So thereâs this guy, right? He sits in a room by himself, with a computer and a keyboard full of Chinese characters. He doesnât know Chinese, though, in fact he doesnât even realise that Chinese is a language. He just thinks itâs a bunch of odd symbols. Anyway, the computer prints out a paragraph of Chinese, and he thinks, whoa, cool shapes. And then a message is displayed on the computer monitor: which character comes next?
This guy has no idea how the hell heâs meant to know that, so he just presses a random character on the keyboard. And then the computer goes BZZZT, wrong! The correct character was THIS one, and it flashes a character on the screen. And the guy thinks, augh, dammit! I hope I get it right next time. And sure enough, computer prints out another paragraph of Chinese, and then it asks the guy, what comes next?
He guesses again, and he gets it wrong again, and he goes augh again, and this carries on for a while. But eventually, he presses the button and it goes DING! You got it right this time! And he is so happy, you have no idea. This is the best day of his life. He is going to do everything in his power to make that machine go DING again. So he starts paying attention. He looks at the paragraph of Chinese printed out by the machine, and cross-compares it against all the other paragraphs heâs gotten. And, recall, this guy doesnât even know that this is a language, itâs just a sequence of weird symbols to him. But itâs a sequence that forms patterns. He notices that if a particular symbol is displayed, then the next symbol is more likely to be this one. He notices some symbols are more common in general. Bit by bit, he starts to draw statistical inferences about the symbols, he analyses the printouts every way he can, he writes extensive notes to himself on how to recognise the patterns.
Over time, his guesses begin to get more and more accurate. He hears those lovely DING sounds that indicate his prediction was correct more and more often, and he manages to use that to condition his instincts better and better, picking up on cues consciously and subconsciously to get better and better at pressing the right button on the keyboard. Eventually, his accuracy is like 70% or something â pretty damn good for a guy who doesnât even know Chinese is a language.
One day, something odd happens.
He gets a printout, the machine asks what character comes next, and he presses a button on the keyboard andâ silence. No sound at all. Instead, the machine prints out the exact same sequence again, but with one small change. The character he input on the keyboard has been added to the end of the sequence.
Which character comes next?
This weirds the guy out, but he thinks, well. This is clearly a test of my prediction abilities. So Iâm not going to treat this printout any differently to any other printout made by the machine â shit, Iâll pretend that last printout I got? Never even happened. Iâm just going to keep acting like this is a normal day on the job, and Iâm going to predict the next symbol in this sequence as if it was one of the thousands of printouts Iâve seen before. And thatâs what he does! He presses what symbol comes next, and then another printout comes out with that symbol added to the end, and then he presses what he thinks will be the next symbol in that sequence. And then, eventually, he thinks, âhm. I donât think thereâs any symbol after this one. I think this is the end of the sequence.â And so he presses the âENDâ button on his keyboard, and sits back, satisfied.
Unbeknownst to him, the sequence of characters he input wasnât just some meaningless string of symbols. See, the printouts he was getting, they were all always grammatically correct Chinese. And that first printout heâd gotten that day in particular? It was a question: âHow do I open a door.â The string of characters he had just input, what he had determined to be the most likely string of symbols to come next, formed a comprehensible response that read, âYou turn the handle and pushâ.
One day you decide to visit this guyâs office. Youâve heard heâs learning Chinese, and for whatever reason you decide to test his progress. So you ask him, âHey, which character means dog?â
He looks at you like youâve got two heads. You may as well have asked him which of his shoes means âdogâ, or which of the hairs on the back of his arm. Thereâs no connection in his mind at all between language and his little symbol prediction game, indeed, he thinks of it as an advanced form of mathematics rather than anything to do with linguistics. He hadnât even conceived of the idea that what he was doing could be considered a kind of communication any more than algebra is. He says to you, âBuddy, theyâre just funny symbols. No need to get all philosophical about it.â
Suddenly, another printout comes out of the machine. He stares at it, puzzles over it, but you can tell he doesnât know what it says. You do, though. Youâre fluent in the language. You can see that it says the words, âDo you actually speak Chinese, or are you just a guy in a room doing statistics and shit?â
The guy leans over to you, and says confidently, âI know it looks like a jumble of completely random characters. But itâs actually a very sophisticated mathematical sequence,â and then he presses a button on the keyboard. And another, and another, and another, and slowly but surely he composes a sequence of characters that, unbeknownst to him, reads âYes, I know Chinese fluently! If I didnât I would not be able to speak with you.â
That is how ChatGPT works.
