What you will learn Why your daily work did not change when AlphaGo won, and why it is different now. Once that question has an answer, everything else gets easier.
Lee Sedol, a 9-dan Go player, played five games against AlphaGo in Seoul. From March 9th to the 15th, for a full week, an entire country watched.
People who had never played Go watched too. Not because they were curious about Go, but because they were curious about whether a machine could beat a human.
The result was AlphaGo 4, Lee Sedol 1.
After three straight losses, Lee Sedol won game four. Move 78, cutting into the middle of the board. Commentators called it "God's Touch."
By AlphaGo's own calculation, the chance a human would play that move was one in ten thousand. AlphaGo blundered on move 79 and did not realise it was losing until move 87.
That game remains the only official record of AlphaGo losing to a human.
Shock. The conventional wisdom was that Go had far too many possible positions for a machine to win any time soon. Headlines about "which jobs disappear now" were everywhere.
Let us be honest for a moment.
Since March 2016, has the way you work changed because of AlphaGo?
Almost certainly not. The news was hot, but the office stayed the same. You still wrote your own email, built your own reports, and read your own long documents.
Why?
AlphaGo played Go better than any human. And that was all it did.
Off the board, it could do nothing. To anyone who did not play Go, AlphaGo was "impressive, but not my problem." So when the news cooled off, daily life went back to normal.
HyperTeamsgraph TD
A["AlphaGo · 2016"] --> B["Extremely good at Go"]
B --> C["Can do nothing off the board"]
C --> D["No use to anyone<br/>who does not play Go"]Something called ChatGPT appeared. This time the reaction was different.
It passed 100 million users in two months. For comparison:
| Service | Time to 100 million |
|---|---|
| about 2.5 years | |
| TikTok | about 9 months |
| ChatGPT | about 2 months |
Analysts who had covered the field for years said they had never seen a curve like it.
Not because the technology was more impressive. Because the set of people who could use it was different.
| AlphaGo | Today's AI | |
|---|---|---|
| Good at | Go | language |
| Who can use it | people who play Go | anyone who works with words |
| Preparation needed | you must know Go | you must be able to talk |
| Relationship to you | a spectacle | a tool you can use today |
Think about what you actually do at work. Writing email, producing reports, organising material, capturing what was said in a meeting — nearly all of it is language.
What AlphaGo was good at (Go) was irrelevant to most people. What today's AI is good at (language) is relevant to almost every desk job. That single difference explains the difference in reaction.
This part matters. AlphaGo and today's AI do not merely play different games — they fail differently.
| AlphaGo | Today's AI | |
|---|---|---|
| Problem type | rules are exact (the rules of Go) | no rules (what is "good writing"?) |
| Right answer | exists (you win or you lose) | none (only better answers) |
| How it succeeds | calculates moves to the end | imitates what people have written |
| How it fails | a miscalculation | it invents something plausible |
Remember that last row. Today's AI does not fail the way AlphaGo failed. It fails in a confident tone of voice. We give that its own chapter later — things to watch out for.
The most common misconception starts right here If you think "AlphaGo beat Lee Sedol, so AI must be smarter than people," you will use today's AI badly. Today's AI is less a genius and more a very tireless assistant — fast, never bored, and never to be trusted without a check.
So far this has been an office story. In the same period, something happened on the research side too.
| 2024 | Awarded to | For |
|---|---|---|
| Nobel Prize in Physics | John Hopfield, Geoffrey Hinton | foundational discoveries and inventions that enable machine learning with artificial neural networks |
| Nobel Prize in Chemistry | Demis Hassabis, John Jumper (Google DeepMind), David Baker | DeepMind's share was for protein structure prediction with AlphaFold |
In a single year, both Physics and Chemistry went to AI-related work. One prize could be called a coincidence; two is something else. It is a symbolic moment — AI moved past being a tool that assists research and into the place where the discovery happens.
Two things worth adding:
Which is why this product exists Taking AI that is already proven in world-class research and making it usable in an organisation's everyday work is what HyperTeams Connect does. Predicting a protein structure and tidying up meeting notes differ in scale, but the technology underneath is the same family.
If the Nobel Prizes were about awarding results that AI helped produce, September 2026 was something slightly different. AI took a problem people had failed to solve for close to a century, and came back with an answer of its own.
