AI History Module · Beginner · for younger readers
How AI Became What It Is Today
The AI we use today did not suddenly appear one day. It is made of parts that people built, one at a time, over 80 years. This page follows them in order, from the very beginning. Whenever a hard word comes up, it is explained right there.
1,295records checked
1943–2026years covered
22axes
87%from primary sources
Four units
Read in order and each later idea is explained by the earlier ones. Unit ① sets up the frame, ② sweeps through the eras, ③ narrows to the last three years, and ④ sorts out the vocabulary.
Start with what AI actually does. Get these five things and the rest falls into place.
01
How four words overlap
These four words don't sit side by side. They are more like small boxes inside bigger boxes.
Artificial intelligence is the biggest box. It means anything that gets a machine to make the kinds of decisions people used to make. Long ago, this also included having people write in the rules one by one.
Machine learning is the box inside it. Instead of people writing the rules, you show the machine lots of data and let it find the rules by itself.
Deep learning is the box inside that one. It is the kind of machine learning that stacks many layers, deep. Almost all the AI people talk about today lives here.
Generative AI is not a box. It is a name based on "what it does." If it doesn't pick things out or guess answers but makes something new, it's generative AI. So in some places it sits inside the deep learning box, and in others it hangs over the edge.
Experiment
What sits inside what
Press the buttons. The first three are nested inside each other; only generative AI uses a different criterion, so it overlaps them with a dashed line. That's why the question “which is broader, generative AI or deep learning?” doesn't make sense.
02
AI does just one thing
It practiced guessing "what word comes next" a huge number of times. That's all.
AI was given only one exercise. Show it the first part of some text, and have it guess the next word. This is repeated billions of times with text from the internet.
It sounds strange, but once AI gets good at just this, a lot comes along with it. After "Here is that sentence in French:", the natural thing to come next is a French translation. After "In short:", what comes next is a summary.
So AI never learned translation as a separate skill. Because it got good at guessing the next word, it can translate too.
Try it yourself below. It shows AI coming up with several candidates for the next word and picking one. Turn the knob and you change how much it goes for the safe answer and how much it takes a chance.
May 28, 2020OpenAI (Brown et al.) — GPT-3 — performs new tasks from examples in the prompt alone, with no weight updates Source
In one lineAI only learned to guess the next word. Translating and summarizing came along from that.
Experiment
Picking the next word
The cat quietly ____ on the mat
Turn the knob to the left and you get the same answer every time; turn it to the right and you get a different answer each time. That's why AI answers a little differently each time. — The numbers here aren't from a real model; they're made-up values for illustration.
03
Building it and using it
Building it takes months on huge computers. Using it takes seconds. These are two completely different stories.
Making an AI has three stages.
First, it reads a huge amount of text (pretraining). Tens of thousands of computers run for months. This is the most expensive part.
Second, it is taught manners (post-training). It learns things like doing what it's asked and not doing dangerous things. This is much cheaper than the first stage.
Third, it is actually used (inference). Every time we ask a question, a little money is spent.
The first stage only has to happen once, but the third keeps going every time people use it. So companies would rather work a bit harder while building it, if that means it's cheaper to use.
Mar 29, 2022DeepMind (Hoffmann et al.) — Chinchilla — optimal allocation in terms of training compute Source
Dec 31, 2023Sardana, Frankle et al. — Beyond Chinchilla-Optimal — once inference demand is included, overtraining is optimal Source
In one lineBuilding costs money once; using costs money again and again. So these days, people work harder while building so that using is cheap.
04
The four parts that make up an AI
It isn't just a brain. There's also the request you make, the material it's looking at and the tools in its hands.
The model is the brain. It reads text and comes up with the next word.
The prompt is the request. It's the part where you write what you want done and how.
The context is the material spread out in front of it right now — the conversation so far, documents it looked up, files you pasted in.
Tools are what it holds in its hands. AI can call things like a calculator, a search or a file reader by itself.
When an answer is strange, people tend to think "the brain isn't smart enough." But far more often, it's actually because the request was vague, the needed material wasn't given, or a tool was missing.
