TowardSingularity
A running record of how far frontier AI has got, measured one domain at a time on each domain's own measure, with the source and the date on every figure.
One question,
asked the same way.
How far has frontier AI got in a domain where somebody publishes dated measurements, whether that is a task length or a score on a fixed test?
Most writing about AI progress gives you a headline number with nothing behind it, or a forecast that quietly disappears once it turns out wrong. This site is built for the reader who wants to check. Every figure carries the source that published it, the units it was published in and the date it was read, so the reasoning can be argued with and not just the number.
Today 3 of the 10 domains carry a forecast year. The rest are on the site with the same weight, saying plainly that there is nothing to fit yet. A domain gets a date when the data allows one, never because it would look better with one.
The forecast is a real statistical object rather than a mood. It has something it is estimating, a spread around it, and a set of assumptions written down where you can attack them. Change an assumption on the forecast page and the curve moves in front of you, labelled as yours.
- What it tracks
- Frontier AI, domain by domain
- Domains with a date
- 3 of 10
- Published by
- Vuelance
- Contact
- contact@vuelance.com
Three things, and no fourth
Readings
Four public readings, each in the units its own source uses, each carrying the date it was fetched. Nothing is averaged into a single score, because averaging would need a common scale, and picking that scale is a choice rather than a measurement.
Domain forecasts
Where a domain has a dated series and a declared bar, the site fits a trend and publishes the year it clears that bar, with an 80% interval around it. Every domain runs the same layers and the same equations.
The gaps
Most domains have no dated series to fit, so they carry no year. They are listed beside the ones that do rather than left out, because the shape of the coverage is part of the answer.
The method is public
Every equation behind a date is written up in full, with each field it needs and where you would get that field. That includes the two fields that are judgement calls rather than measurements, which the page marks as such. Each pipeline run appends a row to the changelog, so a figure you quote stays attached to the inputs that produced it.
Who publishes it
TowardSingularity is built and published by Vuelance, a small studio that makes data products which show their working: open inputs, a published method, and a record that stays up even when it turns out wrong. The studio's habit and the site's rules are the same rules.
Questions, corrections and source suggestions: contact@vuelance.com
See what it
measures today.
Ten domains, the readings behind each one, and the gaps stated plainly.