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

Source-grounded literature notes. Citations point to logical sections of the conversation, e.g. [§ Where value accrues]. Own words unless quoted.

Four questions [Adler frame]

Q1 — What is it about? Independent analyst Benedict Evans (ex-Andreessen Horowitz) on his twice-yearly deck AI Eats the World. The argument is deliberately deflationary: AI is as big a deal as the internet or mobile — and only as big — and we are at “1997”, when most of what will matter hasn’t been built and no one knows how it will work. Around that he works through where value accrues in the AI stack, the jobs debate, anti-AI sentiment, and distribution as the durable moat.

Q2 — How is it argued? By analogy and historical base-rate. Almost every claim is anchored to a prior technology cycle: VisiCalc and the accountants, Excel and the bankers, e-commerce and “the SKU”, the recorded-music U-curve, telecoms margins, and repeatedly the 1997 internet. He explicitly refuses prediction — his deck is “80 slides of saying we don’t know” — and elevates humility (“it depends”, “it’ll probably be okay”) to method.

Q3 — Is it true? Strong where it leans on historical pattern (commoditisation logic, the lump-of-labour record, diffusion lag); weaker where it hardens into a confident negative forecast — that model labs will have no pricing power — which is genuinely contested by people closer to the labs (Dario Amodei argues frontier AI is a differentiated Cournot oligopoly with durable margins, see Model Commoditisation). His jobs optimism rests on the lump-of-labour argument, which holds historically but which he concedes is cold comfort to someone entering an exposed profession now. He is an analyst, not an operator, and openly discounts argument-from-authority — including Dario’s on labour economics.

Q4 — What of it? A frame for staying sane amid hype and doom: presume radical uncertainty; ask whether a role is a task or a job; expect value to move up the stack to apps and distribution while models commoditise; and, individually, dive in and learn rather than resist for moral comfort.

Glossary

The “1997” frame — AI is a genuine platform shift on the scale of the internet/mobile, but at an early, pre-paradigm stage: most things don’t work yet, the winners and use cases aren’t built, and adoption is jaggedly distributed. Asking “is it 20% or 100% bigger than the internet?” is unproductive.
Task vs job — whether AI automates a role depends on whether the hard part is the visible task or the surrounding judgment. See Task vs Job.
Jevons paradox / lump-of-labour — making a task cheaper usually means doing far more of it (Excel → more, not fewer, bankers); the belief that there is a fixed quantity of work to be automated away is a fallacy.
Model commoditisation — foundation models show no winner-take-all network effects, so competition persists and pricing power erodes; value accrues up-stack. See Model Commoditisation.
Value up the stack / cloud-not-Windows — like AWS (undifferentiated infrastructure, value above it) rather than Windows (platform lock-in), the leverage sits with apps and distribution, not the model.
Distribution as moat — when the product is a commodity, distribution decides; incumbents (Google, Meta, Apple) spraying an adequate model everywhere can beat a marginally better one.
Jagged frontier — capability and adoption are uneven and hard to predict in advance; see Jagged Intelligence.
Radical uncertainty — Evans’s epistemic stance: “it depends”, and have the humility to say what you can’t know.

Claims by section

§ The 1997 frame

Evans’s “most controversial opinion”: AI is as big as the internet or mobile, and only as big — a deliberate rebuke to both the “bigger than the industrial revolution” camp and the “you don’t get how big this is” camp. Pushing the internet analogy, we are at 1997: very exciting, most stuff doesn’t work yet, the things people will do haven’t been built, and how any of it works isn’t clear. Those who have “taken the pill” wrongly imagine everyone is already there; in reality adoption is a wide spread — even among 13–18-year-olds, ~15–20% are daily and ~20% weekly users, while ~60% don’t use it at all. This maps to the jagged-frontier question (where does it work, can you tell in advance, can you tell after?). Arguing “is it 20% or 100% bigger than the internet?” is unproductive; it’s a fundamental change whose mechanics are unknown — his new deck is, by one reader’s account, “80 slides of saying we don’t know.”

§ Consultancies and forward-deployed engineers

Why are the frontier labs buying consultancies and hiring forward-deployed engineers? Because reimagining a company’s workflows is a project needing people, not a prompt — “an Accenture outsource software developer who works in San Francisco”. Companies don’t keep architects or analysts on staff; they hire Bain/McKinsey or Accenture when they need a burst of judgment and execution. This is the task vs job point: McKinsey’s deliverable is a deck, but you pay them to walk the company, read the politics, and talk to customers — not for the PowerPoint, which Claude will make a crappy version of. See Task vs Job.

