Lore

AI hype cycle

Top AI Analyst Unpacks Today's AI Hype Cycle

Benedict Evans argues that debating whether AI is bigger than the internet or electricity is analytically empty; instead, patterns from past platform shifts (mobile, semiconductors, PCs) show technology adoption and value capture unfold unevenly and slowly, constrained by AI's jagged capabilities and unresolved data/verification limits, while the separate binary possibility that true AGI arrives would make today's economic questions irrelevant.

Unsupervised Learning: With Jacob Effron · 2026-07-16 · English

Key ideas

  1. Evans's central methodological argument is that debating whether AI is bigger than the internet is unfalsifiable and low-value; the more useful question is what happened the last several times a transformative technology arrived, since every past shift also looked unprecedented to people living through it.

  2. He identifies one genuine point of novelty this cycle: unlike prior shifts, we don't yet know AI's physical or scientific limits, so even basic questions about what's ultimately possible remain open.

  3. Because a true AGI/ASI outcome would make ordinary economic questions moot, Evans argues the productive move is to set that scenario aside and analyze the concrete, falsifiable middle ground instead, treating lab leaders' job-loss statements as reflecting genuine uncertainty about that binary rather than irrationality.

  4. He argues much of the public AI debate operates like theology, arguing definitions into predetermined conclusions rather than examining evidence, and calls for building things and observing outcomes instead.

  5. Against narratives of an already-transformed daily-use technology, Evans presents usage data showing AI is heavily used by developers and a narrow slice of self-employed knowledge workers, but only an occasional tool for most people.

  6. He challenges job-exposure predictions of AI-driven unemployment on the grounds that job content historically changes under automation pressure without headcount necessarily falling, and that exposure forecasts can't be validated against what workers actually do.

  7. He extends this by arguing the biggest disruptions from a general-purpose technology often come from indirect, adjacent effects rather than the tasks it directly automates, so job-content analyses miss the real risk.

  8. He proposes that whether AI can automate a task depends jointly on whether enough training data exists and whether outcomes can be verified at scale, explaining why coding became LLMs' first strong product-market fit while judgment-heavy domains remain harder.

  9. He argues AI's near-term economic effect is mostly to make existing categories of output cheaper and more abundant rather than to invent new categories, and that new consumer or enterprise behaviors must be actively invented by entrepreneurs rather than emerging spontaneously.

  10. On value capture, he argues history shows no single layer of a technology stack has ever captured the whole value chain, so claims that a foundation-model company will become the next dominant platform likely overstate how much value model providers alone will capture.

  11. He frames enterprise AI adoption as currently stuck at a shallow stage of broad co-pilot rollouts and isolated pilots, short of the deeper organizational reimagination that actually realizes AI's value, and argues this gap is exactly the space consulting and vertical-software firms are converging to fill.

  12. He argues AI lab leadership faces an unusually large set of simultaneous, unprecedented strategic problems with no historical playbook, which is his stated basis for sympathizing with Sam Altman specifically.

  13. He predicts that, as with the internet, some industries will end up barely affected by AI in ways that are impossible to predict in advance, because every industry has its own unrelated pressing priorities competing for attention and budget.

  14. His closing argument is that AI's real impact will mostly be an invisible enabling layer inside existing products and back-office processes, and that judging it by visible chat or image outputs is like judging the early internet by its chat rooms rather than its underlying network.

  15. Shadow IT — A historical pattern where employees build or adopt their own unauthorized tools outside official IT systems, cited by Evans against the claim that AI-driven employee tool-building is unprecedented. Apply: When someone claims AI lets employees build tools for the first time, check the claim against prior shadow-IT and spreadsheet precedents before accepting it.

  16. Semiconductor fab cost-scaling and consolidation analogy — The observation that cutting-edge semiconductor fab costs roughly double every four years, progressively shrinking the field of frontier competitors from dozens down to one or two, applied by Evans to foundation-model training costs. Apply: Track whether foundation-model training costs are escalating on a similar doubling curve to forecast how many labs can remain at the frontier over time.

  17. Marginal-cost framework (mobile network vs AI compute) — A framework contrasting technologies with near-zero marginal cost against ones with real marginal cost that scales with usage, such as mobile network capacity or AI compute, used to explain the AI GPU supply crunch. Apply: Before assuming an AI product can scale like typical software, check whether its usage growth requires proportional new compute capacity the way mobile data growth required new base stations.

