现在是 AI 的 1997 年：Benedict Evans 为什么说它很大，但只和互联网一样大

Evans says his most controversial view is that AI is as big a deal as the internet or mobile and only as big. He says some in tech think it is more like the industrial revolution, but he points out that the internet and smartphones were already very big, so the comparison is not small. He also thinks it is not productive to argue about whether AI is 20% bigger than the internet or 100%.

If you take the internet comparison seriously, he says, we are in 1997. Most things do not work yet, most of what people will build has not been built, and it is unclear how it will work when it does. Adoption is also very uneven: some people in tech bought clusters of Mac minis and stopped using Google, while most people outside tech use this every week or two at most.

He also warns against calling the race between model labs early, comparing it to asking in 1997 whether it would be Excite or Yahoo, when the answer was neither. He says his new presentation, about 80 slides long, is half-jokingly described as saying "we don't know." To explain the uneven moment, he uses the analogy of VisiCalc: accountants seeing the first spreadsheets found it mind-blowing, while lawyers or journalists thought it clever but not for them. Software developers are the accountants of today, he says, and survey data suggests that even among 13 to 18 year olds only about 15 to 20% use AI daily.

Lenny asks why the leading AI labs are investing in consultancies and forward-deployed engineers. Evans jokes that a forward-deployed engineer is an Accenture outsourced developer who lives in San Francisco. More seriously, he says companies don't keep people idle waiting to run big projects, which is why they hire consultancies, ad agencies or architects. Rethinking internal workflows around AI is exactly this kind of project, needing five or ten people for a month or two, and then more work to implement it.

His broader point is to separate the task from the job. He uses an elevator attendant as an example of a job that really was a task, and then the Jevons paradox, which he describes as price elasticity: when something becomes cheaper, you may do more of it. Excel did not shrink the hours of junior bankers, and software developers now write a fraction of the code they used to without the number of engineers falling to a tenth.

Amazon, he says, gets you the SKU, but knowing which SKU you want is another job; similarly, Claude Code can write code, but someone still has to decide what to build. People who say they made a McKinsey deck with Claude miss that clients pay Bain to walk through the enterprise, learn the politics and talk to customers, and the slide deck is just the task. He adds that newspapers and record labels got hit when the physical part of their business decoupled. He also points to charts showing the number of accountants kept rising through adding machines, mainframes, spreadsheets and cloud.

Evans notes that even the most advanced AI companies keep increasing headcount. On entry-level job warnings from AI lab leaders, he says he dislikes arguments from authority, and a lab CEO's views on model progress are relevant while his theories of labor markets are not. He says he does not believe the cynical view that such claims are just stock-pumping.

His core argument is historical. Since 1800, when around 90% of people were peasants, technology has automated jobs and created new ones that nobody could have foreseen, such as railway engineer. Each round brings frictional pain, lost jobs and hollowed-out towns, but also higher GDP and richer societies. Most jobs lost are, in retrospect, poor jobs. He says that AI might be different only if it is much faster, and while adoption is quick because it stands on the internet, the internet itself was faster than what came before.

The other part is how the world works. He says doomers act as if every big company will buy ChatGPT tomorrow and fire staff two weeks later, and he calls those people morons. In his view an enterprise software sales cycle is about 18 months if you are lucky, and nobody will tear out SAP overnight. Even many SaaS companies founded just before ChatGPT could have been started 15 years earlier; it took people years to see which problem they could solve with which technology.

Lenny says Evans's view is comforting: this is huge, but we've been through such transformations before. Evans points to a slide with an IBM ad from the 1950s promising an electronic calculator the size of a fridge that is like having 150 extra engineers, which he says is basically the pitch of Claude Code. He also describes how long it took to build a chart on supermarket SKUs before the web: days on the phone, expensive calls and two weeks of work, against two hours on Google now. We forget how big the internet was, he says, because it is simply the world we live in.

