How AI Evolved

The three waves: rules, machine learning, deep learning

AI has reinvented itself roughly three times. Each wave is a different answer to the question "how do we make a machine do something smart?"

1. Symbolic AI (1950s to 1980s): write the rules. Humans hand-coded explicit rules: "IF the patient has a fever AND a rash, THEN suggest measles." It worked for narrow, tidy problems and broke the moment the real world got messy. You cannot write a rule for every sentence in English.

2. Machine Learning (1980s to 2010s): learn the rules from data. Instead of writing rules, we showed the machine thousands of examples and let it find the patterns itself. Spam filters and recommendation systems came from here. Powerful, but it still needed humans to hand-pick which features mattered.

3. Deep Learning (2010s to now): learn the features too. Using neural networks with many layers, the machine learns not just the rules but which features matter straight from raw data: pixels, audio, text. This is the wave behind image recognition, voice assistants, and every modern chatbot.

A useful note for later: these are nested, not replaced. Deep learning is a kind of machine learning, which is a kind of AI. Each new wave wrapped around the last rather than throwing it away.

You will map that nesting precisely in the next lessons.

It helps to picture what building each wave actually felt like. In the symbolic era, progress meant rooms of experts sitting beside programmers, translating hard-won know-how into ever-longer lists of rules; the systems were sharp right up until the exceptions piled up and the list became impossible to maintain. The machine-learning era rewrote the job description: instead of writing rules, teams collected and labelled data and argued over which signals were worth measuring. The deep-learning era then quietly handed over even that last human job — deciding which features matter — to the machine itself, so the work shifted again, this time toward gathering data and compute at scale. Notice the trajectory across all three: with each wave, humans hand more of the thinking to the system, and the bottleneck moves from human cleverness to sheer scale.

Three milestones: Deep Blue, ImageNet, the Transformer

If the three waves are the "how," a handful of landmark moments are the "when" — the events that flipped AI from a lab curiosity into something you feel every day.

• 1997 — Deep Blue beats Garry Kasparov. IBM's chess machine defeated the reigning world champion, the first time a computer beat one under tournament conditions. It won by brute force, checking about 200 million positions a second, not by understanding chess. A symbolic-era triumph, and an early taste of "superhuman but utterly narrow."
• 2012 — the deep-learning spark. A deep neural network crushed the ImageNet image-recognition contest, proving that learning features from raw data beat hand-crafting them. This lit the fuse on the modern boom.
• 2017 — the Transformer. Google researchers published "Attention Is All You Need," introducing the architecture behind essentially every modern chatbot: ChatGPT, Claude, Gemini. Crucially, it was far easier to scale than what came before, which set up everything that followed.

Three moments, three ideas: brute-force rules, then learning from data, then the architecture that made today's chatbots possible.

One thing those tidy dates hide is the gap between them. Five years passed between the Transformer in 2017 and ChatGPT reaching the public in 2022 — years spent scaling the idea up and teaching it to follow instructions and stay helpful. Breakthroughs rarely arrive as products; they arrive as papers, and the real-world impact lands years later, once someone has scaled and packaged them into something a person can actually use. Worth holding onto the next time a research headline promises to change everything by next quarter.

Why "this time is different"

Earlier AI booms came with big promises and then went quiet. Twice — in the mid-1970s and again in the late 1980s — the field over-promised, failed to deliver, and watched its funding collapse. These busts were so recognizable the field gave them a name: AI winters. (If that sounds like the hype cycle from the last lesson, it is exactly that. AI has ridden the curve from peak to trough before.) So why believe this wave is different?

Because generative AI crossed a line the others never did: it handles language and images, the open-ended, human stuff, well enough to be useful to ordinary people, not just specialists. When the technology landed in a chat box anyone could type into, adoption stopped being a research story and became an everyone story.

The scale of that shift is easy to miss. When ChatGPT opened to the public in late 2022, it reached an estimated 100 million users in about two months — at the time the fastest-growing consumer app on record. Nothing about the underlying math was magic; what changed was that the technology finally spoke plain language, so anyone could use it without a manual.

The engine: scaling data, compute, and parameters

Here is the single most important idea in this lesson. Modern AI improves through scaling:

More data + more compute (computing power) + more parameters (the model's adjustable knobs) tends to produce a more capable model.

