ten tasks, the year of each, the system, the figure — and the human it was measured against
Built as a memo rather than a table: the thing worth hanging it up for reads across the room, with the full reference in smaller type below it.
Three sizes in one archive — from a sheet on the fridge to a poster on the wall.
A dot marks the year of a milestone on a shared axis. Every row names its source — a paper or an official result. Checked in August 2026.
The best-known line on this list is the 5.1 % error on image recognition that machines crossed in 2015. It is not the score of humanity, nor an average over a sample: it is one annotator who practised on 500 photographs and then labelled 1500 test ones. The paper the figure comes from says so plainly.
The same paper has a second annotator. He practised on 100 photographs and made errors on 12.0 % of the images — more than twice as often as the first, on the same task. The «human level» here therefore runs from 5.1 to 12.0 per cent depending on how long the person practised.
That is why every row of this sheet names who the machine was measured against: a world champion, professional transcribers, a whole ranked ladder, 630 olympiad schoolchildren — or no human at all, but other programs. Without that name the year means nothing.
Why the numbers can be trusted. Every countable number here is computed from data rather than typed in, and the sheet must pass 7 independent checks before it is allowed to build — how many items there are, the order they come in, and the claim in the highlighted box itself. If one fails, there is no file.
who published each standard, when, what it does to the sound, and how many bits per second it allows
two scales · 1000 and 1024 · why a drive looks smaller
main axis and cross axis · which property belongs to which
10 shapes, and one algorithm drawn with them