Who he is
Affiliation: Independent technology analyst, London. Role check run 2026-09-21 (dated job-change / funding search): no new employer, no fund launch, no lab role — but three standing commercial seats the v1 file never listed. Venture Partner, Entrepreneur First (appointed 2020-05-19, TechCrunch; current status unverified this run — EF's team page did not resolve). Venture Partner, Mosaic Ventures (one day a week per the same 2020 report; Mosaic's own site still ran his May-2026 deck under its banner on 2026-05-22; its team page could not be read this run, so treat as live-but-not-reconfirmed). Advisor to Edelman's global technology practice (announced 2024-10-16; he presented Edelman's AI trust polling with them in November 2025). Before that: Partner, Andreessen Horowitz, 2014–2019 — six years, San Francisco (Chartwell bureau bio; v1's "~2014–2020" was a year loose); Enders Analysis; earlier telecoms equity research, Orange, Channel 4, NBC Universal (pre-a16z dates not independently verified this run — his about page and his Wikipedia entry both returned errors). Weekly newsletter, "150,000–200,000 subscribers" by his own count, "more than 175,000" per the bureau; the "AI eats the world" deck, twice yearly since 2025; "Another Podcast" with Toni Cowan-Brown, still weekly in 2026.
One-line position: AI is a platform shift as big as the web and mobile — "and only as big" — so the questions that matter are the product, the buyer, the distribution and who keeps the margin, not the benchmark. His provisional 2026 answer: "so far, models do seem to be commodities" (slide headline, May 2026), value migrates up the stack to distribution and to apps nobody has built yet, and the capital pouring in has the structure of a bubble even though the technology is real.
Discipline & technical bet
Market-structure analyst (someone who works out who ends up with the money in an industry, and why). He does not build models, fund them or run money; the wager is analytical, and it is concrete. (1) Frontier models converge — his May-2026 deck puts the best-to-average benchmark gap at roughly five points and closing within months — and they have no network effects (a product that gets better because more people use it), so (2) pricing power at the model layer decays once today's supply crunch ends, and (3) the durable value lands in distribution and in applications that don't exist yet. Corollary: the hyperscaler capex wave — about $700B guided for 2026 by Microsoft, Alphabet, Amazon and Meta, with capex running 30–55% of sales versus 2–5% in 2010 (deck, May 2026) — is individually rational actors spending into an aggregate with the shape of a bubble. And the unit of work AI touches is the task, not the job, so job-by-job exposure scoring is, in his words, fooling yourself. Layer focus: app — every conclusion he draws routes value up the stack; infrastructure is analyzed as a cost line, never as an investable bottleneck (the --investing twin's standing complaint about him).
Key claims (Says)
- Empirical / interpretive: AI is a real platform shift, comparable to the PC, web and smartphone — not a fad. Held — restated with a ceiling he now states out loud: "the biggest platform shift since the iPhone, but it's only that... not a civilization-altering technology like electricity" (Knowledge Project, 2025-08-28); "as big a deal as the internet or mobile — and only as big" (Lenny's Podcast, 2026-05-31). The stated upper bound makes this more gradeable than v1 had it.
- Analytic / normative: The hard question is "what's the product and who pays." Held — the OpenAI essay is four versions of that question (2026-02-19); the newest essay moves it inside the firm: "The hard part is knowing that you need a tool for this in the first place, and then knowing what the tool should do" (2026-09-03).
- Empirical, contested: Diffusion is measured in a decade-plus, not quarters. Held at the market level — 19.8% of US businesses were using AI as of 2026-05-03, three and a half years after ChatGPT (Census BTOS, 2026-05-26); the Fed's own roll-up puts firm adoption at 18% at end-2025, with 41% of workers using generative AI at work and 12% daily (FEDS Notes, 2026-04-03); his own deck shows most enterprise categories still majority-pilot (software development 40% in production, customer service 32%, Q3 2025). Valence flag: reach is the fastest of any platform — ChatGPT 900M+ weekly users, Gemini app 950M monthly (Alphabet Q2 2026, 2026-07-22) — but his claim is about depth and workflow change, and the data splits exactly along the line he drew.
