Bad at Magnitudes

AI success on open-ended problems jumped from 25% to 76% in eight months. That’s the topline of Anthropic’s recent post. The deeper number, buried two paragraphs down: the doubling time on task length is itself shrinking, from seven months to four. The slope isn’t just steep. It’s steepening.¹

Most people are bad at magnitudes. They’re worse when the magnitude is moving.

A million seconds is 11 days. A billion is 31 years. Most people hear “millionaire” and “billionaire” as two flavors of the same thing. They’re not. A million is the upper bound of what disciplined wage labor can buy you over a lifetime. A billion is an outside shot for almost everyone, and almost nobody gets there as labor. To earn a billion at $1M a year takes 1,000 years. You can’t get there by working harder. You get there by taking a bet (founder equity, capital ownership) or by allocating other people’s capital (carry on a fund).² The shape of the income determines the ceiling, not the hours.

The lesson: linear thinking misses the gap. The shape matters more than the snapshot. Get the shape wrong and your conclusions about the snapshot are wrong by construction.

This happens to entire technological eras. Moore’s Law has been declared dead in the 1980s, 1990s, and 2000s. It outlasted every prediction of imminent stagnation by 30 years.

DNA sequencing is the better comparator for AI velocity. It outpaced Moore’s Law. The first human genome (2003) cost about $3 billion. Today a genome costs under $1,000. That’s a million-fold cost reduction in 20 years. The Carlson curve³ (sequencing) compounds at about 6.5% per month; Moore’s Law compounds at about 2.85% per month. Sequencing was faster, by a lot. And at every step, serious people declared the next drop impossible. “$10,000 genome can’t happen.” “$1,000 genome is a fantasy.” Each prediction outlived its premise by about three years.

Pattern: when a curve is moving, the people predicting stagnation tend to be wrong on timing more often than they’re right on direction.

The Anthropic post says Claude went from completing 4-minute human tasks (March 2024) to 12-hour tasks (May 2026). That’s a 180x time-horizon expansion in two years. The shape is familiar.

• Gutenberg’s printing press (1440s): a single scribe could copy 1 to 3 pages a day. A single press could produce around 3,600. Roughly 1,000x to 3,600x productivity gain per worker, depending on how you count.
• Ford’s Model T moving assembly line (1913-1914): chassis assembly went from 12 hours 28 minutes (stationary, September 1913) to 93 minutes (moving line, April 1914). An 8x speedup in seven months.
• Audio transcription, for the UX researchers reading this: a 60-minute interview used to take 4 to 6 hours of human labor to transcribe verbatim. AI does it in roughly 10 minutes at 90%+ accuracy. A 30x labor-time reduction in two years. Every researcher reading this has done some version of this math.

Each was an order-of-magnitude step-change in time per output. None were anticipated by people running the math from inside the prior regime. AI is doing the same thing on a much wider class of tasks, except the curve hasn’t stopped moving.

Yes-but: maybe it plateaus at 80%. Maybe the train doesn’t continue. We don’t know the asymptote. The point isn’t predicting where the curve ends. It’s noticing that anyone arguing about a single frame is making a magnitude bet without knowing they’re betting.

Get the order of magnitude right and the conclusion follows. Get it wrong and the conclusion is wrong by definition. The slope is the story. The doubling time is the punchline.⁴

Also read: Grand Theory of AI Skepticism (In UX Research).