My wife thinks someone is replying
The other day my wife was chatting with AI.
So let’s overlook the actual impossibilities and find out.
Before we do the maths, people have been trying versions of this for 256 years.
Nearly all of them shut down or got caught. None of them came close to today’s scale.
So what would it actually take?
Let’s be generous. Assume everyone types at world-record speed, 200 words a minute, non-stop. Everyone is happy to wait 10 minutes for a reply. And it’s a remote company spread across every time zone, so someone is always awake.
The average ChatGPT user sends about 4 messages a day, and a reply is roughly 500 words now. With over a billion people using the big AI chatbots, that’s about 6 billion messages every day.
10.4 million people typing at world-record speed, round the clock, with no sleep, no lunch, and no days off.
Okay, let’s run it like a real company instead. Which means it needs a CEO. Me.
11.8 billion internal messages a day, just to keep the place running. Oh, and we still have to answer the 6 billion messages people actually sent us.
All in, you’d need 77.4 million staff just to keep 10.4 million people typing at any given moment.
If you want to check my maths: a 40-hour week, time off from a 2025 survey of UK call centres, 12.1 people per manager (the US average), and 2 HR people per 100 staff (SHRM). People quit at the US rate, earn the US median for customer service, and get 153 chat messages a day, like the average Microsoft Teams user. The biggest Zoom call holds 1 million.
More people won’t work. But what about a guessing game from 1950, played by a mathematician called Claude Shannon? Yes, Claude.
He picked a line from a book, hid it from his wife Betty, and got her to guess it one letter at a time.
Betty didn’t know the sentence, she just knew the patterns of English.
My wife does this to me every day. She’s had years of practice.
So what if we replaced Betty with a computer?
For more than 60 years, the answer was no. Giving machines more to read barely helped, because they could only see the last couple of words. Try guessing what comes after “speak fluent…” with no other clues.
In 2017, eight people at Google let it see every word at once and weigh the ones that matter, like France. They called that attention, and the machine a transformer. It’s the T in ChatGPT.
Suddenly, giving it more to read actually worked. More text, bigger brain, better guesses.
Writing a reply is the same thing. It reads what you said and guesses the next word, or piece of a word (called a token), sticks it on the end, looks back at everything again, and guesses the next one.
It’s read websites, books, forums, fan wikis, and code. A top model trains on more than 32 trillion tokens. Reading that non-stop would take you about 190,000 years.
It doesn’t store those pages to look up later like a search engine. It just keeps the patterns.
So I explained all of that to my wife.
Nobody is sat there typing. Just millions of us, over years, answering all at once.
