All podcasts / Lenny's Podcast / Summary

The most rational take on AI you'll hear this year | Benedict Evans

2026-05-31 - 80 min - source - Read full transcript

Lenny Rachitsky (host)Benedict Evans

Key insights

AI is exactly as big a deal as the internet or mobile were, no bigger and no smaller, and we are at the AI equivalent of 1997.

Benedict rejects both the dismissive view (AI is a fad) and the Industrial-Revolution-scale view. He compares the current moment to the internet in 1997: clearly transformative, but most applications haven't been built yet, adoption is wildly uneven (some people run local models obsessively, most people use AI weekly at best), and predicting winners now is like betting on Excite vs. Yahoo.

ai-hype-cycle

The right question for job impact is not 'can AI do this task' but 'is the task the job.'

Benedict's central framework: an elevator attendant's job was literally the task of pressing a button, so it fully automated away. But a McKinsey partner's job is not producing a slide deck; it is navigating internal politics, talking to customers, and building judgment, so AI producing a mediocre deck doesn't replace what clients are actually paying for.

ai-job-impact

AI labs are ironically hiring more professional-services and forward-deployed-engineer talent, not fewer, even as their products are pitched as automating that exact work.

Anthropic and OpenAI are investing in consultancy-style talent because deploying AI inside a large enterprise (mapping which of dozens of internal workflows can be automated, then rebuilding them) is itself a multi-month project requiring the same expertise that Bain, McKinsey, and Accenture already sell - companies don't have spare headcount sitting around to do that work themselves.

ai-job-impact

Cheaper labor does not mechanically shrink headcount; historically it has expanded total output and value, per the Jevons Paradox.

Benedict cites Excel not eliminating junior investment-banking jobs (associates still work long hours) and total US accounting employment rising through a century of adding machines, spreadsheets, ERP, and cloud software as evidence that automation of a task usually increases the addressable scope of the job rather than eliminating the role.

ai-job-impact

Foundation model labs may not have durable pricing power, because there's no winner-take-all dynamic among 3-6+ competing frontier labs.

Benedict argues the current era of loss-making token pricing is a temporary disequilibrium, not the steady state. Once demand and supply for compute normalize, competing commodity model providers should see pricing and margins compress toward a utility-like structure (he compares this to the global mobile industry, whose revenue grew 1,500x on data usage since 2010 while stock prices went nowhere).

value-chain-economics

Distribution, not product quality, is becoming the dominant competitive moat as AI products commoditize.

Because the underlying models are increasingly interchangeable to normal users, incumbents with existing distribution (Google via Search/Android, Meta via its apps, Apple via the iPhone) can win with a merely-adequate AI product. Benedict draws the direct analogy to Internet Explorer beating the technically superior Netscape purely through Windows bundling, even though winning the browser war ultimately mattered less than expected because value moved further up the stack.

value-chain-economics

The much-discussed AI water usage backlash is not supported by the data; other backlash concerns (energy, job loss) are more grounded but still unproven at scale.

Citing a Lawrence Livermore Lab study, Benedict states US data-center water consumption is about 0.017 percent of total US water use, calling the water narrative 'completely fake' in aggregate (while acknowledging real, localized harm where a data center dominates one small town's water supply). He's more circumspect on jobs: economists disagree on whether AI is driving the current slowdown in youth employment, since the pattern holds across degree-holders and non-exposed fields alike.

anti-ai-backlash

There is essentially no reliable public data on how many people actually use AI or what impact it's having, which lets both hype and backlash run ahead of evidence.

Benedict notes AI labs release no daily-active-user numbers, only selectively framed usage studies, leaving most real analysis to academic economists reverse-engineering government labor statistics (like the US census and BLS surveys) or paid consultancy surveys of a few thousand people - a strikingly thin evidence base for such a consequential debate.

epistemic-humility-on-ai

Terms like 'AI' and 'AGI' are moving targets that get redefined the moment a capability starts working, which makes many public debates about them largely semantic.

Quoting AI scientist Larry Tesler ('AI is whatever machines can't do yet'), Benedict points out that people once called image recognition and sentiment analysis 'not really AI,' and that AGI is now informally being redefined downward to mean 'does some percentage of economically valuable work' rather than anything resembling general human-level reasoning.

epistemic-humility-on-ai

You cannot reliably predict in advance which jobs are exposed to AI disruption; attempts to score professions by percentage of automatable tasks are fundamentally flawed.

Benedict likens government datasets that try to assign an 'AI exposure percentage' to each profession to the failed 1980s expert-systems approach to AI: breaking a real-world skill into discrete logical steps never captures how the job actually works. He cites taxi drivers (assumed internet-proof in 1997, then upended by Uber) and personal trainers (assumed AI-proof, now plausibly replaceable by a phone camera and an LLM) as cases where intuition about exposure was simply wrong.

ai-job-impact

The only actionable career advice Benedict offers is to engage directly with the tools rather than retreat into moral objection.