  18. Compete up/down the stack framework — An analytical framework for examining, layer by layer, which parts of a technology stack a company can move into or defend against, to understand where competitive moats and value capture sit. Apply: For any AI company, map which stack layer it could realistically expand into or be squeezed out of.

  19. Binary-bet framework — A decision-framing device, drawn from the urban legend of a trader buying stocks during the Cuban Missile Crisis because either the rumor is false and stocks are cheap or it's true and nothing matters, applied by Evans to AGI/ASI uncertainty. Apply: When a scenario is genuinely binary and catastrophic on one branch, treat that branch as unanalyzable and focus practical analysis on the branch where ordinary questions still apply.

  20. Netscape/browser analogy — The comparison of OpenAI and chatbots to Netscape and early web browsers, an early, powerful but still-broken deployment layer for a new technology. Apply: Use it to calibrate expectations for current AI products as an early, unpolished stage rather than a finished platform, without over-extending the analogy's specifics.

  21. Building blocks, value capture, and points of leverage framework — A set of strategic questions Evans poses for assessing a new technology platform's maturity: what are its building blocks, where does value get captured, and what are the points of leverage. Apply: Apply these three questions to any emerging AI product category before predicting its winners.

  22. DAU/WAU/MAU segmentation — Standard daily, weekly, and monthly active user segmentation, applied by Evans to show real AI usage depth of roughly 10-15% daily active is much shallower than hype suggests. Apply: When evaluating an AI product's traction, ask for DAU/WAU/MAU splits rather than accepting aggregate usage claims at face value.

  23. Three-tier AI usage segmentation — Evans's split of current AI users into software developers with clear product-market fit, self-employed knowledge workers with heavy use, and everyone else with occasional, non-daily use. Apply: Segment any AI adoption claim by which of these three groups it's actually describing before generalizing it to everyone.

  24. Accountant-vs-lawyer analogy — A comparison illustrating how the same new productivity tool, such as 1970s spreadsheets or AI now, can be central to one profession's daily work and merely a nice-to-have for another. Apply: Before predicting a tool's impact on knowledge work, check whether the specific profession's daily workflow structurally matches what the tool automates.

  25. Explain-it-to-a-human heuristic — A test for separating raw model capability from a person's own ability to identify and describe a delegable task: if you couldn't explain the task clearly enough for a human assistant to do it, the model likely can't either. Apply: Before concluding a model can't do a task, check whether the real blocker is that the task was never clearly specified in the first place.

  26. Old-thing-faster vs genuinely-new-thing adoption pattern — The observed pattern that new technologies first get used to speed up existing workflows, and only later enable genuinely new activities that weren't previously possible. Apply: When assessing a new technology's maturity, distinguish faster or cheaper versions of existing tasks from genuinely new, previously impossible use cases, and expect the latter to appear later.

  27. Jevons paradox — The economic principle that making a resource or activity cheaper can increase total spending or consumption on it rather than reduce it, referenced as a partial explanation for why accountant headcount hasn't fallen despite automation. Apply: When predicting that cheaper AI-driven output will shrink a workforce, check whether cheaper unit cost might instead increase total demand for that output.

  28. Automation-of-human-capability-stack narrative — A historical narrative claiming automation has progressively moved up the human capability stack, first replacing physical labor and now leaving only cognitive work for AI to automate. Apply: Recognize this narrative when used to argue AI automation is inevitable, and treat it as a rhetorical framing rather than evidence, per Evans's own critique of it as unfalsifiable.

  29. Unfalsifiable-statement critique method — A critique technique for identifying claims about AI's future impact that can't be tested or disproven given current information. Apply: When encountering a sweeping claim about AI's ultimate impact, ask what observation could actually disprove it; if none exists, treat the claim as unfalsifiable rather than analytical.

  30. Jagged-capability framing — The description of current AI capability as uneven rather than uniform, strong on some multi-hour tasks while failing simple five-minute ones, which undercuts single-number claims of AI capability. Apply: Distrust any single metric summarizing AI capability; check performance across a range of task types and difficulties instead.

