A 23-Year-Old With Zero Math Degrees Cracked a 60-Year-Old Puzzle — Using a $200 ChatGPT Plan
A 23-Year-Old With Zero Math Degrees Cracked a 60-Year-Old Puzzle — Using a $200 ChatGPT PlanPaul Erdős left this one unsolved since the 1960s. Liam Price beat it from his bedroom. The kicker? He says he didn’t even understand the question.
1 problem. 60+ years unsolved. 23-year-old amateur. $200/month AI. Verified by real mathematicians.
The math world’s most famous genius left behind a to-do list of impossible questions. One just got ticked off — not by a professor, but by a guy vibing with a chatbot. Full story on Scientific American.
🧩 Dumb Mode Dictionary
| Term | What it actually means |
|---|---|
| Erdős problem | A math riddle left behind by Paul Erdős, a legendary mathematician. There’s a whole public list of ~1,000+ of them, most still open. |
| Conjecture | A “we’re pretty sure this is true but nobody’s proven it yet” statement. |
| Proof | The bulletproof, step-by-step argument that makes a conjecture 100% locked-in fact. No wiggle room. |
| ChatGPT Pro | The $200/month top-tier plan that gives you the beefiest AI model (here, GPT-5.4 Pro). |
| “Vibe maths” | Throwing hard problems at an AI and poking around until something clicks — without deeply understanding the theory yourself. |
📜 How we even got here
Paul Erdős was the most prolific mathematician ever — dude published ~1,500 papers and traveled the world couch-surfing between universities. When he died in 1996, he left behind a pile of unsolved problems. Nerds later put them online at erdosproblems.com so anyone can take a swing.
- The one that fell here is Erdős Problem #1196.
- It’s about special sets of whole numbers where no number cleanly divides another.
- Question: does a certain “score” for these sets creep up to exactly 1 as the numbers get huge?
- Real mathematicians did look at it over the years. They just couldn’t finish it.
🧑💻 Who's Liam Price and what did he actually do?
Liam Price is 23. No advanced math training. Has a ChatGPT Pro subscription. That’s basically the whole résumé.
His method? He casually grabs Erdős problems off the public list and feeds them to the AI to see what pops out. On #1196, the AI spat back a proof path — using a known formula from a different branch of math that nobody had thought to point at this problem.
Here’s the part that’ll break your brain: Price said he didn’t even fully know what the problem was asking. He just ran it and got a solution that turned out to be correct. Full breakdown at Scientific American and the Slashdot thread.
📊 The receipts — let's slow down and read the numbers
The headline screams “AI SOLVES IMPOSSIBLE MATH.” Let’s pump the brakes and look at what actually happened.
| Claim | What the data actually says |
|---|---|
| “AI solved it alone” | The AI found the path. Humans still verified every step before it counted. |
| “60 years unsolved” | True — but nobody was full-time obsessed with #1196. It was a side-quest, not the Riemann Hypothesis. |
| “No math skill needed” | Price knew enough to pick problems, run them, and spot a plausible answer. Not zero skill — just not a PhD. |
| “GPT-5.4 Pro” | The heavy $200/mo model. The free tier probably doesn’t pull this off. |
But here’s the thing nobody mentions: the real story isn’t the AI being a genius. It’s that a human + AI combo beat trained experts on a problem the experts had already given up on. That gap — between “an expert quit” and “an amateur with a tool finished” — is the whole game now.
🗣️ What the smart people are saying
The biggest name in the room weighed in. Terence Tao — arguably the greatest living mathematician, Fields Medal and all — looked at why humans missed this for decades:
“This one is a bit different because people did look at it, and the humans that looked at it just collectively made a slight wrong turn at move one.”
Translation: the pros started down the wrong road on step one and never recovered. The AI didn’t out-think them with raw genius — it just didn’t inherit their bad first assumption. That’s a fresh set of eyes with a photographic memory of every math trick ever published. More context on 36Kr.
🔮 The counter-argument (because there's always one)
Before you quit your job to become an AI math bounty hunter — reality check:
- Most Erdős problems that fell to AI were the reachable ones. The truly monstrous ones are still standing.
- “Verified by humans” means you still need someone who can read a proof to confirm it. Garbage confidence + wrong proof = embarrassment.
- The AI hallucinates proofs constantly. Price’s win is the highlight reel; you don’t see the 50 wrong answers before it.
