Changelog

Questions, faculties and scoring will evolve; new modules and add-ons are added as the research progresses. Item versions are tracked, so old answers stay comparable.

1. 2 November 2026v0.1

   • SiteTimed items now open on a warning screen that says how long you have and what happens when time runs out. The problem and the clock appear together only when you press Start.

2. 28 October 2026v0.1

   • ItemsThe full battery now includes held-out items, about a third of each module. They are never published, their answers aren't shown, and they are scored on our server, so AI models can be compared on items they can't have seen.

3. 27 October 2026v0.1

   • SiteThe full battery is open: six modules, about 35 minutes, covering all ten faculties with two tasks per sub-faculty. Take them in any order and resume if interrupted. Your results show each faculty against the human range once 30 people have taken it.
   • SiteSigning up and signing in now include a quick automated check that you are a person, so the data stays human.

4. 26 October 2026v0.1

   • ItemsThe seven "Spot the fake" sets were rewritten with everyday questions. The old ones leaned on science and medical facts, so they tested what you know rather than whether you notice a false premise.
   • ItemsThe "Read between the lines" items now say what the best reply is judged on: handling the moment without making it awkward. How directly people say things varies between cultures, and "best" was left to guess.
   • SiteClearer feedback on the card-turning puzzles, and a preview of your result image before you share or download it.

5. 21 October 2026v0.1

   • DataAI figures updated against the latest leaderboards: ClockBench (best model now 83.3%, up from 13.3% at launch), ARC-AGI-2, SimpleBench and GPQA Diamond. We now track published results in a database, including Epoch AI's benchmark data, so stale figures get caught.
   • FacultiesFaculty pages say exactly what each human and AI figure measures, and how we arrived at the two human figures that are our own estimates. A new page explains how rule induction, learning from feedback, deductive reasoning and abstract rule induction differ.
   • SiteThe homepage shows how AI passed the standard benchmarks one by one since 2012, from the Stanford AI Index 2026.

6. 11 October 2026v0.1

   • ItemsSix items on reading text precisely (counting letters, spotting a repeated word, finding swapped letters), mirroring the CUTE benchmark.
   • ItemsSix working-memory items: digits or letters shown one at a time, then typed back in order or reversed, mirroring Huang et al. (2025).
   • ItemsItems for the remaining sub-faculties: number problems with a tempting wrong answer, theory of mind, analogies, Stroop colour words, block-stacking plans, two-machine choices, one-minute word lists and a long digit stream. Each mirrors a published study or a classic test.
   • FacultiesEach faculty page lists its sub-faculties and which ones our items cover so far.

7. 10 October 2026v0.1

   • SiteYour result shows your time per item, and you can rate the test or report an item. New feedback, contribute and support pages.
   • DataOptional LinkedIn profile in the questionnaire, kept apart from research data. All data rights are self-service from your account.

8. 9 October 2026v0.1

   • SiteEach human and AI figure links to the test it comes from. Your account shows your cognitive profile.

9. 7 October 2026v0.1

   • ItemsFirst version: the 5-minute test, seven items drawn from a bank of 90, across the ten faculties of the Google DeepMind framework.