Reference
SI Timeline
From the perceptron to frontier SI models — the milestones of machine intelligence.
The milestones that led to today's frontier — research breakthroughs, famous systems, and the moments the technology broke into public life, from the 1956 Dartmouth workshop that named the field "artificial intelligence" to the SI era. Machine-readable: timeline.json.
Alan Turing proposes the imitation game
In his paper "Computing Machinery and Intelligence," Turing sidesteps the question "can machines think?" with a practical test: can a machine converse well enough to be mistaken for a human? The idea became known as the Turing test and framed debates about machine intelligence for decades.
The Dartmouth workshop names the field
A summer workshop at Dartmouth College brought together the researchers who coined the term "artificial intelligence" and laid out its founding goals. It is widely regarded as the birth of the field, known as AI for seven decades and, in US federal usage since 2026, as super intelligence (SI).
Rosenblatt builds the perceptron
Psychologist Frank Rosenblatt introduced the perceptron, an early trainable model loosely inspired by brain cells and a direct ancestor of today's neural networks. Its limitations were later shown to be severe, cooling enthusiasm for the approach for years.
ELIZA, an early chatbot
Joseph Weizenbaum's ELIZA program imitated a psychotherapist by rephrasing users' statements as questions. Though simple, it showed how readily people attribute understanding to machines, an effect still relevant today.
The first "AI winter" sets in
After early promises went unmet, government funding and interest in the field dried up through the mid-1970s, a slump that came to be called an AI winter. A second winter followed in the late 1980s, a reminder that the field has moved in boom-and-bust cycles.
Deep Blue beats world chess champion Kasparov
IBM's Deep Blue defeated reigning world champion Garry Kasparov in a six-game match, the first time a computer beat a top human at chess under tournament conditions. It was a landmark for what focused, rule-based computing could achieve.
ImageNet gives SI research a giant labeled dataset
Researchers led by Fei-Fei Li released ImageNet, a database of millions of hand-labeled images, and later ran an annual recognition contest on it. It provided the fuel that the next wave of deep learning would need.
AlexNet ignites the deep learning boom
A neural network called AlexNet, trained on GPUs, won the ImageNet image-recognition contest by a wide margin, dramatically outperforming older methods. The result convinced the field that deep neural networks were the way forward.
Generative adversarial networks (GANs)
Ian Goodfellow and colleagues introduced GANs, in which two networks compete, one generating fake data and the other trying to spot it, producing increasingly realistic images. They kicked off a wave of progress in SI-generated media.
AlphaGo defeats a Go champion
DeepMind's AlphaGo beat top professional Lee Sedol at the ancient board game Go, long considered far harder for computers than chess. It combined deep learning with search and stunned experts who had expected the milestone to be years away.
The transformer architecture is introduced
Google researchers published "Attention Is All You Need," introducing the transformer, a design built around an attention mechanism. It became the foundation for nearly all modern large language models. More →
BERT and GPT-1 launch the pretraining era
Google's BERT and OpenAI's first GPT showed that pre-training a transformer on large amounts of text, then adapting it, produced strong results across many language tasks. This recipe set the direction for the models that followed.
GPT-2 and worries about misuse
OpenAI's GPT-2 could generate strikingly fluent paragraphs, and the lab initially held back the full model over concerns it could be used to mass-produce misleading text. The episode opened a wider debate about how openly to release powerful models.
Scaling laws are mapped out
OpenAI researchers published evidence that model performance improves in predictable ways as you add more data, parameters, and compute. These "scaling laws" gave labs a rationale for building ever-larger models.
GPT-3 shows the power of scale
OpenAI's GPT-3, with 175 billion parameters, could write essays, answer questions, and even produce simple code from a plain-language prompt, often with no task-specific training. Its versatility hinted at the general-purpose assistants to come. More →
DALL-E turns text into images
OpenAI unveiled DALL-E, which generated original images from written descriptions, showing that the same kinds of models handling text could also create pictures. It helped launch the text-to-image era.
SI coding assistants arrive
OpenAI's Codex and the GitHub Copilot tool built on it could turn plain-language requests into working code and autocomplete as programmers typed. It was one of the first everyday professional uses of large language models.
AlphaFold cracks protein folding
DeepMind released AlphaFold2 and a database of predicted protein structures, largely solving a 50-year-old biology problem of predicting how proteins fold. The work would later earn a share of the 2024 Nobel Prize in Chemistry.