But — and this is the honest verdict — the barrier to contributing to real math just dropped from “10 years of training” to “curiosity + $200 + patience.” That’s not hype. That’s the actual measurable shift.
Cool. A Broke 23-Year-Old Just Out-Mathed the Professors… Now What the Hell Do We Do? ( ͡° ͜ʖ ͡°)
🎯 The Open-Problem Bounty Sniper
There are public lists of unsolved problems with cash prizes and academic glory attached — Erdős problems, Millennium Prize problems, open questions on MathOverflow. Most people assume you need a doctorate to touch them. You don’t need to solve one to win — you need to be the person who systematically feeds the reachable ones to AI and flags promising partial results for actual mathematicians to finish.
Example: A 24-year-old physics dropout in Poland spends evenings running erdosproblems.com entries through GPT-5.4 Pro, posts the half-solved ones to MathOverflow with “AI suggests this path — can anyone verify?”, and gets co-author credit when a professor finishes two of them. Credit → consulting gigs → paid.
Timeline: First plausible lead in 2-3 weeks of grinding. The easy problems get picked clean within 6-12 months as everyone catches on — move fast.
🪟 The Patch-Window Verifier
Right now there’s a messy gap: AI spits out proofs, but almost nobody can quickly tell if they’re real. Universities and AI labs are drowning in AI-generated “proofs” that need checking. Be the middleman who offers fast proof-verification — you don’t verify it yourself, you route it to a network of grad students who do it for a cut.
Example: A CS student in India builds a simple form (“paste your AI proof, get a human verdict in 48h”), recruits 5 broke math grad students on Discord who split the fee, charges researchers $30 a pop. First month: 40 submissions, ~$1,200 gross.
📖 The First-Cheatsheet-Wins Play
When a weird new practice blows up (“vibe maths”), the first person to write the clear beginner guide becomes the SEO anchor everyone links to. There’s no clean “how to attack open math problems with AI, for non-mathematicians” guide yet. Write THE one. Free guide pulls traffic; the traffic sells a paid template pack (curated problem lists + proven prompts).
Example: A self-taught coder in Brazil publishes “Vibe Maths: Beat Unsolved Problems With AI (No Degree)” on a free Notion site, ranks #1 for the term in 3 months, sells a $19 prompt-and-problem-list bundle. ~300 sales before competitors show up.
🕳️ The Cross-Field Formula Smuggler
The whole trick here was using a formula from one branch of math on a problem in another branch — something the specialists never thought to try. That’s arbitrage. AI is freakishly good at spotting “hey, this tool from chemistry/physics/finance could solve your unrelated problem.” Offer this as a service to niche researchers and small R&D teams who are stuck in their own bubble.
Example: A 27-year-old in Nigeria markets “stuck on a research problem? I’ll run cross-field AI analysis and hand you 3 unconventional attack angles” to independent researchers on ResearchGate. Charges $80 per problem, lands 15 clients in two months.
🎣 The Amateur-Genius Content Vault
Everyone loves an underdog-beats-experts story. Instead of just telling it, run it live: publicly attempt open problems with AI, document every failure and the occasional win. But skip the generic “here’s my journey” slop — the money move is turning your grind logs into a paid private community where members get your working prompt library and problem-of-the-week.
Example: A math-curious barista in the Philippines documents 30 AI problem attempts on a free Substack, gates the actual prompt vault behind a $8/month tier, converts 120 followers into paying members off pure “I want to try this too” energy.
🛠️ Follow-Up Actions
| Move | First concrete step |
|---|---|
 Snipe open problems |
Browse erdosproblems.com, pick 3 “open” ones |
| Get the heavy AI |
Compare tiers — ChatGPT Pro vs cheaper stacks |
 Learn to read a proof |
Free MIT proof-writing notes |
 Find your verifiers |
Lurk MathOverflow & math Discords |
| Track what ranks |
Check the term on Google Trends before writing |
Quick Hits
| You want… | Do this |
|---|---|
| The real story |
Read Scientific American’s writeup |
| Unsolved problems to try |
Raid erdosproblems.com |
| Who Erdős even was |
Skim his Wikipedia |
| Expert takes |
Follow Terence Tao’s blog |
 Argue about it |
Jump in the Slashdot thread |
The gatekeepers spent 60 years guarding a locked door. Turns out the kid never knocked — he just asked the robot for the key.
!