The Navier-Stokes equations describe how fluids — water, air — move. They are used in aircraft design, weather forecasting, blood flow analysis. Navier wrote his version in 1822 and Stokes refined it in 1845, so the equations are 200 years old; yet nobody had proved whether their solutions in three dimensions always stay smooth. In 2000 the Clay Mathematics Institute made it one of the seven Millennium Prize Problems, with a $1 million prize attached.
On 8 September 2026, OpenAI published a result obtained by putting about ten thousand AI agents on it.
| Thrown at it | roughly 10,000 AI agents |
| Time taken | 88 hours from the first agent (plus 17 hours to formalise) |
| Messages exchanged | about 2.7 million on this problem alone |
| Result | a solution whose velocity blows up to infinity in finite time — a counterexample to "always smooth" |
1. One machine did not solve it; many split the work. The agents were divided into groups — some pushing toward "a solution always exists", others toward "a singularity forms" — and at intervals they shared results so the promising lines could be carried forward. That is how a human research lab works. It is also exactly what AI agents, later in this guide, are. Only the scale differs.
2. The grading was done by a machine, not a person. OpenAI translated the proof into Lean, a proof-checking language. An AI producing text that looks like mathematics and a machine verifying every line of it are entirely different events. Think of it as the strictest possible form of what this guide keeps saying: don't trust it, make it show its grounds (Grounding).
3. And it is still not "solved". OpenAI said it does not intend to claim the prize, because peer review and journal publication are still ahead. The accurate phrasing is not "solved it" but "presented a solution and a formalised proof." It is also worth knowing that prior work preceded it: in August 2026, researchers showed that singularities can form in related fluid equations and formalised that result in Lean.
Nor did this arrive out of nowhere. The list below includes only results that cleared a check — the competitions were graded by people, and the rest had their proofs inspected by machine.
| When | What happened |
|---|---|
| July 2024 | DeepMind's AI reaches silver-medal standard at the International Mathematical Olympiad (proofs in Lean; later published in Nature) |
| July 2025 | Gold-medal standard at the same competition (Google and OpenAI both) |
| January 2026 | Erdős problem #728 solved essentially unaided by AI — checked by the mathematician Terence Tao |
| August 2026 | Solutions to ten problems open for a decade or more, every one of them passing a Lean check |
| September 2026 | Navier-Stokes |
Two years took it from a high-school competition to a Millennium problem. Read the table down the page and you see the speed; read it across and you see that what made the speed trustworthy, every time, was verification.
The third point is the one closest to your own work The announcement is spectacular; the verification is not finished. That is not unique to Millennium problems — the same is true of the report you asked an AI to write. An answer arriving and an answer passing review are two different events. Things to watch out for, later on, is about this.
A little over three years have passed since 2022. We are past the novelty stage and into the actually-used-at-work stage. Published surveys sketch roughly this picture:
The numbers vary a lot by survey and by date. Take the direction (going up, results starting to be measured) rather than the figures, and check the original source if you need a specific number.
The important part is this: unlike the AlphaGo era, this one is your problem.
This part (New to AI) exists to get you using it today. Unlike Go, it takes no special talent and no prior knowledge.
The next chapter explains what this thing actually is, by comparing it to search.
1. Why did ordinary work barely change in the six years after AlphaGo?
Because AlphaGo was a machine that only played Go. Off the board it could not write an email or read a document. To anyone who did not play Go it was "impressive, but not my problem."
2. Explain ChatGPT reaching 100 million users in two months — without referring to the technology.
The range of people who could use it was different. Go, which AlphaGo was good at, matters to a small group. Language, which today's AI is good at, is nearly all of desk work. And there is nothing to prepare — if you can talk, you can start.
3. Why is "AlphaGo beat Lee Sedol, so AI is smarter than people" a dangerous thought?
The two are different in kind. AlphaGo solved a problem with exact rules and a right answer by calculation; today's AI handles problems with no right answer by imitating human writing. So they fail differently — today's AI invents things in a confident tone. Treat it as a genius and hand it work without checking, and that is exactly where the accident happens.
4. Why is it inaccurate to say OpenAI's Navier-Stokes result "solved a Millennium Prize Problem"?
Because the verification is still outstanding. OpenAI presented a solution and a proof formalised in Lean, but it has not gone through peer review or journal publication, and OpenAI did not claim the prize. An answer arriving and an answer passing review are two different events.
Next we look at what this tool actually is, and how it differs from search → What AI is