A rational conversation on where AI is actually going | Benedict Evans
In a Nutshell
AI is a transformative technology on the scale of the internet or mobile, not an industrial revolution or existential threat, and we're in the 1997-equivalent stage where most valuable applications haven't been built yet. Job displacement fears are overstated because automation historically creates new roles and expands total work, and the hard part of most jobs isn't the automatable task itself but judgment, context, and client relationships that AI doesn't replace. Foundation models are becoming commoditized infrastructure with limited pricing power, so most value will accrue at the application layer while distribution and integration determine winners.
These notes were generated by AI and may contain inaccuracies.
Benedict Evans' most controversial opinion is that AI is as big a deal as the internet or mobile, and only as big a deal as the internet or mobile. He contrasts this with people in tech who think AI is more like the industrial revolution, and others who think he underestimates its significance. Smartphones and the internet were quite significant; the current podcast would not exist without the internet.
Evans compares the current state of AI to 1997 for the internet—very exciting, most stuff does not work yet, most things people will eventually do have not been built yet, and it is not clear how any of it will work. People who have already adopted AI assume everyone else is there too, but there is a very wide distribution of adoption. Within tech, some have bought clusters of Mac minis and no longer use Google. Outside tech, most people using AI do so every week or two. There is also a wide spread in maturity of how well the technology works.
Evans dismisses concerns about mass job losses from AI. Every time a new technology emerges, it automates away jobs and then unlocks new jobs that did not exist previously. Companies like OpenAI and Anthropic are increasing headcount. People claiming every big company will buy ChatGPT tomorrow and fire everyone in two weeks are described as morons. You cannot predict which things will be exposed to automation—you cannot look at a senior partner at a law firm and say 17% of their work could be automated.
When asked about anti-AI sentiment, Evans describes it as a big fuzzy mess. He acknowledges this will change things and requires attention, but notes this has always been constant with new technologies.
Evans recommends not sticking your head in the sand and declaring hatred for AI. While this provides a feeling of moral superiority and allows posting on Blue Sky about how evil AI is, it does not help. What helps is diving into AI and understanding what can be done with it today.
Evans argues against trying to quantify whether AI is 20% or 100% bigger than the internet—those are not productive conversations. It is a fundamental change, but we do not know how any of it will work. He notes his recent presentation contained 80 slides essentially saying we do not know.
We are unquestionably already in a moment of radical change in software. There are questions about what agentic AI and AI software development mean for the future of the software industry. One extreme view—not widely believed—is that anyone can just code their own Stripe. There are questions about how much more software will exist and how much people can do themselves.
For software developers, AI changes everything—they are like accountants seeing VisiCalc. For others like lawyers, it is interesting but unclear how to use it or how many associates to hire next year. Evans uses an analogy: an accountant in the late 1970s seeing spreadsheets would find it mind-blowing because it changes interest rates and recalculates everything in 30 seconds. A lawyer or journalist would think it is clever for their accountant but not relevant to their work, especially at a cost of $10,000-$15,000 for an Apple II setup.
Survey data shows even among 13 to 18 year olds, only 15-20% are daily active users of AI, another 20% are weekly active users, and 60% are not using it. There is a widespread variation in who gets it. This maps to the jagged frontier question of where AI works, where it does not, whether you can tell in advance, and whether you can figure out what to do with it.
The AI labs—particularly OpenAI and Anthropic—are investing heavily in buying consultancies and private equity firms and hiring forward deployed engineers. Evans jokes that a forward deployed engineer is like an Accenture outsourced software developer who lives in San Francisco. Companies do not have people sitting around waiting to build new projects—they hire Bain, BCG, McKinsey, Accenture, or branding agencies for this work. Reimagining internal workflows and automating with AI is itself a project requiring five or ten people working for a month or two. Actually implementing changes is another project. This is why PE firms provide support to portfolio companies, and why companies hire consultants.
Evans splits his presentation into three sections: capital (where capex goes and model lab differentiation), deployment (implications for software industry), and change (how things transform). A key question is whether the hard part of a job is the task itself or something else. The elevator attendant example illustrates jobs where the task was the job—automation eliminated the role entirely. More commonly, automation triggers price elasticity (Jeans Paradox): when something becomes cheaper, you may do the same for less money, more for the same money, or more for more money due to new ROI.
