Benedict Evans on AI: As Big as the Internet, and Only as Big
Lenny's Podcast
Benedict Evans spent years as the in-house analyst at Andreessen Horowitz and has since worked as an independent analyst, publishing a large presentation on technology trends every six months. The latest is titled AI Eats the World. In this conversation with Lenny Rachitsky, the central question is what people still aren't pricing in about AI's effect on work and life. Evans's position is that AI is a fundamental platform shift. It is also one whose outcomes we can barely predict yet, and one that history suggests we will probably come through. His short version is that it will "probably be okay."
"As big a deal as the internet or mobile, and only as big"
Evans opens with what he calls his most controversial opinion: AI is as big a deal as the internet or mobile, and only as big a deal. People in tech who compare it to the industrial revolution push back on the second half. Evans responds that smartphones and the internet were enormous. This conversation couldn't happen without the internet, he notes.
He then pushes the internet comparison further. In his view we are in about 1997. It's very exciting, most things don't really work yet, most of what people will eventually build hasn't been built, and it's unclear how any of it will work once it does. People who have "already taken the pill" tend to assume everyone else is where they are. Evans sees a very wide distribution instead. Some people in tech have bought clusters of Mac minis and stopped using Google. Outside tech, he says, most people who use AI at all use it every week or two.
He describes the uncertainty as fractal. At the top level, AI will change everything, but he doesn't think arguing whether it is 20% or 100% bigger than the internet is productive. At the next level down, questions like whether the model labs have pricing power, whether OpenAI has won, or whether Anthropic is ahead this week look like asking in 1997 whether Excite or Yahoo would win. The answer then was generally neither. One comment on his new presentation said it was "80 slides saying we don't know." Evans calls that slightly facetious but also kind of true.
The accountant seeing VisiCalc
Asked how far we are from things being radically different, Evans says software development is already in that moment. That raises open questions about the software industry: how much people will build for themselves, and how much more software will exist. Almost nobody, he says, seriously believes you'll just vibe-code your own Stripe.
At the other extreme is a law firm. It finds all this interesting but wants to know how to use it, how to avoid becoming the next story about a filing full of hallucinations, and how many associates to hire next year.
His main analogy is an accountant seeing the first software spreadsheet in the late 1970s. Change the interest rate and every other number updates, doing a week of work in 30 seconds. For an accountant this was mind-blowing. A lawyer or journalist watching the same demo would think it was clever and that their accountant should see it, but it wasn't what they did. They might use it for a timesheet if it didn't cost $10,000–15,000 in today's money to buy the Apple II, monitor and printer. They needed a word processor, which came shortly afterwards. Evans puts software developers in the accountant's position today, with a "before VisiCalc / after VisiCalc" split that he maps to "before Claude Code / after Claude Code." Many others are using AI to varying degrees but remain slightly puzzled.
He cites survey data from the presentation showing that even among 13-to-18-year-olds, roughly 15–20% are daily users and another 20% are weekly users, while about 60% say they don't use it. He links this wide spread to the "jagged frontier": where does it work, can you tell in advance, and can you work out for yourself what to do with it.
Why the AI labs are investing in consultants
Rachitsky raises a trend Evans has written about. The labs, especially OpenAI and Anthropic, are investing in professional services, consultancies and forward-deployed engineers, even though many expected AI to wipe out consultants.
Evans jokes that a forward-deployed engineer is like an Accenture outsourced developer who works in San Francisco. His serious point is that companies don't keep spare people waiting for big new projects. They don't have idle staff to redesign their stores, figure out where stores should be, or work out why churn is too high. That's why they hire Bain, BCG or McKinsey on one side, Accenture or Infosys on the other, or branding agencies and architecture firms. A company has no reason to keep 15 architects on staff when it can hire a firm.
Reimagining every internal workflow to see what AI could automate is a project in itself: five or ten people for a month or two. Implementing it is another project, involving plugging vertical systems into horizontal ones, building new workflows and training people. Someone has to do that work. Some PE firms provide that support to their portfolio companies, and otherwise companies hire consultants or integrators.
Is it the task or the job?
This leads to the theme Evans uses to organize the "change" section of his presentation, which also has sections on capital and deployment. Is the hard part of a job writing code line by line, or making the PowerPoint, or is it something else?
Sometimes the task really is the job. Evans lives in a building with a manually operated elevator, driven with a lever by an attendant. Those jobs were automated after the 1950s and replaced by a button. More often, he argues, the relevant dynamic is price elasticity, which is what the Jevons paradox amounts to. When something gets cheaper, do you do the same for less money, more for the same money, or more for more money because new returns open up? He recalls an old joke of his: before Excel, junior investment bankers worked brutal hours, and thanks to Excel, Goldman associates now finish at lunchtime on Fridays. That obviously didn't happen. Similarly, Apple writes something like 90% of an iPhone app's code (modem drivers, graphics drivers, file system), yet we didn't end up with a tenth as many engineers.
His e-commerce analogy makes the same point. Amazon gets you the SKU if you already know which SKU you want. If you don't know which microphone to buy, Amazon probably isn't where you start. Claude Code can write the code, but which code do you want, which features, for which customer, and how will you take it to market? A McKinsey-style 75-slide deck made with Claude is, in his words, a crappy version of the real thing. Even a good one wouldn't be what clients pay for. They pay Bain to walk around the company, understand the politics, work out why something wasn't done, and talk to customers. The PowerPoint is the task, not the job.
He also looks at industries the internet hurt because two things could be separated. Record companies didn't think of themselves as manufacturers of small plastic discs, and newspapers didn't think of themselves as printing and trucking businesses. Once those functions were decoupled, they were in trouble. Often, though, the pieces can't be decoupled, or that wasn't the real problem. His presentation includes charts showing the number of accountants rising through the 20th century and again in the 21st, despite adding machines, punch cards, mainframes, databases, ERP, spreadsheets and cloud software. His conclusion is that it has to be more complicated than automation. Rachitsky adds that even the most advanced AI companies, and companies like Every whose founder Dan Shipper recently appeared on the podcast, are increasing headcount.
The "job apocalypse" and 200 years of automation
Asked about warnings such as Dario Amodei's that entry-level jobs will disappear, Evans first objects to argument from authority. If you use it, he says, the authority should be relevant. He's interested in Amodei's view of where models will go in the next six to twelve months, but not especially in his theories of labor markets and comparative advantage. Evans says he doesn't believe the cynical view that such statements are just stock pumping.
His substantive argument is historical. Every new technology automates some jobs, and through price elasticity and what it enables, it creates new ones. Around 1800, he says, roughly 90% of people were peasants worried about crop failure. Ever since, we have been automating jobs and creating new ones. You can always see the job that's going away, but not the new one, because it doesn't exist yet and often sounds silly ("railway engineer — what's a railway?"). Each round brings frictional pain: people lose jobs and towns get hollowed out. On the other side, Evans argues, we're all richer. Typesetters, telephone operators and typists largely disappeared, and he says the jobs that vanish usually look like bad jobs in retrospect.
So is AI different? One answer is speed. Evans agrees adoption is faster but attributes it to standing on the shoulders of giants. ChatGPT could reach around 900 million users because that many people were already online. When Marc Andreessen launched Netscape in 1993–94, there were only 50–100 million PCs in the world. Each wave builds on the ones before, so the internet was fast too.
Why companies won't fire everyone in two weeks
His other answer is about how organizations actually work. He calls the doomers who imagine big companies buying ChatGPT tomorrow and firing their staff two weeks later "morons." In his view their basic failure to understand the world was the starting point of all their other mistakes. Enterprise sales cycles are 18 months if you're lucky, longer than the gap between a startup's funding rounds, and even slower in sectors like aerospace or healthcare. Nobody is going to rip out SAP next week. Over three, five or ten years, sector by sector, the whole landscape may look radically different and jobs will have changed, but it will take that long.
Figuring out what to do with a new technology also takes time. He recalls Frame.io, a video collaboration company a16z looked at, though he concedes it's a weak example because it relied on cutting-edge web technology. His broader point: take ten random SaaS companies founded the day before ChatGPT launched and ask how many could have been founded at any time in the previous 15 years. The delay was someone noticing a problem inside an industry and seeing how to solve it. That didn't all happen the day after Google Docs, and it took 10 to 20 years.
"Completely different, just like everything else"
Near the end of his deck, Evans has a slide saying roughly that this will be completely different from everything else, just like everything else. It is followed by a 1950s IBM ad showing rows of men in white shirts holding slide rules, with the slogan that an IBM electronic calculator (the word "computer" wasn't used yet) is like having 150 extra engineers. Evans notes that's basically the pitch for Claude Code.
To show how much we forget, he describes making a chart of how many products American supermarkets have stocked since the 1950s. The chart's point is that barcodes let supermarkets track, and therefore stock, far more items. To make it he had to discover that the Food Marketing Institute existed, that it published an annual SKU count, and that it went back to the 1950s. In 1994 you wouldn't have known any of that. You'd have spent days and money on long-distance calls looking for a library that held the reports, or paid the Institute for copies, and two weeks later you might look at the chart and decide it wasn't interesting. For Evans it took two hours on Google. Enormous changes become invisible once they are simply how the world is.
AGI as a moving target
Rachitsky asks whether AGI or superintelligence is what makes this time different. Evans says he doesn't know. We have no theory of human intelligence, no theory of why these models work so well, and no theory of how much better they'll get, so everyone is "vibes forecasting." He's dismissive of late-night philosophizing about consciousness.
What he thinks can be observed is redefinition. He cites Larry Tesler's line that AI is whatever machines can't do yet, and once they can, it's "just software." People say image recognition or sentiment analysis "isn't AI," in the way jet airliners were once "technology" and now aren't. AGI, in his view, is being redefined to mean what works now, such as doing a certain percentage of economically valuable work. He points out that a 1975 IBM mainframe could do a meaningful share of work previously done by people. He can't tell whether superintelligence now means more or less than AGI, given how the terms have shifted. As with arguments over whether crypto is blockchain or the reverse, there's no correct answer. What matters is being clear about what you mean.
Whether we get human-level intelligence, he says, we have no way of knowing. His serious point is that you don't need to believe any of it. If models stopped improving tomorrow, this would still be an incredibly useful, world-changing technology rolled out over the next decade.
Bigger markets, and the margin structure of utilities
Rachitsky relays Marc Andreessen's view that companies can now grow far larger than before, with many more trillion-dollar companies to come. Evans ties this to Andreessen's "software is eating the world" thesis: the addressable market keeps growing as software reaches more of the economy. He gives rough install bases for each platform: peak mainframes perhaps 70,000–80,000 units, 50–100 million PCs when the internet took off, perhaps 1–1.5 billion PCs today (maybe 700–800 million consumer), and about 5.5–6 billion smartphones. A few years ago people argued we'd run out of new users. Evans says that was the wrong model, because growth is now coming from automating large new areas of the economy. He compares it to electricity spreading into factories and then into everything.
That comparison leads to his "capital" argument. He quotes Sam Altman saying AI intelligence would be sold on a meter like water or electricity, and responds that someone should explain utility margins to him. TV makers don't pay a share of your bill to the electricity company, and Bosch doesn't pay utilities a cut of each washing machine.
