BlogAI Just Crossed a Line: It’s Discovering New Math Now

The big theme: AIArtificial Intelligence (AI) is a very large and broad spectrum of technologies which most people would be familiar with through... is starting to do real frontier work in math, not just solving textbook problems, but finding new bounds, new functions, and new approaches… fast. And layered on top of that? The messy human layer: model races, rumours, lawsuits, and the very real pull toward military and geopolitical use.

If you only take one idea from this week, it’s this:

We’re moving from “AIArtificial Intelligence (AI) is a very large and broad spectrum of technologies which most people would be familiar with through... helps with math” to “AIArtificial Intelligence (AI) is a very large and broad spectrum of technologies which most people would be familiar with through... discovers math.”

AI Is Producing New Math, Not Just Answers

Math is becoming one of the clearest proving grounds for AIArtificial Intelligence (AI) is a very large and broad spectrum of technologies which most people would be familiar with through... capability because it’s crisp, verifiable, and novelty is measurable. That’s why this week’s stories landed so hard.

Grok 4.20 and the Bellman function claim

One of the biggest moments this week was the claim that Grok 4.20 discovered a Bellman function.

If that’s real and reproducible, it’s not “cute math.” Bellman functions can represent deep structure in optimisation and bounds — the sort of thing that can matter for real-world systems over time.

What grabbed attention wasn’t just the claim — it was the implication: AIArtificial Intelligence (AI) is a very large and broad spectrum of technologies which most people would be familiar with through... behaving less like a calculator and more like a system that can search the space of possible ideas and land somewhere humans didn’t.

The Aeros problem and GPT 5.2 as a co-researcher

In the other video, we saw a human + model pairing tackle the Aeros problem, with Neil Smani and GPT 5.2 working together on something “very few people on earth” could touch.

The part that matters isn’t only the result — it’s the workflow:

• AIArtificial Intelligence (AI) is a very large and broad spectrum of technologies which most people would be familiar with through... proposing directions
• helping verify steps
• accelerating iteration
• operating like a co-researcher, not just an assistant

“Intelligence Explosion” as a Practical Concept (Not Sci-Fi)

One video frames this directly as “intelligence explosion” — the idea that if AIArtificial Intelligence (AI) is a very large and broad spectrum of technologies which most people would be familiar with through... can help build better AIArtificial Intelligence (AI) is a very large and broad spectrum of technologies which most people would be familiar with through... (or better tools), progress can compound.

Cautious about the rhetoric, but I’m less cautious about the underlying pattern:

When AIArtificial Intelligence (AI) is a very large and broad spectrum of technologies which most people would be familiar with through... shortens iteration cycles, you get:

• more shots on goal
• faster refinement
• a greater chance of real breakthroughs

The key question is whether this is robust — or whether we’re watching highlight reels. Because math has a way of humbling everyone.

Algorithms Matter: Small Wins Become Platform Shifts

One of the most underrated points this week: the next wave of advantage may come from algorithms, not just bigger models.

A strong example: Google’s Alpha Evolve improving matrix multiplication after ~50 years.

That’s not just cool trivia. Matrix multiplication is infrastructure math. If you make core operations cheaper, it scales everything downstream:

• training
• inference
• research velocity
• deploymentDeployment is the process of releasing your application or software out onto the specific environment where it will need to... cost

People obsess over “which chatbot is #1,” but the deeper story is: the stack gets reshaped quietly by improvements like this.

The Less Fun Part: Power, Lawsuits, and Military Gravity

While the math is inspiring, the incentives around AIArtificial Intelligence (AI) is a very large and broad spectrum of technologies which most people would be familiar with through... are getting sharper — and more militarised.

This week also pivoted into:

• Elon vs OpenAI / Sam Altman drama (which is also signalling to regulators, investors, and talent)
• and the tone shift around xAI partnering with the US “Department of War” as the video frames it

One phrase that sticks with me is the “unfiltered and rapid response” angle in high-stakes contexts — because unfiltered + fast + military is a volatile combination unless governance is exceptionally tight.

If we had to summarise the week in one sentence:

Frontier math suggests capability spikes — and military partnerships suggest where those spikes might get aimed.

The Model Race & Rumour Mill: Slateflow, Tidewisp, and the Arena Meta

This week included rumours of new Grok variants — “Slateflow” and “Tidewisp” — appearing in the LM Arena ecosystem, positioned around speed and distinct operating approaches.

Rumours are cheap, but the pattern is real:

• release cadence is compressing
• evaluation is increasingly public
• teams get pushed toward “vibes + quick wins”

We’re more interested in a harder set of questions:

• Can it do research reliably?
• Can it produce verifiable artifacts (proofs, functions, algorithms)?
• Can it stay stable under pressure?

The Reality Tax: Compute, Fine-Tuning, and Cost Transparency

Even when intelligence leaps, the same wall shows up:
GPUs, cloud    When referencing computing, the term 'cloud' refers to the generalisation of interconnected computers and services running on the Internet,... bills, orchestration, and deploymentDeployment is the process of releasing your application or software out onto the specific environment where it will need to... pain.

This week also touched on the infrastructure market responding to that demand — for example:

• fine-tuning workflows that are simpler
• cost structures that are more predictable (token-based transparency)
• less DevOpsDevOps is a term that describes development and operations as they apply to software development. Prior to the use of... overhead
• faster iteration loops

Where this lands for builders

If you’re building with AIArtificial Intelligence (AI) is a very large and broad spectrum of technologies which most people would be familiar with through... right now, this is what I’d keep in mind:

• Math breakthroughs are a signal because they’re verifiable
• Algorithmic improvements change the game quietly
• Incentives shape outcomes — especially when defence and geopolitics enter the picture
• Real-world delivery still comes down to infrastructure, cost, and repeatable workflows

If you watched either of these videos too, we’d love to know:
Do you think we’re seeing true research-grade AIArtificial Intelligence (AI) is a very large and broad spectrum of technologies which most people would be familiar with through... now — or are these cherry-picked wins?

Here are the videos

• GROK 4.20 cracked the code
• AI just solved one of the hardest math problems… (INTELLIGENCE EXPLOSION)

Work with Aerion Technologies

At Aerion Technologies, we help teams move from “AIArtificial Intelligence (AI) is a very large and broad spectrum of technologies which most people would be familiar with through... ideas” to AIArtificial Intelligence (AI) is a very large and broad spectrum of technologies which most people would be familiar with through... that actually ships — with real systems, real workflows, and real outcomes.

And if you’re building software (AIArtificial Intelligence (AI) is a very large and broad spectrum of technologies which most people would be familiar with through... or otherwise) and want a clearer path from idea → plan → delivery, that’s exactly what DevReady is designed for.

FAQs