- Empirical: Reliability / error rates are the central productization bottleneck — a tool that is wrong an unpredictable fraction of the time needs a use case that tolerates that. Drifted — the mechanism held ("Are better models better?", 2025-01: builders wrap probabilistic models inside deterministic software and manage the error rate with tooling and UX; "The Deep Research problem," 2025-02-18: "another amazing demo, until it breaks"), but by 2026 he had moved the bottleneck himself — to organizations and process: "Writing code isn't the hard part" (deck, May 2026); "Most people are not tool builders" (2026-09-03). And the first killer app landed exactly where errors are cheapest to catch: coding, at $3B of annualized enterprise spend, six times the next category (deck). Reliability is still a bottleneck; it is no longer the bottleneck in his own telling.
- Empirical, contested: Incumbents frequently capture platform shifts; the assumption that startups win the AI wave is unproven. Drifted — graded mechanism and valence separately. Mechanism held: "The incumbents have matched the tech and are leveraging their product and distribution" (his Threads post, 2026-02); Gemini's 950M monthly users ride Google distribution. Valence not held so far: the two largest revenue lines of the wave are entrants — OpenAI ~$40B annualized (Bloomberg, 2026-08-13), Anthropic $47B confirmed by the company (Series H, via Willison, 2026-05-29) and reportedly above $65B (Bloomberg, 2026-08-17) — and the largest app-layer company, Cursor, sold for $60B to SpaceX, not to a software incumbent (closed 2026-08-14). Both halves stay live; on the observable he would pick — who has the revenue — the entrants lead.
- Epistemic: "Nobody knows yet" is the honest position on end-state use cases — like not knowing in 2008 what the iPhone would become. Unfalsifiable — rewrite or retire. He still says it, and more strongly: scoring exposure is "mostly impossible" (2026-05-24); "the stuff that actually mattered was the stuff that wasn't even possible before and that no-one even imagined" (2026-09-03). A rewritten check (ours, not his) sits in the predictions list as P2.
- Analytic: The right metric is real-workflow adoption and usage, not leaderboards or downloads. Held — and he built the metrics: "AI metrics" (2025-06-09), DAU/WAU of 5–15% ("GenAI's adoption puzzle," 2025-05-25), prompts-per-user from OpenAI's own Wrapped data (2026-02-19). Behavior matches talk: the decks are built on these numbers.
- Empirical / interpretive: We overestimate near-term impact and underestimate long-term impact. Unfalsifiable — rewrite or retire (v1 already conceded it can't be scored). Correction: v1 called this an "Amara's-law framing"; no instance of him using that name was found — the frame is his, the label was ours.
- Descriptive of his stance: Brackets x-risk, alignment and distributional ethics as outside his lane. Held, sharpened — it is not a neutral bracket. "Ways to think about AGI" (2024-05-04) calls the doomer argument circular and expects AGI to be "just more software and more automation"; earlier: "It will just be software. Unless, of course, it goes to AGI and kills us all," and the AGI conversation is "a lot of very cultish people talking to each other" (Contagious, 2024-01-09). A dismissal is a position; the file now records it as one.
- NEW 2026 — empirical: Models are commodities; value goes up the stack. Live — mechanism held, valence untested. Mechanism: "And so far, models do seem to be commodities" (deck, May 2026); "It doesn't have unique tech... no network effect" (2026-02-19); "every dynamic we can see points to the latter," i.e. low-margin commodity infrastructure, with inference gross margins of 40–50% before retraining costs ("Ways to think about token pricing," 2026-07-09). Valence: the model layer is currently producing the fastest revenue ramp in software history (the OpenAI and Anthropic run-rates above; private valuations of $850–900B each, per his own deck). But his claim is about margin after the supply crunch, not revenue during it, and he says so: "we're in a supply crunch, and this is unstable." Not gradeable until the crunch ends — check P7.
- NEW — empirical: Consumer use is "a mile wide and an inch deep." Held on his data — 900M+ weekly users, about 5% paying, 80% of users under 1,000 prompts across all of 2025 (OpenAI Wrapped, cited 2026-02-19 and in the deck). Watch: Gemini app daily actives tripled year on year (Alphabet Q2 2026) — depth is rising from a low base.
- NEW — interpretive: The capital has "the structure of a bubble" while each actor is individually rational; "every bubble is different — and they are generally still bubbles" (SuperAI talk, as reported by Forbes, 2026-05-19 — the article itself could not be opened this run; quoted as reported). No timing content. Unfalsifiable as stated — per-branch checks in the predictions list. Conflict note: a machine summary of the Lenny's page reads him as "rejecting the bubble narrative"; the deck and Forbes have him calling the capital structure a bubble. The deck wins: the two statements are about different objects — the technology is real, the financing is bubble-shaped.