Each ingredient pulls its own weight. Data is the raw experience: the text, images, and code the model learns from. Compute is the raw horsepower: the chips (mostly GPUs) and the weeks of training time, which is why serious AI now runs in warehouse-sized data centres. Parameters are the adjustable knobs: more of them let the model capture more subtle patterns. Grow all three together and capability climbs; starve any one and the other two cannot make up for it.

It helps to have a rough sense of the magnitudes involved. Early models learned from tidy, hand-curated datasets a small team could inspect by eye. Today's frontier models train on a sizeable slice of the public internet — a body of text no human could read in a thousand lifetimes — using thousands of specialised chips running continuously for weeks or months inside a single data centre. The cost of one training run has climbed from trivial to something only the largest organisations on earth can afford. That is what "scaling" means in practice: not a tweak, but an industrial effort in data, chips, and electricity. It is also why AI has quietly become as much an infrastructure business as a software one, and why the names that dominate it are the ones that can afford the buildings.

This is why a stronger model seems to ship every few months. It is not (usually) a sudden genius breakthrough. It is often the same recipe run bigger: more text, more chips, more knobs, more training time. That predictability is exactly what makes AI a strategic force rather than a lucky streak, and it is why companies are pouring billions into data centers and chips (a theme you will revisit in The 5-Layer AI Stack).

To feel what scaling bought, walk one family forward: GPT-2 (2019) could stay on topic for a sentence or two before drifting; GPT-3 (2020) wrote coherent paragraphs; GPT-4 (2023) handled professional-level exams and multi-step problems that earlier versions flunked; the 2025 frontier reasons through problems step by step. The same basic recipe, each time run bigger, and each step unlocked things the one before could not do.

Why it all came together now

If scaling is the recipe, an obvious question follows: why did it start working around 2012 and not in 1990? The answer is that scaling needs three ingredients at once — data, compute, and algorithms — and for most of AI's history at least one of them was missing.

• Data arrived because the internet quietly assembled the largest library in history: decades of digitised text, images, and code, all in a form a machine could actually read.
• Compute arrived from an unlikely place. GPUs — chips originally built to render video-game graphics — turned out to be almost perfectly suited to the repetitive math inside neural networks, and years of gaming demand had already paid to make them fast and comparatively cheap.
• Algorithms arrived with deep learning in 2012 and the Transformer in 2017: designs that, crucially, kept getting better the more data and compute you poured into them, instead of plateauing.

For decades AI had at most one or two of these at a time, which is a large part of why the winters happened — the ambition was there, the fuel was not. Only in the late 2010s did all three land together. That convergence, not a lone genius or a single eureka moment, is why the boom happened when it did. And it is why the progress feels durable rather than fragile: the three ingredients are still abundant and still growing, so the engine has more room to run.

Where the frontier is heading

Two directions dominate the current edge:

• Reasoning models that "think" for longer before answering, working through a problem step by step instead of replying instantly. This trades speed for better answers on hard problems.
• Agents: AI that does not just answer but takes actions, using tools, browsing, and completing multi-step tasks on your behalf. You will meet agents hands-on in the Generative AI for Everyone course.

Both directions are really the same scaling story pointed at a new axis. Once ever-bigger models began to plateau on raw knowledge, the fresh gains moved to letting a model spend more effort at the moment you ask — thinking longer, checking its own work, calling tools — rather than only being trained bigger beforehand. In other words, the frontier is learning to scale thinking time, not just training size. That is the shift worth watching, because it changes where the next capability jumps will come from.

Why this matters: plan for a moving target

History is not the point; leverage is. Because progress runs largely on scaling, it is predictable rather than a lottery, and that changes how you should act:

• Plan for a moving target. The model you judge today will likely be beaten within months. Build around "models keep improving," not around one fixed snapshot of what AI can do.
• Don't over-invest in today's limits. A weakness that blocks you now may quietly vanish with the next generation. Re-test rather than assume a limit is permanent.
• Watch the frontier, not the noise. Reasoning and agents are where the next real gains are landing. That is where to point your attention, and your first small experiments.

You do not need to predict the future to use this. You only need to expect steady improvement and stay ready to move when it arrives.

What you can now do with AI's evolution

• Name the three waves of AI: symbolic, machine learning, deep learning
• Explain why generative AI broke through to the mainstream when earlier waves did not
• Use scaling to explain why a stronger model keeps arriving every few months
• Point to where the frontier is moving next: reasoning and agents

What's next: How Machines Learn

You now know AI learns from data rather than hand-written rules. But what does "learn" actually mean? Next: How Machines Learn, with zero equations.