Intelligence is foundation
Subscribe
  • Luma
  • About
  • Sources
  • Ecosystem
  • Nura
  • Marbl Codes
00:00
Contact
[email protected]
Connect
  • YouTube
  • LinkedIn
  • GitHub
Legal
Privacy Cookies Terms
  1. Home›
  2. Featured›
  3. Artificial Intelligence›
  4. An AI Just Proved Theorems Mathematicians Haven't Solved Yet
Artificial Intelligence Sunday, 19 April 2026

An AI Just Proved Theorems Mathematicians Haven't Solved Yet

Share: LinkedIn
An AI Just Proved Theorems Mathematicians Haven't Solved Yet

Google's Aletheia system solved 6 out of 10 novel mathematical problems in the FirstProof challenge. No human guidance. No hints. Just an AI agent working through research-level proofs that professional mathematicians are actively trying to crack.

This isn't incremental progress. This is autonomous discovery.

What Aletheia Actually Did

The system uses Google's Gemini 3 Deep Think model - a reasoning engine built specifically for complex problem-solving. It tackled problems from the FirstProof challenge, a set of open mathematical questions that don't have known solutions yet.

On top of that, Aletheia scored roughly 92% on IMO-ProofBench, a benchmark based on International Mathematical Olympiad problems. These are the questions that make teenage maths prodigies sweat. Aletheia worked through them like routine exercises.

The difference between solving known problems and discovering new proofs is enormous. One is pattern-matching against existing solutions. The other requires genuine mathematical creativity - forming conjectures, testing approaches, abandoning dead ends, and constructing novel arguments from first principles.

Aletheia did the second thing.

Why This Feels Different

We've had AI that assists mathematicians for years. Tools that suggest lemmas, check proofs for errors, search databases for relevant theorems. Those are powerful, but they're co-pilots. The human is still flying the plane.

Aletheia doesn't need the human. It formulates its own approach, explores its own proof strategies, and produces complete solutions without intervention. The mathematician shows up at the end to verify the work, not to guide it.

This is what fully autonomous agentic research looks like. The system has goals (solve the problem), constraints (formal mathematical rules), and the capacity to iterate independently until it succeeds or exhausts its options.

For researchers, this changes the game. Instead of spending months on a single proof attempt, you could set Aletheia loose on ten different problems overnight. Some will fail. Some might succeed. Either way, you wake up to results, not just more open questions.

The Practical Implications

Mathematics underpins everything from cryptography to physics to machine learning itself. Breakthroughs in pure maths often take decades to find practical applications - and then they become foundational.

If AI systems can generate novel proofs, the pace of mathematical discovery could accelerate dramatically. Problems that have been open for years might get solved in sprints. New conjectures could be tested faster than human researchers can propose them.

But there's a subtler shift here. Mathematics has always been a deeply human discipline - intuition, creativity, aesthetic judgement about which approaches feel promising. Aletheia doesn't have intuition in the human sense. It has computation and search at a scale no human can match.

That creates a strange dynamic. The proofs it produces might be correct but utterly unintuitive to human mathematicians. We might end up with a growing body of proven theorems that we struggle to understand conceptually, even though we can verify them mechanically.

What Comes Next

The next step is obvious: point Aletheia at the big unsolved problems. The Riemann Hypothesis. The P vs NP question. The Birch and Swinnerton-Dyer conjecture. These are the kind of problems that define careers and sometimes entire fields.

Will Aletheia crack them? Maybe not tomorrow. But the fact that we're even asking the question tells you how far this has come.

For now, the key result is this: AI systems can do original mathematical research. Not assist with it. Not speed it up. Actually do it, from scratch, without a human in the loop.

That's a different world from the one we were in six months ago.

More Featured Insights

Quantum Computing
Quantum Light at Room Temperature Changes the Economics of Everything
Web Development
How to Stop AI Agents From Draining Their Own Wallets

Today's Sources

InfoQ
Google's Aletheia Advances the State of the Art of Fully Autonomous Agentic Math Research
ScienceDaily – Artificial Intelligence
Artificial neurons successfully communicate with living brain cells
TechCrunch
AI chip startup Cerebras files for IPO
Wired AI
Schematik Is 'Cursor for Hardware.' Anthropic Wants In
TechCrunch
Tesla brings its robotaxi service to Dallas and Houston
TechCrunch AI
Anthropic's relationship with the Trump administration seems to be thawing
GeekWire
The $100-a-month workforce: How an entrepreneur uses AI bots to bootstrap a delivery startup
InfoQ
Meta Reports 4x Higher Bug Detection with Just-in-Time Testing
AI Business News
The Real AI Shift Isn't New Models. It's Control.
ScienceDaily – Artificial Intelligence
Quantum AI just got shockingly good at predicting chaos
BBC Technology
What is Claude Mythos and what risks does it pose?
Hugging Face Blog
Building a Fast Multilingual OCR Model with Synthetic Data
Phys.org Quantum Physics
Bright quantum light emission achieved at room temperature in 2D semiconductors
Phys.org Quantum Physics
Universal quantum protocol extracts maximum work without knowing a system's state in advance
Scott Aaronson
Three greats who we've lost
Quantum Zeitgeist
Lattice Surgery Optimises Quantum Computation, Reducing Failure Rates by 16 Percent
Quantum Zeitgeist
Neutrino Force Skirts Standard Model
Quanta Magazine
Quantum 'Jamming' Explores the Truly Fundamental Principles of Nature
Dev.to
Telegram Bot Approvals: Mobile-First Transaction Signing for AI Agents
InfoQ
AWS Announces General Availability of DevOps Agent for Automated Incident Investigation
Dev.to
Scaling DNS in Multi-Cluster Kubernetes with ExternalDNS (AWS Route 53)
InfoQ
Pulumi Adds Full Bun Runtime Support
Hacker News
The RAM shortage could last years
Dev.to
Making dart:ui_web Compile in Flutter Tests - The Conditional Import Pattern
Cloudflare Blog
Introducing the Agent Readiness score. Is your site agent-ready?
Dev.to
The Chip Account: What It Costs the Planet to Make One Chip
GitHub Blog
Bringing more transparency to GitHub's status page
Cloudflare Blog
Shared Dictionaries: compression that keeps up with the agentic web
Dev.to
I built a 13-database ORM for Node.js because Prisma doesn't support Oracle
Dev.to
What Actually Happens When You Click a Button - From Browser to Database and Back

About the Curator

Richard Bland
Richard Bland
Founder, Marbl Codes

27+ years in software development, curating the tech news that matters.

Subscribe RSS Feed
View Full Digest Today's Intelligence
Richard Bland
About Sources Privacy Cookies Terms Thou Art That
MEM Digital Ltd t/a Marbl Codes
Co. 13753194 (England & Wales)
VAT: 400325657
3-4 Brittens Court, Clifton Reynes, Olney, MK46 5LG
© 2026 MEM Digital Ltd