- NEW — empirical: Coding is the first killer app, and "writing code isn't the hard part." Held — $3B annualized enterprise coding spend, six times the next category (deck); Cursor at $3B ARR in May 2026 then a $60B exit (timeline); Anthropic's ramp is widely described as coding-led — the mix has not been verified from a primary source this run.
- NEW 2026-05-24 — analytic: Job-exposure scoring is "mostly impossible" — "you're fooling yourself, because you do not actually know what those jobs are today." Drifted — the Canaries update (Stanford Digital Economy Lab, August 2026; ADP payroll through June 2026) finds workers aged 22–25 in AI-exposed occupations 19% below their less-exposed peers, with declines concentrated where AI use substitutes for tasks and flat-to-rising employment where it complements. That is an exposure score predicting a real hiring margin. His mechanism (jobs are bundles of tasks; substitution versus complement) is the one the data validates; his valence ("can't measure it") is contradicted at the entry-level margin. No economy-wide displacement — that part he has right.
Notable predictions — with falsifiable checks
- P1 (2023) ChatGPT/LLMs are a platform shift comparable to web and mobile. Held so far. Check: at 2028-12, Census firm-level AI use above 50% → Held; below 30% with hyperscaler AI revenue growth back on the pre-2023 trend → Wrong.
- P2 (2023–24) Use cases and value capture stay unsettled; no killer app beyond coding/assistants. Held as of 2026-09 — coding is the one, consumer beyond chat is still thin, and his 2026-09-03 essay maintains it. Rewritten check for the "nobody knows" claim: a non-coding, non-chatbot AI-native product with ≥$5B annualized revenue by 2027-12 → Wrong.
- P3 (2023–24) Incumbents are well placed to capture much of the wave. Drifted (see the claim above). Check at 2027-12: an incumbent leads on consumer reach on a like-for-like metric and either OpenAI or Anthropic has been absorbed → Held; both labs independent and one of them leads on reach → Wrong.
- P4 (recurring) Near-term impact overestimated, long-term underestimated. Unfalsifiable — rewrite or retire.
- P5 (2010s) Smartphone installed base and "mobile is eating the world." Held — historical; the 2013–2017 mobile decks are on his presentations index. Not re-scored this run.
- P6 (2026-02-19) OpenAI must convert brand and users into distribution, infrastructure and enterprise contracts "before model parity makes the core product irrelevant." Live — the "no unique tech" half is playing out: OpenAI at ~$40B annualized was overtaken on run-rate by Anthropic (Bloomberg, August 2026). Check at 2027-12: OpenAI revenue growth below Anthropic's for four straight quarters and ChatGPT weekly-user growth under 10% year on year → Held; OpenAI retakes the revenue lead with rising DAU/WAU → Drifted.
- P7 (2026-07-09) Once the supply crunch ends, models trend to low-margin commodity infrastructure; "the market will get shaken out over the next few years." Live. Check: when either lab publishes audited financials (both reported pre-IPO, Bloomberg, August 2026), inference gross margins printing ≥60% and holding four quarters through 2027-12 → Wrong; ≤45% or falling → Held.
- P8 (2026-09-03) A professional-services boom helping enterprises deploy AI; "18-month sales process"; bundling and unbundling continue. Live. Check at 2027-12: the largest IT-services firms (Accenture, TCS, Infosys) report AI-related revenue growing faster than total revenue for four straight quarters. No baseline in this file yet — the first check sets it.
- Bubble, per-branch (2026-05-19): (a) hyperscaler 2027 capex guidance cut more than 20% from 2026 actuals by 2027-12 → the deflation branch resolves; (b) lab revenue doubles again through 2027 with capex still ≥30% of sales → "rational actors, bubble-shaped aggregate" stays live and unresolved; (c) neither → retire the framing at the 2028 review.
Revealed behavior (Does)
- Time: one long essay roughly per quarter in 2026 (February, May, July, September) plus the May deck, the weekly newsletter and the weekly podcast. All of it maps; none of it calls. Zero dated forecasts in the review window.