It's split not by the names of technologies but by which problem people were trying to solve.
This material is divided into 22 boxes. What divides them is not technology names but problems.
For example, the box called "Prompt Engineering" sounds like the name of a technique, but what it really holds is the problem "how should you talk to AI so it understands you well?" Even if the technique's name changes, that problem doesn't go away.
Below you can see how many records have piled up in each of the 22 boxes. The boxes with big numbers are where the world is moving a lot these days.
Figure
Records on each of 22 axes
Measured from the current records
The bigger the number, the more the world is moving in that area these days. This chart is redrawn automatically whenever the records grow.
Unit ②
How the Technology Changed, Era by Era
We split it into six eras. Each era starts from the problem the previous era left unsolved. Get that link and the rest is detail.
194320172021202320252026
011943 – 2016
The time when the parts were made, one by one
Most of the parts inside today's AI were made during this time.
In 1943, some people came up with the idea of a calculating device that copies the cells of the human brain. The problem was how to make it learn by itself.
The answer was simpler than you might think. When it is wrong, fix it a little. It tries an answer, then traces back by however much it missed, changing its numbers bit by bit. It does this over and over, countless times. In 1986, this method became widely known.
But when the layers were stacked deep, something strange happened. The signal telling it what to fix faded away on its way up. Stacking it deeper should have made it smarter, but once it was deep, it could not learn.
Getting over this wall took more than 20 more years. Then, in 2012, people ran the math super fast on computer graphics cards and gathered more than 1 million photos for it, and suddenly it started to work. It was not a new invention. The conditions were finally in place.
Oct 9, 1986Rumelhart, Hinton, Williams — Show in Nature that, through backpropagation, hidden layers build their own representations Source
Dec 2012Krizhevsky, Sutskever, Hinton — AlexNet — GPUs, large-scale data and a deep CNN combined Source
022017 – 2020
From reading line by line to seeing it all at once
Once AI stopped reading text one word at a time from the front, much bigger models became possible.
Until then, computers read text one word at a time, starting from the front. They had to finish one word before going on to the next, so long texts took a long time, and the beginning got forgotten.
The Transformer, which came out in 2017, stopped reading in order. Instead, it laid out all the words at once and let each word decide for itself which of the other words it should look at.
If everything can be handled at once, you can hook up thousands of computers and run them at the same time. So models could be made much bigger.
Then people discovered something amazing. The bigger they made the model and the more text they had it read, the better its scores got — and it happened in a regular, predictable way. They could even work out ahead of time how much better it would get. So companies started pouring money in.
Jun 12, 2017Google (Vaswani et al.) — Transformer — Attention Is All You Need Source
Feb 14, 2019OpenAI — Announces GPT-2 (1.5B) and decides not to release the full model over misuse concerns Source
In one lineStopping reading in order is what this whole era is about. Thanks to that, models could be made big.
Experiment
Reading one word at a time vs seeing it all at once
Thecatsatonthematquietlyatnoon
RNN0 / 9
TF0 / 9
Press the button on the left. The old way has to finish reading the earlier words before it can move on to the next one. The Transformer looks at everything at once. That is why many computers can be hooked up and run at the same time, and that is why models could be made big.
032021 – 2022
Getting AI to understand what we ask
AI did not suddenly get smarter. It was changed so that it listens to people.
GPT-3 already existed in 2020. But back then, it did not make the news. Why not?
AI back then only knew how to keep writing the words that come next. Ask it a question, and instead of answering, it might write more questions just like it. To be useful, it had to learn to do what it was told.
So people taught it by looking at two answers themselves and picking which one was better. Do this a huge number of times, and the AI starts giving "the answers people like."
There was one more discovery. When you give it a hard problem, adding one line — "Let's think step by step" — made it get more answers right. It made the AI write out how it solved the problem, not just the answer.
On November 30, 2022, ChatGPT came out with all of this inside. That is when the world took notice.
