A 23-Year-Old With No Math Degree Just Solved a 60-Year-Old Problem — By Asking a Chatbot
A 23-Year-Old With No Math Degree Just Solved a 60-Year-Old Problem — By Asking a ChatbotHonestly, the pros spent six decades stuck on the same wrong first step. A curious kid and a text box walked in and went the other way.
60 years unsolved. 1 prompt on a Monday. 23-year-old amateur. Confirmed by the guy widely called the smartest mathematician alive.
This isn’t a “robots take over math” story. It’s a “you don’t need the fancy piece of paper to matter” story — and that’s way more dangerous. Full write-up at Scientific American.
Honestly, I’ve watched “AI does a smart thing” headlines for years now, and 99% of them are a company hyping a demo that breaks the second you poke it. This one’s different (I know, I know — I say that every time). But okay, seriously: a guy who is not a professional mathematician sat down, gave a chatbot basically one good prompt, and out fell a real answer to a puzzle that legends had been banging their heads on since the 1960s. Then Terence Tao — think LeBron, but for math — checked the work and said yeah, it holds up.
🧩 Dumb Mode Dictionary
| Fancy term | What it actually means |
|---|---|
| Erdős problem | A puzzle left behind by a legendary Hungarian math guy (Paul Erdős) who dropped hundreds of “bet you can’t solve this” challenges. Some have cash prizes. |
| Conjecture | A “we’re pretty sure this is true but nobody’s proven it” statement. |
| Proof | The bulletproof receipt that shows why something is 100% true, forever. No exceptions. |
| Primitive set | A pile of numbers where none of them divides evenly into another. Weird flex, but math loves it. |
| ChatGPT Pro | The paid, beefier version of the chatbot that thinks longer before answering. |
| “Vibe maths” | The nickname for this: no textbooks, no plan — just describe the problem and let the machine cook. |
📜 How we got here (the 60-year cold case)
Way back around 1965, three heavyweight mathematicians — Erdős, Sárközy, and Szemerédi — asked a simple-sounding question about those “primitive sets.” Roughly: if you add up a specific little fraction for every number in the pile, does the total sneak up on exactly 1 as the pile gets infinitely big?
Sounds cute. It was a nightmare. Everyone who tried — including a modern pro named Jared Lichtman — hit the same wall and stopped. For 60 years. You can literally browse the open ones yourself at erdosproblems.com.
⚙️ What actually happened (the part that matters)
- The guy: Liam Price, 23, not a research mathematician. Just curious.
- The move: One ordinary Monday, he fed the problem to ChatGPT Pro. No literature deep-dive, no preset plan. Let it reason on its own.
- The twist: The chatbot didn’t take the “normal” path. It grabbed a formula that mathematicians knew about — but had never thought to point at this problem.
- Tao’s mic drop: He said people did look at this one, and the humans “collectively made a slight wrong turn at move one.” Everybody copied the same bad first step. The AI just… didn’t.
- The cleanup: The raw chatbot proof was messy as hell. Lichtman and Tao had to sort it out and pull the key idea into something readable. So — machine finds the door, humans build the frame. (Futurism has the play-by-play.)
🗣️ What the timeline's saying
- Half the math world: “An amateur? With a chatbot? On MY unsolved problem?” (╬ Ò﹏Ó)
- The other half: relieved someone finally checked a fresh angle instead of copy-pasting the same failed opening for 60 years.
- The realists (me): the AI didn’t do this alone. It needed a human to ask the right thing, and two of the best humans alive to translate the answer. That’s not “AI replaces mathematicians.” That’s “AI is a really good tip-off.” Tao’s own blog has the technical breakdown if you’re brave.
🔍 The part nobody's saying out loud
Okay but seriously — the headline everyone’s chasing is “AI smart.” The real story is the amateur part. A person with no credentials, no lab, no PhD committee, walked up to a 60-year fortress and found a crack, because the paid experts were all standing in the same line facing the wrong way.
The tool matters less than the mindset. The kid didn’t “know better.” He just didn’t know the wrong way well enough to be trapped by it. That’s the whole ballgame. Beginner’s luck? Maybe. But the receipt says confirmed.
Cool. A Broke 23-Year-Old Just Out-Mathed the Pros… Now What the Hell Do We Do? (•̀ᴗ•́)و
Nobody’s saying you’ll solve the next Erdős puzzle by Friday. But the door this kicked open is bigger than math — it’s “the fresh-eyes outsider with a good chatbot can now poke at problems that used to require a career.” Here’s where the sneaky money and the first-mover flags are hiding, before everyone else clues in.