Evans began his career as a telecoms analyst and uses that industry as a warning. By his figures, global mobile revenue is about $1 trillion a year, mobile capex about $200 billion (around $300 billion for all telecoms, 15–20% of revenue), and mobile data consumption is roughly 1,500–2,000 times its 2010 level. Yet the stocks have gone nowhere in 25 years. He describes it as an ex-growth, low-margin commodity utility: an extraordinary piece of infrastructure where all the cool things are built by someone else. The carriers once thought they would build what people did on their phones, and in the end neither they nor Apple did most of it.
Models versus applications: cloud, not Windows
Evans sees the key question for foundation models as whether the model or chatbot does everything. The "Claude for X" products look to him like Excel's templates under File → New, while the real equivalents are billion-dollar companies. If everything has to be apps, the labs can't build all of them, just as Microsoft didn't. Then the question is whether models have leverage up the stack as Windows did, or whether this looks like AWS. A law firm buying software doesn't care which cloud it runs on, and developers don't standardize on one cloud because customers do.
Crucially, he says, models don't appear to have network effects, so there's no winner-takes-all dynamic and competition should continue indefinitely. Without radical differentiation, why would the labs have pricing power? He expects the result to look more like cloud than Windows, with three to six companies selling a commodity near marginal cost and value accruing further up the stack. He separates today's "radical price disequilibrium" from the eventual steady state. He mentions the OpenClaw developer reportedly spending $1.5 million on tokens in a month and compares it to someone getting a $50,000 mobile data bill in 2010: temporary. He acknowledges semiconductor analysts disagree, admits he is exaggerating their view, and says he looks forward to being proven wrong.
He stresses that this could be completely wrong. Anyone discussing mobile in 2000 would have missed that a "has-been PC company from Cupertino" and a search company would win. As a junior analyst in 1999, he visited a UK dot-com selling computer components online. On the train back from Birmingham, a senior banker, David Tate, summed it up: you can say "dot-com" all you like, it's a low-margin reseller with one-time sales. Evans thinks the labs are in a similar position: lots of science, but so is flat-panel manufacturing, which has won Nobel prizes and is still a low-margin commodity.
Asked what he'd invest in, Evans says he was a poor sell-side analyst, partly because he wasn't interested in share prices. He emphasizes that being right isn't the same as being early, and the right company isn't the same as the right price. Google, Apple, Meta and Amazon face questions but it's hard to see a real problem for them. He points to mobile, where some companies missed the shift entirely, like Yahoo Mail and possibly eBay, while for Google and Amazon little changed and for Meta it was great, with cameras and notifications always on hand. He says he's massively oversimplifying. He cites Steven Sinofsky's observation that incumbents always try to make the new thing a feature, and sometimes they're right.
Distribution as the moat
Rachitsky raises a recurring theme on the podcast: as software gets easier to build, distribution becomes the key asset, which favors incumbents. Evans agrees, joking with the Drake meme ("I don't like GPT wrappers, I do like harnesses"). If the product is a commodity, distribution is what matters.
His comparison is web browsers. The rendering engine can be better or worse, but the browser product is a thin wrapper around it: an input box and an output box. The last real innovation was tabs 20 to 25 years ago. Microsoft used distribution to win browsers, and then winning browsers turned out not to matter because the value was further up the stack. Today, Evans says, Google is using distribution to push Gemini. A normal user can't tell it apart from ChatGPT. Meta, which people in tech have written off, showed up in survey data between ChatGPT and Gemini even before its newest efforts, because it had "sprayed it on every surface" and the product was adequate. He reads OpenAI's late-2025 strategy of trying many things at once as a search for a flywheel and distribution before Google, Meta and Amazon make their offerings the defaults.
Apple, he says, is the last shoe to drop. Rewatching WWDC 2024, the Apple Intelligence half remains the most compelling vision of a personal AI assistant he has seen: agentic, on-device, tool-using, with no prompt injection or hallucinations and a standardized intent system across 10,000 apps. He isn't surprised Apple couldn't ship it, since nobody else has either. He wants to see what Apple shows at WWDC, possibly powered by Gemini. If so, Gemini-based features on Android and Gemini-powered Apple Intelligence on iOS would be different products, because the model is "the dumb thing underneath" and the product decisions and distribution sit on top. Apple has a billion devices able to run things on-device. Evans mocks Google's recent Android intelligence launch for coming only to its most powerful devices, meaning, as he puts it, "the five people who bought a Pixel phone." He says the launch getting so little attention shows how little people now watch Android.
The anti-AI backlash: a "big fuzzy mess"
Rachitsky notes surveys showing AI is unpopular, local opposition to data centers, and Eric Schmidt being booed at a commencement speech whenever he mentioned AI. Evans sees a mix of issues with very different merits.
Higher electricity bills are real, he says, though only in a small number of places. The water argument he calls essentially fake. Data centers mostly use closed-loop cooling, and he cites a Livermore lab study from late 2024 estimating US data-center water use at about 0.017% of national consumption. If a small town's single well is handed to a data center, that's a planning problem, not a data-center problem. By his estimate, data centers use about 5% of US energy and might grow by about one percentage point a year.
On jobs, he says economists don't know yet. Some charts suggest an effect and others don't. Employment of 18-to-24-year-olds has slowed, but apparently similarly for people with and without degrees, and in fields that look exposed to AI and fields that don't. He also points to a broader lack of data. The labs give no meaningful usage figures, not even a daily active user number for ChatGPT, which he calls "crazy." So evidence comes from academics working through BLS surveys and from consultancies and agencies surveying 20,000 people. Politically, though, none of that matters to a student who can't find a job, whatever the cause.
There are also narrower fights. Evans is, by his own admission, deliberately unkind about illustrators of young-adult romance covers upset that such images can now be generated. He also mentions culture wars among novelists and ebook writers, "AI slop," and the figure that 30–40% of new podcasts are AI-generated. He compares all this to the social media backlash, compressed in time: some true, some partly true, some false. His example of the false category is the belief that "Facebook sells your data," which he calls untrue but firmly held, invoking Jonathan Swift's line about not being able to reason someone out of a position they didn't reason themselves into. He doesn't think Trump's proposed executive order on dangerous models, driven by concerns like cyber risk, is what drives the Main Street backlash.
Every technology gives new ways to ruin lives
Evans attributes his calm partly to being British and says someone from Iran might feel differently. His view is that every technology brings change we need to worry about, and that this is constant. He recalls a wave of late-1970s books panicking about databases, half of which was true. Putting all police and government records online really was different.
He doesn't dismiss the harms. On deepfake nudes, the dumb response is "haven't you heard of Photoshop?" A 15-year-old couldn't use Photoshop to make explicit images of every girl in their school and send them around in an afternoon, and now they can. Social media promised that the only gay kid in a village could find their tribe, but so could the only Nazi or pedophile. Connecting everyone connected all the bad people and our worst instincts, and he expects AI to produce its own versions of this.
His cautionary example is the UK Post Office scandal. Post offices are mostly franchises run by small business owners, often second-generation Indian immigrants running pharmacies. A Fujitsu-built point-of-sale system had bugs that showed cash shortfalls. The Post Office treated them as theft. Hundreds went to prison, there were suicides and bankruptcies, and people lost their homes, while Post Office and Fujitsu staff testified that the system had no bugs. Evans calls it essentially 1970s technology. His lesson is that every wave of technology can ruin lives, deliberately (he cites Chinese mass surveillance) or accidentally, and we need to be conscious of that without panicking.
Kids and careers
On parenting, Evans says it depends on timing. A child entering the job market in the next year or two faces a situation where everything is up in the air. In five years, things will probably have settled down in unpredictable ways. His son is in his early teens. Many questions predate ChatGPT: the collapse of gatekeepers, trusting TikTok influencers, where you get your understanding of world events. Quoting George Carlin's line about drivers (anyone slower is an idiot, anyone faster is a maniac), he says every parent thinks they're in the middle. He has no systematic plan and would settle for his son not breaking his Chromebook again.
He won't name jobs to steer toward or away from, noting his own path from equity analyst to industry to consulting. His only framework is to find the intersection of skills you have, work those skills make you good at, and things people will pay for, and to get at least two of the three.
The question not enough people ask
Asked what question is under-asked, Evans names two. First, whether model labs will have pricing power, which he thinks many simply assume. Second, the task-versus-job question and how it plays out. He ends his deck with global recorded music revenue, which follows a U shape. It roughly halved from 2000 to about 2015, then recovered to about 75% of the inflation-adjusted peak thanks to streaming. The first half answers "what if I don't have to pay $15 for a CD to get that track?" The second answers a different question: "what if $15 a month gets you all the music there is?" New technologies start by doing the old thing, more of it, in a new place, like putting Flickr on mobile or printing out emails. Then they enable things only possible with the new technology, and eventually they redefine the question. Spotify isn't an online music store. You often only know what the question was after someone has built the billion-dollar answer, and Spotify, Uber and Airbnb all looked crazy at first.
Rachitsky notes that coding, once thought the last profession to be automated, has been transformed first. Evans uses this to criticize efforts built on the US government's O*NET dataset that score each occupation as X% exposed to AI. He calls them "deluded horseshit" for two reasons. First, it repeats the expert-systems mistake: you can't decompose a senior law partner into "17% automatable" any more than you could build a cat detector from 700 hand-written steps. Second, exposure is inherently unpredictable. In 1997 people said newspapers would benefit from lower printing bills and taxis had nothing to do with the internet. He recently saw personal trainers listed as unaffected by AI, then imagines propping up a phone and having an AI design a routine and correct his form. He concedes that might be nonsense, but that's the point.
Impact also varies. Uber, in Evans's telling, demolished the taxi business in many cities while making it much bigger. Airbnb's effect on hotels, by the numbers, is marginal: it created a separate business and perhaps slowed hotel growth. His wife flies to Milwaukee next week, arrives at 8 p.m., wants room service and a gym at 6 a.m., and drives to a client at 7. There's no chance she'll stay in an Airbnb, and half of hotel demand is business travel. Someone once complained that Evans's answer to everything is "it depends." He agrees that it does.
What to do about it
Evans first qualifies the reassurance. As Keynes said, in the long run we're all dead, and averages hide individuals: nobody died in World War I "on average," but a 19-year-old in 1914 had perhaps a one-in-three chance of not coming back. For associates and people hoping to become associates, the future of the professional-services pyramid is very unclear.
His only advice is not to bury your head in the sand. Declaring that you hate AI gives you a feeling of moral superiority and lets you shout on Bluesky about how evil it is. He says he's happy for you, but it won't help. What helps is immersing yourself, understanding what you can do with the technology and how it changes things, and figuring out how to be a great hire. That might still not be enough. But if a law firm is cutting associate hiring from 100 to 50, announcing in the interview that you'll never use AI is probably the wrong approach. He compares it to how he and Rachitsky had to absorb mobile and the internet.
"I'm the lawyer looking at ChatGPT"
Asked how he uses AI, Evans says he struggles with the question because he's in the lawyer's position. What he'd want automated is precise information retrieval, which he says is what AI is currently worst at. He calls that an observation, not a criticism. He uses it for proofreading, images, and redecorating his apartment, where it worked "fantastically well": repaint this room, add this light and table and rug, change the rug's color. He quotes someone's line that AI is good at what computers are bad at and bad at what computers are good at. He acknowledges that his job, synthesizing large amounts of material into new ideas at a desk all day, is unusual.
