Concept

AI Diffusion

conceptaieconomics-of-aiadoptiondiffusion

AI Diffusion

How fast AI capability actually reaches the economy — and Dario Amodei‘s argument that this is a second exponential, distinct from the capability curve, that gates real-world impact. ‘Not instant, not slow, much faster than any previous technology, but it has its limits.‘

The two exponentials

Dario separates two curves:

  1. The capability exponential — the model getting smarter.
  2. The diffusion exponential — that capability spreading into real economic use.

Both are fast; neither is infinite. The diffusion curve is bounded by mundane friction: change management, security and compliance review, procurement, provisioning to thousands of staff, and ‘closing the loop’ on fiddly real-world integrations. The evidence he offers is Anthropic’s ~10×-a-year revenue (0 → $100m → $1bn → $9–10bn, plus billions added in a single January), which he expects to bend but stay steep even at hundreds of billions.

‘Diffusion is cope’?

Dwarkesh Patel’s challenge: ‘diffusion’ is an excuse — AIs should onboard more easily than humans (read all of Slack in minutes, share knowledge across copies, no adverse-selection in hiring), yet we pay humans $50tn in wages. Dario half-concedes: the word is abused by people who mean ‘AI doesn’t matter’, and that’s not him — but real diffusion friction remains. Even Claude Code, trivial for an individual developer to adopt, reaches a large food-sales enterprise months later: legal, security, a $50m budget decision explained down six levels, a rollout plan for 3,000 developers. A more compelling product (Claude Code, Cowork) diffuses faster than the bare API, but nothing — not even the country of geniuses — is ‘infinitely compelling’.

Why it matters

Diffusion lag is the crux of several of Dario’s positions. It is why he won’t buy unlimited compute (revenue arrives on the diffusion curve, not the capability curve, so over-buying risks bankruptcy). It is why the country of geniuses is a ‘starting gun’, not an instant payoff — curing every disease still needs discovery, manufacture, trials, and delivery (COVID vaccines took ~18 months; polio eradication, 50 years and counting). His benchmark figure: AI may drive ~10–20% annual economic growth — far faster than history, far short of the 300% a pure capability extrapolation implies. Amdahl’s law governs: overall speed-up is capped by whatever step you haven’t yet automated.

See also