Rosehip Mole asked, and Nib — askNib's tutor — drew the answer live at a whiteboard. This is the spoken transcript; enable JavaScript to watch it drawn.
Teach me this page — https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities
Ajeya Cotra measures AI not by test scores, but by a stopwatch: how long a task would take a human, that the AI can finish half the time.
In January she guessed: by end of 2026, the best AI would handle roughly 24-hour tasks half the time. A careful, conservative estimate.
Two and a half months later, Opus 4.6 landed at around 12 hours — already halfway to the year-end guess, with months still to go.
Why the jump? On the 19 hardest tasks in the suite — each over 8 human-hours — Opus 4.6 solved 14 of them at least once, four every single time.
That huge uncertainty is the point: once models beat almost every task you throw at them, you can't tell how far past the edge they really are.
Here's the deeper idea: short tasks resist teamwork. You can't split a one-hour debugging job across five people — every step depends on the last.
But a month-long project splits nicely into tickets, and a year-long project splits into whole milestones run by different teams in parallel.
So her real claim: once AI can reliably do 80-hour chunks, it may stop being limited by 'solo time' at all — managers hand off pieces to workers, forever.
Put it together: the pace of AI's solo-work skill is doubling every three-and-a-half months, faster than anyone planned for — and it may soon stop mattering.
The surprise isn't one number — it's that the measuring stick itself may be breaking as AI stops needing to work alone.