A Brief History of AI

Welcome to Day 4. You can now place any system in the AI/ML/DL hierarchy. Today we add the missing dimension — time. Where did all this come from, and why did it suddenly work after decades of false starts?

The headline most people carry is wrong: AI did not begin in 2022 with ChatGPT. It's a seventy-year-old field with a dramatic boom-and-bust history. Understanding that arc is the single best thing you can do to read the present clearly — to tell durable progress from the hype that has burned this field twice before.

Here's the whole story compressed into one shape: AI moves in cycles. A genuine breakthrough sparks excitement, the excitement curdles into overpromising, reality fails to match the promises, funding and faith collapse into what the field literally calls an 'AI winter' — and then a new idea thaws it and the cycle restarts.

This has happened more than once, which is why veterans are wary of grand claims. What's different now is magnitude: today's boom is by far the biggest, best-funded, and most productive yet. But the cyclical pattern is the frame to keep in mind for the rest of this post — it's the lens that makes every milestone and every winter make sense.

This timeline marks the beats of the story so the cycles become visible at a glance. 1956: the field is named at Dartmouth. The 1960s–70s: early optimism and the first programs. The mid-70s: the first winter. The 1980s: the expert-systems boom and its bust. 2012 onward: the deep-learning explosion that's still going.

Laid out on a single line, the rhythm is unmistakable — peaks of hype, troughs of disappointment, each peak higher than the last. The visual does something prose can't: it shows you that 'overnight' AI success is really a seventy-year staircase, and that the current moment is one more (very tall) step, not a clean break from everything before.

The term 'Artificial Intelligence' was coined in 1956 at the Dartmouth Summer Research Project — a small workshop of mathematicians and scientists who believed that human-level intelligence could be largely solved within a generation. They proposed to make significant progress in a single summer.

They were off by roughly seventy years, and counting. But the detail that matters isn't the miss — it's that the miss set a pattern. Almost every generation since has predicted human-level AI was just around the corner, and almost every one has been wrong by decades. That founding burst of over-optimism is the very first instance of a habit you'll see repeat through the entire history, right up to today's confident timelines.

The four beats of the cycle, dated. 1956–74: early optimism, the first AI programs, generous funding on the strength of bold predictions. 1974–80: the first AI winter, when the systems underdelivered against the hype and the money dried up.

The 1980s brought a second boom around 'expert systems' — programs encoding specialist knowledge as rules — which also overpromised and busted by the early 90s. Then, from 2012, the combination of deep learning, abundant data, and GPU compute reignited the field into the boom we're living in. Notice the shape repeats: a technical breakthrough, a wave of money and promises, a collision with reality, a retreat. The technology changes each time; the human pattern around it doesn't.

If the pattern always ends in winter, why hasn't this one? The honest answer is in what each boom ran on. Past booms ran on clever hand-crafted rules — and clever rules eventually run out of road. You hit problems no one can write rules for, results plateau, and disappointment sets in.

This boom runs on something with a different growth curve: data plus compute plus learning methods that keep improving as you scale them up. For the first time, pouring in more data and more compute reliably yields more capability instead of diminishing returns. The results compound rather than plateau. That's the core reason this boom has lasted and hasn't (yet) busted — though 'yet' is doing real work in that sentence.

Two specific moments lit the current fire, and they're worth knowing by name. In 2012, a neural network called AlexNet used GPUs to win an image-recognition contest by a stunning margin, proving that deep learning actually works when you give it enough data and compute. That result ended the long skepticism about neural nets almost overnight.

In 2017, a paper titled 'Attention Is All You Need' introduced the Transformer — an architecture that scales gracefully to enormous datasets, effectively the whole internet. Every modern large language model, ChatGPT included, is a descendant of that 2017 architecture. AlexNet proved deep learning could see; the Transformer let it read and write. Together they're the ignition of the era you're living through.

This chart plots the compute used to train top models over time, and the story is in the slope. It doesn't rise gently — it explodes, often on a logarithmic scale, with each headline model consuming dramatically more compute than the last. That curve is the physical engine behind the capability gains you keep reading about.

The takeaway is double-edged. On one side, it explains why models keep getting more capable: we keep feeding them more compute and they keep rewarding us for it. On the other, it's the seed of a possible future winter — that exponential can't rise forever, and if the cost of the next leap outruns its value, the economics that have kept this boom alive could finally bite.

The most important lesson the winters teach is calibration. Every single generation of AI researchers believed human-level AI — AGI — was just around the corner, and every single one was wrong. The hype outran the reality, the funding evaporated, and real progress stalled for a decade or more each time.

The right stance, then, is excited but skeptical. The capabilities in front of you are genuinely real — that's not in doubt, and dismissing them is its own mistake. But the confident timelines layered on top of those capabilities have a seventy-year track record of being fantasy. Believe the demo; discount the prediction about what next year holds. That two-part posture is the whole inheritance of the AI winters.

So AI is a seventy-year cycle of breakthrough, hype, and winter — and today's boom is different mainly because it runs on something that scales. Hold both truths at once: the capability is real, the timelines are usually fantasy. Save this for the next time a headline declares AGI is months away.

Tomorrow we come back to the present and the concrete. Day 5 is real-world AI applications — the systems already shipping in products you use, so you can point to exactly where all this history has landed.