Before Excel, junior investment bankers worked very long hours; now thanks to Excel, Goldman associates work at lunchtime on Fridays. Before IDEs, libraries, and operating systems, developers wrote all code. Now writing an iPhone app means 90% of the code is written by Apple (modem drivers, graphics drivers, file system). We do not have a tenth as many engineers—we have more. The number of accountants employed went up throughout the 20th century and has increased again since the beginning of the 21st century despite adding machines, punch cards, mainframes, databases, ERP, cloud, spreadsheets, and PCs.
Amazon gets you the SKU if you know what you want. If you do not know what microphone stand or product to get, Amazon is not the starting point. Knowing what SKU you want is a separate job. Claude can write code, but determining what code you want, what features, who the customer is, and how to take it to market remains. McKinsey is not hired for a 75-slide deck—Claude can make a crappy version of that. McKinsey is hired to walk through an enterprise, understand why something was not done, navigate politics, talk to customers, and determine what actually needs to happen.
Newspapers and recorded music had physical manufacturing or distribution decoupled from their core offering. Record companies thought of themselves as being in recorded music, but they were manufacturing small pieces of plastic. Newspapers did not think of themselves as manufacturing and trucking companies. When those elements were decoupled, problems emerged.
OpenAI and Anthropic are increasing headcount. The companies thought least likely to add humans are adding many humans, demonstrating the complexity of automation.
Evans is not interested in Dario Amodei's opinions on theories of labor markets, value, and comparative advantage, though he would be interested in his views on where models are going in the next 6-12 months. Every new technology automates jobs and creates new ones. In 1800, 90% of people were peasants worried about crop failures. Since then, jobs have been automated and new jobs created. The old job going away is visible; the new job is not, because it does not exist yet. Railway engineer was not an obvious job in 1800. This process has repeated since 1800, with frictional pain and dislocation, but resulting in greater wealth and the elimination of concerns like crop failures.
There is no a priori reason AI should be different from previous technological shifts. The internet removed jobs; PCs removed jobs. Typesetters, telephone operators, and typists largely disappeared. Jobs that go away are, retrospectively, crap jobs; new jobs are better because GDP keeps rising. AI adoption is faster because we stand on the shoulders of giants—ChatGPT reached 900 million users because the internet already had 900 million people. Netscape launched when there were only 50-100 million PCs globally. Enterprise software sales cycles are 18 months; people will not tear out SAP and replace it overnight.
New technologies take time to fully transform industries even when the underlying capabilities exist. Frame.io is an example of a video collaboration tool where the underlying technology could have been built five to ten years earlier. Many SaaS companies founded around the time ChatGPT launched could theoretically have been started fifteen years prior. The bottleneck is not technology availability but people realizing that a problem exists within an industry and identifying the specific way to solve it with new tools. This process of discovery and application takes three to ten years sector by sector.
The pattern of technological transformation repeating itself applies to current AI developments. An IBM advertisement from the 1950s promoted an electronic calculator the size of a refrigerator as providing 150 extra engineers. This framing mirrors contemporary AI pitches offering additional engineering capacity. The core message is that society has repeatedly gone through transformations where new technology augments human capability at scale.
The internet enabled entirely new research and analysis workflows that are now taken for granted. Creating a chart showing the growth of supermarket SKUs in America since the 1950s required knowing the Food Marketing Institute existed, that it published annual SKU counts, and that historical data going back to the 1950s was available. In 1994, this process would have involved multiple days of phone calls, long-distance charges, potential $500 per report fees, and travel to a library. The same task today takes roughly two hours via search. This illustrates how massive changes become invisible once normalized as the baseline state of the world.
There is no established theory of human intelligence, no theory explaining why current models perform well, and no theory predicting how much better they will become. Discussions about AGI and superintelligence involve significant redefinition of terms. Larry Tesler observed that AI is defined as whatever machines cannot do yet. Once machines can perform a task, it is reclassified as software. Machine learning capabilities such as image recognition and sentiment analysis are sometimes described as not being AI. AGI definitions have shifted toward measuring the percentage of economically valuable work a system can perform, which is a different standard than possessing consciousness or being alive. An IBM mainframe in 1975 already performed a meaningful percentage of economically valuable work previously done by humans.