Rachitsky mentions a Pete Holmes joke that we wanted AI to clean the streets and instead it wants to make art. Evans says the general desire is for AI to do the boring things, not the fun ones. He links this to his view that a chatbot is a blank screen with a jagged edge, and the fix is wrapping it in use cases. Part of it is AI disappearing into things. He dictates most of his writing as voice memos transcribed automatically by the iPhone's built-in Notes app. He wonders whether that's still AI or just voice recognition. There's probably an LLM in there, he says, but at some point it's just automation.
He also notes that consumer AI has no breakout apps yet. He thinks the main reason is marginal cost: you can't make an app free, reach 50 million users, and then find a business model.
Convergence, and pushing past what you already know
In the lightning round, Evans recommends Jerome K. Jerome's Three Men in a Boat, which he treats as having a story for every situation. He also recommends William Cronon's Nature's Metropolis, a history of Chicago that he says reads like a book about broadband: standardization, logistics, channel conflict, network neutrality, and the point at which it became cheaper to ship cattle from New York to Chicago for slaughter and back than to slaughter them in New York. He urges tech people to read books other than The Lord of the Rings. He recently watched The Seventh Seal and found it brilliant.
He owns 20 or 30 old phones, including an Ericsson from around 1998 and Japanese phones from 2001 with a color screen and a camera, which clients lined up to see and which still charge. Before the iPhone, he says, phone design varied wildly, like cars before wind tunnels, because everyone was differentiating around a tiny gray screen. People imagined some would carry pocket PCs, some keyboards, some folding devices, and nobody expected convergence on one form.
Asked for a motto, Evans offers "it depends," then "it'll probably be okay." Rachitsky notes the "probably." He closes by describing his own challenge: the moment you really understand something is the moment you have to move on. He thinks he spent too much of last year insisting that models still hallucinate, which they do if you push them a little. But that doesn't mean they aren't useful, and the task he sets himself is to keep asking different questions.
My most controversial opinion is that I think that AI is as big a deal as the internet or mobile and only as big a deal as the internet or mobile.
What's your just the coming job apocalypse?
Every time we have a new technology, it automates away a bunch of jobs and then that automation unlocks a bunch of new jobs and you don't know the new job because it doesn't exist yet. We've had that process over and over again.
Even just looking at the most advanced AI companies, Anthropic, OpenAI, everyone's increasing headcount.
You talk to these doomers on Twitter and they would act like every big company is going to buy ChatGPT tomorrow and then in two weeks' time they'll fire all their staff. These people are morons. You can't predict which things are going to be exposed. You can't look at a senior partner at a law firm and say, "Well, 17% of their work could be automated." This is horseshit.
I'm curious if you're following the anti-AI sentiment.
It's a big fuzzy mess. Yes, this will change a bunch of stuff and we'll need to worry about it, but that's kind of a constant. We've always had that.
What would be a couple things you recommend people do to be more successful in this future?
Don't stick your head in the sand and say, "I hate all of this stuff." That gives you a great feeling of moral superiority and you can go on Bluesky and shout at everybody about how evil AI is. Like, great, I'm happy for you. But that's not going to help. What helps is you diving into this and coming out understanding what you can do with today.
My guest is Benedict Evans. Benedict was a longtime partner at a16z as their in-house analyst and resident thinker. Before that, he was a longtime equity researcher. And for the past six years, he's been an independent analyst tracking the most important tech trends and sharing what he's learning. Most recently, as you'd expect, he's spending all his time on how AI is changing our lives. And in his words, AI is eating the world. In this conversation, we go deep on what we're still not pricing in on the impact that AI is going to have on our lives and our work, the rise of anti-AI sentiment, the impact on jobs, where in the value chain most of the value will accrue, and tons more. If you are worried about AI or just confused about where things are heading, this conversation will teach you a lot and also make you feel better.
Before we get into it, don't forget to check out lennyspass.com for a year free of some of the most amazing, hottest, most well-crafted AI products in the world, available exclusively to Lenny's Newsletter subscribers. With that, I bring you Benedict Evans.
Benedict, thank you so much for being here. Welcome to the podcast.
Thank you for inviting me.
You just put out this deck called AI Eats the World. I want to ask you kind of the flip side of this. We all know it's a big deal. Knowing that, what do you think people are still not fully pricing in when they think about the change that they're going to experience to their lives and their work?
An interesting way of thinking about it: I did a podcast last year with someone where I said, you know, my most controversial opinion is that I think that AI is as big a deal as the internet or mobile and only as big a deal as the internet or mobile, because clearly there's a bunch of people in tech who think no, this is more like the industrial revolution or something. And there are a whole bunch of people underneath saying, well, he thinks this is just as big as, does he not understand how big this is? And I'm like, smartphones were quite a big deal. The internet was quite a big deal. We wouldn't be doing this if it wasn't for the internet.
So there's like one layer of, but then if you dig into that, like if you're going to make the internet comparison, it's like we're in 1997. Like it's very exciting. Most stuff kind of doesn't work yet. Most of the stuff that people are going to do hasn't been built yet and it's not really clear how any of it's going to work when it does work.
And the people who have already got it, who have already taken whichever pill it is, I forget which, sort of imagine that everybody in the world is already there. And the truth is you've got this kind of very wide distribution. So there's people in tech who bought their cluster of Mac minis and, you know, don't use Google anymore. And then you look outside tech and, setting aside the idiots who think that this isn't real, you know, most people who are using this are using this every week or two maybe. So you've got that kind of spread of adoption and that spread of maturity of how well this works.
And then within that you can make sort of specific points about, well, how are the models going to work, and do the model labs have pricing power, and where's the value going to be, and, you know, has OpenAI won the whole thing, or, you know, has Anthropic got it this week? And so then you can kind of get into calling those races where again it's like being in 1997 and saying, well, is it going to be Excite or Yahoo? And the answer was no, generally.
So there's a sort of a fractal point here. There's the sort of super high level that this is going to change absolutely everything. I don't think it's particularly productive to say, well, is it 20% bigger than the internet or 100%? Those aren't productive conversations. But it's one of those fundamental changes, but then you don't know how any of it is going to work. In fact, I just published this. I do a presentation every six months and I just published one yesterday, and one of the comments was, "Benedict, this is 80 slides saying 'we don't know'," which is slightly facetious but also kind of true.
This episode is brought to you by our season's presenting sponsor, WorkOS. What do OpenAI, Anthropic, Cursor, Vercel, Replit, Sierra, Clay, and hundreds of other winning companies all have in common? They are all powered by WorkOS. If you're building a product for the enterprise, you've felt the pain of integrating single sign-on, SCIM, RBAC, audit logs, and other features required by large companies. WorkOS turns those deal blockers into drop-in APIs with a modern developer platform built specifically for B2B SaaS. Literally every startup that I'm an investor in that starts to expand upmarket ends up working with WorkOS. And that's because they are the best. Whether you are a seed-stage startup trying to land your first enterprise customer or a unicorn expanding globally, WorkOS is the fastest path to becoming enterprise ready and unblocking growth. It's essentially Stripe for enterprise features. Visit workos.com to get started or just hit up their Slack where they have actual engineers waiting to answer your questions. WorkOS allows you to build faster with delightful APIs, comprehensive docs, and a smooth developer experience. Go to workos.com to make your app enterprise ready today.
So, if we're in this 1997 timeline for AI, I know so much of your message is we don't know where it's going exactly yet. Do you have a sense of just the timeline to, okay, now things are going to be radically changing? Like, where are we in that cycle? You talk about all these different cycles we've been through. How far are we from just, wow, it's all different now?
Well, unquestionably we're already in that moment in software. And then there's a conversation about, well, what does agentic and AI software development, two separate things that merge together, mean for the future of the software industry? You know, there's one extreme which no one really believes, which is, you know, hey, you'll just vibe code your own Stripe, and no one actually believes that. But clearly there's a whole bunch of questions about what this means for the software industry and how much stuff you'll be able to do yourself or how much more software there will be, and that's one whole conversation.
But the other extreme is, you know, if you're in a law firm, this is all very interesting, but how exactly do we use this? And how do we work out how not to be the next story that we've submitted something with hallucinations in it? And how many associates are we going to hire next year? What does this mean for us?
One of the analogies I used in the presentation is imagine you're an accountant seeing the first software spreadsheets in the late '70s. This is mind-blowing. You know, you change the interest rate here and all the other numbers change and it does a week of work for you in like 30 seconds. And we can talk about what that meant for the accounting industry, but clearly if you're an accountant, this is obviously mind-blowing. But if you were a lawyer looking at that or a journalist looking at that, you'd think, well, that's very clever and my accountant should see this, but that's not what I do. I might use it for my time sheet next week if it didn't cost 10 or $15,000 to get the Apple II and the monitor and the printer to run it, which is what it cost if you adjust, but that's not what I do. And you need a word processor, which actually came very shortly afterwards.
And so that's sort of the moment that we're in. There's some people, like software developers, who are the accountants seeing VisiCalc, like, oh my god, this changes everything, like before VisiCalc and after VisiCalc, before Claude Code and after Claude Code. A lot of other people are picking it up, using it to varying degrees, but slightly puzzled. So there's a bunch of survey data that I put in the presentation that, even if you look at like 13 to 18 year olds or something, it's still kind of 15, 20% of people are daily active users and another 20% are weekly active users, and then the other 60% of the people in that demographic say they are not using this.
So there's a sort of very wide spread of who gets it, which I think also maps, and this is kind of almost a separate point, to the sort of jagged frontier question of where does this work, where does it not work? Can you tell where it's going to work? Is it intuitive to know where it would work? Can you tell after it worked? Can you work out for yourself what you would do with this? And all of those intersect. If you're a software developer, a lot of other people are having a moment, or they're not, or again we're in that kind of 1997 moment of, okay, what is this?
Along those lines, something you've been writing a bit about is this unexpected investment in professional services, consulting services, forward deployed engineers. All the AI labs, at least the two big ones, OpenAI, Anthropic, are investing in, buying massive consultancies and PE firms. Talk about just what's happening there, why that's happening.
Well, it's funny, I was kind of groping for a joke last night when I wrote my newsletter and couldn't quite get it to land, but something like, you know the joke that a machine learning scientist is a statistician who lives in San Francisco? And there's something in there of like a forward deployed engineer is an Accenture outsourced software developer who lives in San Francisco, or works in San Francisco.
I mean, you know, joking apart, if you have any experience of professional services, companies do not have lots of people sitting around waiting to build a big new project or do a big new piece of analysis or build a big new piece of technology or a new product, or work out how they're going to redesign their stores, or work out where the stores should be, or try and work out why the churn is too high. All of those kinds of questions are reasons why you hire Bain, BCG, McKinsey on one side, or Accenture, Infosys, whoever on the other, or you hire a branding agency, or you hire a firm of architects or whatever. And it's always like, well, we could hire some architects, but why on earth would we want to have 15 architects on staff when we just go and hire an architecture firm? We just go and hire an ad agency.
