A short history of AI

AI is nearly seventy years old, and it has gone through the same cycle more than once: an idea shines in demos, promises outrun results, and funding freezes. Knowing that cycle helps you tell what actually changed from what was merely renamed.

We will ask one question of each wave: what did it change about the work of building software? This lesson covers four turns: from hand-written rules to learning from data, the design behind today's language models, the step from research model to product, and why today's work is about controlling models that act.

From rules to data

In 1956, a workshop at Dartmouth launched AI as a field. Its proposal, written by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, argued that every aspect of intelligence could be described precisely enough for a machine to simulate it. So people wrote rules and facts, and programs reasoned with them: symbolic AI.

In 1958, Frank Rosenblatt published the perceptron, a machine that got no rules: it adjusted a set of internal numbers, its weights, after each example until its answers improved. In 1969, Minsky and Papert's book PerceptronsPerceptron Rosenblatt's 1958 single-layer model that learns weights from examples; it can only tell apart categories that a straight line (or flat plane) can separate. showed that such a simple machine could not learn some basic patterns, and interest cooled. In 1973, the Lighthill report told the British government that AI had not kept its promises. The cuts that followed are the first AI winterAI winter A period of sharply reduced funding and interest in AI after promises outran results; the two main ones began in the 1970s and in 1987..

The 1980s brought expert systemsExpert system A 1980s-style program that encodes specialist knowledge as hand-written if-then rules; it does not learn.: thousands of rules written with specialists, such as XCON, which DEC used to configure its computers. They could not learn, so every exception was one more rule to maintain. When the market for the specialized machines they ran on collapsed around 1987, the second winter arrived.

In 1986, Rumelhart, Hinton and Williams popularized backpropagation, a way to train networks with several layers. In the 1990s, statistical methods such as support-vector machines took over. The engineer's job moved from writing rules to collecting data and choosing by hand which traits of it, the features, the model should look at.

Then came images. ImageNet (2009) supplied millions of labeled photos. In 2012, AlexNet, a many-layered network trained on graphics cards (GPUs), won the ImageNet competition by a wide margin. From then on, the network found the useful features by itself. That is deep learning, and its basics are in Machine learning in one lesson.

From language to the transformer

For language, words first had to become numbers. word2vec (2013) turned each word into a list of numbers, placed so that words with similar meanings sit close together. It is the ancestor of the embeddingsEmbedding A list of numbers that captures the meaning of a text, so texts with similar meanings get numbers that sit close together. in Embeddings and RAG.

In 2014, seq2seq translated whole sentences by reading them word by word, and Bahdanau, Cho and Bengio added attentionAttention The mechanism that lets a model decide, for each token, how much to draw on every other token in the text.: at each step, the model decides which parts of the input to focus on, like a translator glancing back at the relevant words. In 2017, "Attention Is All You Need" (Google) kept attention and dropped the word-by-word reading. This design, the transformerTransformer The 2017 neural network design behind today's LLMs, built around attention and able to process all the tokens of a text at once., looks at all of a text at once. That suits GPUs, which do many calculations in parallel, and made trainingTraining Adjusting model weights from data so its predictions improve. on far more text practical.

A new pattern followed: pretrain once, adapt many times. One organization trains a large general model on a huge amount of text; many others adapt it. OpenAI introduced GPT in 2018, and Google released BERT the same year. In 2020, scaling laws showed that these models improve predictably with more size, data and computing power.

Also in 2020, GPT-3 showed that, given a few examples inside the request, a model could handle new tasks without any retraining. This is in-context learningIn-context learning A model's ability to perform a new task from instructions or examples in the prompt, without updating its weights., and it changed the interface: many tasks moved from training a model to writing the right text for it, the promptPrompt Everything sent to the model in one call: instructions, conversation messages and any tool definitions.. More in LLMs as software components.

