A Short History of AI
From symbolic rules and machine learning to LLMs. What each paradigm shift replaced.
Don't memorize the years. The one thing to see: every generation of technology appears because the stretch still done by humans became unsustainable. That line keeps moving right, but it never disappears.
Signs you'll recognize:
- You treated today's generative AI as something that appeared overnight.
- You heard about "AI winters" and assumed the field shut down completely.
- You compared expert systems, machine learning, deep learning, and LLMs as four parallel products.
A line that keeps moving
Every era of AI is made of two parts: the stretch done by humans, and the stretch the machine learns on its own. All of technological progress is this boundary shifting right.
The motive is always the same: the human part became too much to do. When rules couldn't be written fast enough, machines learned from data. When feature design couldn't keep up, networks extracted features themselves. When training a separate model per task became endless, we trained one foundation and changed its use with prompts.
This view beats memorizing years because it answers a more practical question: each time the line moves, the skills that pay well change too.
Four shifts
Shift one: humans write rules. Early researchers believed that if knowledge were represented as symbols and enough reasoning rules were written, machines would show intelligence. Expert systems genuinely worked in well-bounded domains—medical diagnosis, equipment troubleshooting—where rules are finite and experts can articulate them.
They hit the knowledge-acquisition bottleneck: real life has too many exceptions, the rulebook keeps growing, and changing one rule ripples through everything. Maintenance cost exploded with scale. The so-called "AI winters" were investment and expectations retreating, not research stopping—the underlying work kept going.
Shift two: machines learn from samples. Machine learning stopped asking humans to write every judgment rule. Instead, you provide labeled data and the algorithm finds statistical relationships between inputs and answers. The bottleneck moved from "writing rules" to "preparing data and designing features."
This era had the most tedious human work: to classify spam, someone thought up features like "how many exclamation marks" or "is the sender in your contacts" and fed them in one by one. Feature quality decided the result; that craft was valuable back then.
Shift three: networks learn representations. Deep learning used multi-layer neural networks to extract features from raw data layer by layer—no more deciding which features to look at. AlexNet's 2012 result on large-scale image classification made "GPU + big data + deep networks" the mainstream recipe. alexnet
But a problem remained: every task needed its own model. One for image classification, one for translation, one for sentiment, each with its own labeled data. Tasks are endless; models couldn't be trained one by one forever.
Shift four: train one foundation, change its use. The 2017 Transformer made large-scale language training easy to parallelize, and scaling finally took off. transformer The recipe became: pre-train one general-purpose base on massive text so it learns the patterns of language itself, then use the same model for different jobs just by changing the prompt—no retraining.
Today's LLMs are the result of that path reaching scale. The human part has been compressed to its shortest, but it's still there, just with a different job: define the task, prepare the context, verify the result.
Six milestones
If you need a timeline to drop into a document, these six suffice.
| Year | What happened | What actually changed |
|---|---|---|
| 1950 | Turing proposes the imitation game turing | Turns the philosophical question "can machines think?" into an observable, discussable behavioral question. |
| 1956 | Dartmouth workshop dartmouth | "Artificial intelligence" becomes a formal research field with a name and boundaries. |
| 1970s–80s | Expert systems rise, then cool off | Proves narrow-domain rules have value, and exposes the ceiling of knowledge acquisition and maintenance. |
| 2012 | AlexNet alexnet | The data-GPU-deep-network combination establishes overwhelming dominance on perception tasks. |
| 2016–17 | AlphaGo and the Transformer alphagotransformer | The former shows complex decision-making; the latter builds the parallel-scalable foundation for LLMs. |
| 2022 onward | Conversational LLMs go mainstream | Natural language becomes a universal interaction layer; AI moves from a single capability into workflows. |
These milestones mark shifts in engineering method and product entry, not "intelligence levels." Baseline: August 2026, following the original papers and institutions' public materials.
Three reminders for builders
One: capability progress and reliability progress are not the same. Every wave starts with dazzling demos, and the evaluation, process, and safety engineering arrive only later. The impressive demo you saw today is usually still a full phase away from shippable. Judge an AI feature by its maturity tier, not its demo.
Two: old paradigms don't disappear; they settle into their proper roles. Rules still fit permissions, billing, and risk control—places that demand one hundred percent certainty. Statistical models still fit stable classification tasks. LLMs fit open-ended natural language. Mature products mix all three, putting the deterministic parts in rules instead of handing everything to the model.
Three: each expansion trades in one bottleneck for another. The rule era lacked knowledge; machine learning lacked labels; deep learning lacked compute; the foundation-model era lacks high-quality context, verifiable feedback, and controllable action boundaries.
All three are product definition problems, not model capability problems. That's why the most valuable people today aren't the ones who tune parameters—they're the ones who can describe the task and define the acceptance criteria. The red line has moved to your feet.