On AGI, Evans says nobody has a theory of human intelligence, of why these models work so well, or of how much better they will get, so everyone is forecasting by vibes. He quotes the line, attributed to AI scientist Larry Tesler, that AI is whatever machines can't do yet. Terms keep shifting: AGI is now defined as doing a share of economically valuable work, which a 1975 mainframe could also claim, and superintelligence is similarly slippery.

He thinks arguing over definitions is like arguing whether crypto is blockchain. The practical point is that even if models stopped improving tomorrow, this would be an incredibly useful technology that gets rolled out over ten years. On Marc Andreessen's point that companies can now grow far larger, Evans agrees that the addressable market keeps expanding as software moves into more of the economy, from tens of thousands of mainframes to billions of smartphones. He adds that the lump-of-labor fallacy applies to the idea that all the money goes to a few AI leaders.

Evans takes issue with Sam Altman's line about selling intelligence on a meter like electricity. He asks the utility question: when you watch TV, the TV company does not pay a share of your bill to the electricity company. He starts from a telecom analogy: the global mobile industry earns about a trillion dollars a year and spends about $200 billion a year on capex, mobile data consumption is up roughly 1,500 to 2,000 times since 2010, yet the carriers' stocks have gone nowhere in 25 years because the cool applications are built further up the stack.

He then asks whether the chatbot will do the whole thing or whether it all has to be apps. If all the apps have to be built by other companies, do foundation models have leverage like Windows, or is this more like AWS, where a law firm buying software doesn't care which cloud runs it? He argues that models seem to lack network effects and are commodities as far as users can tell, so there should be ongoing competition and lower margins, with value further up the stack.

He says the current price situation, like the person spending $1.5 million on tokens in a month, is like someone getting a 50 grand mobile data bill in 2010, which is temporary. He admits his thesis may be wrong, since in 1997 most people would have missed almost everything about the next decade, and says we should presume we do not know. Asked about investing, he says he was a mediocre sell-side analyst, is not an investor, and notes that being right is different from being early or having the right price. He adds that in mobile, some companies missed it completely while for others, such as Google, it changed little.

Lenny raises the idea that distribution is becoming a bigger moat as software gets easier to build. Evans agrees that if the product is a commodity, distribution is what matters. His analogy is the web browser: the browser product is a thin wrapper around a rendering engine, with no real innovation since tabs, and Microsoft used distribution to break in, but winning browsers did not matter because the value was further up the stack.

Now, he says, Google is using distribution to push Gemini, and for a normal user there is little difference. Meta, which people in tech had written off, was already between ChatGPT and Gemini in survey data after it had put its AI on every surface. He describes OpenAI's late-2025 strategy as trying everything to find a flywheel before Google, Meta and Amazon make their own the default.

On Apple, he says the 2024 WWDC vision of a personal AI assistant was the most compelling he has seen, even if nobody could ship it. He notes that the model is the dumb thing underneath, and different companies will make different product decisions on top, with Apple able to run things on a billion devices at the edge. He adds that Google's new Android intelligence will reach only a few buyers of the Pixel phone, which shows how little attention goes to Android and iPhone.

Lenny asks about the growing anti-AI sentiment, which polls place below even ICE, and communities blocking data centers. Evans calls it a big fuzzy mess. Electricity bills rising is a real concern in a small number of places. The water issue, he says, is largely fake: a Livermore lab study at the end of 2024 estimated US data center water use at about 0.017% of national consumption, and a local well being capped is a planning problem. Data centers are around 5% of US energy and might grow by about one percentage point a year.

On jobs, he says the honest answer is that we don't know yet. There is a slowdown in employment for 18 to 24 year olds, but it looks the same for people with and without degrees and in fields exposed or not exposed to AI. He adds that the labs give almost no usage data, not even a daily active user number for ChatGPT, so the data comes from academic economists and consultancies. Politically, he notes, a student who can't get a job will blame AI whether or not it is the cause.