From model to product

InstructGPT (2022) turned a text predictor into an assistant with RLHFRLHF Reinforcement learning from human feedback: tuning a model with human preference judgments so it follows instructions. (reinforcement learning from human feedback): people rated answers, and the model was tuned toward the ones they preferred. OpenAI launched ChatGPT, a sibling of InstructGPT, on November 30, 2022, bringing that ability to the general public.

In 2023, Meta released LLaMA under a research license and then Llama 2, whose license allows commercial use. With open-weight modelsOpen-weight model A model whose trained weights are published for download, under a license that may still restrict use., anyone can download a trained model and run it on their own machines, usually shared through the Transformers library and the Hugging Face HubHugging Face Hub A platform that hosts models, datasets and demos, each with its own documentation, versions and license.. See Models, open weights and Hugging Face.

The latest wave connects models to the world:

• Outside knowledge. RAG (2020) looks up relevant documents when a question arrives, instead of expecting the model to have memorized everything.
• Actions. Research such as ReAct and Toolformer taught models to use tools. In 2023, OpenAI shipped function calling: the model says which function it wants run and with which values, and your program decides whether to run it. See Tool calling and harnesses.
• Code. GitHub Copilot (2021) suggested code as you typed. In February 2025, Anthropic introduced Claude Code, an agentAgent A system in which a model, in a loop, chooses the next action until it decides it is done or hits a stop condition. that reads a codebase, edits files and runs commands. The same month, Andrej Karpathy coined vibe codingVibe coding Karpathy's 2025 term for building software by accepting AI-generated code without reading it. for accepting generated code without reading it.
• Standards. In November 2024, Anthropic published the Model Context Protocol (MCPMCP The Model Context Protocol: an open standard that lets AI applications connect to tools, data and prompts through one common interface.), an open standard for connecting assistants to the systems where data lives. See MCP.

Putting it into practice

• Build on what improves by itself. Base modelsBase model A model after its first, general training, which only continues text and has not yet been tuned to follow instructions. get better each generation; hand-written workarounds for today's limits tend to go stale.
• Don't mistake a demo for a system. Ask for tests on real data, measured costs and a plan for switching to another model or provider.
• Keep the fast-changing layer replaceable. Models turn over every few months. Your data, tests and control layer last longer, so effort there pays off.
• Treat governance as real work. Once a model can act, permissions, approvals and audit become central design decisions.

The course's three case studies are open source and belong to this wave. Adapta, a prototype, combines RAGRAG Retrieval-augmented generation: finding relevant passages in your documents and adding them to the prompt before the model answers. and LoRALoRA Low-rank adaptation: a cheaper way to fine-tune that trains small add-on matrices instead of all of the model's weights. to customize models on your own infrastructure (Embeddings and RAG, Fine-tuning, LoRA and QLoRA). docket, in beta, is a runtime that governs agents with policies and approvals (Governance and human-in-the-loop). Tack is a project manager that, on its development branch, coordinates agent execution (Durable agent execution).

Sources

• Symbolic era and winters: Dartmouth proposal, Rosenblatt 1958, Lighthill report, second winter
• Learning from data: backpropagation, SVMs, ImageNet, AlexNet
• Language: word2vec, seq2seq, attention, transformer, GPT, BERT, scaling laws, Chinchilla, GPT-3
• Product: InstructGPT, ChatGPT, LLaMA, Llama 2, Transformers
• Tools and agents: RAG, ReAct, Toolformer, function calling, Copilot, Claude Code, MCP, Karpathy on vibe coding

Key takeaways

• AI has been through two winters; both came when promises outran what the technique could support.
• The long-term shift was from writing rules by hand to learning them from data.
• The transformer made it practical to train huge language models once and adapt them to many tasks, often with just a prompt.
• Human feedback turned text predictors into assistants; open weightsOpen weights A model whose trained parameters can be downloaded and run anywhere, regardless of whether its code or data are published. and function calling turned them into components you can host and connect.
• Today's work is about control: what a model may do, who approves it, and how its actions are checked.

Check yourself