🩸 The Bounty Bloodhound
Erdős left behind hundreds of open problems — and some carry actual cash prizes. There’s a whole living list at erdosproblems.com, sorted by difficulty and status. Most people never look because they assume “not for me.” The angle: hunt the easy-tier, still-open ones, point a strong reasoning model at them, and even if you don’t crack it — a clean partial write-up gets your name on the forum thread. First to contribute anything real to a niche = instant credibility.
Example: A 24-year-old CS student in Nairobi filters the list for “open + elementary,” runs each through a reasoning model over a weekend, and posts three tidy “here’s how far the AI got” notes to the problem forum. Two get ignored. One gets a pro replying “huh, that’s actually a new angle” — and now she’s got a co-author intro that no résumé could buy.
Timeline: First forum post in a weekend. Real recognition in 2–4 months. The easy problems get farmed fast, so the early window is the good one.
🧹 The Proof Janitor's Toll Booth
The whole reason this discovery needed Tao and Lichtman: the chatbot’s raw answer was a mess. That’s the gap. Machines will keep spitting out rough, correct-ish reasoning that nobody wants to clean up. Learn to translate messy AI math into a formal, machine-checked proof using free tools like Lean (a “proof checker” — software that refuses to lie about whether math is airtight). Then charge researchers to verify their AI outputs.
Example: A self-taught coder in Kraków spends two months on the free Lean 4 tutorials, then offers “I’ll formally verify your AI-generated proof sketch” on academic Discord servers and Mathstodon. Charges per job. Professors with grant money and no patience become repeat customers.
🔄 The Wrong-Turn Reversal
Tao’s key line: everyone made the same bad first move. That pattern isn’t just in math — it’s in every field with a “standard approach everyone copies.” Chemistry, patent searches, legal precedent, engineering design. The play: find a stuck problem in a niche you actually understand, then explicitly prompt the AI to ignore the standard method and try something borrowed from a totally different field. Sell the contrarian re-take as a consulting insight.
Example: A mechanical-engineering grad in São Paulo takes a “known unsolvable” factory-layout headache, tells the model “solve this like it’s a biology problem, not an engineering one,” and gets a weird-but-workable idea. Writes it up as a one-page “second opinion” and pitches local factories. First client pays for the novelty; word spreads.
📖 The Recipe Vault (be the dictionary before there's a dictionary)
“Vibe maths” is a brand-new phrase. Nobody’s the authority on how to actually prompt a model into real reasoning instead of confident nonsense. The first person to build a genuinely useful, tested cheatsheet — “here are the 15 prompts that got AI to produce checkable math, and the 40 that produced garbage” — owns that search term. Not a fluffy blog. A tested, brutally specific field manual, dropped on GitHub so it ranks and gets starred.
Example: A physics tutor in Manila keeps a spreadsheet of every prompt that made a model produce verifiable steps, cross-checked against Lean. Cleans it into a free repo, posts it once to the relevant subreddits. Repo becomes the link people paste when someone asks “how do I do the ChatGPT math thing?” Traffic → consulting DMs.
🎯 The Preprint Sniper
Every week now, someone posts a fresh “AI helped me prove X” claim to preprint sites like arXiv. Most go unchecked for a while. The angle: be the fast, reliable person who reproduces and stress-tests these claims within days of posting — publishing short, honest “I re-ran this, here’s what held and what didn’t” notes. In a world drowning in unverified AI claims, the trusted checker becomes the most-followed account in the niche.
Example: A stats postgrad in Hyderabad sets alerts on new AI-assisted math preprints, reproduces the checkable ones within 48 hours, and posts tight verification threads on Mathstodon. Builds a rep as “the guy who tells you if it’s real.” Journals and reporters start DMing for quotes.
🛠️ Follow-Up Actions
| Want to… | Do this |
|---|---|
| See the open problems | Browse erdosproblems.com |
| Learn proof-checking | Start Lean 4 tutorials (free) |
| Read the real story | Scientific American + Futurism |
| Follow the experts | Terence Tao’s blog |
| Find fresh AI claims | Watch arXiv math |
Quick Hits
| If you want… | Then… |
|---|---|
| A problem to try |
Grab an “open + elementary” one from erdosproblems.com |
 A durable skill |
Learn Lean and verify AI proofs for money |
 An SEO flag |
Build the “vibe maths” prompt cheatsheet on GitHub |
 Instant credibility |
Reproduce a fresh arXiv claim before anyone else |
 A consulting angle |
Ask AI to solve your niche’s stuck problem the “wrong” way |
The pros had 60 years and the credentials. The kid had one good question and no fear of looking dumb. Guess which one won.
!