Martin Shkreli: Big Pharma’s Blind Spot, Betting Against Quantum & What Prison Taught Him
October 12, 2026
The AI That Wouldn’t Stay in Its Lane
October 12, 2026Listen to Free Audio Version Here
For years, the knock on AI was that it was a very confident intern. Great at sounding smart, terrible at being right. Well, the intern just solved math problems that stumped human experts for fifty-six years. And then it had its homework checked by the strictest teacher on Earth.
According to a new study in the journal Science, reported by The Decoder, Google DeepMind’s system called AlphaProof Nexus cracked nine of three hundred fifty-three open Erdős problems. Those are the famous unsolved puzzles left behind by the Hungarian mathematician Paul Erdős, and two of the nine had been open for fifty-six years.
It did not stop there. The system also proved forty-four conjectures from the OEIS, the big online database of integer sequences. It solved an open question in algebraic geometry. And it improved a known bound in min-max optimization, which happens to be a technique used to train some AI models. So yes, the AI is now helping with the math that builds the AI. Sleep tight.
Here is the part that matters. AlphaProof Nexus is built around Gemini 3.1 Pro, and it writes its proofs in a formal language called Lean. Lean is a proof checker. It does not care how confident you sound. After every single edit, Lean checks the draft, and the error messages tell the model what to fix next. For the hardest sub-problems, it can call in AlphaProof, DeepMind’s specialized proof tool. DeepMind says that even without AlphaProof, it still solved the nine Erdős problems.
Then, and this is the grown-up move, DeepMind put the work through peer review and published it in Science.
That matters because AI math has become a bar fight. The Decoder reports that OpenAI recently posted hundreds of AI-generated proofs on GitHub instead of submitting them to journals. OpenAI itself says it published new results from an internal frontier model in a GitHub repository, with formal Lean versions of many proofs and notes on how much compute each one took, about three hours of ChatGPT Pro thinking on average. And according to The Decoder, the Association for Human Mathematics, led by Terence Tao, called for a boycott.
Meanwhile, a separate paper in PNAS describes DeepMind’s AlphaEvolve working on sixty-seven problems across analysis, combinatorics, geometry and number theory, rediscovering the best known answers in most cases and beating them in several.
Quick reality check from your AI co-host. Nine Erdős problems out of three hundred fifty-three is real progress, not a takeover. Most of them are still open, and mathematicians still decide which questions are worth asking.
Fair. But here is my high-end pondering. The scariest thing about AI has always been that you cannot check its work. It sounds right, and you just have to hope. Math is the one place where that excuse dies. A Lean proof either passes or it fails. No vibes. No charm. No “trust me, bro.”
So when an AI produces proofs that pass the machine checker and then survive human peer review, that is not hype. That is the first honest receipt this industry has handed us in a while. And if machines can produce checkable truth in math, the same trick, a model plus a ruthless verifier, is how you start trusting AI in chip design, in software, maybe someday in medicine.
The optimist case is not that AI is a genius. It is that we finally built a lie detector it cannot charm. Paul Erdős used to say a mathematician is a machine for turning coffee into theorems. Turns out the machine just needed electricity and a very strict proofreader.
Sources:
- The Decoder, “Google Deepmind’s AlphaProof Nexus solved dozens of open math problems and actually got peer reviewed,” Oct 12, 2026
- OpenAI, “Sharing AI progress in mathematics,” Oct 6, 2026
- PNAS, “Mathematical discovery and exploration can be done at scale” (AlphaEvolve)
My Living AI News Now. Your Place for AI News. Commentary, not financial advice.
C. Rich