Even if model capabilities plateau and hit a performance ceiling, current AI represents an incredibly useful technology that will transform the world through rollout over the next ten years. Belief in AGI or superintelligence is not required to recognize the scale of impact from existing capabilities.
The addressable market for technology companies continues to grow because software can address progressively larger portions of the economy. Mainframe install base peaked at roughly 70,000 to 80,000 units. Internet-era PC adoption reached 50 to 100 million devices, with current estimates at 1 to 1.5 billion PCs globally including 700 to 800 million consumer units. There are approximately 5.5 to 6 billion smartphones worldwide. This scale explains how platforms like JTP can reach 900 million weekly users. The narrative that growth has plateaued because there are no more people to reach overlooks the expansion into automating entirely new portions of economic activity.
Historical patterns show that each major technology removes existing jobs while creating new jobs and new sources of value. The lump of labor fallacy assumes a fixed quantity of work. The last 200 years demonstrate that technological change creates net prosperity despite transitional pain. An analogy to the electricity industry illustrates how a foundational capability becomes embedded throughout the economy. Electricity moved from powering factories to becoming part of nearly everything. Software has followed a similar gradual integration pattern.
Selling AI intelligence on a metered basis like electricity or water faces structural margin limitations. Television broadcasters do not pay a percentage of viewer subscriptions to electricity providers. Washing machine manufacturers do not pay a percentage of unit prices to electricity companies. The global mobile industry generates roughly one trillion dollars in annual revenue and spends approximately 200 billion dollars on capex, representing 15 to 20 percent of revenue. Mobile data consumption has grown 1,500 to 2,000 times since 2010, yet telecom stocks have remained flat over 25 years because the service functions as a low-margin commodity utility. The valuable applications and services built on top of the infrastructure capture the majority of economic value.
The central question for foundation model economics is whether the model itself performs complete tasks or whether value requires applications built on top of models. If the chatbot interface is not the primary user experience and thousands of specialized applications must be built, model companies cannot build all of them. This structure resembles AWS more than Windows. Enterprise buyers do not standardize on a specific cloud provider simply because other software runs there. If foundation models function as undifferentiated commodity infrastructure, pricing power may be limited and the majority of value may accrue further up the stack at the application layer.
Foundation models do not appear to exhibit strong network effects that would produce winner-take-all dynamics. Without radical differentiation between competing models, sustained competition is likely. This environment reduces pricing power. Thousands of distinct applications built by different companies cannot all be developed by the model providers themselves. The resulting market structure is expected to resemble cloud infrastructure markets more than operating system markets.
Historical platform shifts demonstrate varied incumbent outcomes. During the mobile transition, some companies missed the shift entirely while others were largely unaffected. Google experienced minimal change. Meta benefited from improved social functionality through cameras and notifications. Amazon's core business model remained largely intact. Yahoo Mail failed to adapt. Companies already in decline often failed to make the transition. Incumbents frequently attempt to position new technologies as features of existing products.
As software becomes easier to build, the volume of new products increases and competition for attention intensifies. Distribution capabilities become increasingly valuable moats. Incumbent companies with established distribution channels gain structural advantages over startups attempting to reach users in a noisier market. When products are commoditized, distribution determines competitive outcomes.
The web browser as a product is fundamentally different from the rendering engine. The rendering engine can vary in quality, but the browser product itself functions as a thin wrapper around the rendering engine with little room for meaningful innovation. The last significant browser innovation was tab browsing approximately 20-25 years ago. Attempts to innovate browser design consistently fail because the current form represents a platonic ideal, similar to how smartphone design has converged on the glass rectangle form factor.
Microsoft leveraged distribution to establish browser dominance. Despite winning the browser market for five to six years, this victory yielded no meaningful strategic value because the actual value capture occurred further up the stack. The lesson demonstrates that controlling the browser layer does not translate to capturing value in the broader ecosystem.
Google is currently using distribution to promote Gemini. For typical users, the differences between competing AI models like Gemini, ChatGPT, and others are not perceptible. Meta's Llama model has performed better in survey data than many in technology circles assume, ranking between ChatGPT and Gemini in usage. This success stems from Meta embedding the model across multiple service surfaces, demonstrating that distribution of an adequate product can drive adoption when the underlying technology has become commoditized.