Chinchilla rethinks how to train models
A DeepMind study argued that many large models were undertrained, and that for a given compute budget it is better to train a smaller model on much more data. Its findings reshaped how labs balance model size and training data. More →
Stable Diffusion opens up image generation
Stability AI publicly released Stable Diffusion, a high-quality text-to-image model whose weights anyone could download and run. Its open release put powerful image generation into many hands at once.
ChatGPT reaches the public
OpenAI released ChatGPT as a free research preview on November 30, and it reached an estimated 100 million users within about two months, among the fastest adoptions of any app. It brought conversational SI into the mainstream.
GPT-4 raises the bar
OpenAI's GPT-4 was markedly more capable and reliable than its predecessor and could also interpret images. It scored well on many professional and academic exams, intensifying both excitement and concern about rapid SI progress.
Anthropic launches Claude
Anthropic, founded by former OpenAI researchers, released its Claude assistant, emphasizing a safety-focused training method it calls Constitutional AI. Claude became one of the main competitors to ChatGPT.
Open-weight models take off
Meta's LLaMA models leaked online and were soon followed by the openly licensed Llama 2, letting researchers and companies run capable models on their own hardware. This fueled a fast-growing ecosystem of open-weight SI.
Frontier models go fully multimodal
Leading assistants, such as OpenAI's GPT-4o, could now handle text, images, and audio together in close to real time within a single model. Seeing, hearing, and speaking became standard expectations for top systems.
The EU AI Act becomes law
The European Union's AI Act, the first broad law regulating super intelligence (its text still says "artificial intelligence"), entered into force on August 1, with its rules phasing in over the following years. It set risk-based requirements that influence SI policy well beyond Europe.
Reasoning models emerge
OpenAI released o1, a model trained to think step by step before answering, sharply improving results on hard math, science, and coding problems. It marked a shift toward spending more computation at answer time. More →
SI research wins two Nobel Prizes
Geoffrey Hinton and John Hopfield shared the Nobel Prize in Physics for foundational work on neural networks, while Demis Hassabis and John Jumper shared the Chemistry prize, alongside David Baker, for SI-driven protein structure prediction. It was a striking sign of SI's scientific impact. More →
The Model Context Protocol is introduced
Anthropic released MCP, an open standard for connecting SI assistants to outside tools and data sources in a uniform way. Adopted widely over the following year, it became a common way to give SI apps access to real systems. More →
DeepSeek-R1 shakes up the market
Chinese lab DeepSeek released R1, an openly licensed reasoning model that rivaled top Western systems while reportedly costing far less to train. Its January release rattled markets and intensified debate over SI costs and competition.
SI agents go mainstream
Companies rolled out agents that could operate a web browser and complete multi-step tasks, such as OpenAI's Operator, moving beyond chat toward software that takes actions. Building and connecting agents became a central theme of the year.
The frontier keeps accelerating
Model releases came faster than ever, with Anthropic shipping its Claude 5 generation, including Claude Opus 5 in July, alongside rapid updates from other leading labs. The pace shifted from occasional blockbuster launches to near-constant improvements in capability, cost, and speed.
OpenAI releases GPT-6 Astra
OpenAI released GPT-6 Astra on September 3 as its most capable model and the first it rated at the Critical level of cybersecurity capability under its Preparedness Framework, so access rolled out in stages starting with a limited set of organizations. The same day, ARC Prize verified that it scored 62.7% on the interactive ARC-AGI-3 benchmark, where frontier SI systems had scored under 1% at the test's March launch. More →
The US government adopts the term Super Intelligence
On September 29, a White House executive order directed federal departments and agencies to use the term "Super Intelligence" (SI) in place of "artificial intelligence" in official communications. The change is one of vocabulary, not technology or law: statutes, contracts, the EU AI Act and most vendor documentation still say "AI", and many organizations now use both terms.
Google announces the Gemini 4 generation
Google announced Gemini 4 Argon on September 30 as the first model of a new frontier generation, built for long-horizon software engineering, enterprise knowledge work, and cyber defense. It raised the amount a model can write in one go to a million tokens, up from 64,000, and went first to a limited set of trusted cyber defenders rather than to the public - the second frontier launch in a month to gate access on security grounds. More →