A brief history of AI

A summer workshop at Dartmouth College, organized by John McCarthy and others, uses the term "artificial intelligence" and launches it as a research field.

Frank Rosenblatt builds a machine that learns to recognize simple patterns by adjusting weighted connections, an early ancestor of today's neural networks.

Neural networkMachine learning

Joseph Weizenbaum's ELIZA mimics a therapist with simple pattern matching. People confide in it anyway, which troubles its creator.

ChatbotAnthropomorphismNatural language processing

The UK's Lighthill report concludes AI has failed to deliver on its promises. Funding is cut on both sides of the Atlantic for years.

Companies spent heavily on hand-coded rule systems through the 1980s. When the specialized hardware market collapses, a second AI winter follows.

IBM's chess computer defeats the reigning world champion in a six-game match, mostly through brute-force search rather than learning.

Fei-Fei Li's team releases a dataset of millions of labeled images. Its yearly competition becomes the benchmark that sparks the deep learning boom.

Training dataComputer vision

IBM's Watson beats two champions on live TV by parsing tricky clues. Later that year, Apple's Siri puts a voice assistant in millions of pockets.

Natural language processing

Ian Goodfellow introduces generative adversarial networks, in which two networks compete to create and spot fakes. It's an early step toward generative AI.

Generative AIDeepfake

DeepMind's system wins at Go, a game long thought too intuitive for computers, combining deep learning with search and self-play.

OpenAI's first GPT and Google's BERT show that a model pretrained on lots of text can be fine-tuned for many language tasks.

LLMFine-tuning

With 175 billion parameters, GPT-3 can do new tasks from a few examples in the prompt. Bigger models, it turns out, keep getting more capable.

LLMParametersPrompt

DeepMind's AlphaFold 2 predicts protein structures with near-lab accuracy at the CASP14 competition, a breakthrough for biology.

GitHub Copilot, built on an OpenAI model, starts suggesting code as developers type. It's the first AI tool many programmers use every day.

Agentic coding

OpenAI releases a chatbot tuned with human feedback. It reaches 100 million users in about two months, and hallucination becomes a household word.

ChatbotLLMRLHFHallucination

OpenAI's GPT-4 and Anthropic's Claude launch in March, Meta releases Llama 2 with open weights in July, and Google's Gemini arrives in December.

Frontier modelMultimodalOpen-weight model

The first broad AI law sorts systems by risk and phases in obligations over several years, from banned uses to rules for high-risk systems.

BiasGuardrails

OpenAI's o1 spends extra time reasoning step by step before answering, opening a new way to improve models beyond making them bigger.

Reasoning modelTest-time computeChain of thought

In October, Anthropic shows Claude operating a computer by clicking and typing. In November, it releases MCP, an open standard for connecting AI to apps and data.

Computer useMCPTool use

Andrej Karpathy coins vibe coding, and agentic coding tools like Claude Code start planning, editing, and testing code across whole projects.

Vibe codingAgentic codingAI agent

OpenAI releases GPT-5 in August and Google releases Gemini 3 in November, each pitched as a step up in reasoning and agent work.

Frontier modelReasoning model

Anthropic introduces Agent Skills in October. In December, MCP moves to the new Agentic AI Foundation under the Linux Foundation, co-founded by Anthropic, Block, and OpenAI.

SkillsMCPContext engineering

Researchers coin workslop for polished but hollow AI output at work, and Merriam-Webster names slop its 2025 Word of the Year.

SlopWorkslop

AI data centers buy up so much memory that chip shortages raise prices for laptops, phones, and AI services alike.

RAMageddonCompute