User:Rongzhou/Transformer is a super-charged next word prediction algorithm

Think of a Transformer model as a super-charged version of the next-word-prediction algorithm built into your cellphone’s keyboard: Instead of predicting the next word by looking at the 3-5 words you have typed just before, this super algorithm looks at the entire conversation and beyond; instead of always predicting your next word based on a fixed mathematical formula, this algorithm “learns” by adjusting its formula over time with your feedback.

For example, when you are discussing the details of your Sunday evening dinner plan with your date via SMS, the algorithm looks at every message sent between you and your date up until that point, as well as all your other SMS conversations, and calculate the statistically most probable candidates for the next word in your conversation. It bases its calculation not only on how often you use certain words, but also on the positions of a sentence you have used them in, and on the context you used them in, i.e.,what was said right before/after?

At first, the predictions are quite bad. You feel like the algorithm would have done better rolling a dice, so you do not click on the suggestions at all. Instead, you type your own responses. The algorithm sees this, and it adjusts its calculations slightly to make the word you typed rank higher in its predictions. The algorithm patiently does this for hundreds of thousands of times, getting ever-so-slightly better each time.

And you see the difference: the next words the algorithm suggests become increasingly relevant, and you increasingly click on one of its suggestions instead of typing your own word. This makes the algorithm happy: its effort is paying off. The algorithm is motivated to continue learning:now also from the words that you clicked. Those words already ranked high in its predictions, but not necessary the highest. So the algorithm adapts its math to boost them higher, ideally to the first two positions. The algorithm also starts learning from the words that you do not click: it adapts its math to suppress a word ever so slightly each time it appeared in the algorithm’s top predictions but was not clicked by you. Logically, the algorithm continues to learn from the words you typed outside its predictions.

Some time later, the algorithm got quite good and you got quite busy in life. One day you received a message out-of-the-blue from an aunt you have not seen for years: “Hi, how is it going? The weather these days are quite good, ain’t they?”. You never liked the aunt that much, and the message was as boring as it gets. So, you decided to entrust your super next-word-prediction algorithm to respond to her. For the first word, the algorithm suggested “Yes” as the top candidate. That’s boring as hell as a response, you think to yourself, but who cares? So you click on it. The algorithm then suggests “!”, “it” and “.”. Whatever, you are responding to a boring-as-hell aunt that sent you a boring-as-hell message out-of-the-blue, so you click on the first one “!”. The algorithm then suggests “The”, “Weather”, and “I”. Whatever, you click on “Weather”, the second choice, just to vary the recipe a bit. The algorithm then suggests “these”, “here”, and “is”. At this point you stoped caring, as the top 3 suggestions of the algorithm is always good enough to continue the sentence anyways, so you started clicking randomly: sometimes you click the first one, sometimes you click the second, and occasionally the third. Very occasionally, just to spice it a bit, you click on a random word out side the top 3.

The message came out as this: “Yes! Weather here these days is also good. I am doing alright. You?”

Not bad, you thought to yourself. Boring response for a boring aunt. So you click send.

Your aunt responded some time later back:“I am visiting Paris in a couple months from now. Would you mind me staying at your place? I looked up the prices for hotels. They are all quite expensive.”

My place is not a free hotel, you thought to yourself. You have been through this before. Relatives messaging you out-of-the-blue to sleep at your place. You don’t even want to bother inventing up an excuse. So, you look towards your super algorithm once again. You click on the message input box, and the algorithm immediately spits out its suggestions for the first word: “Sorry” “I” “No”. Whatever, you say to yourself, let’s see what the algorithm got: so you started once again clicking randomly, mostly on the first two suggestions, but occasionally the third, and every once-in-a-while a word outside the top 3.

In the end, the message came out:

“Sorry, I will not be at home around that time. I hope you find a nice hotel. Have a great trip!”

Not bad at all, you think. A tiny white lie — a polite fiction that lets everybody save face. Better than any excuse you would have mustered on your own, honestly. You click send and forget about it.