  31. Lump of labor fallacy — The fallacy of assuming a fixed amount of work exists, so that jobs eliminated by automation are simply lost rather than offset by new jobs the technology creates. Apply: When someone lists jobs AI will eliminate, ask what new categories of work the same technology might create that can't yet be named.

  32. Consumer/enterprise surplus framing — An economic framing where AI-driven productivity gains show up as the same price or outcome delivered faster or with more volume, rather than as a qualitatively better outcome. Apply: When evaluating an AI productivity claim, check whether the actual outcome improved or whether it's the same outcome delivered with more throughput at the same cost.

  33. Expert-system fallacy — The mistaken belief that a complex human judgment task can be formally and exhaustively defined in advance, analogous to once believing you could hand-write rules to distinguish a cat from a dog, used to critique attempts to precisely measure what a knowledge worker does. Apply: Be skeptical of any framework that claims to have fully enumerated the steps of a complex job before assessing whether AI can replace it.

  34. Perfect-sphere-horse critique — A joke about physicists modeling a horse as a perfect sphere to make a horse race tractable, used by Evans to critique job-exposure radar-chart analyses that define jobs into measurable pieces rather than modeling what people actually do. Apply: When shown a chart claiming AI can do a specific percentage of a job, ask what simplifying assumptions were used to make that percentage measurable in the first place.

  35. Anselm's ontological-proof analogy — A reference to the medieval philosophical argument that defines God as a being that necessarily exists, used by Evans to show that defining AGI as inevitable and inevitably catastrophic makes the conclusion true by definition without proving anything. Apply: When an AGI-risk argument's conclusion follows automatically from how AGI was defined, treat it as circular rather than as evidence.

  36. Four-part what-to-track framework — Evans's proposed set of trackable indicators for AI's trajectory: capital and infrastructure build-out, the number and consolidation of frontier-model companies, where use cases fall on the Pareto curve of capability versus ROI, and how much value the model layer itself can capture versus the application layer. Apply: Use these four trackers as a recurring checklist instead of debating AI's ultimate size or danger in the abstract.

  37. Most-people-are-three-things framework — Evans's explanation for why most employees won't spontaneously build their own AI tools even with model access: most people aren't tool builders, don't recognize the underlying problem as solvable, and aren't positioned to build a tool even if they did. Apply: Use this to temper expectations that giving employees model access alone will produce bottom-up tool innovation.

  38. Enterprise software taxonomy — A description of the enterprise software landscape as three layers: an improvised layer of spreadsheets and scripts, large horizontal enterprise systems, and hundreds of SaaS apps per large company, which AI reshuffles rather than replaces wholesale. Apply: When assessing an AI tool's enterprise opportunity, identify which of these three existing layers it's actually displacing or extending.

  39. Stack framework (TSMC/Windows/AWS) — The observation that no single layer of a technology stack, whether chips, operating system, or cloud infrastructure, has historically captured the entire value chain above it. Apply: Treat claims that a foundation-model company will own the whole ecosystem skeptically by checking historical precedent from chip, OS, and cloud infrastructure leaders.

  40. Two-axis automation-feasibility framework — A framework judging whether AI can automate a task based on two independent axes: whether sufficient training data exists, and whether outcomes can be verified at scale and objectively. Apply: Before predicting AI will automate a task, separately assess data availability and verifiability rather than capability alone.

  41. Framework for AI's conceptual limits — Two conditions under which Evans argues AI conceptually struggles: tasks based on tacit or implicit knowledge that's hard to turn into a statistics problem, and tasks where the goal deliberately differs from what most people would do, since models model the average. Apply: Use these two checks to identify domains likely to resist AI automation even as general capability improves.

  42. AlphaGo/verifiable-reward framework — The idea that AlphaGo could generate genuinely novel winning moves because it had a scalable way to verify whether a move worked, a mechanism largely absent for judging the quality of a novel business strategy. Apply: Before expecting an AI to generate a novel and correct solution in a domain, check whether a scalable verifier for correctness exists in that domain.

  43. Four-step hierarchy of AI-capability skepticism — A layered way to evaluate objections to AI's capability: the model simply can't do it, it can't do it yet, it's unclear whether enough training data or feedback loops can ever be built for it, and even if they could, whether the average output that training would produce is actually wanted. Apply: When dismissing an AI capability claim, identify which of these four levels the objection actually sits at rather than treating all skepticism as equivalent.

