A rational conversation on where AI is actually going | Benedict Evans
Benedict Evans, independent tech analyst and former a16z partner, argues that AI is as transformative as the internet or mobile—but only as transformative—and that we are roughly at a '1997 moment' where most applications haven't been built yet. He challenges both AI doomers and hype maximalists, arguing that jobs will transform rather than disappear en masse, that foundational model companies will likely lose pricing power over time, and that the application layer is where most value will accrue.
Summary
Benedict Evans joins Lenny's podcast to discuss his presentation 'AI is Eating the World,' offering a grounded, historically-informed perspective on where AI is actually headed. His central framing is that AI is as big a deal as the internet or mobile—a view he calls his 'most controversial opinion'—because it dismisses both those who think AI is trivial and those who think it exceeds all prior technological revolutions. He situates the current moment as analogous to 1997 on the internet timeline: exciting, real, but with most of the important applications not yet built and most of the important questions not yet answered.
Evans draws heavily on historical analogies to contextuate AI's impact on jobs. He references the introduction of VisiCalc (the first spreadsheet), which transformed accounting but didn't eliminate accountants—in fact, the number of accountants kept rising through every wave of automation including adding machines, mainframes, ERP, and cloud software. He argues that the 'lump of labor' fallacy underlies most jobpocalypse thinking: new technologies automate some tasks, lower costs, enable price elasticity, and unlock entirely new categories of work that didn't previously exist. He is skeptical of Dario Amodei's claims about eliminating entry-level jobs, noting that being an AI lab CEO doesn't confer expertise in labor economics.
On enterprise adoption, Evans dismisses the notion that companies will rapidly fire staff after adopting AI, pointing out that enterprise software sales cycles run 18+ months, that tearing out legacy systems takes years, and that working out how to reorganize internal workflows is itself a major project requiring outside help—which explains why AI labs are investing in consulting capacity rather than eliminating it. He notes the irony that the companies you'd expect to need the fewest humans (AI labs) are increasing headcount the fastest.
Evans spends considerable time on the question of where value will accrue in the AI stack. He argues that foundational model companies (OpenAI, Anthropic, etc.) are unlikely to retain pricing power long-term because the models appear to lack network effects, competition is intensifying, and open-source alternatives exist. He draws an analogy to the mobile telecom industry: despite building objectively remarkable global infrastructure, telcos have seen their stocks go nowhere in 25 years because they sell a commodity at marginal cost while all the value is created further up the stack. He predicts AI will look more like AWS (a commodity infrastructure layer) than Windows (a platform with strong network effects and lock-in), meaning most value will flow to application-layer companies.
On distribution, Evans argues that as models commoditize, distribution becomes the key moat. He notes that Google and Meta are using their massive distribution to push Gemini and Meta AI respectively, and that even if these products are slightly inferior, adequacy plus ubiquity beats excellence plus obscurity. He sees OpenAI's 'everything everywhere' strategy from late 2024 as a recognition of this dynamic—a race to establish default status before incumbents lock in their user bases.
Evans addresses the anti-AI backlash, characterizing it as a 'big fuzzy mess' of legitimate concerns (deepfake abuse, employment anxiety, energy use) mixed with factual errors (data center water usage is approximately 0.017% of US water consumption, per Livermore Lab data). He compares it to the social media backlash, where some concerns were valid, some partially valid, and some simply false—but all were politically potent regardless of accuracy.
On AGI and superintelligence, Evans is deliberately agnostic, noting that we have no theory of human intelligence, no theory of why LLMs work as well as they do, and no reliable basis for forecasting how much better they will get. He observes that 'AI' is a moving target—once something works, people say 'that's just software'—and that AGI definitions keep shifting to match whatever is currently achievable. He argues you don't need to believe in AGI to believe AI is a giant deal; even if models stopped improving tomorrow, the current technology is transformative enough to reshape the economy over the next decade.
For practical advice, Evans tells listeners not to stick their heads in the sand or adopt performative anti-AI stances. He argues that the only productive response is to immerse yourself in the technology, understand what it can and cannot do, and figure out how it changes your specific role—because that understanding will be essential for employability regardless of how the technology evolves.
Key Insights
- Evans argues that AI is exactly as big as the internet or mobile—no more, no less—and that comparing it to the industrial revolution reflects a failure to appreciate how transformative smartphones and the internet actually were.
- Evans claims we are at a '1997 internet moment' with AI: the technology is real and exciting, but most applications haven't been built yet, most things don't fully work, and the people who will win haven't been identified.
- Evans contends that the number of accountants kept rising through every wave of automation (adding machines, mainframes, ERP, cloud), which he uses to argue that the relationship between automation and employment is far more complex than simple displacement.
- Evans argues that enterprise software sales cycles of 18+ months, combined with the complexity of redesigning internal workflows, make the scenario of companies buying ChatGPT and firing staff within weeks 'a complete failure to understand how the world works.'
- Evans claims foundational model companies are unlikely to retain pricing power because models appear to lack network effects, meaning competition will persist indefinitely and margins will compress—analogous to how mobile telcos built extraordinary infrastructure but captured little value.
- Evans predicts AI infrastructure will resemble AWS (commodity layer, value accrues up the stack) rather than Windows (platform with network effects and lock-in), meaning application-layer companies will capture most of the economic value.
- Evans argues that the real question for any profession is not 'what percentage of tasks can be automated' but 'what is the task vs. what is the job'—noting that what clients actually pay McKinsey for is organizational diagnosis and political navigation, not PowerPoint slides.
- Evans observes that distribution becomes the dominant competitive moat as models commoditize, pointing to Google and Meta's ability to spray adequate AI across billions of existing users as a more powerful strategy than building a marginally superior model.
- Evans characterizes Dario Amodei's predictions about entry-level job elimination as an inappropriate application of authority, arguing that running an AI lab does not confer expertise in labor economics or comparative advantage theory.
- Evans notes that the anti-AI backlash around data center water usage is largely factually incorrect—Livermore Lab estimated US data center water consumption at approximately 0.017% of total US water consumption—but acknowledges this doesn't make it politically irrelevant.
- Evans argues that the 'jagged frontier' problem—the difficulty of knowing in advance which specific tasks AI will handle well—makes systematic job exposure analyses like the O*NET scoring methodology 'ridiculous deluded horseshit,' because professions cannot be accurately decomposed into automatable sub-tasks.
- Evans claims that even if AI model improvement stopped entirely tomorrow, the existing technology is already transformative enough to reshape the global economy over the next decade, meaning belief in AGI or superintelligence is not a prerequisite for taking AI seriously.
Topics
Full transcript available for MurmurCast members
Sign Up to Access