An open atlas of human progress

Progress across dozens of fields

Since 2023, are things improving faster than they used to?

How far the typical trend sits from the path it followed in 2003–2022, in units of its normal year-to-year wobble. 0 means right on track.

AI-exposed trendsAll trendsmiddle half of AI-exposed trends

Long-run measures of human progress, from crop yields and world records to solar prices and solved maths problems, each checked against its own history. Is anything speeding up where AI should matter?

AI trends vs all trends

For every trend we drew the straight-line path it followed from 2003 to 2022, then asked whether it has been running ahead of that path since 2023. Trends where AI could plausibly speed things up are in blue; everything else is in grey.

Loading trends…

Every trend, one dot each

How far each trend has moved from its old path since 2023. Dots right of the line are clearly ahead. Tap a dot to open that trend.

AI-exposedeverything else

The biggest movers among AI-exposed trends

Each tile shows the trend's real values, with its old path as a dashed line.

How to read this, and what to be careful about

The test. For each trend we fit a straight line to 2003–2022 (on a log scale where growth compounds, like prices or records), extend it forward, and measure how far the data since 2023 sit from that line, in units of the trend's own normal noise. More than two units ahead counts as "ahead of trend". Values are signed so that ahead always means faster progress: a cost falling faster than before counts as ahead.

Chance. Trends wander. Running the identical test at earlier dates shows how often a trend gets this far ahead by luck alone, which is the "by chance" figure above.

What "AI-exposed" means. A blind panel of four independent AI raters (two Claude Opus 5.5, two Claude Fable 5.1) judged every trend from its definition alone: title, field and unit, with no data, no results and no access to files or the web. Each asked how plausibly AI could directly speed up progress in that measure within a few years. A trend counts as AI-exposed when the panel's consensus is "high". Open any trend to see the four votes.

The catch. The raters agree well with each other, but they cannot forget what they already know: several of the standout trends (Erdős problems, Google Translate, AlphaFold-era protein structures) are famous AI stories. Leaving out every trend a rater said it knew the outcome of, the gap shrinks and becomes uncertain. So read the AI-vs-all comparison as strong evidence in a few well-known areas, not yet as a broad shift.

Who built this. The trends were gathered, checked and cross-checked by a team of AI research agents (Claude) working from public sources, with a human editor. Every trend links to its source. Please check the source before citing a number.

Explore every trend

Search, filter and open any trend to see its full history, how it compares with its old path, and exactly where the numbers come from.