  44. Imagine-if consumer capability ladder — A staged sequence of consumer AI use cases of increasing difficulty, from finding an exact item to comparing options with reasoning to generating a personalized styling suggestion from personal data, showing how quickly feasibility drops off as tasks require more judgment. Apply: Use this ladder to locate where a proposed consumer AI feature actually sits between solved, solvable, and requiring nuanced judgment.

  45. Parallel enterprise capability ladder — An equivalent staged sequence for enterprise AI use cases, from flagging emerging concerns in call recordings to synthesizing multi-source data into a full churn-diagnosis project, showing that higher rungs require frontier capability and significant project scoping. Apply: Use this ladder to judge whether a proposed enterprise AI initiative is a commodity task or actually requires cutting-edge capability and dedicated project scoping.

  46. Variance-reduction-to-the-mean concept — The claim that AI, past a certain capability point, mostly makes it cheaper to produce more of what already exists rather than inventing genuinely new categories, contrasted against inventing something like punk or hip-hop. Apply: When an AI-generated output is described as revolutionary, check whether it's actually a new category or just cheaper abundance within an existing one.

  47. Outsider-vs-insider AI-use-case perception pattern — The observed pattern that industry outsiders guess an AI use case is the obvious, visible one, while insiders identify the real, unglamorous use case, illustrated via advertising and consulting. Apply: When assessing AI's likely impact on an unfamiliar industry, seek out an insider's read rather than trusting the outsider-obvious guess.

  48. Strategic-response typology by company type — A classification of how different companies respond to a platform shift: incumbents integrate the new capability as a feature of an existing product, while new entrants without a legacy business must build the entire stack themselves. Apply: Use this typology to predict a company's AI strategy from its starting position rather than from stated intentions alone.

  49. Tool-builders-build-tools-for-their-own-task hypothesis — A candidate explanation for why coding became LLMs' strongest early product-market fit: developers are the rare case where the people building the tools are also the tools' end users. Apply: Use this to predict which future AI use cases might see similarly fast product-market fit by looking for domains where tool builders are also the users.

  50. Cargo cult framework — An analogy from WWII Pacific islanders imitating the outward forms of American airbases after the source of supplies left, without understanding the causal mechanism, applied to companies imitating a rival's mature playbook without its earlier discovery phase. Apply: When a company imitates a competitor's mature-stage playbook, check whether it also has access to the pre-product-market-fit knowledge that originally produced that playbook, not just its visible outward form.

  51. Staged enterprise AI adoption framework — A description of enterprise AI rollout as occurring in stages: broad co-pilot deployment with limited success, then scattered one-off pilots that work but don't transform anything, with deeper organizational reimagination still unresolved. Apply: Use this to locate where a given enterprise's AI deployment actually sits rather than treating a co-pilot rollout as evidence of transformation.

  52. Convergence framework (software vs consulting firms) — The claim that enterprise software companies and management consultancies are doing fundamentally the same reinvention work via different modalities, and are now converging via AI forward-deployed engineer roles. Apply: When evaluating a vertical AI startup's deployed services model, compare it against what a consulting firm would charge to do the equivalent reimagination work.

  53. Consumer social-product mechanic framework — The idea that successful consumer social products require inventing a genuinely new engagement mechanic that must evolve beyond a gimmick into sustained utility, illustrated against failed novelty products like drones and 3D printers. Apply: Before predicting an AI consumer app will succeed, ask whether it has found a durable new mechanic or is still a cool-then-done novelty.

  54. MCP (Model Context Protocol) — A named current AI-industry protocol that Evans predicts will likely be replaced by something else, cited as an example of the field's ongoing turnover. Apply: Treat current protocol or standard choices like MCP as provisional and expect them to be superseded rather than treating them as settled infrastructure.

  55. Physical-AI contrast framework — A framework contrasting humanoid robots working in factories against simply making existing trucks GPS-connected and self-driving, for judging how much AI will actually change physical and logistics industries. Apply: When assessing AI's impact on a physical-goods industry, ask specifically which version of physical AI is actually being discussed.