- Money: takes venture-partner seats (Mosaic since 2020; EF, 2020) and a PR-firm advisory seat (Edelman, 2024-10) rather than running money or joining a lab; sells framing through a paid speaker bureau. No fund, no disclosed positions, and no published personal position on any AI name that this run could find.
- Hires: none — a one-person shop by design, plus a podcast co-host.
- Cadence tell: the "annual" deck became twice-yearly in 2025 (Spring/Autumn, per the presentations index). The one place his output accelerated is the AI deck — where the audience and the speaking money are.
- Says vs Does on the central axis: the "models are commodities, value goes up the stack" analyst sits inside two venture firms whose portfolios are built on top of those models. Thesis and incentive point the same way. Not evidence he is wrong — evidence the thesis is comfortable.
Feels
Wants to be the calm one in a room full of enthusiasts and doomers; the mobile-era credential is the thing to protect, so the fear is being the analyst whose analogy engine smoothed over a real discontinuity a second time (the 2019 "ML is the new SQL" near-miss — carried from the --investing twin, not re-verified this run). The hope is that "things will probably be okay" (Lenny's, 2026-05-31) — and that it stays a product story he can map.
Hears
Operator and investor data before opinion: OpenAI's own disclosures (Wrapped, weekly users), hyperscaler capex lines, TSMC and Nvidia numbers, a16z-network product people, Mosaic and EF founders, Edelman's AI trust polling, Ben Thompson (Stratechery interview, 2026-02-05, paywalled). Downstream of the installation-versus-deployment view of technology waves and of Christensen's disruption vocabulary ("What kind of disruption?", 2025-03-14). Not downstream of the alignment literature or of labor economics — no reference to the Canaries payroll work turned up in the May-2026 jobs essay summary.
Sees
Twenty years of platform-shift base rates with receipts: he watched the smartphone supply chain become the component base for everything, and watched mobile operators spend "$200 billion a year" while "all the value was captured by other people further up the stack" (token-pricing essay). From two venture seats he also sees the deal flow of companies being built on commodity models — the app layer forming before anyone counts it.
Incentive map
Re-derived from scratch this cycle. Who pays him: the newsletter audience (as an asset), paid corporate speaking (bureau-listed; the --investing twin's client list — Alphabet, Amazon, LVMH, Visa — was not re-verified this run), Edelman's technology practice, and two venture-partner seats whose terms are undisclosed. What he sells: framing and "the questions to ask" to boards, PR clients and LPs. What he can't say: (1) "the app layer won't capture value" — both venture seats are app-layer bets; (2) "this is crypto again" — his speaking market is corporate strategy teams that need AI to be real; (3) a dated call — his product is the map, and a missed date would reclassify him from map-maker to pundit. What he can say that lab voices can't: "OpenAI has no moat," "the capital structure is a bubble" — no lab equity was found. Rule-25 discount, corrected: v1's "independent analyst with no current lab or fund position — lower commercial-interest discount" is Wrong: he has held venture-partner seats since 2020 and a PR-firm advisory seat since 2024-10, all public at the time v1 was written. Revised weighting — low discount on his lab-critical claims (labs don't pay him); moderate discount on "value goes up the stack" (the stack does); and a structural discount on anything that requires him to be surprised, because "nobody knows" pays him either way. The --investing twin flagged all of this on 2026-07-19; this file never caught up.
Theories aligned with
- Great Stagnation — adjacent: shares the "AI is big but slow to diffuse" view, though he is agnostic on whether it breaks the productivity slowdown; still no TFP position in 2026.
- Augmentation Thesis — loosely; treats current AI as a tool layered into existing work with value capture unresolved. His task-versus-job frame (elevator operators versus accountants) is augmentation-shaped, but his 2026-05-24 essay says the split can't be scored in advance — the Canaries data says it partly can.
- Jevons' Paradox — consistent with his "cheaper capability expands the market in unexpected directions" framing. v1 TBD resolved: no instance of him invoking Jevons by name was found (search 2026-09-21); his own version is the barcode move — when tracking inventory became free, the number of products multiplied (carried from the --investing twin; the barcode essay was not re-located this run).
- Deployment-versus-invention phases — his structural frame for where we are ("a 1997 internet moment," Lenny's 2026-05-31); "History doesn't repeat, but it rhymes" (deck slide).