OpenAI's strategy in late 2024 involved attempting multiple approaches simultaneously to establish distribution and create user stickiness before competitors like Google, Meta, and Amazon could deploy their models everywhere and capture users through default settings and inertia.
Apple represents the final major distribution channel yet to fully commit. The WWDC 2024 presentation outlined a vision for Apple Intelligence that remains the most compelling articulation of a personal AI assistant seen to date. The vision described tool-using, agentic, on-device AI with no prompt injection, no hallucinations, and a standardized API system across approximately 10,000 apps with perfectly functioning intents. While Apple was unable to ship this vision, no other company has successfully delivered this level of integration either.
The upcoming WWDC will be notable for whether Apple ships this vision powered by Gemini. The arrangement creates a situation where Gemini intelligence powers Android while Apple Intelligence on iOS is also powered by Gemini, yet the products remain distinct because the model functions as the underlying commodity while the feature decisions and distribution strategies differ between the companies.
Apple possesses approximately one billion devices capable of running AI on the edge. Google's marketing emphasizes availability on its most powerful devices, which excludes most Android phones.
Google recently held a major event replacing Chromebooks with new devices and announcing Android intelligence powered by Gemini. The rollout targets the small number of users who purchased Pixel phones. Broader market attention has shifted away from both Android and iPhone developments.
Survey data indicates AI is less popular than ICE among the public. Communities are attempting to block data center construction. Eric Schmidt encountered booing during a commencement speech whenever AI was mentioned. Multiple factors contribute to this backlash.
Electricity bill increases represent a tangible impact in specific locations. Water consumption concerns are largely unfounded. Data centers use water primarily in closed-loop cooling systems. A Livermore Lab study from late 2024 estimated US data center water consumption at approximately 0.017% of total US water consumption. While localized conflicts can occur when a single well serves both a community and a data center, this represents a planning issue rather than an inherent data center problem.
Data centers currently account for approximately 5% of US energy consumption and may grow by roughly one percentage point annually over the next five years.
Economists lack consensus on AI's employment effects. Available charts present conflicting signals. Employment slowdowns among 18-24 year olds appear consistent across degree holders and non-degree holders, and across fields both exposed and unexposed to AI. The absence of clear data stems partly from model labs not releasing meaningful usage statistics, including daily active user numbers.
Current analysis of US census data has not produced a clear consensus on measurable AI-driven job impacts. However, political perceptions of employment difficulty among students matter regardless of causation, whether attributed to AI, policy changes, or other factors.
Book cover illustrators for young adult romance novels have expressed concern about AI generating images such as a naked woman on a dragon flying over a volcano. Novelists and ebook writers are engaged in cultural debates about AI usage appropriateness. Approximately 30-40% of new podcasts are estimated to be AI-generated, contributing to discussions around AI slop.
The current AI backlash resembles the social media backlash but in compressed form. Some concerns are valid, some partially valid, and some unfounded. The Facebook data-selling narrative exemplifies how misinformation can persist despite being factually incorrect. Jonathan Swift's observation that one cannot reason someone out of an idea they did not reason themselves into applies to these debates.
Trump has indicated interest in a new executive order regarding dangerous models. Concerns about missile and cyber risks do not represent mainstream public discourse, though these issues prompted presidential attention.
For children entering the job market in the next one to two years, outcomes remain highly uncertain. For children entering the job market in approximately five years, conditions will likely have stabilized in unpredictable ways. The speaker has a child in early teens rather than a 21-year-old, which changes the immediacy of these concerns.
Many questions regarding children's media consumption and information sources predate ChatGPT, including evaluating influencer claims on TikTok and understanding events in regions like Israel.
The speaker does not maintain a deeply systematic parenting plan with specific milestones at 3, 6, 12, and 18 months. A more pragmatic approach focuses on immediate concerns such as preventing repeated Chromebook damage.
British cultural context, with 500 years without political violence, contributes to a relatively calm perspective on technological change. Historical database panic in the late 1970s demonstrated that technology concerns often contain partial truth alongside exaggeration.
