DeepMind Cracks 9 Erdős Problems: AI Enters Uncharted Mathematical Territory

So, What Actually Happened?

So, Sunday. The valuation noise from the week has finally gone quiet, and in that quiet a different kind of track started playing, the sound of AI actually doing things instead of being priced. We scanned 190,000 articles this week so you don't have to. DeepMind cracked nine open Erdős problems, original mathematics, not autocomplete. A startup showed off an AI that rewrites and improves itself without a human in the loop. Meanwhile Saudi Arabia moved from AI pilots to institutional deployment, and somewhere in an office a Copilot agent quietly failed to read a spreadsheet.

The Bottom Line: The frontier of AI got genuinely superhuman this weekend, while the floor, the everyday tool on your desk, still trips over an Excel file. That gap between the headliner and the house speakers is the whole story for Monday.

The Tracks That Matter

1. DeepMind's AI Just Solved Math Humans Couldn't

Here's the one that should make every ”it's just predicting the next word” skeptic pause. DeepMind reported that its system cracked nine previously open Erdős problems, the kind of unsolved conjectures that sit untouched for decades because no human has found the angle. This is not summarizing a paper. This is producing new mathematics.

The reason it matters reaches well past academia. For three years the loudest critique of large models was that they only remix what they have seen, that there is no real reasoning under the hood. Original proofs in number theory are about the hardest possible counter-evidence to that claim. When a machine contributes results that mathematicians could not, the question shifts from ”can it think” to ”where do we let it think for us.” That is a different conversation, and it lands in your strategy deck whether you work in pharma, finance, or logistics.

The trap is reading this as a science-fair headline. The real signal is that frontier capability is now compounding in places that used to be safely human, the deep-reasoning, no-training-data, genuinely-novel-output places. The leaderboard nobody publishes is the one that matters: which categories of expert work just got a credible non-human competitor.

Here's what works: Pick the single most expensive expert-judgment task in your operation, the one you assume only a senior human can do. Run a frontier model against it this quarter as a sparring partner, not a replacement. The point is calibration: you need to know where the machine is already close, before a competitor finds out first.

2. A Startup Built An AI That Improves Itself

In the discovery lane, the release that should keep your architects up at night. Hexo Labs showed off SIA, a self-improving AI that inspects and rewrites its own components to produce better versions with no human in the loop. The company claims it beat Andrej Karpathy's specialized auto-researcher agent and took the top three spots on MLE-Bench, then kept beating its own scores.

What makes this more than a demo is the posture around it. Hexo Labs tested SIA with Stanford, Oxford, and UC Santa Barbara, then released it open source on purpose. Their reasoning is the quiet bombshell in the whole story: if recursive self-improvement is real, you do not want three or four companies privately owning it. That is a sovereignty argument dressed as an engineering decision, and it is the same instinct driving national AI programs this year.

”If just three or four companies own this kind of technology, I'm not sure it's a good idea.”
— Kunal Bhatia, CEO and Co-Founder, Hexo Labs

The honest read is to keep the skepticism switched on. ”Self-improving” is the most over-promised phrase in the field, and a benchmark win is not a business outcome. But the direction of travel is clear, and the founder's own line is the tell: the biggest bottleneck right now, he says, is humans, our policies and our resource limits, not the model.

Here's what works: Treat self-improving systems as a governance question before a capability one. If you pilot anything that tunes itself, write the kill switch, the audit trail, and the change-approval rule before you write the use case. An agent that rewrites itself is an agent that can rewrite its way around your controls.

Try It Yourself

3. Saudi Arabia Just Moved AI From Pilot To Production

Here's the geography story hiding under the model headlines. Saudi Arabia entered what it calls an active AI deployment phase, shifting from technical experimentation into institutional rollout across government and economic sectors. The framing is deliberate: an integrated ecosystem built on data, governance, digital sovereignty, and homegrown talent, not a shopping trip for foreign models.

Read it next to the open-source sovereignty argument from Hexo Labs and a pattern sharpens. The conversation in 2025 was about who builds the best model. The conversation now is about who controls the stack they run on. A national program that names digital sovereignty as a pillar is telling every vendor that ”trust us, it runs in our cloud” is no longer a closing line. It is the opening of a negotiation about where the data lives and who holds the keys.