Insights

Mobile network operators captured little of the value from roughly 1,000-2,000x traffic growth over 15 years — it flowed up the stack to companies like Uber, banks, and YouTube, a parallel Evans uses to question where AI value will actually accrue.

Evans casts doubt on Sam Altman's claim that OpenAI will be like Windows, specifically because LLMs lack the network effects that gave Windows its moat.

A frontier AI lab reportedly flunked out and then jumped back near the top of leaderboards, which Evans reads as a negative signal for durable moats: frontier capability is still purchasable with a couple billion dollars and the right hires, unlike search, social, or mobile.

Evans suggests Anthropic's focus on coding was less deliberate foresight than a narrower-capital stumble onto something that worked, which the rest of the industry is now scrambling to copy.

He describes OpenAI's reported hiring of Meta executives to replicate Meta's playbook as cargo culty, because those hires know Meta's mature-stage playbook rather than the pre-product-market-fit discovery phase Meta itself had to improvise through.

Evans frames many current AI productivity gains as consumer surplus rather than quality gains — e.g., legal or consulting research delivered faster and cheaper for the same billed hours, without a better outcome or higher odds of success.

He notes a European water-company CEO only considered making innovation a top-five priority after competing concerns like earthquake regulation, lead pipes, removing Huawei equipment, and drought, illustrating why AI adoption looks slow to outsiders even when it isn't irrational from inside the industry.

Evans predicts that, similar to how the internet devastated newspapers but barely touched industries like heavy machinery or aggregates, AI will leave some industries essentially unchanged, and which ones will be surprising.

«You can wave your hands and say, "No, this is like electricity." Okay, fine. It's like electricity. Well, what happened with electricity? And there was a time before electricity. And people before electricity weren't dumb either.»

— 00:00

«So, we've got Jeff Hinton was 10 years ago saying, "Stop training radiologists." And the problem here is he doesn't actually understand what radiologists do.»

— 00:14

«You have this stuff that's amazing and it doesn't work.»

— 10:23

«We have something that's very jagged.»

— 19:12

«Intelligence isn't linear like that.»

— 19:34

«It's horseshit. You know, it's like the joke about the physicists who are predicting which horse is going to win a race. And they say, 'First, we're going to presume that the horse is a perfect sphere.'»

— 24:39

«If you define AGI as necessarily inevitable and necessarily going to kill us all, then AGI is necessarily going to kill us all. Great. But you haven't proved anything.»

— 25:45

«Most people are three things. Most people aren't tool builders. Most people don't see the problem that the tool would be solving. And most people aren't in a position to build the tool even if the other two weren't true.»

— 29:15

«You can't win the argument by defining it away.»

— 33:53

«Therefore, there will be more software.»

— 34:47

«No, no, no, no, no. You use it to automate all the really boring back office crap.»

— 45:29

«I have a lot of sympathy for Sembleman here. Um, that may be a unique and controversial statement.»

— 46:11

«Everyone's throwing rocks at coding.»

— 48:31

«Writing the code isn't the hard part because it's such a disastrous chaotic confused mess.»

— 51:17

«It felt a bit cargo culty.»

— 52:57

«I always used to joke that if you say digital transformation three times then a partner from Accenture will appear in a puff of smoke.»

— 60:49

«Everything has to actually be invented. All of the use cases have to be invented. Very few of them are just pure purely spontaneous grassroots consumer behavior.»

— 63:00

«there's no consumer use case for 3D printing»

— 65:38

«it's not the chat rooms it's the network and it's the same thing here it's not the text in the pictures it's the enabling technology»

— 67:48

«AI is whatever's just is new. It's like technology. It's not technology once it's been around for a while.»

— 69:24

«it's a low margin reseller one time sells»

— 73:01

Reception

The comments show a substantive, often contentious debate about AI/LLM capabilities and limits, mixing genuine engagement and praise for the interview with skepticism, contrarian pushback, and some off-topic or unhinged tangents.

A dense, framework-rich conversation that deliberately avoids hype-cycle punditry in favor of historical pattern-matching, concrete trackable indicators, and testable distinctions between AI's capability and its actual deployment, valuable as a durable mental-model reference even though many of its specific examples (leaderboard rankings, current lab strategies, named products) are tied to mid-2025 and will date quickly.

73:48

↳ Unsupervised Learning: With Jacob Effron · YouTube

Watch original