10 Math & CS Breakthroughs Quietly Rewriting the Rules
In this article
> **Bottom line:** Between 2023 and 2025, mathematicians and computer scientists quietly cracked problems that had resisted proof for decades — the 3D Kakeya conjecture, a new time-space tradeoff for algorithms, a lower matrix-multiplication exponent, and AI systems (DeepMind's AlphaEvolve, Google's Gemini Deep Think) that discovered faster algorithms and hit gold-medal-level performance on IMO problems.
None of this trended the way a new ChatGPT model does.
But these results are already reshaping how chips get scheduled, how proofs get verified, and how fast your database can run a shortest-path query.
If you write software or make decisions about AI investment, the story isn't the headline model — it's the math underneath it.
I spent three years assuming pure math was the one corner of tech immune to AI hype.
Then I watched a DeepMind system called AlphaEvolve rediscover Strassen's 1969 matrix multiplication trick in an afternoon — and then beat it.
That's when I realized I'd been reading the wrong headlines.
We've all been staring at the wrong scoreboard.
Why Nobody's Watching the Real Scoreboard
Every week there's a new model benchmark: who's beating whom on MMLU, who has the longer context window, whose agent can book your flight. It's a legitimate story.
But it's also a distraction from a quieter, weirder story happening one layer down — in the math and theoretical CS departments where the actual *rules* of computation get rewritten.
Here's why that matters right now, in 2026, more than it has in decades.
**AI systems are no longer just consumers of algorithms — they're producers of them.** Google DeepMind's AlphaEvolve spent 2025 discovering new matrix multiplication constructions and shaving real compute off Google's own data center scheduling, reportedly trimming resource usage by a fraction of a percent that, at Google's scale, translates into meaningful energy and cost savings.
That's not a chatbot answering trivia. That's an AI finding a genuinely new mathematical object.
Meanwhile, human mathematicians proved things in 2025 that textbooks said might take another generation.
Hong Wang and Joshua Zahl closed out the three-dimensional Kakeya conjecture — a geometry problem about how thin needles pointing in every direction can be packed together — which has quietly underpinned harmonic analysis and number theory for 40 years.
Ryan Williams dropped a paper showing any algorithm running in time *t* can be simulated using roughly √(t log t) space, a result so far outside expectations that theorists spent weeks re-checking it before believing it.
These aren't ivory-tower curiosities. They're the substrate everything else runs on.
Everyone's Celebrating the Wrong Half of This Story
Here's my contrarian take: the tech press's obsession with "AI beats human mathematicians" is missing what's actually interesting — and in some ways, it's backwards.
The breakthrough isn't that AI systems are replacing mathematicians.
It's that **AI and human mathematicians started collaborating as peers**, and the collaboration itself became a new kind of research infrastructure.
Terence Tao's "Equational Theories Project" didn't use AI to skip the hard part.
It used a mix of human insight, automated theorem provers, and mass online collaboration to knock out thousands of open questions about algebraic structures in a matter of months — a pace that would've been unthinkable for a solo mathematician or a lone AI.
The narrative that "AI solved math" also undersells how much of this is still algorithm design, not pattern matching.
When AlphaEvolve found a faster way to multiply 4×4 matrices, it wasn't guessing — it was running an evolutionary search over program space, guided by the same kind of rigorous verification a human referee would demand.
The "AI did math" framing makes for a punchier headline than "AI ran a very well-designed search," but the second one is closer to true, and it's the one that tells you what's actually reproducible.
**If you're impressed by chatbots and unimpressed by algorithms, you have the ratio backwards.** The chatbot is a demo. The algorithm is infrastructure that outlives the demo by decades.
The Three-Layer Stack: How to Actually Think About This Wave
I've started sorting every one of these breakthroughs into what I call **the Three-Layer Stack** — because "10 unrelated math facts" isn't a mental model, and mental models are what actually stick.
Layer 1: Proof Infrastructure
This is the layer where AI and formal verification tools (Lean, primarily) are changing *how* math gets checked, not just what gets discovered.
