Companion to Cheap Composition · Draft 3

The Personal Singularity

The singularity for an individual isn't a date. It's the day the thing you're best at becomes cheap.

For each skill below, two dates. When a machine could do it, and when the people who do it felt it in their income. The gap between the two is the real story. It's set by law, liability, and whether the work needs hands, not by how clever the model is.

A date with a source is documented, and the source is linked in its row. Every other date is one person's estimate with a range, whichever side of today it falls, and the estimates get scored in September 2027 and again in September 2028.

Where are you on it?

What you're best at This names your row. The three answers set the dates.

One unit of that work usually takes This sets the capability year. Longer tasks arrive later.

How much of it is hands-on, physical, in places that aren't predictable?

10%

noneall of it

How much of it needs a licence, a signature, or a specific accountable human?

20%

noneall of it

The timeline

What a machine can do it felt in income the gap you

How sure solid: documented, source linked dashed, with a range: estimate

Tap or hover a row for its evidence, what gets cheap and what stays.

How the dates are made

Documented dates link to their evidence. For a capability date that is a benchmark result, a study or a deployment doing the work at scale. For a felt date it is a measured fall in pay, work or hiring. A date without that stays an estimate, even when it is already behind us.

Estimated capability years come from task length. METR measures how long a task AI can finish at 50% reliability. Its January 2026 update put the doubling time at about four months since 2023 (down from seven over the longer run), passing one hour in early 2025. In May 2026 it reported a model at 16 hours or more and warned that its tasks cannot measure reliably beyond that. So the longer a unit of your work, the later the machine gets there. Hands-on physical work in unpredictable places adds up to eight years, on the assumption that robots trail language by about a decade.

Estimated felt years add a lag: one year as a base, up to eight more for trust (law, licensing, liability, and whether people need a specific accountable human), and up to one more for hands-on work. Those three numbers are assumptions, not measurements. Radiology is the calibration point: by 2019 a model could do one of its reading tasks, spotting lung cancer on low-dose CT screening scans, as well as or better than radiologists, and radiologists are still being hired. The calculator above shows every term for your answers.

The narrow models that read scans or drove cars arrived before general models did, which is why a few capability dates sit earlier than the task-length rule alone would give.

Sources

METR, Measuring AI Ability to Complete Long Software Tasks, 2025; Time Horizon 1.1, January 2026; and its time horizons page, updated May 2026.
Brynjolfsson, Chandar and Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Stanford Digital Economy Lab, 2025, updated August 2026.
Anthropic Economic Index, 2025: real usage by occupation and task.
Eloundou, Manning, Mishkin and Rock, GPTs are GPTs, 2023: task-level exposure estimates.
The evidence for each documented date is linked in its row and in the table below.

All rows as a table, with their evidence

Skill	Group	A machine can do it	Felt in income	Evidence

Made in September 2026 with Claude. Draft 3. The model behind "place me" is three answers, a doubling rate and a few stated assumptions, not a forecast of your life.