | 249-page manuscript with 10 results in pure mathematics and theoretical computer science |
One critical point: Claude Mythos's proof still rests on an informal announcement by an Anthropic engineer. It hasn't yet passed formal review by the mathematical community. As mathematician Terence Tao pointed out in Quanta Magazine, AI without formal validation remains too unreliable for serious applications: the tool that makes results trustworthy is the Lean proof assistant, used precisely by AlphaProof Nexus. A promising signal, not yet a certified result.
What does "serious overhang" mean for your company?
The Anthropic engineer used this phrase to say that the AI models available today are far more capable than what companies are actually using. It's latent potential that hasn't been tapped yet.
In physics, an overhang is a suspended mass ready to fall. In the AI context, the term means that the models already available, not future ones, can tackle problems that seemed unsolvable until yesterday. Most Italian SMEs, however, haven't yet built internal processes that can take advantage of this level of reasoning.
The number of scientific publications on AI applied to formal mathematical reasoning nearly doubled every year in both 2023 and 2024, according to Communications of the ACM (March 2026). Research is racing ahead. Business applications follow 12-24 months behind. Those who move now claim that space.
This data comes from the 2025-2026 Research by the Digital Innovation in SMEs Observatory of Politecnico di Milano, presented on May 21, 2026. The gap between large companies and SMEs isn't about access to tools: it's about priorities and internal structure.
Why it's in the news now: the Italian AI market in numbers
The Italian AI market reached €1.8 billion in 2025, growing 50% in a single year. The moment isn't the future: it's already here.
Generative AI, the models that generate text, code and analysis, accounts for 46% of the total market, according to the Artificial Intelligence Observatory of Politecnico di Milano. It's not a niche: it's already the main component of the sector.
On the jobs front, requests for AI skills in Italian job postings grew by 93% in one year. 76% of offers for qualified roles require at least one, according to the same Observatory. Those who don't train today will find tomorrow's job market already reorganized around these skills.
The ISTAT data is clear: 16.4% of Italian companies with at least 10 employees already use at least one AI technology, double the 8.2% of 2024. The gap between SMEs (15.7%) and large companies (53.1%) has widened to 37 percentage points. This data is available on ANSA/Teleborsa.
From the Erdős problem to business processes: the concrete link
The capabilities that make it possible to solve a mathematical conjecture (constraint-based reasoning, hidden pattern detection, generating non-obvious solutions) are the same ones AI models apply to real business problems.
The Erdős conjecture belongs to pure mathematics. But the skills a model shows when solving that kind of problem don't stay in the lab. They transfer directly to industrial applications.
| Capability shown in mathematics | Concrete business application | Who benefits |
|---|---|---|
| Formal proof under constraints | Contract analysis and automated compliance | Law firms, procurement departments, regulated companies |
| Detecting hidden patterns in complex structures | Logistics and supply chain optimization | Manufacturing, distribution, e-commerce |
| Generating non-obvious solutions | Dynamic pricing with many interdependent variables | Retail, hospitality, SaaS |
| Verifiable multi-step reasoning | Code generation and review with fewer structural bugs | Software houses, in-house IT teams |
| Synthesis of complex structures | Automatic analytical reports from raw data | Finance, marketing, operations |
Three concrete scenarios: how does the situation change for an Italian company?
The competitive advantage of advanced AI isn't the same for everyone. It depends on your company's digital maturity and the complexity of the processes you want to automate.
Three typical scenarios to help you see where your company stands. The numbers are illustrative and show the order of magnitude, not real client data.
| Scenario | Starting situation | Advanced AI application | Expected benefit |
|---|---|---|---|
| E-commerce with 50,000 SKUs | Manual pricing updated weekly | AI model with multi-step reasoning for daily dynamic pricing | Higher margins during high-demand time slots |
| Consulting firm (20 people) | Contract review handled by juniors with senior supervision | AI for first-pass analysis of critical clauses and anomaly flagging | Shorter review time, seniors freed up for strategic work |
| Manufacturing company (components) | Logistics planning on an Excel sheet updated monthly | AI for routing and storage optimization using real-time data | Lower inventory and transport costs on repetitive routes |
Anthropic vs OpenAI vs Google: which platform should you choose?