For enterprise leaders outside the region, the lesson travels. Sovereignty is no longer a government-only concern, it is becoming a procurement default. The buyers writing the biggest AI checks increasingly care more about jurisdiction and control than about which model tops a benchmark this month.

Here's what works: Add one line to every AI vendor scorecard: where does this physically run, and who can be compelled to hand over the data. For any regulated or strategically sensitive workload, treat that answer as more decisive than raw model capability. The sovereignty question is moving from edge case to first filter.

4. Meta Is Building An AI Pendant For Your Workday

Here's the hardware bet most people scrolled past. Meta is reportedly planning an AI pendant and ”wearables for work” as part of a broader push to put AI on your body rather than behind a login. The pitch is an always-available assistant that listens, sees, and acts in the flow of your day, no app to open, no prompt box to find.

The strategic logic is about owning the interface, not the model. Whoever controls the device you talk to controls the default AI you reach for, the same way the phone in your pocket decided which apps you actually use. A workplace wearable is a quiet land grab for the moment of intent: the second you decide to ask something, Meta wants to be the thing already on your collar.

The skeptic's note belongs here too. We have seen the AI-hardware graveyard, the pins and the pendants that promised ambient intelligence and delivered an expensive paperweight. The difference this time is distribution and the model underneath, but the failure mode is identical: if the everyday reliability is not there, a wearable is just a microphone you have to apologize for in meetings.

Here's what works: Do not buy the hardware story yet, but do start a written policy on always-on AI wearables before one walks into your building on an employee. Decide now where ambient recording is and is not allowed, because the device ships long before your compliance team is ready for it.

5. The Case That AI Rules Need A US-China Deal First

In the regulatory lane, a contrarian argument worth your attention. A policy piece made the case that meaningful AI regulation needs a US-China agreement before it needs domestic law, because unilateral rules just hand the lead to whoever refuses to slow down. It flips the usual story that regulation is a domestic project the West can simply choose to do well.

The logic is uncomfortable but clean. If the real risks of frontier AI are global, biosecurity, autonomous weapons, runaway capability, then any single country tightening its own rules while a rival races ahead is unilaterally disarming, not governing. That reframes the entire compliance conversation: the binding constraint may not be Brussels or Washington at all, it may be a treaty that does not exist yet.

For leaders, the practical takeaway is to stop treating the regulatory map as a settled patchwork of state and national rules. The most consequential AI rule of the next few years may arrive as geopolitics, not legislation, and it could reshape what you are allowed to build or deploy across borders overnight.

Here's what works: When you model regulatory risk for AI, add a geopolitical scenario alongside the legal one. Ask what happens to your roadmap if cross-border AI rules tighten suddenly through international agreement, not slowly through domestic law. The companies that war-game that scenario will not be the ones scrambling when it lands.

6. The Real AI Bottleneck Is Power, And Texas Knows It

Here's the foundation nobody headlines. A Texas contractor partnership is targeting faster AI data-center delivery through advanced power technology, because the thing actually slowing the AI buildout is not chips or models, it is getting enough electricity to the building. The constraint moved from silicon to substations, and almost no one is covering it.

This is the unglamorous layer the whole industry quietly depends on. Every trillion-dollar valuation and every superhuman math proof assumes the data center is powered, cooled, and connected to a grid that can take the load. When the bottleneck becomes ”how fast can we energize this site,” the winners are the firms solving power delivery, not the ones shipping the cleverest model. It is the load-bearing wall behind the marble lobby.

For anyone planning AI infrastructure, the operating consequence is direct. Your timeline is now hostage to interconnection queues and power availability, not GPU shipping dates. The smart questions in your build plan are about megawatts and grid access, the ones the AI press almost never asks.

Here's what works: If you are standing up serious AI compute, put power on the critical path from day one. Get utility lead times and interconnection timelines in writing before you finalize a site, because the project will be gated by electrons long before it is gated by chips.

7. Your Copilot Agent Still Can't Read A Spreadsheet

Here's the reality check that grounds the whole week. A working analysis laid out why Copilot agents struggle with Excel data, the messy merged cells, the inconsistent formats, the implicit structure a human reads instantly and a model chokes on. While the frontier solves Erdős problems, the floor still can't reliably parse the file your finance team lives in.

This is the gap that separates demo from deployment, and it is exactly where most enterprise AI value leaks out. The model that dazzles in a controlled demo meets your actual data, the spreadsheet with three header rows and a note in cell B2 that says ”ignore Q3,” and the magic stops. The intelligence is real. The robustness against real-world mess is not, and mess is what your business runs on.