- "Models are commodities" — his 2026 flagship; the mirror image of the frontier-matters camp (Altman, Amodei) and, oddly, a cousin of Sutskever's "age of scaling is over" — both say scale no longer differentiates; Evans says nobody differentiates, Sutskever says new ideas will.
What he's reacting against
- "This changes everything" hype that skips from a demo to economic transformation without specifying the product or the buyer
- The inverse reflex — pattern-matching AI to crypto/metaverse and dismissing it as another hype cycle
- Benchmark-centric discourse that treats model capability as if it were market adoption
- Forecasts that overweight the next 18 months and underweight the next decade
- Tech-determinism in both directions: that the technology alone dictates the outcome, rather than products, distribution and incumbents
- NEW (2024–26): AGI discourse itself — circular definitions and "cultish people talking to each other" (2024-01, 2024-05); and, from 2026, the "everyone becomes a tool-builder / apps are dead" story (2026-09-03) and job-exposure radar charts (2026-05-24)
- NEW (2026): "chat is a terrible UX; general use needs apps" — the chatbot as the end state (deck, May 2026; carried from the twin's reading of the same deck)
Where he overlaps / splits (with Rich)
- Overlaps with Tyler Cowen on "AI is big but slow"; both are diffusion-measured and skeptical of overnight transformation. Splits: Cowen theorizes (O-ring, high-agency dividend); Evans maps markets and stays with product and business questions.
- Overlaps with David Autor on gradualism and the gap between capability and labor-market effect.
- Splits with Marc Andreessen (his former a16z colleague) on register — Evans is analytical and skeptical of "everything changes," Andreessen programmatically optimistic. New in 2026: Andreessen's infrastructure-scarcity bet and Evans's "infrastructure is a cost line" are opposite reads of the same capex.
- Splits with Leopold Aschenbrenner, Sam Altman, Dario Amodei on timelines — Evans expects much slower economic impact than their capability forecasts imply. He agrees with Aschenbrenner on the input facts (unprecedented capex, power binding) and reasons from demand where Aschenbrenner reasons from capability — Evans is the hedge at the monetization joint (twin).
- Partial overlap with Erik Brynjolfsson on the productivity J-curve (gains lag adoption); Evans is less committed to a specific payoff. New split: Brynjolfsson's own Canaries data scores exposure at the hiring margin — the thing Evans says can't be scored.
- Vindicates Carl Benedikt Frey by accident: Evans's "no economy-wide displacement, jobs change" is compatible with Frey's displacement-first entry-level story, and the Canaries −19% is evidence for Frey that Evans's "fooling yourself" framing would have declined to look for.
- Overlaps with Ed Zitron on "OpenAI has no moat" and on the capex-to-revenue gap; splits on valence — Zitron reads the whole thing as fraud-shaped, Evans as a real technology with bubble-shaped financing.
- Overlaps with Adam Tooze on the financing air pocket; Evans's adoption-depth metrics are the demand-side clock for Tooze's balance-sheet worry.
- Overlaps with Mary Meeker as a chart-driven ex-analyst; splits on commitment — she dates things from data, he maps and refuses (twin).
- Overlaps with Peter Thiel in effect: the telecom-operator warning ("dominance ≠ value capture") is the analytical version of Thiel's sold-the-supply-layer trade (twin).
- Splits with Ilya Sutskever on what ends the scaling era: Sutskever says a new science of generalization, Evans says nothing in particular — the models converge and the money moves upstairs.
Track record
- Strong, widely cited mobile/smartphone analysis at Enders and a16z (2010s) — installed-base and "mobile is eating the world" framing held up; the "smartphone supply chain becomes the component base for everything" call is Rich's demand-stacking thesis run one cycle earlier (twin).
- Public skepticism toward crypto and metaverse hype cycles — vindicated, unchanged from v1.
- Historical near-miss, carried from the --investing twin and not re-verified this run: "The deployment phase of machine learning" (2019) called ML "the new SQL" and under-called the 2022 LLM discontinuity — the analogy engine's known failure mode.
- 2025–26 essay sequence — adoption puzzle (2025-05) → AI metrics (2025-06) → OpenAI has no moat (2026-02) → exposure can't be scored (2026-05) → token pricing (2026-07) → tools and transformation (2026-09) — is internally consistent and increasingly scoreable; this is the first cycle in which he has a position hard enough to be wrong (models-as-commodities).