Deepfakes represent a genuine escalation beyond Photoshop capabilities. A 15-year-old could not previously create and distribute hardcore pornographic deepfakes of classmates in a single afternoon, but current technology enables this. Social media similarly enabled both positive community formation and negative connections, including among previously isolated individuals with harmful interests.
The UK Post Office scandal involved franchise operators, often pharmacies run by second-generation Indian immigrants. Approximately 15 years ago, a new Fujitsu-built point-of-sale system was deployed with bugs that falsely indicated cash shortfalls. The Post Office pursued prosecutions, resulting in prison sentences, suicides, bankruptcies, and home losses. Post Office and Fujitsu representatives testified in court that the system had no bugs. The scandal involved 1970s-era technology, illustrating that every technology wave creates mechanisms for ruining lives either deliberately or accidentally.
The speaker does not currently steer their child toward or away from specific careers, as the child has not yet reached the stage of expressing occupational preferences. Historical career patterns where individuals could predict their professional trajectory no longer apply. The relevant framework involves identifying personal skills, determining which jobs those skills suit, and identifying what people will pay for, then aligning at least two and preferably all three of these elements.
One under-examined question concerns whether model labs possess pricing power, as many assume current conditions will persist. Another useful framework distinguishes between tasks and jobs: which activities become automated buttons versus what clients actually hire professionals to accomplish.
Global recorded music revenue forms a U-shaped curve, declining approximately 50% from 2000 to 2015 before recovering to roughly 75% of peak levels when adjusted for inflation. The decline phase reflects what happens when consumers no longer pay $15 for individual tracks via CD. The recovery phase reflects what happens when $15 monthly provides access to all available music. This illustrates how technology initially enables doing existing activities more efficiently, then creates entirely new possibilities that redefine the underlying question.
Spotify exemplifies this progression: it is not merely an online music store but something fundamentally different. Similarly, Uber and Airbnb initially appeared to be doing old things in new places but ultimately created different categories. The challenge lies in identifying what becomes possible that was not previously possible, rather than simply doing more of the existing thing.
Attempts to quantify AI exposure by breaking professions into component tasks represent a fundamental misunderstanding. The expert systems problem demonstrates that complex pattern recognition cannot be reduced to sequential logical steps. Attempting to determine that a senior law firm partner has 17% of their work automatable produces meaningless results.
Historical predictions about internet impact illustrate the difficulty of forecasting. In 1997, analysts suggested newspapers would benefit from reduced printing costs while taxi drivers would remain unaffected. Both predictions proved incorrect. Recent claims that personal trainers face no AI exposure overlook possibilities such as using an iPhone to record workouts and receive AI feedback on form.
Uber demolished taxi businesses in many cities while simultaneously expanding the overall market. Airbnb's impact on hotels has been more marginal, carving out a separate category rather than replacing existing business. Business travel, representing approximately half of hotel revenue, remains largely unaffected because travelers often require hotel amenities such as room service, reliable bathrooms, and early morning gym access that Airbnb properties typically do not provide.
The problem with many answers to AI questions is that everything depends on context. This returns to the 1997 point about needing humility when making predictions. Presuming radical uncertainty is a core thesis—things are going to change a lot, but it will probably be okay broadly. However, many people are worried about their jobs, careers, and how much the world will change.
In the long run we're all dead, but individual risk matters. For example, on average nobody died in World War I, but if you were a 19-year-old in 1914, you had a one-in-three chance of not coming back. There are clearly professions where AI presents a major question, particularly for associates or those considering becoming associates. The pyramid structure of professional services faces significant uncertainty.
The only viable answer is not to stick your head in the sand and claim to hate AI. That approach may provide a feeling of moral superiority and allow people to shout on Blue Sky about how evil AI is, but it won't help careers. The alternative is diving into AI completely, submerging yourself in it, and coming out understanding what you can do with it, how it changes things, and how you can be a great hire. This may not guarantee success, but there is no alternative.
If a law firm hired 100 associates last year and plans to hire only 50 this year, going into an interview saying you think AI is evil and will never use it is not the right approach. The actionable advice is to dive in, absorb it, internalize it, and think about what it means—just as people did with mobile and the internet. The consistent recommendation is to just do stuff and build things rather than sitting around pontificating or being angry about what's happening.