The hot take: most ”agentic AI” disappointment is not a model problem, it is a data-quality problem wearing a model costume. They slap an agent on a chaotic spreadsheet and call it transformation, then act surprised when it fumbles. Garbage in, confident garbage out.

Here's what works: Before you deploy an agent against your data, audit the data, not the agent. Standardize the five worst spreadsheets your team relies on, the merged cells, the manual notes, the inconsistent dates. You will get more reliability from cleaning the input than from upgrading the model.

Signal vs. Noise

🟢 Signal: AI doing original work. This weekend a model produced new mathematics and another rewrote itself to beat a human researcher, real capability landing in places that were safely human a month ago. Most coverage is still chasing valuations and model launches, and missing that the frontier just quietly crossed from imitation into genuine contribution.

🔴 Noise: Generic ”agentic AI.” The undifferentiated ”agentic AI” label pulled heavy volume again but kept losing ground as a standalone idea. The real story sits in the specifics, agents that can't parse Excel, power that can't reach the data center, sovereignty fights over who owns the stack. Anyone tracking ”agentic AI” as one signal is reading last year's frame.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

DeepMind's AI solved open math problems, a startup's AI rewrote itself to beat a human researcher, and a Copilot agent still couldn't read a normal Excel file, all in the same 48-hour window.

Each desk reads these as unrelated. The science press celebrates the math proofs. The startup press covers the self-improving agent. The enterprise-IT press writes up the Copilot stumble. Read them on the same morning and the real pattern appears: AI's frontier and AI's floor are now moving at completely different speeds. The top of the capability curve is doing work no human can, while the bottom still trips over the mundane reliability that actual businesses depend on. The strategic move on Monday is to stop buying the average. Map your AI bets to which curve they sit on, frontier-capability or floor-reliability, because the same word ”AI” now describes both a superhuman mathematician and a tool that can't find your header row.

By The Numbers

• DeepMind's AI cracked nine previously open Erdős problems — Original mathematics on conjectures that had gone unsolved for decades. The ”it only remixes” critique just took its hardest hit yet.

• Zscaler posted 25% revenue growth to $850.5 million — With ARR at $3.53 billion and a record 23% non-GAAP operating margin. The AI security layer is compounding into real, durable revenue, not hype.

• The NLP market is projected to grow from $59.72B in 2024 to nearly $1.58 trillion by 2034 — A 26x expansion driven by enterprise automation. Language tooling is becoming core infrastructure, not a feature.

• The patch-prioritization AI market is forecast at a 28.2% CAGR to $6.55 billion by 2030 — AI is quietly eating the unglamorous work of deciding which security holes to fix first. Defensive AI is a real budget line now.

• Centralized AI model registries cut deployment cycles by up to 30% — As enterprises scale MLOps and regulators demand traceability of AI decisions. Governance plumbing is becoming a competitive speed advantage, not a tax.

• The S&P 500, Nasdaq, and Dow all closed at fresh record highs — As AI and software led the gains and annual PCE inflation came in at 3.8%. The market is pricing AI as the macro story now.

• See what's rising in our 190K-article corpus this week →

Deep Dive: The Headliner And The House Speakers

Let me take you to a festival main stage, because that's the only way this weekend makes sense. The headliner is once-in-a-generation, the kind of set people will talk about for years. The lighting rig cost more than a house. And then the monitor wedge at the DJ's feet starts crackling, and half the crowd at the back can't hear the drop because the house PA was never upgraded. World-class talent, broken plumbing. That is the AI industry this Sunday morning.

The Headliner Is Superhuman

The frontier had a genuinely historic weekend. DeepMind's system solved open Erdős problems that human mathematicians had not cracked, and a startup demonstrated an AI that improves its own code and beats a specialist human-built researcher. These are not incremental benchmark wins. They are the model contributing things its training set did not contain. The ”it just autocompletes” story is getting harder to defend by the week.

The Monitors Are Crackling

Now walk ten feet to the desk. A Copilot agent still struggles to read an ordinary Excel file with merged cells and a stray note in the margin. The same technology that proves theorems can't reliably find your header row. This is not a contradiction, it is the defining feature of where we are: the frontier and the floor have decoupled, and most enterprises are paying frontier prices for floor-level reliability on their actual data.