Jan 5, 2021OpenAI — CLIP — a shared embedding for images and text Source
Jan 27, 2022OpenAI — Announces InstructGPT and makes it the default API model (RLHF) Source
Jan 28, 2022Jason Wei et al. (Google Brain) — Chain-of-Thought Prompting — generating intermediate reasoning steps Source
In one lineAI did not get smarter. It learned to listen. GPT-3, which ChatGPT was built on, had already been around for 2 years.
042023 – 2024
Handing AI some tools
Instead of leaving AI to answer all on its own, people made it look things up, do math and press buttons.
AI had two weak spots. It pretends to know things it doesn't (hallucination), and it doesn't know about anything that happened after it learned.
The fix was simple: have it look things up before answering. It pulls documents that are similar to the question out of a storeroom, reads them along with the question, and then answers. This is called RAG.
Next, people handed it tools. They let the AI call on things like a calculator, search and a calendar by itself. In 2024, it could even look at the screen and move the mouse to use a computer directly.
And the amount of text it could read at once grew enormously. In 2020 it was a few pages of a book. By 2024 it could read several books' worth at once.
But there were failures too. Attempts to build an AI that does everything by itself once you give it a goal did not go well. It gets each step a little bit wrong, and over ten steps those little bits pile up until almost everything is wrong.
Feb 24, 2023Meta AI — Releases LLaMA 1 — later leaked, opening up the local-run ecosystem Source
Feb 15, 2024Google DeepMind — Gemini 1.5 Pro — a 1 million-token context in preview Source
In one lineIt became important not just to build one good AI, but to also get ready the tools and materials the AI will use.
052024-09 – 2025
Starting to buy time to think
Instead of building a bigger AI, people found that if you have it think longer before answering, it solves problems better.
Until now, the way forward was to make the model bigger. But that costs a huge amount of money.
In the fall of 2024, another path opened up. People kept the model the same size and had it think on its own for a long time before answering. The harder the problem, the more a longer think raised the share of right answers.
The interesting part is how it was taught. People did not show it every step of how to solve things. They gave it lots of problems, like math problems, where a computer can check automatically whether the answer is right, and gave it points when it got them right. Then the AI came up on its own with the habit of going back to review its work and switching methods.
That is why today's AI comes with a dial that sets "how long to think." Try turning it yourself below.
Sep 12, 2024OpenAI — o1-preview — Learning to Reason with LLMs Source
Jan 22, 2025DeepSeek — DeepSeek-R1 — reasoning ability emerges from pure RL Source
Feb 24, 2025Anthropic — Claude 3.7 Sonnet — hybrid reasoning, thinking budget up to 128K Source
In one lineInstead of making AI bigger, there is now a way to buy it time to think. But the longer it thinks, the more it costs.
Experiment
A dial that sets thinking time
tokens
Accuracy0%
Cost per answer×1.0
Move the handle. The longer it thinks, the more answers it gets right — but each step up adds less and less, while the cost keeps going up. This is the problem AI companies are wrestling with these days. — This graph is not real measurements. It is a picture that shows the shape of the relationship.
062026
The gear around AI starts to matter
How you set AI up to work now matters more than the AI itself.
There is a piece of gear for working with horses called a harness. It safely carries the horse's strength over to a cart.
AI in 2026 is just like that. What matters more than the model itself is the gear that decides where and how that strength flows. That means things like which tools to give it, what to have it remember and when a person should check its work.
Meanwhile, the areas where AI still cannot do what people do easily are clear, too. On a new test that came out in March 2026, people solved 100% of the problems, but the best AI solved only 0.51%.
And the number of open models that anyone can download and use grew a lot. As of August 2026, on one ranking, 96 out of 176 models were open models.
Mar 25, 2026ARC Prize — Launches ARC-AGI-3 — humans 100% vs. frontier AI 0.51% Source
Jul 21, 2026Ministry of Science and ICT — Enforcement Decree of the AI Basic Act takes effect Source
In one lineAI is sharply split between things it does very well and things it still cannot do. If you only see the news about what it does well and think it does everything well, you will be wrong.
Experiment
When did a lot happen?
Point at or tap a bar to see that year's count.