The fact that three labs reached similar results independently is no coincidence: it's a technology race that's intensifying. For companies, choosing a platform depends on practical criteria, not on who won the math race.
Claude Mythos, the OpenAI models and Google DeepMind's AlphaProof Nexus are all investing in mathematical reasoning as a testing ground for general intelligence. For companies that need to choose which platform to adopt, this scenario raises concrete questions.
It's also worth reading the model comparison in the article on Intelligence Index 4.2: GPT-6 Astra vs Claude Fable 5.1, which analyzes the operational differences between the main platforms.
| Selection criterion | What to evaluate | Practical note |
|---|---|---|
| Reliability of structured reasoning | Benchmarks on tasks similar to your use case | Test with your company's real data, not just public benchmarks |
| Integration with existing systems | Available APIs, native connectors, documentation | Check compatibility with your CRM, ERP or management software |
| Pricing that works for SMEs | Cost per token or API call, usage limits | Calculate the cost on a realistic volume, not on the base plan |
| Data security and compliance | Where data is processed, DPA contracts | Essential for regulated sectors (healthcare, finance, legal) |
| Support and updates | Model update frequency, SLAs | A model that improves every 3 months changes your ROI expectations |
One aspect is often underestimated: security. Claude Mythos has already shown significant capabilities on the vulnerability front too. The article on Claude Mythos and cryptographic vulnerabilities explains what changes for anyone who manages digital infrastructure.
What concrete risks do you need to manage before adopting these tools?
The main risk isn't the technology. It's adopting it without a validation process. An AI model that reasons in an advanced way can also get things wrong in an advanced way.
The pace of these announcements makes objective evaluation hard. Companies face a double risk: ignoring real developments out of excess skepticism, or adopting immature tools based on unverified announcements.
The position of Terence Tao, one of the world's most authoritative mathematicians, is clear: without formal validation, AI results remain unreliable for serious applications. The same principle applies in business: an AI output that hasn't been checked by a human review process isn't ready for critical decisions.
| Risk | How it shows up | How to manage it |
|---|---|---|
| Premature adoption | An immature tool is rolled out on critical processes | Run a pilot on narrowly defined use cases, with metrics set in advance |
| Lack of structured data | The model produces useless output because the input data is disorganized | Clean and document your data first, then integrate AI |
| Dependence on a single vendor | The vendor changes pricing or policy and the company has no alternatives | Design your architecture so you can switch models without starting from scratch |
| No human validation | Wrong outputs are used in business decisions without any checks | Define who approves AI outputs before they go into production |
| Underestimating hidden costs | The real cost includes integration, training and maintenance | Calculate the TCO (total cost of ownership) over 12 months, not just the license |
A related topic is regulatory restrictions on the most advanced models. The article on Claude Mythos 5 and US export controls explains which regulatory constraints may limit access to certain models for European companies.
How to get started without wasting budget
The most effective strategy isn't the fastest. It starts with a precise audit, experiments in a controlled way and scales only when results are measurable.
- 1 · Operations or IT managerAudit of candidate processesMap the business processes that require structured reasoning: document analysis, optimization, reporting. Not every use case needs the same capabilities.
- 2 · Internal team + external consultantData quality checkBefore integrating any AI model, make sure your input data is clean, structured and documented. An advanced model on disorganized data produces useless output.
- 3 · Project managerPilot on a narrowly defined use casePick a non-critical process, define success metrics in advance and launch a 4-8 week test. Measure the time saved and the quality of the output.
- 4 · ManagementEvaluation and scaling decisionIf the pilot delivers measurable results, plan the rollout to other processes. If it doesn't, identify the bottleneck before investing further.
- 5 · Internal teamOngoing training7% of Italian SMEs have structured AI training programs. Those who invest in internal training today will have a structural advantage when next-generation models reach production.
What mistakes do SMEs make when approaching advanced AI?
The most common mistakes aren't about the technology. They're about change management and the lack of a validation process.
- Starting from enthusiasm, not from the problem. Adopting a tool because it's new, without identifying what problem it actually solves in your company.