§ Jevons and the lump-of-labour

Automating a task usually expands the work rather than eliminating the worker (the Jevons paradox = applied price elasticity). Before Excel, junior bankers worked brutal hours; after Excel, they still do. Apple now writes ~90% of an iPhone app’s code via libraries and OS, yet there are more developers, not fewer. The count of accountants rose across a century of adding machines, mainframes, ERP, and spreadsheets. The jobs that vanish (typesetters, telephone operators) are, in retrospect, the crap jobs; GDP keeps rising.

§ Where value accrues

The deck’s “capital” section. Evans’s structural argument that foundation models look like a commodity utility: they show no winner-take-all network effects, so competition persists indefinitely, so there’s no radical product differentiation, so there’s no pricing power — “have you got like three or four companies all selling the same thing?” He skewers Sam Altman’s “selling intelligence on a meter like water or electricity”: utilities are low-margin commodities — when you watch TV the broadcaster doesn’t pay the electricity company a cut. His telecoms career is the cautionary tale: global mobile data is ~1,500–2,000× its 2010 volume, the industry spends ~15–20% of revenue on capex, and the stocks went nowhere for 25 years because “all the cool stuff is made by you” — value sits up the stack. The open question: does the chatbot become the UX (models capture it), or does it have to be “apps” the labs won’t all build — making this cloud, not Windows? He distinguishes today’s “radical price disequilibrium” (the OpenClaw user burning $1.5m of tokens in a month, like a $50k mobile-data bill in 2010) from the steady-state equilibrium, where margins compress. See Model Commoditisation.

§ The jobs debate

On the “job apocalypse”: Evans invokes 200 years of automation creating new work you can’t name in advance (“railway engineer — what’s a railway?”). Is AI different? Adoption is faster, but only because it stands on the shoulders of giants (900m can use ChatGPT because 900m are already online), and the internet was faster than its predecessors too. The “doomers” imagine every company buys ChatGPT tomorrow and fires everyone in a fortnight, ignoring 18-month enterprise sales cycles — no one rips out SAP overnight; it takes 2–10 years, sector by sector. He is scathing about “X% of profession Y is automatable” analyses — the expert-systems fallacy (you can’t recognise a cat by hand-coding edge/ear/eye detectors, and you can’t decompose a senior partner’s job that way: “this is horseshit”) — and about confident predictions of which jobs are exposed (the 1997 “taxis are safe” surprise → Uber; his own examples: personal trainers, video editors, coding itself). The honest caveat: on average no one died in WWI, but a 19-year-old in 1914 had a one-in-three chance — averages are cold comfort to those in exposed roles.

§ Anti-AI sentiment

A “big fuzzy mass”. Some is real (a few localities’ electricity bills); some is “completely fake” (data-centre water use is ~0.017% of US consumption — a planning problem, not a data-centre problem). The employment data is genuinely thin: the labs publish no daily-active-user numbers, so economists back claims out of BLS surveys with no consensus yet on AI’s job impact. Some is niche culture-war (cover artists, novelists, “AI slop”). He likens it to the social-media backlash but “much more compressed” — some true, some not, much a fuzzy middle — and notes deepfakes are a genuinely new harm (a 15-year-old couldn’t mass-produce them with Photoshop), as every technology brings new ways to ruin lives (the UK Post Office/Horizon scandal — 1970s tech ruining hundreds of lives).

§ Distribution as moat

“I don’t like GPT wrappers, I do like harnesses.” When the product is a commodity, distribution decides. Browsers are the precedent: the product is a thin wrapper on a rendering engine, Microsoft used distribution to win, and then it didn’t matter because value moved up-stack. Now Google drives Gemini and Meta sprays an adequate model across its surfaces; OpenAI’s “everything, everywhere, yesterday” is a scramble for distribution before the incumbents’ defaults lock in. Apple’s 2024 Apple-Intelligence vision (on-device, tool-using, agentic, no hallucinations, standardised intents across 10,000 apps) was the most compelling — and shippable by no one yet. The model is “the dumb thing underneath that powers a feature”.

§ Advice and radical uncertainty

For the worried: don’t stick your head in the sand for “a great feeling of moral superiority” — “that’s not going to help.” Dive in, submerge yourself, understand what you can do with it; going into a shrinking-associate-pool interview saying “I’ll never use AI” is the wrong move. His own AI use is thin because his job (synthesising novel ideas) is precisely what the models are worst at — “I am the accountant looking at the spreadsheet” — though he dictates most writing now (is that AI, or just automation that disappeared?). The closing posture: presume radical uncertainty, keep pushing past whatever you’ve just understood, and retain the humility to say “it depends”.