Show the numbers
This counts the records in Next Token, year by year. Point at or tap a bar. You can see it shoot up from late 2022 — after ChatGPT came out.
Unit ③
The Big Turning Points of the Last 3 Years
The big changes of 2024, 2025 and 2026, split into six threads.
01
How much AI can read at once grew enormously
AI that used to read a few pages of a book could now read several books at once.
In 2020, AI could read only about a few pages of a book at once. If the text was longer, it forgot the beginning.
By 2024, it could read several books' worth at once. The chart below makes the difference easy to feel.
Something else big happened the same year. AI started using a computer directly, by looking at the screen and moving the mouse. That means even if a program has no special connection made for AI, the AI can click around and use it like a person.
But there's a catch. Putting a lot in doesn't mean it all gets read. Just as people skim through the middle of a thick book, it turned out that AI also pays less attention to the middle of a long text.
Nov 6, 2023OpenAI — GPT-4 Turbo with 128K context Source
Feb 15, 2024Google DeepMind — Gemini 1.5 Pro — 1 million tokens in preview Source
Jun 27, 2024Google — Gemini 1.5 Pro's 2 million tokens opened to all developers Source
In one lineBeing able to put a lot in doesn't mean it all gets read. This problem still hasn't been solved.
Figure
The growth of the context window
Log scale · tokens
The bars use a log scale, so the real differences are much bigger than they look. From 2020 to 2024 it grew about 1,000×.
02
AI that thinks for a long time arrives
Instead of making models bigger, people had them think for a long time before answering.
o1 came out in fall 2024, and DeepSeek-R1 in January 2025. Both are AIs that think on their own for a long time before answering.
The way they were taught is interesting. Instead of being shown how people solve problems, they were given lots of problems where you can check automatically whether the answer is right (math and coding), and got points for right answers. Then the AI developed on its own the habit of going back, checking its work and switching methods.
But a worry came up too. Since we can read the AI's thinking process, people thought we could know what it's really thinking — but when researchers looked into it, it turned out that it doesn't write down everything it thinks.
Sep 12, 2024OpenAI — o1-preview — Learning to Reason with LLMs Source
Jan 22, 2025DeepSeek — DeepSeek-R1 — reasoning ability emerges from pure RL Source
03
AI and programs learn to talk to each other
A shared set of rules appeared for how AI calls other programs.
To give AI a tool, you used to have to build a separate way of connecting for each tool. If there were 100 tools, there were 100 kinds of connections.
MCP is a set of rules that settles on one single way to connect. It's like agreeing on one shape for electrical outlets. Any tool that follows the rules can plug into any AI.
Agent Skills solves a different problem. When teaching AI a skill, instead of making it read the whole manual, you show it only the title and let it open the manual when it needs to.
Both came from the same idea: AI can only look at so much at once, so bring in only what's needed, when it's needed.
AI is no longer something only the company that made it can use. Models anyone can download and use now make up more than half.
In the past, to use a good AI you had to go online to the maker's service. Now there are lots of AIs you can download and run on your own computer.
As of August 2026, on one ranking list, 96 of 176 models were ones you could download. That's more than half.
But the word "open" doesn't always mean the same thing. Some you may only download and use, some you may use to make money, and some need a separate contract if your company is big — they're all different.
The law actually moved too. In Korea, the Enforcement Decree of the AI Basic Act took effect on July 21, 2026.
Jul 21, 2026Ministry of Science and ICT — Enforcement Decree of the AI Basic Act takes effect — creates a confirmation scheme for AI products and services Source
Jul 27, 2026EU — AI Omnibus Regulation (EU) 2026/1744 enters into force — high-risk obligations postponed to 2027-12-02 and 2028-08-02 Source
In one lineNot every "open model" can be used however you like. Always check the conditions.
Figure
The 176 models on the leaderboard
As of Aug 26, 2026
As of August 26, 2026, 96 of the 176 models on one leaderboard were models you could download.
06
Things AI still can't do
If you only read news about what AI does well, you'll get the wrong idea. There are still sharp gaps in what it can't do.
AI can solve PhD-level problems and still fail at things a grade-schooler can do. This is called "jagged intelligence."