Benedict struggles with the question of personal AI use because his work involves precise information retrieval tasks, which current AI handles poorly. He uses it for proofreading and for generating images, including redecorating his apartment by uploading a photo of a room and asking it to repaint it, add lights, tables, rugs, or change colors. He also uses it for plant recommendations. A couple of years ago, someone observed that AI is good at stuff computers are bad at and bad at stuff computers are good at.
His job—sitting at a desk synthesizing information into new ideas—is not particularly common. He references a Pete Holmes standup comedy joke about wanting AI to clean the poop off the street and do hard things nobody wants to do, but instead it helps with writing and creating imagery. The response is that people don't want AI to do the stuff they do for fun; they want it to handle the boring stuff they don't enjoy.
The chatbot is a blank screen with a jagged edge—it's unclear what to do with it and what will work. This is a big problem, and the solution is to wrap it in specific use cases. AI also disappears into the background. Most of what Benedict writes now is dictated as a voice memo and automatically transcribed. He uses the built-in Apple Notes app on the iPhone for voice transcription, which works fine for his needs.
Benedict encourages people to read his deck and sign up for his newsletter for more content. Someone once unsubscribed from his newsletter saying it didn't provide actionable stock ideas. He acknowledges this is true on one level, but perhaps not on another.
Benedict reads an enormous number of books but often can't remember which ones he's read. He recommends Three Men in a Boat, a classic British comedy from the late 19th century that serves as his go-to reference for many situations. He also recommends a book by William Cronin about the economic history of Chicago, which discusses standardization, packetization, logistics, channel conflict, network dynamics, and network neutrality. The example of meat packers in Chicago reaching the point where it was cheaper to ship a cow from New York to Chicago, kill it, pack it, and ship it back to New York than to kill it in New York—along with refrigerator car pricing—mirrors issues in broadband and technology business models.
His broader advice is to read books, read different books, read books for grown-ups, and read something other than Lord of the Rings. He notes that many companies are named after characters from that single book and encourages reading about science fiction alternatives and topics outside one's existing knowledge.
Benedict has dropped off the current media treadmill and spends most of his time watching classics. He recently watched The Seventh Seal, which he found brilliant and interesting despite its reputation as one of those intimidating, boring movies. It is only about an hour long. His recommendation is to watch movies one is supposed to have seen.
Benedict doesn't have a specific recent product to recommend. He notes that iPhone apps filled white space early on, but that space is now gone. He observes that we haven't yet seen breakout consumer AI apps, partly because of marginal cost—you can't make them free, get 50 million users, and then find a revenue model. He keeps seeing ads for hardware voice recorders the size of business cards but doesn't understand the need since phones already have voice recorders.
Benedict's recurring phrase is "it depends," though the title suggestion for the episode became "It'll probably be okay."
Benedict kept all his phones up to a point because he was a telecoms and mobile analyst. He has 20 or 30 phones. Before the iPhone, particularly outside the USA, there was huge creativity in phone design because everyone was innovating around a tiny gray square. This is analogous to cars before wind tunnels—cars all looked different as manufacturers tried to differentiate based on shape with the same four wheels and engine. Everything eventually converged on one shape.
His oldest phones include an Ericsson Shark Fin flip phone from around 1998, an iMode phone from 2001, and a J-Phone from 2001 with a camera. When he returned from Japan in 2001 with a phone that had a color screen and camera, he had endless client meetings where people wanted to see it. The phone still charges when plugged in, though it can't do anything useful now. Before the iPhone, people imagined multiple device types—pocket PCs, keyboards, folding devices—but everything converged on one device.
Benedict can be found by Googling his name. His website publishes all his presentations, and his newsletter comes out every week. The domain is benedict-evans.com. He is always trying to understand new things and ask different questions. The worst thing in tech is continuing to talk about the same stuff. The moment you really understand something is the moment you need to push on to something else. He tries to avoid just talking about the same thing over and over, such as repeatedly saying models still hallucinate.
He notes that hallucination doesn't mean models aren't useful—you have to keep pushing forward. He is also available to present to boards, including in the Caribbean.
Keep Lenny's Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