- His measured stance is still hard to falsify in either direction — a strength for credibility, a weakness for scoring; the checks above are this file's attempt to force grades by end-2027.
- Influence is via framing and vocabulary (S-curves, "what's the product," installed base, "mile wide and an inch deep") more than dated calls.
- File corrections this cycle: v1's "annual 'big presentation'" is twice-yearly since 2025; v1's "'The new gravity' series" does not exist — the decks are "AI and Everything Else" (2023) and "AI eats the world" (2024, Spring 2025, Autumn 2025, Spring 2026).
Empirical vs normative
- Empirical / analytic: market structure, adoption curves, business-model viability, who captures value — his core register; in 2026 extended to capex-to-sales ratios, model convergence and prompts-per-user.
- Normative: light — he maps what is happening rather than prescribing; deliberately avoids policy and ethics prescriptions. The one recurring prescription is methodological: count depth, not reach.
- Bracketed: x-risk, alignment, labor justice, distribution — not denied, outside his stated lane (CONVENTIONS rule 26 — separate questions, not a tribal split). Note the bracket is loaded: he calls the AGI conversation cultish and circular.
- Conflict note (CONVENTIONS rule 25), corrected this cycle: not unconflicted — venture-partner seats at Mosaic Ventures and Entrepreneur First since 2020, Edelman advisory since 2024-10, bureau speaking. Weight as informed industry analysis with a stack-side tilt, not as a neutral referee.
Weak spots / open questions
- The "measured" stance is hard to falsify — almost any outcome can be read as "diffusion takes time, as I said." Mitigation: P1–P8 above now carry dates.
- Low forecast resolution: few dated, falsifiable calls to score against (contrast Amodei / Aschenbrenner). His 2026 models-as-commodities thesis is the first exception; it is the one to watch.
- Risk of the "this time is the same" trap (CONVENTIONS rule 24) — pattern-matching AI to prior waves could under-call a genuine discontinuity if agents compound faster than past general-purpose technologies. Sharper now: what observable would make him concede a discontinuity rather than re-narrate it as slow diffusion? He has never stated one (twin's open question, still open).
- Industry-structure lens can miss the societal, labor and safety dimensions the rest of the roster centers — confirmed by the jobs essay declining to score exposure three months before payroll data scored it.
- Analyst, not theorist — strong at describing the present, lighter on a predictive model of where it ends.
- Open: does the "incumbents capture it" thesis hold if agentic AI collapses distribution advantages? Partial answer in 2026: the entrants lead on revenue while the incumbent leads on reach — the question has split, not resolved.
- Open (from the twin): if models commoditize, does margin flow down to silicon and power or up to apps and distribution? He says up — but his own TSMC example shows a scarce physical layer keeping fine margins. Which layers are TSMC-like?
- Open: "usage is a mile wide and an inch deep" is measured on chatbots; if agents embed inside existing apps, chatbot depth becomes the wrong metric entirely — his own metric may be aging out.
- Unverified this run: the pre-a16z career dates; whether the EF seat is still live; the twin's speaking-client list; the mix of Anthropic's revenue; the barcode essay's location.
Rich's take
- (your synthesis here)
Delta log
2026-09-21 — v1→v2 migration + validation (bulk run)
- Grades: 9 Held / 3 Drifted / 1 Wrong, plus three items marked unfalsifiable (not graded) and four live 2026 predictions with checks attached. Review window: 2026-06-05 (v1) to today; per the learnings file, the June-2026 v1 review is treated as never having run.
- Wrong: the file's own "no current lab or fund position — lower commercial-interest discount." He has been a Venture Partner at Entrepreneur First and at Mosaic Ventures since 2020 (TechCrunch, 2020-05-19) and an Edelman advisor since 2024-10-16 (Edelman). Incentive map re-derived; the discount now has three tiers.
- Drifted: reliability as the bottleneck (he moved it to organizations — 2026-09-03 — and coding routed around it); incumbents capture the wave (mechanism held, valence not — Bloomberg, 2026-08-17; Bloomberg, 2026-08-14); job-exposure scoring "mostly impossible" (Canaries, August 2026).
- Held: platform shift with a stated ceiling (Lenny's, 2026-05-31); product-and-buyer question; decade-plus diffusion at the firm level (Census, 2026-05-26; Fed, 2026-04-03); workflow metrics; x-risk bracketing; mile-wide-inch-deep; coding as killer app; P2 no non-coding killer app; P5 mobile.