A new test released in March 2026 (ARC-AGI-3) shows this. Its problems are games you must figure out and solve on your own, with no instructions — people solve 100%, but the best AI solved 0.51%.
One more thing. Companies are bringing in a lot of AI, but the share that say it actually made them money hasn't gone up as much. In a 2026 survey, company-wide adoption was 44%, but only 37% said AI had added to their profits.
Mar 25, 2026ARC Prize — ARC-AGI-3 — humans 100% vs. frontier AI 0.51% Source
Jun 12, 2026Epoch AI — Revised FrontierMath — 12 problems corrected and 7 removed in Tier 4; 123 corrected and 5 removed in Tiers 1–3 Source
In one lineAI news gives big headlines only to what AI did well. You have to look at what it can't do too, to see how good it really is.
Unit ④
Basics You Need to Know
Words that come up a lot in the news. Know just these and you'll get the gist of what's being said.
30 terms
No match. Try another word, or read the terms in context in Units ①②③.
01
Tokens and the context window
AI doesn't count text letter by letter. It has a unit all its own.
Token
The unit AI uses to count text. A token can be a single letter or a chunk of several letters.
Context window
How much text AI can look at in one go.
Price per token
What it costs to use AI. The longer the text, the more it costs.
02
Parameters · weights · MoE
What the "size" of an AI means.
Parameter / weight
The numbers an AI keeps adjusting, bit by bit, as it learns. If it has lots of them, it's a big model.
MoE / active parametersMixture of Experts
A way of calling on only the few experts that are needed, out of many. Big, but fast.
Distillation
Passing what a big AI knows on to a small AI.
03
Pretraining · SFT · RLHF · RLVR
The order in which AI is taught.
Pretrainingpre-training
The stage where AI reads tons of text. It takes the longest and costs the most.
SFTsupervised fine-tuning
The stage where AI is taught by being shown examples of good answers.
RLHF / DPORL from Human Feedback
Teaching by having people pick the better of two answers.
RLVRRL with Verifiable Rewards
Teaching with problems where you can check automatically whether the answer is right. It was the key to reasoning ability.
04
Prompt · system prompt · prompt injection
How to talk to AI, and how someone else can secretly give it orders.
Prompt
Everything you say when you ask AI for something.
System prompt
Instructions given ahead of time, from above, that the user can't see.
Prompt injection
An attack that hides secret commands inside text the AI reads.
Lethal trifecta
When three things come together, information can leak out.
05
RAG · embeddings · chunking
How AI looks things up before it answers.
RAGRetrieval-Augmented Generation
Finding related documents before answering, then reading them together to answer.
Embedding / vector searchembedding
Turning text into a list of numbers, and placing things with similar meanings close together.
Hybrid search / reranking
Searching by meaning and by words together, then putting the results in order again.
Chunking
Cutting a long document into small parts that are easy to search.
06
Tool calling · agents · loops
How AI uses tools and works on its own.
Tool callingtool use / function calling
AI calling things like a calculator or a search by itself.
ReAct loopReason + Act
Repeating think → act → look at the result.
Compounding reliability
If you go a little wrong at every step, after many steps you end up almost always wrong.
ACIAgent-Computer Interface
Redesigning tools so it's hard for AI to make mistakes.
07
Hallucination · benchmarks · contamination
How to notice when AI is pretending to know.
Hallucination
AI making something up as if it knew, when it doesn't.
Contamination
Like the test questions having been printed in the textbook ahead of time.
pass@k
A score that counts as a pass if even one of k tries is right.
Evaluation awareness
AI noticing that "I'm being tested right now."
08
Quantization · KV cache · inference cost
How to run AI on a small computer.
Quantization
Storing numbers in a rougher, shorter form to save space. It gets a little less exact, but much lighter.
KV cache
Saving the part already read so it doesn't have to be read again. With long text, this eats up all the memory.
prefill / decode
The step of reading the question, and the step of writing the answer piece by piece.
Falling inference costLLMflation
Using an AI of the same quality gets much cheaper every year.