The Tempo: measuring the speed of AI

Every week brings a model release, a funding round or a benchmark result that is described as a turning point. Some weeks really are faster than others, and a reader deserves a way to tell them apart that does not depend on how loud the headlines are.

The Tempo answers one question: compared with its own recent past, how fast did AI move this week? It is a relative measure. It does not say whether AI is fast in some absolute sense, only whether this week was ordinary for the field or not.

Tempo = 100 + 20 × Σ w_i × z_i,  clipped to 40 … 200

Each signal i is counted over a 7-day window. Its count becomes a z-score z_i: how many standard deviations it sits above or below the mean of the same signal over the trailing 52 weeks. Three rules keep the number honest:

• Each z-score is clipped to between −3 and +3, so no single signal can carry the whole number.
• The standard deviation is floored at 5% of the mean, so a very steady signal's small wobble never reads as a shock.
• A z-score needs at least 8 weeks of history; with less, the week is left without a value rather than guessed.

The weighted sum is scaled so that 20 points is one standard deviation of the blend, then centred on 100. A week where every signal sits exactly at its own average reads 100.

The signals were chosen because each is public, counted the same way every week, and measures a different part of the field.

Papers carry a seasonal correction: half of their z-score comes from the trailing baseline and half from the difference with the same week a year earlier, which softens arXiv's conference-deadline weeks. In production, Artificial Analysis' Intelligence Index joins the capability signal; it needs an API key, so the reference run uses METR alone.

The number is named with the tempo markings of music, so it can be said in a sentence: under 70 Largo, 70 to 99 Andante, 100 to 119 Moderato, 120 to 149 Allegro, 150 to 179 Presto, 180 and above Prestissimo. The level is read from the value before it is rounded.

Quakes are kept apart from the number. A quake is a single event with a magnitude set by rule: a new leader on the Intelligence Index or a new METR time-horizon record is 7, a funding round of $10B or more is 6, a major law coming into force is 5, a leadership change at a frontier lab is 4. They are drawn as marks on the Tempo's trace rather than added to it, so one event cannot be counted twice.

In the 90 days before the first reading the reference run found two quakes, both magnitude 6 from the funding feed: a $20.0B round listed for OpenAI on 1 October 2026 and a $15.0B round listed for DeepSeek on 22 October 2026 [5].

• Noise is missing. GDELT keeps only about three months of history and could not be reached from the reference machine. It joins the index once it has a year of its own; until then the other weights are rescaled.
• Epoch lags. Notable models are added to Epoch's dataset weeks after release, so the most recent weeks of the Models signal run low and are revised upward later.
• The funding feed's history is thin. Early weeks carry a median of 1 round a week against 25.5 recently. The baseline is biased low, so the Money z-score runs high; in this reading it was close to its average anyway.
• The newest arXiv week can be under-counted while arXiv is still announcing it, and the Hacker News signal reads at most 1,000 stories a week.
• Weekly, not daily. The reference run uses calendar weeks so a first reading could be taken in one go. Production runs daily on a rolling 7-day window.
• A relative measure. 100 means an average week for the past year, not a slow one. If the whole field speeds up for a year, the baseline rises with it.

The Tempo shown on the Archive's preview pages (127, Allegro) is sample data used to build those pages. It is not a reading. The reading in this paper is 105.

The reference implementation is one Python 3 file with no dependencies beyond the standard library. It pulls each source, caches every response as JSON, and writes the reading and the full weekly series:

python3 -I tempo.py --cache ./cache --out ./out --weeks 64 --end 2026-11-05

With the cache in place a rerun is instant and gives the same numbers. A fresh pull on another day can differ where a source has since revised its counts, most often Epoch and the newest arXiv week. The weekly series in tempo-series.csv holds every raw count and z-score used here, so each figure in this paper can be redrawn from it.

• Daily at 06:00 Amman time, on a rolling 7-day window ending the day before.
• All six signals. A signal that fails to pull keeps its previous z-score and is flagged, instead of being dropped.
• Artificial Analysis joins Capability; GDELT runs on the server.
• The inputs and the output are published as CSV under CC-BY, so anyone can recompute the number.