And so you're supposed to completely reimagine all of the internal workflows of your company and work out which of them could be automated really quickly with AI. That's a project. That's a project that needs like five or 10 people to sit down and spend a month or two working it out. And then actually doing it is another project. Okay, so we need to plug these three vertical systems into these two horizontal systems and build in a bunch of new workflows and train people to do that. Well, guess what? Who's going to do that? Because you don't have a bunch of people sitting around not doing anything. So, on the one side, this is part of the model of some PE firms, which is that they provide support to their portfolio companies to do stuff. And on the other side, that's why you hire, depending on what you're trying to do, you hire Bain or you hire Accenture or you hire publicists to help you work that out.
What's really just funny about this trend is you would think consultants were going to be gone. No, we don't need all these people anymore. AI is going to do their work. Instead, the most cutting edge AI labs are the ones most investing in these folks. I think it's pretty surprising.
Well, one of the strands in my presentation, so I split the presentation into three sections. There's a section on capital, which is basically where is all this capex going and are the model labs going to have differentiation? And then there's a section on deployment, which is basically what does it mean for the software industry? And then the third section is how does this change stuff?
And one of the sort of strands I tried to pull together in the section on change is what's the hard part of the job? Is the hard part of the job writing the code line by line? Is the hard part of the job like giving you the school or making the PowerPoint, or is the hard part of the job something else? Is it the task or the job?
And, you know, pulling that apart, sometimes the task is the job. The classic example is an elevator attendant. I live in a building that has an attended elevator. We have a manual elevator. There's no button. There's a lever in the door and it drives you to your floor. It's a vertical streetcar. It's like one of those trams in San Francisco. They drive you to your floor. And then those all got automated after the '50s and now you get in and you press a button, and pressing the button is a job. So there were some things where the job was a task and the task got automated.
What happens much more, and this is why people talked about the Jevons paradox, is price elasticity, because the Jevons paradox is just applied price elasticity. If you make it cheaper to do something, what happens? Do you do the same for less money, or do you do more for the same amount of money, or do you do more for more money because you've got new ROI?
And if you look at something like the history of accounting or indeed professional services, you know, this is a joke I made on Twitter back when it was Twitter: young people won't believe this, but before Excel, junior investment bankers worked really long hours, and now thanks to Excel, Goldman's associates all work at lunchtime on Fridays. It's like, well, why is that not what happened? You could make the same point in software development. You know, before IDEs and libraries and operating systems, developers had to write all the code. Now, if you write an iPhone app, 90% of the code is written for you by Apple. Like, Apple wrote the modem driver and the graphics drivers and, you know, the file system. You don't need to write any of that. So, we've got like a tenth as many engineers now. Well, no.
And so then you kind of have to look at an industry and work out, well, which is it? And what is the hard part? One of the analogies that occurred to me here is to look at the history of e-commerce, which is that what Amazon does is it gets you the SKU. If you
know what the SKU is, if you know what SKU you want, you want that microphone stand. You know this part number, you can go to Amazon and get it. If you don't know what microphone to get, probably shouldn't start on Amazon. Multiply that by many, many, many product categories. And so what Amazon does is get you the SKU, but knowing what SKU you want is another job. You know, Claude Code can write you the code, but what code do you want? It can make you the features, sure, but what features do you want? Who's your customer? What's the right product for that customer? How you going to take it to market?
And long way of answering your question, why do you hire McKinsey? Are you hiring them to get a 75-slide deck? Well, narrowly Claude Code will make a really, really crappy version of that. And you'll get all these kind of AI grifters on LinkedIn and Twitter and so on saying, "Hey, I made a McKinsey deck with Claude." And you look at it and you think, "Yeah, that's a bunch of dog crap. That's not what you'd get from McKinsey." But even if it was, that's not what you paid them for. What you actually pay Bain to do is to go and walk all over your enterprise, your company, and work out, yes, but why is it that you didn't do that? And how do the politics of this work? And what do you actually need to do? And let's go and talk to your customers and work out what they actually think as opposed to what's on the first page of Google. It's all the other stuff, and the PowerPoint is just the task, but that's not what you hired them for. The same with Amazon versus the retailer, the same with software development. So you've got that kind of split.
The other analogy that occurred to me here was looking at the sort of class of industry that got steamed by the internet because they had those two things and you could split the parts. So you had the physical manufacturing or physical distribution and then you had the other thing, what was the actual thing. Classic examples would be newspapers and recorded music. So record companies do not think of themselves as being in the business of manufacturing small pieces of plastic, but that was what they actually did, and when that went away they were screwed. Same thing for newspapers. Newspapers did not think of themselves as manufacturing and trucking companies. When you decouple that, then that becomes a problem. But often you kind of can't decouple that, or that wasn't really the problem, or you make that thing cheap and then all this other stuff happens as well. And so all of this is just vastly more complicated than saying, well, hey, we're just going to automate the accountants or we're going to automate the consultants. I mean, there's two charts in the presentation of the number of people employed as accountants, which went up right the way through the 20th century and has gone up again since the beginning of the 21st century. So you have adding machines and punch cards and mainframes and databases and ERP and cloud, with spreadsheets and PCs, and the number of accountants keeps going up. And so why is that? Well, it must be more complicated than automation.
Even just looking at the most advanced AI companies, Anthropic, OpenAI, I just had Dan Shipper from Every on the podcast. Everyone's just increasing headcount. Like the companies you would think would be least likely to add humans are adding many, many humans. And to your point, it's really complicated. What's your just kind of gist on the job, the coming job apocalypse? You know, like Dario's talking about all the entry-level people, there are no more jobs. Just like
Yeah. I mean, there's a narrow point here, which is that I don't like argument from authority. And I don't think the fact that you run an AI lab suddenly gives you — or rather, if you're going to use argument from authority, then it should be relevant to the field. So like, I'm interested in Dario's opinions on where models are going to go in the next 6 to 12 months. I'm not particularly interested in opinions on theories of labor and market value and competitive, comparative advantage. Like, yeah, maybe he had a course on that at university. So did I. So I think one needs to be a little bit cautious on, like, well, Dario says. And that's setting aside the cynical view that he's just doing that to pump stock, which I don't believe at all. So this kind of comes back to my point about platform shifts. Every time we have a new technology, it automates away a bunch of jobs, and then that automation, whether it's price elasticity and the enablement of the fact that they became automated, unlocks a bunch of new jobs.
And so you go back to 1800, like 90% of us were peasants and our major concern was, are the crops going to fail, because then we'll all go hungry or worse. And so ever since then we've been automating jobs and creating new jobs, and you can always see the job that's going to go away, and you don't know the new job because it doesn't exist yet, and it's something that sounds dumb anyway, like railway engineer. What's a railway? Why would that be a thing? Who would care? Who would want to go that fast? And so we've had that process over and over again. This is what any first-year economics student would tell you. We've had this process over and over again since 1800, and each time you go through it you get a bunch of frictional pain and dislocation, and a bunch of people lose their jobs, and a bunch of towns get hollowed out, and it all sucks. But when you come through on the other side, we're all richer and we're not worried about the crops failing anymore. And this is the process of the last 200 years.
So then the question is, is there some a priori reason why this would be different to those? Because the internet removed a bunch of jobs, PCs removed a bunch of jobs. There aren't many people working as typesetters anymore, or telephone operators or typists. The internet removed a bunch of jobs, and generally the jobs that go away are crap jobs, seen retrospectively, and the new jobs are better because GDP keeps going up. So is AI different?
And so then there's a couple of answers to this. One theory is, well, this is going to be way quicker, and certainly the adoption of AI is quicker than previous technologies, but this is kind of because you're standing on the shoulders of giants. So you don't need to wait for everyone to buy a piece of expensive hardware, to buy a phone or a PC, or wait for the telco to deploy broadband. It's already there. So of course ChatGPT can get 900 million users, because there's already 900 million people on the internet. Like when Marc Andreessen launched Netscape in, what was it, '93, '94, there were like 50 to 100 million PCs on Earth. So no, you didn't have 900 million users then. But the point is, then he didn't need to wait for phone networks or microchips, and before that you didn't need to wait for electricity, and you didn't need to wait for mass production. So you're always kind of standing on the shoulders of giants. There's always a compounding effect. So yeah, this is faster, but the internet was faster too.
I think the other answer to this, and this kind of comes back to the professional services point, is you talk to these doomers on Twitter and they would act like every big company is going to buy ChatGPT tomorrow and then in two weeks' time they'll fire all their staff. And these people are morons, and this is one of many reasons why doomers were morons, but a complete failure to understand the way the world works, and that was the starting point why they then didn't understand anything else. Typical big company, enterprise software sales cycle — you'll know this better than me — enterprise software sales cycle is like 18 months if you're lucky. You know, this is always the problem. The enterprise sales cycle is shorter than the venture-backed startup funding cycle — longer, rather, longer. It takes you longer to get an enterprise deal than it takes you to go between rounds, and this was always the problem, particularly for sectors like aerospace or healthcare or something. So no, people aren't just going to tear out SAP and replace it with XYZ. Maybe in three, five, 10 years, yes, that whole estate will look radically different and all those jobs will have changed, but it will take three, four, five, 10 years, and it will take time sector by sector, and it will take time for people to work out you could do that thing with this.
One of the companies I always remember that we looked at when I was at Andreessen Horowitz is a company called Frame.io, which is video editing, video collaboration. And there's nothing there that you couldn't have done at least 5 years earlier and maybe 10 years earlier. And actually, that's kind of a bad example, because that relies on a bunch of cutting-edge web technologies. But if you go out and pick 10 random SaaS companies that were started the day before ChatGPT launched, how many of them could have been founded at any point in the previous 15 years? The delay was somebody realizing, oh, that problem exists inside that industry, and oh, this is the way that we would solve it. It didn't all happen the day after Google Docs. It took like 10, 15, 20 years for people to invent all that stuff and work out that you could do that with this. So all of that is a way of saying, well, yes, it is going to be quick, but actually, no, it will kind of take a while for people to work out how to completely change how their business works.
Your view is so comforting because it's basically like, okay, this is a huge deal, but we've been through many transformations before and it's going to be okay.
Well, I have a slide towards the end of the presentation, and I think the title is something like, this is going to be completely different from everything else, just like everything else. And then the next slide is an IBM ad from the 50s, which has got this sea of white men in white shirts and ties, all holding up slide rules. And the slogan on the title of the ad — it's an IBM ad — it says an IBM electronic calculator. This is before it was called a computer. It's an electronic calculator, it's the size of a fridge, is like having 150 extra engineers. Like, how many people listening to this, their company's slogan is basically, we'll give you 150 extra engineers? I mean, isn't that the whole pitch of Claude Code? 150 extra engineers, for free, or not free, that's like a lot of money. And yeah, that's what it gave you. And so yes, we keep going through this over and over and over again. Just to kind of make that tangible, I mean, obviously we couldn't be doing this without the internet. So there's a slide in my presentation, which we could maybe talk about, but it's a slide or chart showing how many products are stocked in supermarkets in America since the 50s. And the point of the slide is to say that barcodes allowed supermarkets to stock way more stuff because they could keep track of it.
