Foundations

Why AI got good so fast

More data, more computing power and one design idea from 2017.

6 min readPlain English

Read it in

In one sentence

AI got good fast because three things arrived at once: vast amounts of digital text, far more computing power, and a 2017 design called the Transformer that could put both to use.

Think of it like…

A builder gets three things in the same year: all the bricks they could want, a fleet of cranes, and a new method that lets many crews work on one building at the same time. Any one would help. Together, they change what can be built.

How it works

No single breakthrough did it. Several ingredients came together, and each one made the others more useful.

1DataDecades of websites, books, forums and code gave models more text to learn from than a person could read in many lifetimes.
2Computing powerGraphics chips, first built for video games, turned out to suit the huge number of simple calculations training needs, and companies built data centers full of them.
3The 2017 designGoogle researchers published the Transformer, which takes in a whole passage at once and weighs how each word relates to the others. It made training on vast amounts of text practical.
4ScaleResearchers found that bigger models trained on more data kept improving in fairly predictable ways, so companies invested to go bigger.
5CoachingPeople wrote and rated example answers, turning a raw text predictor into an assistant that follows instructions. That’s what made chat assistants feel useful in 2022.

Since then, much of the progress has come from models that work through problems step by step, use tools like search, and handle images and voice.

Good at

Writing that reads like a person wrote it
Reading photos, documents and voice notes
Working through multi-step problems

Bad at

Knowing when it’s wrong
Staying current without a search tool
Judgment calls only you can make

Where you’ll see it at work

Voice notes to job summaries. A plumber talks through the job on site and gets a tidy written summary to check and send.
Receipts from photos. A bakery photographs its receipts and gets the supplier, date and total pulled out for a person to check.
Drafts in the house style. A design studio shares three past proposals and gets a new draft that reads like theirs.

Myth vs reality

Myth: One big breakthrough made AI smart overnight.

Reality: It took decades of research, plus data, chips and the 2017 design arriving together. The “overnight” part was when it reached a chat box anyone could use.