Timeline of Artificial Intelligence

• Pan & zoom — scroll to zoom, drag to pan. The X-axis is adaptive: antiquity is compressed while the post-1950 computing era is stretched for detail.
• Click any event dot for a detail panel with description, year, sources and related figures.
• Eras — click a period in the top strip to zoom to it, with a little context on either side; click it again (or press Esc) to return. The ← → keys step from era to era.
• Threads — hover a category to see its line through time. Click the category name on an open event to follow it: ← → then steps through that category's events, one by one.
• Clusters — some key events gather several moments, such as AI's entry into mathematics. Opening one lights up its moments on the timeline and lists them in the panel.
• Share — the address bar always holds the current view, open event and thread, so a copied link opens exactly what you see.
• Search or Surprise me — use the search bar to find an event, model, lab or era; the dice button jumps to a random one.
• View — toggle the year grid, the present-year line, the two AI Winter bands, and the gradient fill below the curves.
• Periodizations — overlay one of seven scholarly historical framings (AI Eras, Industrial Revolutions, Information & Communication, Socio-economic Regimes, Energy Regimes, World History, Existential Moods). Computing Eras can be shown as an independent overlay alongside any of the seven.
• Trends — Moore's Law is the left-axis backbone. Add one companion curve on the right axis: Training Compute, Scaling Laws, METR Task Horizon, Benchmarks vs Human, Context Window, Inference Cost, Datacenter Power, Intelligence Frontier, or Capabilities Frontier.
• Categories — filter event types (mythology, neural networks, game-playing AI, ethics, models, policy, infrastructure, and more).
• Stories — click the Stories button to browse 21 narrated stories woven across AI history.

Training Compute (cyan) charts the exponential growth of the compute used to train AI systems, measured in floating-point operations (FLOP). The cyan line is the frontier envelope — at any moment it tracks the most compute-intensive model published to date, jumping upward each time a record is set. Faint background dots show every notable model in the dataset (521 systems).

Show by Lab (sub-toggle under Training Compute) recolors the curve to reveal which research organizations have driven the trend. Each lab's monotonic record progression appears as its own coloured line; the frontier envelope is segmented by lab. Solo a lab with the S button; the Defaults button restores the five major frontier labs (Google/DeepMind, OpenAI, Anthropic, Meta, xAI).

The long-run pace is famously steep. Before 2010, frontier training compute doubled roughly every 20 months — in step with Moore's Law. Since 2010, it has doubled every ~6 months (Sevilla et al. 2022), a rate more than four times faster than hardware alone would predict, driven by larger budgets, specialised accelerators and massively parallel training.

Moore's Law (blue) — transistor counts in microprocessors from the Intel 4004 (2,300 transistors, 1971) to modern AI accelerators at over 100 billion. Drawn from Karl Rupp's public dataset, itself extending work by Horowitz, Labonte, Shacham, Olukotun, Hammond and Batten.

METR Task Horizon (amber) — the length of real software-engineering tasks (in minutes of expert human time) that each frontier AI model can complete with 50% success. Horizon has roughly doubled every seven months since 2019, rising from ~3 seconds for GPT-2 to ~12 hours for Claude Opus 4.6 (Kwa et al. 2025).

Benchmarks vs Human (multi-colour) reproduces Figure 2.1.1 of the Stanford HAI AI Index 2026. Eleven reference benchmarks — ImageNet, SuperGLUE, MMLU, GPQA Diamond, OSWorld, SWE-bench, VQA, SQuAD 2.0, MATH, MMMU and AIME — are scaled so that the human baseline = 100%. Solid lines track active benchmarks; dashed lines mark benchmarks that have saturated.