The aunt, however, is evidently immune to hints. A few hours later she writes back:

“Oh, that is a shame! What about the week before? Or the week after? I can be flexible, I can work around your schedule!”

You groan out loud. You have been through this before, too: you give a soft “no”, and the other person hears “maybe if I ask again”. You know this will never end unless you stop being nice about it. So once again you turn to the algorithm, and once again you start clicking.

The suggestions roll out one by one. First word: “Actually” — top of the list. Good. Then: “I” “cannot” “accommodate”. Then “you”, “during”, “that”, “time”. And you click — sometimes the first suggestion, sometimes the second, occasionally the third, and every once in a while a word from outside the top 3, just to keep the algorithm on its toes.

In the end, the message came out:

“Actually, I cannot accommodate you during that time. I value my space and my quiet. Please book a hotel.”

You read it twice. Direct. Honest. Borderline rude, but in a way that leaves no room for a third attempt. The aunt, to her credit, takes it surprisingly well:

“Understood. That is very clear of you. I will book a hotel then. Perhaps we can still get a coffee while I am in town?”

The aunt takes it surprisingly well: “Understood. That is very clear of you. I will book a hotel then. Perhaps we can still get a coffee while I am in town?”

Coffee. Coffee is fine. Coffee has a hard limit baked into it: it ends. Unlike the endless negotiation of a guest bedroom, a coffee is a finite commitment. You weigh it for a moment. You kind of respect the aunt now — she took the no like a champion, no sulking, no passive-aggressive “well, I won’t bother you then.” A woman who hears a clear no and simply books a hotel deserves a coffee.

You tap the message input box, and the algorithm spits out its first-word suggestions: “Sure” “Yes” “Maybe”. You flinch at “Maybe”. You have been through this before, too: “maybe” is just “no” with extra steps — a polite fiction in the other direction, one that invites a follow-up. You click “Sure”. The algorithm suggests “why” “let’s” “we”. You click “why”. “not” “don’t” “won’t”. You click “not”. “.” “!” “?” — you click “.”.

“Sure, why not.”

You read it twice. Short. Fine. But something nags at you: you just told this woman you value your space and your quiet, and now you’re handing her a coffee like a door prize for good behavior. The algorithm, sensing your hesitation, offers a second sentence. “When” “I” “Where” — you click “When”. “are” “will” “is” — “are”. “you” “we” “I” — “you”. “in” “coming” “around” — “in”. “town” “Paris” “the” — “town”. “?” “!” “.” — “?”.

“When are you in town?”

Good. A practical question. Coffee requires logistics, and logistics are safe. Nothing about your space, nothing about your quiet — just dates, times, coordinates.

The aunt replies with a window of dates a few weeks out. You glance at your calendar, and the algorithm nudges forward: “That” “Sounds” “Great” — “That”. “works” “sounds” “fits” — “works”. “.” “!” “…” — “.”.

“That works.”

You pause with the cursor blinking. There’s a small window here where the whole thing could be derailed by your own good manners. You can feel the draft of a longer message forming in your head — the “let me know which arrondissement,” the “I know a lovely little place near the canal,” the creeping invitation to expand a coffee into something bigger than a coffee. You have been through this before, too: one friendly clause too many, and suddenly you’ve committed to a weekend walking tour. You catch yourself.

The algorithm, having learned from a hundred previous over-commitments, offers its suggestions: “See” “Great” “Looking” — and then, tucked just below the top three, almost shyly: “Looking forward to it.”

You stare at it. It is a tiny white lie — you are not, strictly speaking, looking forward to anything about this. But it is a friendly, human-sized thing to say, and the algorithm has learned exactly where a polite fiction stops being a lie and starts being lubrication. You click it anyway — third choice, maybe even fourth, just to keep the algorithm on its toes.

The aunt responds with a single thumbs-up emoji. You smile, put the phone down, and return to your evening. Somewhere in the background, the algorithm hums contentedly, recalculating the probability of your next over-commitment and quietly boosting “Looking forward to it” a notch higher in its rankings.

Your super-charged next-word prediction algorithm, is a Transformer.