But making that chart, I had to know there was a thing called the Food Marketing Institute. And I had to have found out that they published a number for how many SKUs there were in supermarkets every year. And then I had to realize they'd been around since the 50s, and if I dug long enough, I might be able to make a whole time series and I could make a whole chart. Now imagine doing that in 1994. First of all, you would have no idea that exists. You'd really need to go and find a library where they publish that number, and then the numbers in that report. You'd have no idea. Then you need to find a library that had them. So you're going to spend like three days on the phone and spend like $50 on long-distance phone calls to find a library that has these. Or maybe you call the Food Marketing Institute and they say, "Yeah, sure, we'll sell them to you for $500 each." So then you're going to get on a — maybe you live in New York, or there's somewhere that has this — and two weeks later you've got the chart and you look at it. And then the other side of this is the life of an analyst: you spend all day making a chart and you look at it and go, that's not very interesting. So you spend two weeks to make the chart and then you look at it and go, yeah, I'm not going to use that. And for me this was like two hours in Google. And so we forget how big a deal the internet was. That's a long way of saying it, but we forget we've had these absolutely enormous changes, and then we don't see it because it's like, that's the world, the world has always been.
What's different potentially this time, even though your quote is, it's different, everything's going to change, just like last time? The big difference obviously is AGI might emerge, and superintelligence, where that could do the work of humans, can do a lot of this stuff for us, can actually replace jobs. Just, thoughts on that element of this transformation we're going through?
I don't know. This is one of the ways I've struggled to write about AI: certainly in 2023, early '24, all the questions were questions you could have asked in December 2022. The questions didn't really change and the strategies didn't really change, and I think the AGI question is kind of the same. I mean, the observation one can make: we have no theory of what human intelligence is, we have no theory of why these models work so well, we have no theory of how much better they will get, so we're all just kind of vibes forecasting as to what will happen. And then you can have the 2 a.m. doped-out philosophy students talking about, hey man, is this consciousness? Maybe we aren't conscious either, we just think we are. Like, yeah, great, thank you.
I think the one thing one can observe today is — so we have no idea, we don't know, we can guess, but we don't really know where this is going to end up — what I think you can say today is that there's a lot of kind of redefinition of terms. So I think a quote I used in my presentation late last year was an AI scientist called Larry Tesler, who said AI is whatever machines can't do yet, because once machines can do it, people say, "Well, that's just software."
And so certainly, I did do a poll on social media every now and then asking, "Is machine learning still AI?" Because I've certainly heard people say, "Oh, that's not AI, that's just image recognition. That's not AI, that's just sentiment analysis." So AI, it's a bit like the word technology. It's like if it's new, then it's technology. But in the 60s, jetliners were technology. Now a jetliner isn't tech. And so there's a sort of sense of AI as a moving target, is whatever just started working. And I think the point here is, now clearly you can see people redefining AGI to mean the stuff that works now. So is AGI — what's the definition now? It's like it can do a certain percentage of economically valuable work. Well, that's a very different thing to it has a soul and it's alive. Because a database can do that. An IBM mainframe in 1975 could do a meaningful percentage of economically valuable work that was previously done by people, and it turned out there was a whole bunch of other stuff that it couldn't do, that we didn't do then, we didn't know existed. So there's a lot of kind of creative redefinition here.
Superintelligence — I'm not sure, is superintelligence more than AGI or less than AGI? Because last year I thought superintelligence was really good but not as good, not actual AGI, and now it's like, oh no, we've already got AGI, but superintelligence, that's really hard. It's like all these terms are like what
What even... it's funny, I was having an argument on Hacker News this morning, which is never a good use of time, but you remember the argument of, you know, people would argue about whether crypto is blockchain or whether blockchain is crypto. There isn't a right answer to that. It's important to understand what you mean when you say that, but there isn't like a correct answer to this. Are we going to get to something that has human-level intelligence? We don't know. I don't think we have any way of answering that question. Maybe, maybe not; you can make arguments either way. Meantime, it does mean in the meanwhile we've got this thing that's clearly kind of a completely transformative technology. And maybe the serious point here is you don't have to believe... even if the models stopped getting better tomorrow, if this is it and we hit a brick wall tomorrow, this is an incredibly useful technology that's going to change the world and get rolled out over the next 10 years. So you don't have to believe in any of that stuff to believe that this is a giant deal.
Something that's definitely changed: I had your former boss Marc Andreessen on the podcast, and we didn't actually talk about this during the conversation. He brought it up before we started recording and I never got to it. He had this insight that the opportunity set for companies now is so much larger. We used to have no trillion-dollar companies. Now we're going to have dozens of trillion-dollar companies. Just the size companies can grow to is going up so much, and valuations also go up along with that. And his point is just people haven't really grokked just how large companies can get now. Like everyone's hitting $100 million ARR in like five months, six months. Just thoughts on that.
Yeah, I mean, this was his whole software is eating the world thesis from, you know, 15 years ago, whenever it was. Yeah, the TAM gets progressively bigger because you can address larger and larger parts of the economy. And so if you think about the classic platform shift framing, I think peak mainframe install base was something like 70, 80,000 units. I mean, slightly fuzzy term: what exactly is a mainframe, and at what point does it become two mainframes as one? But something like that order of magnitude. And then when the internet kicks off there are, as I said, 50 to 100 million PCs on earth. Maybe today there are something over a billion, one to one and a half billion, but obviously a lot of those are corporate. It's like 7, 800 million consumer PCs in the world. There's about five and a half, six billion mobile smartphones in the world, which is why you can have 900 million weekly active users on ChatGPT.
And so there was this narrative like five years ago: well, we've run out of people, so the next thing can't be an order of magnitude bigger. Which was true up to a point, but that was the wrong model, because clearly what's happening now is you're moving in another direction. You're just branching out and automating big new ways of the economy.
Now, back to your job point, you could argue, well, we're just going to replace all the people with AI and all the money will go to Sam Altman, and Marc can buy himself another Gulfstream, add to the fleet. I think the other answer is, it's back to the lump of labor fallacy and the last 200 years: each of these technologies removes a bunch of jobs, creates a bunch of new jobs, creates a bunch of new value, unlocks prosperity for all of us. And that's painful as you go through it, but it always creates more value.
And so here you could certainly make an analogy. The useful analogy to the electricity industry is just saying how electricity became part of absolutely everything, and software has been kind of slowly working its way out. The analogy here would be electricity in factories, and then electricity sort of slowly spreads out. And so that would be the point again, that it slowly spreads out to do more and more things, and so more and more value and a bigger and bigger contribution to the economy.
It also, of course, disappears inside things. And the other side, this is the point of my capital section in the presentation: there's this quote from Sam Altman where he said we're going to be selling AI intelligence on a meter, like water or electricity. And you look at this and think, my dear sweet child, do you need me to explain the margin structure of the utility industry to you? Because guess what, when you watch television, the TV company isn't paying a percentage of your monthly bill to the electricity company. When you wash your clothes, Bosch isn't paying a percentage of the price of the washing machine.
And clearly the much more specific tactical question at the moment is: do we even end up with three giant models, or does it become hundreds of models and open models and local models and so on? And even if we do end up with, say, pick a number, three to six to 10 giant foundation models that cost hundreds of billions of dollars a year, fine. Do they get all the value from that?
Now, I started my career as a telecoms analyst, so I still pay attention to it a bit. The global mobile industry has revenue of about a trillion dollars a year, maybe a bit more now. And it spends about $200 billion a year on capex every year. Total telecoms is about 300; mobile is about 200. It's about 15 to 20% of revenue every year. And if you look at a chart of mobile data consumption, it's an exponential curve, a perfect curve going straight up. And the number now, I think, is about 1,500 to 2,000 times what it was in 2010 globally. And the stocks have gone nowhere in 25 years, because it's an ex-growth, low-margin commodity utility, where they're selling this objectively amazing piece of global technology infrastructure that has enormous complexity and enormous sophistication, but all the cool stuff is made by you. It's made by the people listening to this podcast. It's made by somebody else. This was that pivotal moment where the telcos thought that they would do all the stuff that you did on your iPhone. And not only do they not do it, but Apple doesn't do it either. It's all further up the stack.
And so this is the elemental question right now around foundation models: does the model do the whole thing? Do you just go to the chatbot and get the chatbot to do the whole thing? Can the model companies keep building these Claude for X, Claude for Y things, which to me look very much like what you see if you hit File New in Excel? It's like the templates, but all of those are actually billion-dollar companies as well. And if not, does it all have to be apps, quote unquote, whatever app means? And if it all has to be apps, who builds those? Well, they can't all get built by the model labs, just as they didn't all get built by Microsoft. And so if they're all built by other companies, do the foundation models have leverage up the stack the way Windows did? Or is this more like AWS, where if you're, I don't know, an engineering company or a law firm buying a piece of software, you don't care which cloud it runs on, and you don't have to standardize on AWS because that's where all the software is, and the developers all standardize on AWS because all the customers use AWS. That's not how it works. That's how Windows works, but that's not how AWS works. And so it does sort of seem to me that if the chatbot isn't the UX and it needs to be apps, and the model companies aren't going to build that, and the models themselves are basically commodities, at least as you can see them as users, then why would the model companies have pricing power, and wouldn't all the value be further up the stack? Haven't you basically got three to six companies selling a commodity at marginal cost? Now obviously the SemiAnalysis guys are like, no, no, no, there's going to be infinite pricing power forever. I'm sorry, I'm exaggerating. But I think it's really important to draw a distinction between where we are now, where you have radical price equilibrium, and you've got these... what's the guy, the OpenClaw guy, spent $1.5 million on tokens last month. But that's like somebody getting a 50 grand mobile data bill in 2010. That's temporary. What is the steady-state equilibrium point, where the lines on the chart kind of get lined up and we don't have this weird crazy stuff going on? And then will you have pricing power, or have you got three or four or five companies all selling the same thing? And so then you should have lower pricing and lower margins, and the value goes up the stack.
I am so excited to tell you about this season's supporting sponsor, Vanta. Vanta helps over 15,000 companies like Cursor, Ramp, Duolingo, Snowflake, and Atlassian earn and prove trust with their customers. Teams are building and shipping products faster than ever thanks to AI. But as a result, the amount of risk being introduced into your product and your business is higher than it's ever been. Every security leader that I talk to is feeling the increasing weight of protecting their organization, their business, and not to mention their customer data. Because things are moving so fast, they are constantly reacting, having to guess at priorities, and having to make do with outdated solutions. Vanta automates compliance and risk management with over 35 security and privacy frameworks, including SOC 2, ISO 27001, and HIPAA. This helps companies get compliant fast and stay compliant. More than ever before, trust has the power to make or break your business. Learn more at vanta.com/lenny. And as a listener of this podcast, you get $1,000 off Vanta. That's vanta.com/lenny.
A really interesting takeaway here is that your sense is, over time, the foundational model companies, Anthropic, OpenAI, others, their margins will get squeezed. They will not be as successful as they are today. And the bigger opportunity is in the application layer, the people building on the models, the wrappers.
Yeah. I mean, this is a very sort of deterministic thesis, which is: crucially, what I said is the models don't seem to have network effects. So there doesn't seem to be a winner-takes-all effect where one of these will run away ahead of the others. So you should have competition indefinitely. You have competition indefinitely. You don't have really radical differentiation in what the product is. Then why would you have pricing power? And meanwhile, if you need to have thousands of applications that are all different, built by different people, those can't all be built by the model people. So it should end up looking more like cloud than it looks like Windows.
