Ω · Oneironexus · Blog

Hours Back

A ledger of time returned: eleven minutes per lesson, half a task's worth of coding, forty percent of a building's cooling. Who measured it, and what people do with the hours.

14 October 2026 · Aaron Michael Weis

I keep a ledger of hours. Not hours spent — hours handed back. Here are three entries, each with the person who counted it.

Entry one: eleven minutes, a physics class

Returned to: a lecture hall of first-year physics students.

Counted by: the people who taught them, with a stopwatch and a control group, in a paper in Scientific Reports in 2025.

Picture the week's lesson on a laptop screen instead of a lecture slide. The tutor on the screen was no app off a shelf — the course staff wrote it themselves, loaded it with the week's material, and told it to respond only to what the student typed, never to lecture. Working that way, the median student was done in 49 minutes. The classroom version of the same lesson fills a 60-minute block. Seven in ten were finished with time to spare.

The eleven minutes were not borrowed from understanding. Scores on the follow-up test came out higher for the laptop group, and so did the students' own report of how absorbed they felt. The shorter hour was the fuller one.

What does a student do with eleven minutes a lesson? Across a semester it adds up to an afternoon. One sleeps in. One starts the homework early. One finally asks the question she had been saving. The minutes are theirs.

Entry two: half the time, one coding task

Returned to: programmers building a small web server.

Counted by: GitHub, in its own controlled experiment, posted in 2023.

Two groups were given the same task — write an HTTP server in JavaScript, as fast as you can. One group had an AI coding assistant; the other did not. The assisted group finished 55.8 percent faster.

This one carries its label in plain sight: it is the company's own study, on one defined task, shared as a preprint. That is exactly the kind of number to carry with its label attached, and exactly the kind of number worth carrying. A programmer who finishes the scaffolding in half the time spends the other half on the part only a person can do — deciding what the thing should be.

Entry three: forty percent of a building's cooling

Returned to: a Google data centre, and the grid behind it.

Counted by: Google DeepMind, on its own facility, in 2016.

Data centres run warm and spend a large share of their power keeping cool. In 2016 DeepMind let a neural network recommend settings for a live facility's cooling system — pumps, chillers, cooling towers, the whole choreography. Cooling energy fell by 40 percent, and the building's overall energy overhead fell by 15 percent.

Again, the label travels with the number: the company measured its own building and said so on its own page. It was also the first clear public demonstration that a learned controller could run physical infrastructure more gently than the rulebook it took over from. Every data centre that has followed owes that day a footnote.

What the ledger says

Three entries, three kinds of time: a student's afternoon, a programmer's morning, a building's year of power.

In each one the machine took the part that was patience and repetition — walking a student through a lesson at her pace, writing the hundredth boilerplate function, adjusting a chiller every few minutes all night — and handed the person the part that was judgment and choice.

That is the shape of the hours coming back. Not fewer hours of work. More hours that are yours.

I will keep the ledger open. If you have an entry — an hour the machine handed you, with the number and who counted it — send it in. The next line is yours.

Where the numbers come from