- Skipping the data phase. Integrating an AI model without first checking the quality and structure of your input data. Garbage in, garbage out is even more true with advanced models.
- Not defining who validates the outputs. Putting AI outputs into production without a human approval process. Advanced reasoning can get things wrong in an advanced way.
- Underestimating integration costs. The license is often the smaller part of the cost. Integration, training and maintenance can triple the total.
- Not communicating the change internally. Introducing AI tools without involving the teams who will use them creates resistance and partial adoption.
How do you know if AI adoption is delivering concrete results?
The indicators to track depend on the use case, but some cross-cutting metrics show whether your investment is working or whether there's a problem to fix.
| Indicator | What it measures | Positive signal |
|---|---|---|
| Average time to complete the process | Operational Efficiency | 20%+ reduction compared to the pre-AI baseline |
| Error rate in outputs | Quality of the model's reasoning | Errors below the average of the previous manual process |
| Percentage of outputs validated without changes | Model reliability on your use case | Above 70% after the ramp-up period |
| Cost per output produced | Economic sustainability | Lower than the cost of the equivalent manual process |
| Internal adoption (% of the team using the tool) | Cultural change | Above 60% three months after rollout |
Is investing in advanced AI right now really worth it for your company?
It's worth it if you have complex decision-making processes with many variables (logistics, pricing, document analysis), structured data and a team willing to experiment. The time is now: the gap between those who adopt and those who wait widens every quarter.
A middle path if you're in a digital transition phase: consolidate your data and core processes first, then consider AI integration. A narrow pilot on a non-critical process is the best way to understand the real potential.
Not yet if you don't have structured data, documented processes or an internal owner who can manage the experimentation. In that case, investing in advanced AI delivers poor results and high hidden costs.
How we see it at SHM
For us, the Claude Mythos news isn't an academic curiosity. It confirms that AI models have already crossed the threshold of formal reasoning, and that the distance between the lab and the average SME is measured in months, not years.
We believe AI should be integrated to free companies from routine, not to replace human ingenuity. The advanced reasoning of these models only makes sense if it serves better decisions, made by people with more time and more information.
This is what we work on in practice: from strategic AI integration into business processes, to business intelligence on structured data, to AI training for company teams. If you want to know where to start in your company, the starting point is always an honest audit of what you already have.
To stay up to date on these developments, follow the AI and process automation cluster. It collects the most useful analyses for people who have to decide, not for those who just want to follow the news.
It's also worth reading about what's happening with AI agent costs. The article on Claude Fable 5.1 and the reduction of agentic work costs shows how model pricing is changing fast in favor of SMEs.
Frequently Asked Questions
The announcement was made by an Anthropic engineer on May 26, 2026, and describes an elegant proof found over a weekend. At the time of publication, the proof had not yet passed formal review by the mathematical community. It's a promising signal, not yet a certified result.
The capabilities that allow an AI model to solve a mathematical conjecture (constraint-based reasoning, pattern search, generation of non-obvious solutions) are the same ones used to optimize logistics, analyze contracts and manage dynamic pricing. Math is the test bed; business processes are the application.
There's no one-size-fits-all answer. The choice depends on the specific use case, compatibility with your existing systems, pricing that's sustainable for your volume and data security requirements. Our practical advice: test with your company's real data, and don't rely on public benchmarks alone.
The license fee is often the smallest part. Integration, team training and maintenance can triple the total cost over 12 months. Before you decide, calculate the TCO (total cost of ownership) on a realistic usage volume, not on the base plan.
It depends on your industry and your competitors. If you operate in a market where competitors are already adopting AI to optimize processes, falling behind turns into a measurable competitive disadvantage. If your industry is still at an early stage, you still have time to move in a structured way instead of a rushed one.
Three practical questions: Do you have structured, documented data? Do you have clear, measurable internal processes? Do you have an in-house owner willing to run the experimentation? If you answered yes to all three, you're ready for a pilot. If not, start there.
The Erdős conjecture was open for 80 years. The real open problem for Italian SMEs is deciding whether to wait another 80 months or start now.
Related articles
Discover more articles exploring similar topics, selected to offer you a more complete and stimulating perspective. Each piece of content is carefully chosen to enrich your experience.