Tao's Equational Theories Project lives here, alongside 2025's headline moment: frontier models (Google's Gemini Deep Think and comparable systems from OpenAI) reaching gold-medal-level performance on International Mathematical Olympiad problems, with proofs formally verifiable rather than just plausible-sounding.
This layer matters because it changes the *cost* of rigor — proofs that used to take a grad student a month to check can now get machine-verified in hours.
Layer 2: Algorithmic Speed Records
This is the layer with direct, near-term engineering payoff.
Matrix multiplication's exponent — the number that determines how fast you can multiply two n×n matrices — dropped below 2.3713 in 2024 (Williams, Xu, Xu, and Zhou), continuing a slow march toward the theoretical floor of 2.
AlphaDev's AI-discovered sorting and hashing routines got merged into real production libraries, including LLVM's standard C++ sort — meaning code you've probably already run benefited from an algorithm no human wrote.
Faster negative-weight shortest-path algorithms (building on Bernstein, Nanongkai, and Wulff-Nilsen's near-linear-time breakthrough) are quietly making their way into routing and network-optimization tools.
Layer 3: Structural Limits
This is the layer that tells you what's *possible at all* — the deepest and slowest-moving, but the one that eventually reshapes everything above it.
The Kakeya conjecture's resolution ripples into number theory and PDE analysis.
Ryan Williams' time-space result reopens questions about whether some problems we assumed needed lots of memory secretly don't.
Progress on zero-density estimates for the Riemann zeta function, driven by work from Larry Guth and James Maynard, tightened our error bounds on how primes are distributed — a result that sounds abstract until you remember prime distribution underlies half of modern cryptography.
The stack matters because **breakthroughs at Layer 3 take years to reach Layer 2, and Layer 2 takes months to reach your production code.** If you want to know what's coming, don't watch the chatbot.
Watch the stack.
What This Actually Changes for You in the Next 12 Months
If you're an engineer: expect standard library functions — sorting, hashing, graph algorithms — to keep quietly getting faster underneath you, the way AlphaDev's sort already did.
You won't file a ticket for this. It'll just show up in a language runtime update.
If you're in ML infrastructure or data center ops: the AlphaEvolve story is the one to watch closely.
AI-discovered scheduling and resource-allocation algorithms are already trimming real compute costs at hyperscale, and every major cloud provider is almost certainly running similar internal search-based optimization projects right now, even if they're not publishing about it.
If you're in security or cryptography: the Riemann zeta zero-density improvements and continued number-theoretic progress are worth a bookmark, not a panic.
**Nothing here breaks RSA tomorrow** — but the error bounds on prime distribution are the kind of thing that, shaved down over a decade, eventually forces protocol updates.
If you're a manager evaluating AI tooling: the real signal from 2025's Layer 1 progress isn't "AI can do math now." It's that **verification-first AI workflows — where a model's output gets checked by a formal system before anyone trusts it — are a viable pattern outside of pure math.** Code generation, financial modeling, anything with hard correctness constraints: this is the template.
The Bigger Picture
Here's the thing that actually kept me up after digging into all ten of these: for most of my career, I treated "the algorithm" as a fixed, solved layer — something invented decades ago by people smarter than me, that I just imported and called.
2025 broke that assumption. The floor is still moving.
That's unsettling if you like stable ground.
But it's also the most honest reminder I've had in years that computer science isn't finished — it's a living field, and some of its oldest, most "settled" questions just cracked open in the same 24 months that everyone was busy arguing about chatbot benchmarks.
The mathematicians and theorists doing this work aren't chasing virality.
They're not going to trend on Hacker News the way a new model release does — except, this once, they did, which is probably why you're reading this.
Maybe that's the real story: **the loudest tech news and the most consequential tech news are, increasingly, two different lists.**
Which list are you actually paying attention to — and what's the last "boring" math or CS result you dismissed that you'd bet, in five years, turns out to have mattered more than the AI launch that buried it?
---