Now, that may be completely wrong. One of the points I make in the presentation is, imagine having this conversation about the internet in 1997. What would you have got right? Or indeed having it about mobile in 2000. Most of you would have missed almost all of it. You certainly would not have said that a has-been PC company from Cupertino would win the whole thing; no one would have said that. And a search company with a weird logo? Search, what's that got to do with mobile? No, forget it, you're an idiot. So we should presume we don't know, but there are the basic building blocks of, well, why would they have pricing power?
I don't know. When I was a baby analyst in like '99, I went to see a dot-com company in the UK that was trying to do online selling of computer cards, components, online. They had this whole model and this whole story and the brand and the whole thing. And we went up to see them, and we're on the train back from Birmingham, and this sort of senior banker called David Tate, we're all sitting talking about it, and Tatey says, "It's a low-margin reseller, one-time sales. You can say dot-com all you like. It's a low-margin reseller." And I think that's the crux of this: they're undifferentiated commodity infrastructure providers. There's a lot of science to it, but there's a lot of science in mobile. I mean, what do you pay for a flat panel screen? There are Nobel prizes in flat panel screens. They're still a low-margin commodity. I look forward to being proven wrong, but hey, that's what it looks like now.
This is great. So I know you're not an investor. I know you didn't actually do investing at a16z, even though you were an a16z partner.
Partner. Just sit around and pontificate partner. Would you... are there companies you would invest in? If there are a couple companies you'd invest in now, is there something on that list, or categories even?
I mentioned briefly that I was an analyst. I was a sell-side equity analyst. I was not a very good sell-side equity analyst, partly because I was not interested in talking to clients, partly because I was not interested in share prices, which would seem to be a disqualification to be an equity analyst. And there's a huge difference between being right and being early, and there's a huge difference between the right company and the right price. Now, deterministically, you can look across the market and say, well, it's like the bell curve IQ meme, and the guy with 50 and the guy with 200 are both saying, "Jeff Bezos, smart guy, I buy stock."
And you can certainly overthink all of this, and you can look at Google, Apple, Facebook, Amazon and say it's hard to see a problem for them really with all of this. You can certainly see questions for all of them, and one of them may drop the ball, but it's worth remembering what happened in mobile.
The internet was this big, obvious platform shift. The funny thing about mobile is that some companies missed it completely, and for some of them, it really didn't change anything. For Google, it didn't change anything. For Meta, this was great. This is a way better way to do social than on PC, because you've got a camera and notifications and it's on your phone all the time with you. Amazon, what does this change? It doesn't change everything. I'm massively oversimplifying here, but the point is, meanwhile, Yahoo Mail fails to make the jump. There are companies that were already kind of dying that failed to make the jump. Maybe eBay; you can argue about individual names. The point is that we went through that shift and it didn't change anything for half the internet industry. And so I think you could propose a little bit of that here. What Steven Sinofsky at a16z, who used to run Windows, would always say is, incumbents always try and make the new thing a feature, and sometimes they're right. Sometimes it's a feature.
Actually, along those lines, something I wanted to get your take on. There's this thread that's been happening across a bunch of guests, which is around distribution becoming a bigger and bigger moat, because as software is easier to build, everyone's launching products. Everyone's trying to compete for attention. It's getting harder and harder. It's always been hard to get people's attention, but the noise in the market is just going up like crazy. And to me, that tells me
distribution is becoming a more and more valuable skill and asset. And it also tells me incumbents are going to be a lot more successful because they already have distribution versus a startup that's trying to break through.
Yeah. I mean, there's like a version of, you know, the Drake meme of like he says, "I don't like that. I do like this." It's like, you know, I don't like seeing GPT wrappers. I do like harnesses. [laughter]
So, yeah, I did spend some time talking about this in the presentation I did at the end of last year that if the product is a commodity, then the distribution is what matters. And you know I wrote a thing about OpenAI earlier this year, like how do they compete? Well, there's an obvious comparison here a lot of people made with web browsers.
Fundamentally, the web browser, and there's a distinction here I think between the web browser as product and the web browser rendering engine. The rendering engine can be better or worse, but the browser product is just like a really thin wrapper for a rendering engine. Like there's an input box and an output box and like what else? And which is like, what's the last innovation in browser design? Like tab browsing, which is 20 years ago, 25 years ago, because every now and then somebody tries to innovate in browser design and it never works because you found the platonic ideal. It's like trying to innovate in smartphone design, like you know, it's a glass rectangle, like there's nothing you can do there.
And so what happened of course is that Microsoft uses distribution to break in. Then of course what also happens, setting aside the lawsuit, is that it turns out that winning browsers doesn't matter anyway because the value is further upstack, and so Microsoft wins browsers for like five, six years and it doesn't matter, it doesn't get them anything.
And so clearly what's happening now is that Google is using distribution to drive Gemini, and like what's the difference between Gemini and ChatGPT? And like if you're using this stuff all day then you know, but like for a normal person there's no difference. And the same thing with Meta. Like you look at survey data on which people use, even before the new news thing, like the Llama thing, Meta was up there between ChatGPT and Gemini, which, if you're in tech, people have completely written it off. But it was like they'd sprayed it on every surface and it wasn't that bad. It was fine.
So distribution of an adequate product when the field is basically commodity, distribution and brand become a big deal. You could see that in the OpenAI strategy late last year, was, you know, people called it everything everywhere. And so they were just kind of trying everything to kind of work out how they would get that, like how can we get a flywheel? How can we get distribution? How can we get something that sticks? How can we get something that people use before Google and Meta and Amazon spray it everywhere and get everybody using that one? And then you've got like this inertia and the power of the default, and like why would you switch?
Obviously Apple is kind of the last penny to drop here. There was this sort of slightly weird OpenAI deal and now there's an even weirder story that OpenAI wants to sue Apple. Good luck with that.
The funny thing about the Apple deal thing, just not to go off on a tangent, but like if you go back and watch the WWDC from 2024, like the whole second half of it is Apple Intelligence. That was like the most compelling vision of a personal AI assistant. It's still the most compelling vision I've seen. They then couldn't ship it, but then neither was anybody else. And you watch it again and you're like, "Okay, so you want tool-using agentic on-device AI with no prompt injection and no hallucinations and a completely standardized API system across 10,000 apps with intents that all work perfectly, and like, well, that sounds good to me, but I'm not surprised they couldn't ship it, but nobody else has shipped that." But like that vision was great.
You know, I really want to see what happens at WWDC in a month. Like do they actually ship that now powered by Gemini? But that's also another point, is like, okay, there's going to be the Android intelligence, whatever we call it, Gemini intelligence on Android, and then there's going to be Apple Intelligence on iOS, which is powered by Gemini, but it's not going to be the same set of products. The model is just like the dumb thing underneath, a funny way of putting it, the dumb thing underneath that powers the feature. The model is the commodity that powers different decisions about what the feature should be and different distribution. And in that situation, of course, Apple's got like a billion devices that can run this on edge, and Google has a wonderful marketing slogan, "coming soon to our most powerful devices," meaning it won't work on most Androids. So again, distribution questions.
Interesting. Google I/O's next week, so we'll see what they launched.
Oh no, they launched a... Android just shows how...
Was it today? Yeah.
Well no, they launched it last week. I mean, which is, it just illustrates how much we stopped paying attention to Android and iPhone. Like Google had a whole big thing last week. They're replacing Chromebooks with Google Books, and they've got a new Android intelligence powered by Gemini that will roll out to like the five people who bought a Pixel phone.
You don't work for Google.
Yeah. I want to go in a slightly different direction. Something that I'm curious if you're following is just the anti-AI sentiment that feels like is growing. Feels like if you've seen these surveys, AI is like less popular than ICE. People are trying to stop data centers from being built. I think Eric Schmidt just did a commencement speech and people were booing him every time he mentioned AI. Just like, what do you think is going on? Where do you think this goes over time?
It's interesting and it's a big sort of fuzzy mess of different stuff. I think there is like tangible, like "my electricity bill went up," which applies actually in a very small number of places objectively, but it did, and this is a question.
The water thing is weird because it's just like completely fake, and I should explain what I mean here. Data centers use water for cooling, it's mostly closed loop, but the number of data centers relative to the total amount of water use in the USA is tiny. I actually went and dug into this, and the Livermore lab did a study at the end of 2024 where they estimated US data center water consumption, and it came out at about 0.017% of US water consumption.
Now obviously if you live in a small town and you've got one well and they capped the well and gave all the water to the data center, then you're really pissed off, but that's a planning problem. That's not a data center problem. You know, in generality, yes, data centers are what, like 5% of US energy and might grow at 1% a year for the next five years, one percentage point a year, but the water stuff is just nonsense.
And then you get into more tangible, like, well, what is happening with this? Is it taking jobs away? Where you can watch a bunch of three-hour podcasts of economists talking to each other, and the main answer is we really don't know yet. There's a bunch of charts that kind of say yes and a bunch of charts that kind of say no. And clearly there's a slowdown in employment of, you know, 18 to 24 year olds, but that seems to be the same for people who do and don't have degrees, and the same for people in fields that look exposed to AI and fields that don't look exposed to AI. So there's a lot of econometric argument about this.
And I mean, there's a broader point here, in fact, which is a different point here, that we have very little data on what's going on in AI from anyone. The model labs don't tell us anything, like they don't give us any meaningful usage information. They give us these weird studies of like how many people used this for this and that. They don't give us a daily active use number. We do not have a daily active user number for ChatGPT. This is crazy. And all the data comes from academic economists trying to back stuff out of BLS surveys, or consultancies and marketing agencies spending a whole bunch of money to survey 20,000 people and saying, "What are you doing with this stuff?" Like we don't have good data on what's going on and how many people are really using this.
But to the employment question, hence there's a lot of people looking through all the stuff that the US census collects and trying to work out, well, where can we see this? Can we see productivity? Like what could we see? And the answer right now, I think, is there's no clear consensus that we're seeing an impact on jobs. But of course politically that doesn't matter, like if you're a student and you can't get a job, and that clearly is an issue, whether it's because of AI or whether it's because of Trump and tariffs is a different question.
Then you get kind of niche things, like, you know, people who draw book covers for young adult romance novels are very upset that now you can get a picture of a naked woman on the back of a dragon flying over a volcano without paying them. I'm sorry, I'm being deliberately unkind, but there's a little... and you know, people, particularly novelists, people who write ebooks, there's a huge culture war over whether it's okay to use AI. There's this whole sort of AI slop question, and you know, you saw the number that like 30, 40% of new podcasts are generated by AI. So there's a big fuzzy mass of questions.
Some of this I think is a little bit like the backlash we had around social, but much more compressed. And like social, some of the backlash around social was true and some of it was sort of true and some of it wasn't, you know, always exemplified in the whole "Facebook sells your data" thing, which is just a) not true and b) the people who believe it are absolutely adamant that of course it's true and you're obviously a lunatic for suggesting otherwise. You know, it's like the line from Jonathan Swift that you can't reason somebody out of an idea they weren't reasoned into.
So it was a long way to answer your question, but you got this kind of wide spread of ideas, just as you kind of did with social. There's like 20 different things, some of which are really real and some of which are really not real, and a lot of which are kind of a fuzzy mess in the middle.
