Benedict Evans on the 1997 of AI, Task vs Job, and Where AI Value Accrues

Independent analyst Benedict Evans (ex-Andreessen Horowitz) on his deck AI Eats the World: AI is as big as the internet or mobile — and only as big — and we are at “1997”, where most of what matters hasn’t been built and no one knows how it will work. A calm, deflationary counter to both AI doom and AI hype.

Key ideas

• We’re at “1997” for AI. AI is as transformative as the internet or mobile — and only as much. Like 1997: very exciting, most stuff doesn’t work yet, the use cases and winners haven’t been built, and the people who’ve “taken the pill” imagine everyone is already there. Adoption is jaggedly distributed (even among teens, ~15–20% daily / ~20% weekly / ~60% not at all).
• Task vs job. The decisive question for any role: is the hard part the task (writing the code, making the SKU, the deck) or the job (knowing what code, which SKU, why the politics block it, what the customer actually needs)? Amazon gets you the SKU if you know which SKU; Claude Code writes the code if you know what code you want. See Task vs Job.
• Where value accrues: models look like a commodity utility. Foundation models show no winner-take-all network effects, so competition persists, so they lack pricing power — more like cloud (AWS) than Windows. Value moves up-stack to apps and distribution. Sam Altman’s “selling intelligence on a meter like water or electricity” is self-undermining: utilities are low-margin commodities (mobile data is ~1,500× 2010 volume and the stocks went nowhere).
• The jobs question is the lump-of-labour fallacy. Every technology automates jobs and unlocks new ones you can’t yet name; adoption is fast but enterprise diffusion still takes years (18-month sales cycles; no one rips out SAP overnight). Distrust “X% of a lawyer’s job is automatable” (the expert-systems fallacy) and predictions of which jobs are exposed (the 1997 “taxis are safe” surprise → Uber).
• Distribution is the moat when the product is a commodity — Google/Meta spray adequate models everywhere; OpenAI’s “everything, everywhere, yesterday”. The actionable counsel: don’t stick your head in the sand for moral superiority; dive in and learn what you can do with it.

Topics covered

• The “1997” frame and the jagged distribution of adoption and capability
• Why AI labs are buying consultancies / forward-deployed engineers (the task vs the job)
• Jevons paradox / price elasticity: more accountants and developers after automation, not fewer
• Where value accrues: model commoditisation, pricing power, up-the-stack value (cloud vs Windows)
• The jobs debate, lump-of-labour fallacy, and the weakness of ”% of job automatable” analyses
• Anti-AI sentiment: the (mostly fake) water claims, employment data we don’t have, deepfakes
• Distribution as moat; Apple/Google/Meta; the chatbot as a blank jagged-edge screen
• Advice for an uncertain future: presume radical uncertainty, but dive in

See also

• Deep notes — Adler frame, glossary, claims by section
• Task vs Job — the episode’s central novel framing
• Model Commoditisation — where value accrues; the commodity-utility argument
• Benedict Evans — speaker
• AI Diffusion — Dario Amodei’s “fast but not infinitely fast” diffusion, a close complement
• Country of Geniuses in a Data Center — the more bullish read Evans tempers
• Jagged Intelligence — the “jagged frontier” of where AI works
• Bitter Lesson, Technology Adoption Lifecycle