- Unfalsifiable — rewrite or retire: "nobody knows yet"; near-term over / long-term under; the bubble framing as stated. Rewritten checks attached (P2, bubble per-branch).
- Most surprising delta: the bench's "unconflicted independent analyst" has sat inside two venture firms for six years and a PR firm for two — every fact was public when v1 was written, and the --investing twin had already caught it in July.
- Employer/role check (mandatory): dated search run; no new employer or funding event; affiliations added as above; a16z tenure corrected to 2014–2019 (Chartwell). Post-review dated sources found: essay of 2026-09-03; Bloomberg 2026-08-13/14/17; Canaries August 2026; Alphabet Q2 2026 (2026-07-22); token-pricing essay 2026-07-09.
- v1 TBDs resolved: deck titles (no "new gravity"; "AI eats the world" 2024–2026, twice yearly per the presentations index); Jevons not invoked by name; "Another Podcast" still running (2026); sources now dated and linked.
- Cross-grades: Frey vindicated by the same Canaries evidence that drifts Evans; Brynjolfsson's data-vs-Evans's-frame split logged in both directions; Sutskever/Evans convergence on "scale no longer differentiates" noted.
- Tier: semiannual — he publishes on a twice-yearly deck cadence, holds no positions, and his grades move on slow variables (Census, payroll, capex guidance); wake triggers cover the Autumn deck, any role change, any dated call, and the two events that would finally test the commodity thesis (an S-1 with margins; a capex guidance cut).
- Confidence unchanged at medium: Says is now well sourced from primary essays, but the persona still has few hard calls to score, and two affiliations could not be re-confirmed from the firms' own pages.
Sources
- 2026 primary (his own): "AI, tools and transformation" (2026-09-03) · "Ways to think about token pricing" (2026-07-09) · "Predicting AI job exposure" (2026-05-24) · "How will OpenAI compete?" (2026-02-19) · "AI eats the world," Spring 2026 deck (PDF, May 2026) · same deck, slide transcript · Threads post on OpenAI (2026-02)
- 2024–25 primary (his own): "GenAI's adoption puzzle" (2025-05-25) · "Are better models better?" (2025-01) · "Ways to think about AGI" (2024-05-04) · essay index (for "AI metrics" 2025-06-09, "The Deep Research problem" 2025-02-18, "What kind of disruption?" 2025-03-14, "AI, networks and Mechanical Turks" 2025-11-23) · presentations index (deck titles and years)
- Interviews / talks: Lenny's Podcast (2026-05-31) · Forbes on the SuperAI talk (2026-05-19; not opened this run) · Stratechery interview (2026-02-05; paywalled) · a16z podcast (2025-12-12) · Edelman TrustMakers (2025-11) · Knowledge Project (2025-08-28) · Contagious (2024-01-09) · "Another Podcast" listing
- Roles: TechCrunch — joins EF; Mosaic one day a week (2020-05-19) · Edelman appointment (2024-10-16) · Mosaic Ventures posting his May-2026 deck (2026-05-22) · Chartwell Speakers bio · ben-evans.com (subscriber count)
- Market-level data used for grading: Census BTOS story (2026-05-26) · FEDS Notes, "Monitoring AI adoption" (2026-04-03) · Stanford DEL, Canaries August-2026 update · Canaries paper (PDF) · Bloomberg — OpenAI $40B run-rate (2026-08-13) · Bloomberg — Anthropic >$65B (2026-08-17) · Anthropic $47B, company statement via Willison (2026-05-29) · CNBC — SpaceX to buy Cursor (2026-06-16) · Bloomberg — deal closed (2026-08-14) · Cursor timeline · Alphabet Q2 2026 remarks (2026-07-22) · CNBC on the same call
- Source tier: analyst essays, decks and podcasts — below peer-reviewed or book-length work per CONVENTIONS rule 4; weight as informed industry analysis, not empirical research. Market-level grades above rest on Census, Federal Reserve, Stanford payroll research and company filings/earnings, not on his charts.
- Bench context: --investing twin (reviewed 2026-07-19; read, not modified)
migrated v1→v2 2026-09-21 · backup: _archives/benedict-evans.html.bak-20260921 · converts-from: people/benedict-evans.md · AI & Society domain