All of which means that meanwhile you've got Trump saying he wants a new executive order on dangerous models, which I actually don't think is the thing that drives the backlash. You know, worrying about missile cyber, I don't feel like that's a Main Street America conversation, but that's the thing that got Trump interested in this stuff.
Again, let me go kind of in a tangential direction. Something that I like to ask folks that have kids that come on the podcast, especially people that are thinking so deeply about where things are going. Knowing what you know about just where the world is heading, what AI is going to do to the future, how are you changing the way you raise your kids? Just what are you teaching them differently that might help them in the future?
I don't know. I think there's a curve here in that if you've got kids who are going onto the job market in the next year or two, then everything is up in the air and no one knows how this is going to work. If you've got kids who are going onto the job market in like five years, then who knows? But stuff will have settled down a lot by then, in probably unpredictable ways. So I could be a lot more worried if I had a 21-year-old. You know, I don't. I've got a kid in his sort of early teens. So those questions vary.
Then you've got a lot of the questions that were the same before ChatGPT around, you know, the collapse of gatekeepers, the, you know, should you really believe what that influencer on TikTok says, and where exactly are you getting your understanding of what's going on in Israel, and all of those kinds of social media-ry, internet-y media consumption kinds of questions.
I don't know, there are people who are like super, super intentional about every minute of their child's life. I'm not. I kind of recall the George Carlin line, you know, that anyone who drives faster than you is a maniac and anyone who drives slower is an idiot, and that certainly applies to parenting. So, you know, like everybody thinks they're somewhere in the middle, but I don't have a deeply systematic and widespread and coherent plan for "this is what my child is going to be doing in 3, 6, 12, 18 months' time." I'd settle for him not breaking his Chromebook again.
I like that your just general vibe is it's going to be okay, guys. It's going to be okay.
Yeah. I don't know if you... I think, you know, maybe this is because I'm British and we haven't had political violence in 500 years. And I think, you know, if I came from Iran, I'd have a different attitude to being calm about the future. I think there's a layer of, like, yes, this will change a bunch of stuff and we'll need to worry about it, but that's kind of a constant. We've always had that.
I remember in the whole wave of the panic around social media, I dug up a whole bunch of books in the late '70s about databases. There was a whole panic about databases, and again half of it was true. Like, you know, if all police records and all government records are online, then that's different.
If you think about, for example, the deepfake nudes issue, there's like a dumb reaction to this, which is to say, "Haven't you heard of Photoshop?" Which is true, but a 15-year-old kid couldn't use Photoshop to make hardcore pornographic nudes of every girl in their high school and send them to the whole school in one afternoon.
And turn them into video.
Exactly. Well, yeah, even more. And now they can. So that is different.
It's kind of like, you know, the challenge of social. The thing people would say in the '90s is it's great, you can be the only gay kid in your village and you can find other gay people and you can find your tribe. And guess what? It turned out you could also be the only Nazi in your village, or the only pedophile in your village, or somebody who wanted to look at child porn. And like, yeah, now you can find the other people who like looking at child porn and they'll tell you it's great. So, oops. We connected everybody, and unfortunately that meant we connected all the bad people and all of our own worst instincts and every problem in society. And so that will happen again with AI. You know, deepfake nudes are like the obvious thing we can see now. There will be a whole bunch more of this stuff.
But there's also, and you know, something a kind of technical audience should know about. Do you know about the Post Office scandal in the UK?
Nope.
Okay. So, sidebar here. So in the UK, post offices are mostly franchises run by small business people. So they're run by, like, pharmacies classically. Typically, very often Indian immigrants, second-generation Indian people. And the Post Office, like 15 years ago, rolled out a new point of sale computer system. So they have a separate counter in the back, that's the post office. And so the Post Office rolled out this new computer system built by Fujitsu that had a bunch of bugs in it that showed shortfalls in cash. And the Post
office looks at this and says, "Aha, we knew these people were stealing from us." Hundreds of people get prison. Bunch of suicides, bunch of bankruptcies, people lose their homes. Meanwhile, people from the post office and people from Fujitsu are going to court and swearing there's no bugs in the system and nobody else has had this problem. This is 1970s technology.
And that's really the point that every wave of technology comes with ways that you can ruin people's lives either deliberately or by accident. This is the whole thing of Chinese mass surveillance is deliberate. This is maybe people should go to prison, maybe not. But like we have this with every technology. We have a bunch of ways that you can ruin people's lives and you have to be conscious of that and also kind of not panic about it.
So maybe following that thread and coming back to the kids thing and the jobs thing. Is there like a job you are steering your kid away from and is there a job you kind of think you want to steer them towards?
I don't know about that since it's probably a little bit early yet. He's not quite at the like I want to be a fireman stage. But
That might be a great job.
Yeah. And certainly, you know, if I look at my career, you know, I started as an equity analyst and then I went and worked in industry and then I was a consultant. Like, you know, the days when you kind of knew what your career was going to be are over. You know, there were clearly some people where you want to be an architect, you want to be a software engineer, you know, you want to be X or Y. I don't know. I think, you know, the only kind of thinking I have here is that you slowly work out there's a bunch of skills that you have and there's a bunch of jobs that makes you good at, and then there's a bunch of stuff that people will pay you for, and you want to get at least two of those and preferably all three.
Okay. So, zooming out a little bit, let me ask you a meta question. What's a question about AI that you think nobody's asking yet or not enough people are asking that we should be asking ourselves?
Sure. I mean, we talked about value capture. Obviously this is a whole, everyone is asking, like, you know. I'm not sure how many people are asking whether model labs have pricing power. I think a lot of people are just presuming that the situation today will continue or that of course they will. So I think that's maybe a question that not enough people ask.
I think the question I posed towards the end of my presentation, which we talked about earlier, is like what's the task and what's the job? What is just the thing that becomes a button or make scoo versus what are people actually hiring you for is a kind of a useful way of thinking about this. And clearly there are going to be some jobs where no, that is just a task and that job gets automated away. But there's a bunch where that kind of isn't the question.
The way I actually pulled that together at the end of the deck was a chart of global recorded music revenue, which as you may know is kind of a U-shaped curve, more or less. So, it dropped by about half from 2000 to 2015 or so, and since then has come back to about 75% of the peak, adjusted for inflation. And the way that I look at this is to say, and that's driven by streaming. And I kind of looked at this and said, well, the first half of this chart is saying, what happens if I don't have to pay $15 to get a CD to get that track? And the second half of the chart is saying, what happens if $15 a month gives you all the music that there is? So, it's kind of a completely different sort of question.
And you could, you know, that's the way that you could look at Uber or the way you could look at Airbnb, all these kinds of companies, is that to begin with you do the old thing but more. With every new technology you do the old thing but more of it on the new place. So, you know, you put Flickr on mobile, you print out your emails, and then you make new things that are only possible with the new thing. And then maybe you go a bit further and you kind of completely redefine the question and you make something that isn't that at all. You know, Spotify is not an online music store. It's something else.
And right now, you know, those questions, you only even know what the question is after it's been asked and you built a billion dollar thing that lots of people use, because obviously Spotify looked crazy and Airbnb looked crazy. But I think the way to get at what this means is you have to get past we do the old stuff but more, and you have to get to what do you do that's different that's because of this? What does this change? What wasn't possible before? What gets unlocked, as opposed to just doing the old thing but more of it?
Yeah. Just to support this kind of general theme you have of we don't know what is going to happen, this is unprecedented. If you were to zoom out a few years ago, maybe three years ago, four years ago, the last profession you'd think would be automated is engineering and coding. That feels like the hardest thing, like we're going to need people to build these things. Now it's like the most transformed role of any role. Like you went from writing all your code to 0% of your code is AI. It's almost like you didn't realize it was boring manual labor that could be automated. You thought it was something else.
It's funny. I mean, I was looking at this whole, there's a sort of US government dataset called O*NET or something like that, which tries to kind of analyze every single job, and then people try and kind of score it and they try and say, well, you know, this profession is X or Y% exposed to AI and AI can do Z% of it today. I think this is just the most ridiculous bunch of deluded horseshit.
And there's two reasons for this. The first reason is that this is, ironically, the logical systems problem. The expert systems problem. The problem of expert systems is, for anyone who doesn't know, you try to recognize a picture of a cat and you start building up logical steps. So you make an edge detector and then you make a fur detector and you make an eye detector and you make an ear detector, and 15 years later you've got 700 steps and it doesn't work. And this is what happens when you try and look at a profession and sort of break it down by which bits can be automated and which can't. You can't describe a profession like that, or anyway we can't. You can't kind of look at a senior partner at a law firm and say, well, 17% of their work could be automated. Like, this is horseshit. You can't do that.
I think the other side of the fallacy though is to talk about taxi drivers. So, you know, if we'd been having this conversation in 1997, it's like the Uber test. Imagine we were in 1997: what will be crushed by the internet? Well, newspapers will be fine because they'll save money on the printing bills. This is like a joke, but people said the internet will be great for newspapers. Their printing bills will go down. Well, yes, but no. But the other side is, well, obviously taxi drivers, you couldn't automate that with the internet. It's got nothing to do with the internet. Maybe you'd have internet booking, but no, that's not going to change anything. And of course, it completely changes the whole thing.
And so, the example I saw the other day was things that won't be affected by AI: personal trainers. Okay. So, I take my iPhone and I balance it on the metal piece with the camera pointed at me and I ask an AI to build me a training routine and watch me and tell me if I'm doing it right. Why do I need a personal trainer? Now, that might be complete nonsense. But that's how these things work. Like the stuff that you don't think is ex, you can't predict which things are going to be exposed necessarily, or, you know, a lot of the big companies are things that didn't look like they would work and didn't look like that was exposed.
The other side of this, of course, and this is one of the charts at the end of my presentation, is comparing Uber and Airbnb, because this is like the cliche from Marc Andreessen that Uber doesn't sell software to taxi companies, Airbnb doesn't sell software to hotels. Okay, now let's go and look at the market impact. Well, a whole bunch of cities where Uber demolished the taxi business and made it much bigger as well, made the town became much bigger, and everyone switched. Airbnb's impact on hotels, if you actually go and look at the numbers, is pretty marginal. They carved out this whole other business and maybe they slowed down the growth of hotels a bit.
But, you know, my wife flies to Milwaukee next week. She's going to land at 8:00 at night. She wants to go to a hotel. She wants to have room service. She needs a bathroom, bath. She needs, you know, a gym at 6:00 in the morning, and then at 7 in the morning she's going to drive to the client site. She's not going to stay in an Airbnb. Like, absolutely zero chance she's going to stay in an Airbnb. And half of the hotel business is business travel. And, you know, as soon as you actually get into anything, then it gets complicated.
I remember somebody on social media said the problem with Benedict is his answer to everything is it depends. It's like, yeah, it does. It depends. So, you know, it's back to my 1997 point. You can say some of this. But you have to have that humility.
Yeah. I'm coming back to this phrase you use, presume radical uncertainty, is a nice core thesis here. So knowing all this, it's hard to tell. We don't know exactly where it's going. Things are going to change a lot, but it'll probably be okay broadly. Just, a lot of people listening are pretty worried about their jobs and their careers and how much the world changes. What would be a couple things you recommend people do, knowing what you know, to be more successful in this future?