Scaling Laws (peach) — the scaling laws are empirical power laws: a language model's error falls smoothly and predictably as its parameters, training text and training compute grow. They relate error to size, not to dates, so this curve does not draw them. It draws the decision they govern: how each model split its training budget between model size and training text, as tokens per parameter. Kaplan et al. (2020) concluded that, as budgets grow, most of the extra compute should go into a bigger model, and GPT-3 trained on fewer than two tokens per parameter. Hoffmann et al. (2022), the Chinchilla study, found that size and text should grow in step, at about 20 tokens per parameter across the budgets they tested (the dashed line); that is the split that reaches a given quality for the least training compute. Developers now train far past it on purpose. Counting the cost of running a model as well as building it, a smaller model trained on more text is the cheaper choice once it serves enough users (Sardana et al. 2024). Each dot is one model, larger with a ring for the models Epoch AI counts as the most compute-intensive of their time; the line is the median of each calendar year's models, drawn only for years with at least five. Small models are cheap to over-train, so the median mixes sizes; the large ringed dots show the shift holds at the top too. For mixture-of-experts models, which use only part of the network for each token, the ratio counts the parameters each developer says are active per token.

The Training Data sub-toggle shows instead how many tokens each model was trained on, against the stock of public human text: about 300 trillion tokens (90% range 100 to 1,000 trillion), which Villalobos et al. (2024) estimate the largest training runs will fully use between 2026 and 2032 (80% range) if current trends continue. The stock is adjusted for quality and for how often text can usefully be repeated; reusing text, synthetic text and other kinds of data can stretch it. The line follows the largest training set so far, from GPT-3 (2020), the first model in the set that Epoch AI counts among the most compute-intensive of its time.

Context Window (green) — the maximum input tokens a frontier language model can reason over in a single call, from GPT-2's 1,024 tokens to today's 1–10 million-token systems.

Inference Cost (magenta) — the falling price, in USD per million output tokens, of frontier-class model inference. The cheapest-so-far envelope has dropped roughly 10× per year since 2022 — sometimes called the “other Moore's Law” of AI.

Datacenter Power (vermillion) — the aggregate electrical load of the frontier AI datacenter fleet tracked by Epoch AI: from roughly 50 MW in mid-2023 to ~13 GW today and a projected ~33 GW by 2030. The solid line is historical; the dashed portion is capacity planned or under construction. Landmark dots mark individual mega-campuses (xAI Colossus, Anthropic—Amazon New Carlisle, OpenAI Stargate Abilene, Meta Prometheus, Microsoft Fairwater…).

Intelligence Frontier (pale gold) — the cumulative-maximum of Artificial Analysis's Intelligence Index over time. The bold step line is the record-holder at every date; faint dots beneath are every released model with a score — a reminder that for each new frontier-setter, dozens of also-rans were published the same month. Refreshed monthly.

Safety Thresholds (rose) — every published finding, from Anthropic, OpenAI and Google DeepMind, on whether a frontier model reached one of the capability thresholds its developer set for itself: in biological and chemical weapons, cyber operations, AI research and autonomy. Each mark is one model, coloured by developer: filled for a declared finding, a ring where stronger safeguards were applied without the threshold being established, a diamond for an early-warning alert. The rose step is the highest tier any model had been given by each date: from Baseline in June 2024 to Critical, for cyber, in August 2026. No developer has assigned the Catastrophic tier its framework defines. Underneath, a strip marks when the frameworks were published, the legal compute thresholds (US, EU, California), and the July 2026 evaluation-sandbox escape, kept off the ladder because no framework rates an incident after the fact. Tiers record what a developer declared, not a measured chance of harm, and are not strictly equal across developers. Most developers publish thresholds but no per-model findings, so a missing mark is not evidence of a safe model. The same record, in full, is on the risks page.

Nvidia Market Value (chartreuse) — the stock-market value of the company that designs most of the chips frontier AI is trained and run on, from about $1 billion in 1999 to more than $5 trillion in 2026, on a log scale so the decades before the AI boom stay legible. Dots mark the first reading above each trillion. Market value is a bet on future growth, not revenue or profit. Data: Yahoo Finance, refreshed monthly, shared with the market charts on the risks page.