All Exams Are Pattern Matching

Meeting AI in its second winter — an NLP lab, training data typed by hand, and a professor’s line: every exam is pattern matching.

A look back — long before any of the tools we argue about now.

Looking back from a long way off, the 1990s were the middle of what people call AI’s second winter. The topics with heat then were hardware-software co-design and embedded systems; PC-communication services had barely appeared, and the internet was one breath from exploding. No Yahoo, no Google, no Naver — before StarCraft and Napster. A cell phone was the kind of thing you got as a graduation present.

PL? AI? NLP?

The first time I met the letters “AI” was, I think, in the name of a lab run by Professor Young-taek Kim, who taught our Compiler and Programming Languages courses. It was a topic a long way from the one GPT and the Transformer have all but unified today, and I can’t quite reconstruct whether the lab was called PL, or AI, or NLP — later, with Professor Byoung-tak Zhang, the AI part grew more prominent. The two labs merged and split and merged again. I wonder if the classmates who joined back then foresaw the present we’re standing in.

The NLP work was mostly translation — Korean-English, Korean-Japanese, both directions — and through the department office I took on a lot of part-time jobs typing test data into Hangul, the Korean word processor. I knew a fair amount of Chinese characters; I’d paid good money for a Japanese-capable copy of Hangul, worked at picking up Japanese through anime, and typed printed books page by page, saving each batch to a floppy disk to hand in. The kind of era where forgetting to save was a catastrophe — a continuous run of things hard to imagine now.

Exams Are Pattern Matching

This was before deep learning and machine learning came along to solve problems by sheer force of compute. The algorithms and comparison logic of the time had a small, handmade quality. Even the things we called NLP, without an internet connection, did their best with morphological analysis followed by switch and if-then-else and map — stifling in one sense, but if you read the code through, you could at least mime an understanding of it.

I still can’t tell whether he meant it as a joke, but a line the professor dropped in class has stayed with me: “Every exam is pattern matching.” You have to carve enough comparison cases into memory to construct the right answer — which, in today’s terms, you could read as the importance of data. That it’s still with me says something; at a time when I was stumbling over parsers and string compares, it landed hard, and it’s become a phrase I reach for at work and at home.

AI, After

After that absurdly short encounter, the word all but vanished for me until around 2000, and it wasn’t until the 2010s, working on search at Google, that I met machine learning again — half player, half spectator. Mostly, as the person responsible for a service, my job was the argument over whether we even had to use that black-box method: the metrics were clearly worse, and yet — it’s the trend, so we should use it — that sort of collision.

Then came AlphaGo, CNN and RNN, the Transformer, and the small handmade quality gave way to a dizzying spread of methods. I’m a spectator for the moment, but if I got back on the field, there’s a lot I’d want to try.

P.S. This is from further back, from a more foolish time — something my middle-school Korean teacher told me. In a rebellious stretch when I was asking why I had to study at all, even under threat of a beating, she said it plainly and in earnest: exams, no; but studying — learning — that you do until the day you die, and I hope you keep the joy of it your whole life.

Part of 90s Computer Science Stories — first-person notes on growing up with computers and studying CS in Korea in the 1980s and ’90s. See the full series →

Adapted from my Korean essay on Brunch: brunch.co.kr/@chaesang/46