There are also niche culture wars, such as illustrators and novelists upset about AI art and writing, and concerns about AI slop. Evans compares the mix to the backlash against social media, where some claims were true, some partly true, and some, like the idea that Facebook sells your data, were false but held with certainty. He adds that Trump's interest in an executive order on dangerous models is driven by things like cyber risk, which he doubts is a Main Street concern.

Asked how he is raising his kids, Evans says it depends on timing: if a child enters the job market in the next year or two everything is up in the air, but in five years things will have settled in unpredictable ways. His own child is in his early teens, and he says the questions about media, influencers and gatekeepers predate ChatGPT. He says he lacks a systematic plan and would settle for his son not breaking his Chromebook again.

His general vibe, which Lenny likes, is that it will be okay, though he says that perhaps comes from being British and not having political violence in 500 years. He says we should worry but that worry is a constant, recalling a 1970s panic about databases that was half true. Deepfake nudes, he notes, are a real difference from Photoshop, because a teenager can now make them for a whole school in an afternoon. As with social media, connecting everyone also connected our worst instincts.

He also tells the story of the UK Post Office scandal, where a Fujitsu-built point-of-sale system showed phantom cash shortfalls, hundreds of sub-postmasters were prosecuted, and some lost homes or took their own lives, while officials swore there were no bugs. His point is that every technology offers ways to ruin lives, deliberately or by accident, and we need to be conscious of that without panicking. For career advice, he says to find the overlap between what you are good at, what jobs suit that, and what people pay for, and to get at least two of the three.

Evans says that a question not enough people ask is whether model labs have pricing power, since many assume today's situation will continue. His other question is task versus job. He closes his presentation with a chart of global recorded music revenue, which halved from 2000 to 2015 and has since recovered to about 75% of its peak, driven by streaming. The first half of the chart is what happens when you don't pay $15 for a CD, and the second half is what happens when $15 a month gives you all the music there is.

His lesson is that each technology first does the old thing, then does more of it on the new platform, and then creates something that redefines the question, like Spotify, which is not an online music store. You only know the question after someone has built a billion-dollar product. Lenny notes that engineering, once thought the last profession to be automated, has become the most transformed role.

Evans calls tables of the percentage of each job that AI can do, such as those based on the US O*NET data, deluded. This is the expert-systems fallacy, breaking a thing into hundreds of steps that never add up to the whole. The other fallacy is the taxi-driver one: in 1997 nobody would have expected the internet to disrupt taxis, and now people say personal trainers are safe even though a phone and an AI could watch your form. He adds that Uber transformed taxis in many cities, while Airbnb had only a marginal effect on hotels, since business travelers still want hotels. So his answer to most questions is that it depends.

Lenny asks what people worried about their careers should do. Evans first notes that, as Keynes said, in the long run we're all dead, and an average that is fine can still leave some individuals badly off: a 19-year-old in 1914 faced a one-in-three chance of not coming back. Some professions, especially associates in professional services, face real questions about the pyramid structure.

His advice is not to stick your head in the sand and declare you hate it, which gives a feeling of moral superiority and nothing else. Instead, immerse yourself and learn what you can do with it and how it changes things, just as people did with mobile and the internet. If a law firm cut hiring from 100 associates to 50, he says, telling the interviewer you will never use AI is the wrong approach. Lenny says the advice matches the podcast's recurring theme: build things.

In the AI corner, Evans says he struggles to find uses, as he is like the lawyer looking at ChatGPT: the precise information retrieval he would automate is what AI does worst. He uses it for proofreading and for images, including redecorating his apartment, and he dictates most of what he writes, which may count as AI or may just be automation. He says the chatbot is a blank screen with a jagged edge, and the solution is to wrap it in use cases. He adds that consumer AI has no breakout apps yet, partly because marginal costs prevent a free tier at 50 million users.