← stephan hagemann

An illuminated history of artificial intelligence

Thinking Machines

MCMXLIII · MMXXVI

Eighty years of artificial intelligence: from a 1943 paper that modeled the neuron in logic, through two winters, to machines that write code. For sixty years the resources behind AI grew at the pace of chips. Since 2010 they have grown about five times a year, and since 2019 the length of task an AI agent can finish has doubled roughly every six months. This page plots both, and links every paper along the way.

Plate the first

The Fuel: Training Compute

The computation used to train each notable AI system, in floating-point operations. On the linear scale, seventy years of systems lie flat along the bottom. Switch to the log scale, where each gridline is a thousand times the one below it, and the exponential becomes a straight line.

Notable system Landmark Trend AI winter
Training compute of notable AI systems, 1950 to 2026546 systems from Epoch AI's dataset. Compute grows about 1.5 times a year from 1950 to 2010, from 40 operations for Shannon's Theseus to about 10 to the 18th. From 2010 the frontier grows about 4.8 times a year, reaching about 10 to the 27th by 2026.

Hover over a dot to see the system, and click a named one to read its entry in the chronicle. Switch to the log scale to see every system and the two trends.

Plate the second

The Result: How Long a Task an Agent Can Finish

METR gives AI agents software and research tasks, timed against human experts. A model's time horizon is the length of task, in human time, that it completes half the time.

Model Reported lower bound 2019–2026: doubles every  Since 2024: every 
50% task-completion time horizon of AI models, 2019 to 2026GPT-2 in 2019 completes tasks of a few seconds. GPT-4 in 2023 completes tasks of about four minutes. Claude Opus 4.6 in early 2026 completes tasks of about twelve hours, and an early Claude Mythos Preview at least sixteen hours.

Click a named model to read its entry in the chronicle. Switch to the log scale to see the early models and the doubling trend.

Plate the third

What Happened When

Events in AI history, 1943 to 2026Events in four lanes: ideas and vision, symbolic AI, neural networks and learning, and public milestones. The same events are listed below by era, with links to the papers.

This axis stretches after 2010 to make room for the crowded recent years. Hover over a blossom for details, or click it to jump to the event and its papers below.

About the Data

or, how to measure a thing that keeps changing shape

There is no single measure of capability that covers eighty years, so the page uses two measured stand-ins. Training compute covers the whole history. It measures what went into a system rather than what came out, but across the decades the two have tracked each other closely. The data is Epoch AI's Notable AI Models dataset (CC BY 4.0), as of November 2026. The trend lines are least-squares fits: all systems from 1950 to 2010, and since 2010, each system that was among the ten largest when it was released. They follow the method of Sevilla et al., "Compute Trends Across Three Eras of Machine Learning" (2022).

Time horizons measure what systems can do, but only from 2019 onwards. They come from Kwa et al., "Measuring AI Ability to Complete Long Tasks" (METR, 2025). The values here are refit from METR's public run data, using Time Horizon 1.0 for GPT-2 to GPT-3.5 and 1.1 from GPT-4 on, and land within a few percent of METR's published numbers. The early Claude Mythos Preview figure is METR's reported lower bound: past about 16 hours, METR says its task suite can no longer measure reliably. The tasks are software and research work, so other kinds of work may move at a different pace.

The Chronicle

in nine books, with sources

Index of Sources

every paper, report and book cited above, oldest first