Where did this series end? Draw the line across the shaded region to the ask date — your endpoint is your prediction.
Demo 3 of 17 · Part two · Stage 1
Draw where the series went, then see it.
From 3.5 minutes to five hours, from none of the hardest research problems to all but one: how much did the most-cited AI measurements move, and could you have guessed the size of it?
Before you start. Demo 2 showed a model being measured; here you read published measurements the same way. A benchmark is a test with defined tasks and a scoring rule, and a score is a measurement under conditions — which harness ran the model, how much reasoning effort it was allowed, which task set. Each series below starts with its early points shown. You draw where you think it went; the published points are then revealed with their conditions.
Drag across the chart (or focus it and press Up / Down), then press Reveal. Use Choose a series to jump.
On the tests the field cites most, the change since 2023 is measured in multiples, not points — and every series stops where its published record stops, with its conditions attached.
Sourced published measurements, each point with its conditions in Details
shown measurementsyour estimaterevealed measurementswhere the series stops
Drag across the shaded region, or use arrow keys on the chart.
Sourced a comparison from the same source, on a different scale
On most series your line was below the published result. What does that show?
Demo 3 of 17 · Part two · Stage 2
Every widely cited test gets replaced. Why?
MMLU, ARC-AGI-1, SWE-bench, Terminal-Bench 2.0 — each was the standard test for a while and each has been retired or rebuilt. Were they all simply beaten?
Before you start. A test is saturated when scores compress at the top until it stops separating the leaders — on Terminal-Bench 2.0 the top five sit inside each other's error bars. But a benchmark is retired when it stops measuring what it was meant to measure, and being beaten is only one of four ways that happens; the others are bad tasks, contamination and missing coverage. When a successor arrives it often counts something different, so a rising line across generations is not one trend.
Pick a family before you open it. Was its last retired version beaten, or replaced for another reason?
Press a family under Choose a benchmark family, then read each generation's status and the route chips. Open Three corrections below.
A benchmark is retired when it stops measuring what it was meant to; being beaten is one of four routes, and the replacement often counts something new — so a rising line across generations is not one trend.
Reasoned the status and route chips are the authors' reading of the sources quoted beside each generation; every number and quotation on the cards is sourced and linked
Four benchmark families, generation by generation
Three corrections, and a fourth
Terminal-Bench went 2.0 → 3.0 → 4.0. What does the higher number tell you?
Demo 3 of 17 · Part two · Stage 3
Outside the tests: what has AI done, and who checked?
A Nobel Prize, an Olympiad gold, 2.2 million materials, a Millennium Prize Problem — which of these happened the way the headline says, and how would you tell?
Before you start. A benchmark score is a measurement under conditions; an achievement is an event whose truth is settled by human institutions over months or years. So an achievement card carries two things a score never does: its acceptance status — the rung of the ladder from Nobel Prize down to retracted — and the AI's role: a tool people used, a result the system generated and people verified, or a result produced autonomously under stated rules. Read the rung first, then the headline.
Before you filter: which rung do you think the September 2026 Navier–Stokes proof sits on — and which rung has only one occupant?
Press a rung under Filter by acceptance or a role under Filter by the AI's role. Every card shows both on its face; open Details for the ladder's definitions.
An achievement is an event with an acceptance status and a stated AI role; the rung it sits on is the first thing to read, before the headline — and every rung from Nobel Prize to retracted has a real 2023–2026 occupant.
Sourced every card cites the lab's own page, the journal, the contest body or the institution that owns the problem; one card rests on reporting and says so
Fourteen achievements, by who checked them
Which rung does the September 2026 Navier–Stokes result sit on?
Demo 3 of 17 · Part two · Stage 4
Two points do not choose your forecast.
Given everything above, what can two measurements tell you about next year?
Before you start. In stage 1 you filled a gap in published history: the answer existed. A forecast is different — the answer does not exist yet, and the shape of the curve has to be assumed. An invented score below rises from 20 to 30 in one year. Three rules fit both points exactly: a fixed gain, a fixed multiple, and slowing toward a ceiling. Nothing in the two points chooses between them; only the next measurement does. Demo 4 picks this up: history, not a script, is how you judge a boom.
Eight years on from the second point, how far apart will the three rules be — a few points, or a factor of ten?
Drag Years after the second point and watch the three cards. Then read the boundary card beneath it.
A gap in published history can be estimated; a future cannot — three rules fit the same two points, and only the next measurement chooses between them. Make the claim specific: this system, these tasks, these conditions.
Three rules, one pair of points
Imagine a score rose from 20 to 30 in one year. Three rules fit both points exactly. Move forward in time to see what each implies.
This invented score runs from 0 to 100. Values above 100 expose an assumption that breaks at a ceiling; they are shown as invalid, not clipped.
The boundary, in proportion
Stage 1 led with magnitude because that is what the evidence shows. These are the limits the same sources report, kept to one sentence each.
Make the claim specific.
“This system improved on these tasks, under these conditions” is a claim you can check. “AI can do the job” is not. To plan a change in your own work, choose a representative task and measure quality, time, cost and failures against what you do now.
Fast benchmark progress and slower work in a particular setting can both be true — the 19%-slower randomised trial sits on the same ladder as the Nobel Prize. Neither result decides the other. The next measurement will, and until then the honest forecast names its assumption.