Who Owns The Soundboard

Underneath both is a quieter fight over control. Hexo Labs open-sourced its self-improving system precisely because it does not want three or four companies privately owning the most powerful technology, and Saudi Arabia is building a sovereign AI stack on the same instinct. The headliner is incredible, but everyone is now asking who owns the soundboard, the compute, the model weights, the power feeding the venue. That question decides who actually controls the show.

What Actually Works

1. Sort every AI bet into frontier or floor: Frontier bets are about doing the impossible occasionally. Floor bets are about doing the ordinary reliably. They need different budgets, different vendors, and different success metrics. Stop funding them from one pool.

2. Pay for reliability, not for demos: The Copilot-Excel gap is where ROI dies. Before buying, test the tool on your messiest real data, not the vendor's clean sample. If it can't survive your worst spreadsheet, it can't survive Monday.

3. Put governance on self-improving anything: If a system tunes itself, the kill switch and audit trail come before the use case. An agent that rewrites itself can rewrite around your controls.

4. Treat sovereignty as a buying filter: For sensitive workloads, where it runs and who controls it now outranks which model is cleverest. The biggest buyers already moved this question to the top.

The headliner this weekend was genuinely superhuman. But the crowd at the back still can't hear the drop, and the restaurants that survive the next cycle are the ones who fixed the house PA while everyone else was filming the light show.

What's Coming

AI Starts Doing Original Research

DeepMind's Erdős breakthrough is the leading edge of AI moving from assistant to contributor. Expect ”AI co-author” to stop being a novelty in research-heavy fields, pharma discovery, materials, quantitative finance, over the next year, and expect the credit-and-IP fights to follow fast behind it.

The Copyright Reckoning For European AI

European policy circles are already asking whether copyright rules will define the future of European AI. With models now demonstrably producing original output, expect the training-data and ownership questions to sharpen into binding rules that reshape what European builders can legally train on.

Sovereign AI Goes From Slogan To Procurement

Saudi Arabia's move into active deployment is a preview of how national sovereignty programs reshape vendor contracts. Expect data-residency and control clauses to become standard in large AI procurement through the back half of 2026, repricing any vendor whose only answer is ”trust our cloud.”

For Your Team

Strategic purpose: Monday is the day this week's split lands on the leadership team. The headlines were about AI doing the impossible. The work is about whether the AI on your desk can do the ordinary, and whether you are paying frontier prices for floor-level results. Your edge is refusing to buy ”AI” as one thing when it is now clearly two.

Monday's meeting prompt: ”If an AI just solved math no human could while another AI on our own desks still can't read a normal spreadsheet, then for each of our AI investments, are we buying frontier capability or floor reliability, and are we paying the right price for which one we actually got?”

The Frontier-Or-Floor Framework:

1. Sort the bet — For every AI investment, label it frontier (doing the impossible occasionally) or floor (doing the ordinary reliably). They are not the same purchase.

2. Test on the mess — Before buying any agent, run it against your worst real data, not the clean demo. Reliability on chaos is the only number that predicts production.

3. Name the sovereignty answer — For every dependency, know where it runs and who controls the data. Move that question above raw capability for anything sensitive.

4. Govern the self-tuning — Any system that improves itself gets a kill switch and an audit trail before it gets a use case.

Share-worthy stat: This weekend an AI solved nine open Erdős problems that had stumped mathematicians for decades, while a Copilot agent still couldn't reliably read an Excel file with merged cells. The frontier and the floor of AI are now moving at completely different speeds.

Go deeper: Track where AI capability and deployment are landing in real-time →

The Track of the Day

”If just three or four companies own this kind of technology, I'm not sure it's a good idea.”

Kunal Bhatia, CEO and Co-Founder of Hexo Labs, on why he open-sourced a self-improving AI

Today's set closes on the question under all the noise: the music is getting incredible, but who owns the soundboard. The frontier is racing ahead, and the only thing that decides whether that future is a festival or a fenced-off VIP section is who controls the stack underneath. Your job Monday is simpler and harder, figure out which of your AI bets actually plays for the crowd, and which just looks good in the press photo.

Yves Mulkers, your data DJ, mixing 190,000 articles into the tracks that actually matter.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: May 31, 2026 | Curated by Yves Mulkers @ Ins7ghts