Well, I should just kind of wind back on what you just said, is like, as Keynes tells us, in the long run we're all dead. So, you know, on average, you know, on average nobody died in World War I. Great. But if you're a 19-year-old in 1914, you've got a one in three chance of not coming back. So yes, you know, clearly there's a bunch of professions where this is a major question, and particularly if you're an associate or would have been thinking about being an associate, this is a major question, and it's very unclear how those professions are going to play out. It's very unclear what happens to the pyramid structure of professional services.
The only answer I think one can have is, you know, don't stick your head in the sand and say I hate all of this stuff, because that gives you a great feeling of moral superiority and you can go on Bluesky and shout at each other about how evil AI is. Like, great, I'm happy for you, but that's not going to help. What helps is you diving into this, completely submerging yourself in it, and coming out understanding what you can do with it, how this changes things, how you can be a great hire.
And that may still not help, but, you know, if you're going into a law firm and they're like, well, we hired 100 associates last year and this year we're only going to hire 50, going to the interview and saying, well, I think AI is and I'm never going to use it, is probably not the right mood. So, you know, that may not be particularly comforting, but I don't think there's an alternative. You have to dive into this and absorb it and internalize it and think about what it means, just as, you know, you and I did with mobile and with the internet.
I think that is actually very actionable and very consistent advice on the podcast, is just do stuff. Build it. Don't just sit around and pontificate and be pissed at what's happening.
To close us out, I'm going to take us to AI corner, a recurring corner of the podcast. And the question to you is just, what's one way you've used AI in your work or life that is really interesting, something that other people might be inspired by?
I don't know. I struggle with this question because I'm sort of the lawyer looking at ChatGPT. So, you know, the stuff that I would automate are sort of precise information retrieval tasks, which is precisely the thing that this is kind of worst at. And, you know, that's not a criticism. It's just an observation. The kind of stuff that I would want a machine to do for me is the stuff that AI kind of can't do for me very well at the moment.
I use it for proofreading. I use it, you know, for images. I used it redecorating my apartment. It worked fantastically well at that. Here's a picture of this room. Repaint it. Add this light and this table and this rug. No, change the color of the rug. There's a kind of plants for stuff where it works. But I mean, a couple of years ago, somebody said AI is good at stuff that computers are bad at and bad at stuff that computers are good at. And that's, I struggle to find many examples of those where I need it.
But then, you know, I'm in kind of a unique, weird job. You know, I sit at my desk all day, you know, trying to synthesize a whole bunch of other stuff into a whole bunch of new ideas. That's not particularly common where people spend their time. I struggle to find AI use cases. I am the accountant looking at a spreadsheet and thinking, well, that's very clever and this is clearly going to completely transform everything. But I actually don't make spreadsheets every day.
I went to a standup comedy show of Pete Holmes. I don't know if you know him, and he made this joke that we want AI to clean the poop off the street and do all these hard things that nobody wants to do, but instead it's like, oh, let me help you write, let me help you create imagery. It's like these bohemians, like, no, I don't want to do all these ugly things, I want to be creative, make art.
Yeah, well, I mean, there's, you know, variations of all of this. You know, it's like, I don't want the AI to do the stuff I do for fun, I want it to do the boring stuff that I don't do for fun.
And, you know, finding that mesh, I mean, joking apart, this is going to come back to kind of my chatbot point, that, you know, the chatbot is a blank screen and a jagged edge. Like, what am I supposed to do and what will work? And that's a big problem, and the solution to that problem is to wrap it in use cases.
Part of it is also like AI just disappears. So most of what I write now I dictate. I dictate as a voice memo and that's automatically transcribed. Is that still AI or is that just voice recognition? It's probably an LLM. There's probably an LLM in there. Okay. So maybe that's AI. Well, okay. So what? At a certain point it's just automation.
What do you use for voice transcription?
So I actually find Apple Notes, the app, or the one built into the iPhone, works fine. I mean, I'm conscious that people want others, but like, I dictate it. There it is. It worked. So I'm happy with that.
All right, final question before we get to our very exciting lightning round. Is there anything else that you wanted to share? Anything else you want to leave listeners with?
No, I think, you know, I've monologued plenty and I've gone through a bunch of stuff in the deck. Go read the deck and sign up to my newsletter, and then you will get many more mags of brilliant Benedict Evans wisdom. Some of which may even be useful. Someone unsubscribed from my newsletter and they said, you didn't give me any actionable stock ideas. And I'm like, well, on one level that's completely true. On the other level, maybe not.
Well, with that, Benedict, we've reached our very exciting lightning round. I've got five questions for you. Are you ready?
Sure. First question, what are two or three books that you find yourself recommending most to other people?
It's a tough one for me because I just read an enormous amount of books and then I can't remember which ones I've read. I sometimes often joke that there's a classic British comedy from the late 19th century called Three Men in a Boat, which is like my I Ching. It's like we're having trouble hanging a picture. Well, there's a section about that. You know, we're having trouble doing this. Oh, well, there's a story about that. All of which are hilarious. So Three Men in a Boat is my I Ching.
There's a book by I think William Cronon about the economic history of Chicago which is fascinating and actually very relevant to technology because it's talking basically about standardization and packetization and logistics and channel conflict and network dynamics and network neutrality. So like when the meat packers of Chicago reach the point that it's cheaper to ship a cow from New York to Chicago, kill it, pack it and then ship it back to New York than to kill it in New York. And you know the pricing of refrigerator cars, and it's exactly like reading about broadband. It's all the same kind of those kind of business issues, which is fascinating.
What else have I read? I don't know. Read books. Read different books generally. Read books for grown-ups. Please read something other than Lord of the Rings. If you're going to name another company, like I saw this, and what was the latest, like, Peter Thiel company? I was like, read another book. [laughter] Everything is named after a character from this one book. There's more than one book in the world. If there is more than one book, then all about science fiction. Read about different things. Read about things you don't know about.
Kind of along those lines. You have a favorite recent movie or TV show that you've really enjoyed?
I don't know. I've dropped so badly off the current media treadmill and I just spend most of my time watching classics, which are like always the ones that you're supposed to have seen and that all seem intimidating, and then you watch them and you're like, "Oh, that was actually really good." I watched The Seventh Seal recently, which is like one of those joke Woody Allen terrifying, boring movies, and it was brilliant. It was really interesting, and it's only like an hour. So, go watch one of those movies that you are supposed to have seen or hadn't seen.
Favorite recent product that you've recently discovered that you really love. Could be a gadget, could be an app.
I was speaking at a partner meeting for a company earlier this week. What's today? Monday. No, last week. And met the founder of the company, who has a very famous, the CEO of the company has a very famous name, and admired his shoes and didn't say anything, but then went in Google like half an hour later. Yeah, okay, I'll buy a pair of those.
You want to share the brand or you want to keep it secret?
Okay, we'll keep it secret.
I don't know. I think one comes in waves of new products and you know you get into waves of new things and, like, when's the last time there was a cool app, like iPhone apps? That was, you know, all that white space went. I mean it's partly a function of product shifts, platform shifts, like all the white space went for cool new apps, and now we haven't quite got, actually this is to the earlier point, we don't have breakout consumer AI apps yet, because I think because of marginal cost more than anything else. So you can't make it free and get 50 million users and then have a revenue model. But we don't have those breakout things yet.
For consumer, yeah.
For consumer, no. I just, weird, I keep getting these ads for voice recorders, like somebody's selling like a business card size, like hardware voice recorder. I'm like... But like I didn't get it. Like I've got a voice recorder on my phone.
Yeah. All kinds of cool stuff coming. Okay, two more questions. Do you have a favorite life motto that you find yourself coming back to often in work or in life?
I suppose I've mentioned earlier, apparently I mostly say it depends.
That's going to be the title. [laughter]
No, it'll probably be okay.
Yeah. Okay. That's the vibe I get. I like that. I like that it's probably going to be okay. Not for sure. Okay. Final question. I saw somewhere that you own a lot of old phones. Is that true?
It is. Yes. I kept, I mean, I was a telecoms analyst and mobile analyst and I kept all my phones up to a point. Now they're kind of uninteresting. But as you may remember, like before the iPhone, particularly outside the USA, there was this huge creativity and expansion in what phones look like, because everyone was basically innovating around a little teeny tiny gray square. So everyone was trying to differentiate from everything else, before it kind of result.
It's kind of like cars actually. It's like cars before, like, wind tunnels. Cars all look different and everyone's trying to innovate around, because you've got the same four wheels and the same engine and everyone's trying to like differentiate based on like the shape, and then everything converges on one shape. And it's kind of the same with phones, like everything converged on one shape. Before that there was all this innovation. So yeah, like I have like a whole bunch of PDAs and smartphones and...
How many phones we talking about?
I don't know, like 20 or 30.
Okay. Okay. It's not so crazy. What's like the oldest one? What's the oldest one you got?
So, I have one of those, I should have, you told me I'd have got the box down. I have one of those Ericsson shark fin flip phones from like '98 or something, which is very, again, like hardware design, visual design, trying to differentiate. I've got an i-mode phone from 2001 and a J-Phone from 2001 that has a camera. So, I came back from Japan in 2001 and my phone had a color screen and a camera, and like I just had like endless client meetings and people just wanted to see the phone with a color screen. Like it's mind-blowing. Didn't work outside Japan. I plugged it in the other day. It still charges up. I mean clearly I can't do anything with it.
And like, I mean there's a little bit of an analogy in there as well, in that like we thought there'd be all these different shapes and sizes, and before the iPhone people kind of imagined like, well, some people will have like a little pocket PC and some people have a keyboard and you have like folding, all these different ideas for what it would look like, and we didn't realize it was all going to converge on one device.
Benedict, this was amazing. I learned a ton. I feel better after this conversation. Two final questions. Where can folks find you online? Where do they find this presentation? And how can listeners be useful to you?
If you can Google me, as I always say, my parents had good SEO. So, Google Benedict Evans. And so there's a website where I publish all the presentations that I've done, and sign up for my newsletter, which comes out every week.
Otherwise, how can they be useful to me? Like I'm always trying to understand stuff and I'm always trying to ask different questions. The worst thing in tech is to like carry on talking about the same stuff. It's like, you know, the moment you really understand something is the moment you have to push on to something else. And so I'm always trying to think like, no, am I just talking about the same thing over and over again? Like last year I just spent probably too much time saying, "But these models still hallucinate. Stop telling me they don't hallucinate." And they do. They still hallucinate. You know, you push them a little bit further, any question, and you'll still get like, "Nope, that's not true." But that doesn't mean they're not useful. So you have to kind of keep pushing, keep pushing myself. So that's always the challenge for me, is how do I push? And then yes, if you want me to come and present to your board in the Caribbean, then let me know.
And by the way, the domain is ben-evans.com if folks want to check you out.
And it's ben-evans.com.
Benedict, thank you so much for being here.
Thanks a lot.
Bye everyone. Thank you so much for listening. If you found this valuable, you can subscribe to the show on Apple Podcasts, Spotify, or your favorite podcast app. Also, please consider giving us a rating or leaving a review, as that really helps other listeners find the podcast. You can find all past episodes or learn more about the show at lennyspodcast.com. See you in the next episode.
Article published · Updated