He argues that publicly condemning AI as evil provides emotional satisfaction but no practical benefit; the people who will fare best in a tightening market (e.g., law-firm associate hiring shrinking from 100 to 50 per year) are the ones who deeply understand what the tools can and cannot do, not the ones refusing to touch them on principle.

ai-job-impact

Books referenced

Three Men in a Boat - Jerome K. Jerome - Benedict's go-to comic reference book, calls it 'my I Ching' - has a passage for every situation
Nature's Metropolis - William Cronon - economic history of 19th-century Chicago; Benedict cites its standardization, logistics, and network-dynamics themes as directly relevant to tech and AI business models

Media referenced

The Seventh Seal - movie - Benedict recently watched this Bergman classic and loved it, recommends working through intimidating classics
AI Eating the World - other - Benedict's own biannual AI presentation/deck, referenced throughout as the source of most of his frameworks and charts
WWDC 2024 keynote - other - Apple Intelligence segment called out as the most compelling vision of a personal AI assistant Benedict has seen, though Apple couldn't ship it

Companies

Anthropic - cited alongside OpenAI as investing heavily in professional-services and forward-deployed-engineer talent despite AI supposedly threatening consulting jobs
OpenAI - example of a model lab increasing headcount even as its product is framed as automating white-collar work; also cited over its Apple dispute and 'everything everywhere' distribution strategy
Google - using existing distribution (search, Android, Chrome) to push Gemini, illustrating the value-capture-shifts-to-distribution thesis
Apple - Apple Intelligence vision from WWDC 2024 as the best articulated personal-AI vision yet unshipped; dispute over routing AI features through Gemini
Meta - example of an incumbent whose merely-adequate AI model competes well against frontier labs purely on distribution
Amazon - AWS used as the analogy for what foundation models may become: commodity infrastructure with no vendor lock-in advantage, unlike Windows
Microsoft - Internet Explorer vs Netscape browser-wars analogy for how distribution beats product quality, even though winning browsers ultimately mattered less than expected
McKinsey - used with Bain, BCG, Accenture as the example of what buyers actually pay for (judgment, internal politics navigation) versus the artifact (a slide deck) that AI can now approximate
Frame.io - video-editing/collaboration startup Benedict evaluated at Andreessen Horowitz; used as an example that most SaaS ideas didn't need new technology, just someone realizing the problem existed
Every - Lenny mentions having Dan Shipper from Every on the podcast in the context of AI labs hiring humans instead of automating them away
Spotify - used as the example of a product that isn't just 'the old thing but cheaper' - streaming represents a genuinely new category, not a repackaged music store
Uber - the taxi-driver counterexample: a job people assumed the internet could never touch turned out to be completely restructured by it
Airbnb - contrasted with Uber - its actual measured impact on hotel revenue is much smaller than the popular narrative suggests
Fujitsu - vendor behind the UK Post Office Horizon accounting system in the Post Office scandal, used as a cautionary tale about buggy automated systems ruining real people's lives

Techniques and frameworks

Task vs. job framework - Benedict's central lens for predicting AI job impact: distinguish the automatable task (e.g., writing a slide deck) from the job people are actually paying for (judgment, relationships, navigating politics)
Jevons Paradox / price elasticity - making a task cheaper does not automatically shrink total spend or headcount; it can expand demand and total value instead, as happened with Excel and investment banking
Presume radical uncertainty - Benedict's explicit epistemic stance: nobody, including AI lab CEOs, actually knows where model capability or job impact is heading, so treat confident predictions skeptically

Summary

Lenny Rachitsky's guest is Benedict Evans, the independent tech analyst and former Andreessen Horowitz partner known for his biannual "AI Eating the World" presentation. The conversation is framed around Benedict's self-described "controversial opinion" that AI is exactly as big a deal as the internet or mobile were, no more and no less, and that we're currently living through the AI equivalent of 1997: obviously transformative in retrospect, but with almost none of the actual applications built yet and adoption spread unevenly across a population that ranges from obsessive local-model tinkerers to people who've never touched a chatbot.

The bulk of the episode works through Benedict's "task vs. job" framework for thinking about AI's effect on employment. His argument is that some jobs are literally just a task (an elevator attendant pressing a button), which fully automates away, while most professional jobs bundle a task that AI can approximate (a McKinsey deck, a block of code) with the actual thing being purchased (judgment, relationships, navigating internal politics, knowing what to build). This explains a counterintuitive trend he and Lenny discuss at length: Anthropic and OpenAI, whose products are pitched as automating white-collar work, are simultaneously hiring aggressively into professional-services and forward-deployed-engineering roles, because deploying AI inside a real enterprise is itself the kind of multi-month consulting project that requires exactly the expertise Bain, McKinsey, and Accenture already sell. Benedict pairs this with the Jevons Paradox: cheaper tasks have historically expanded total spend and headcount rather than shrinking them, citing Excel not reducing investment-banking hours and total US accounting employment climbing steadily through a century of automation waves.

A second major thread is value capture: does pricing power sit with the foundation model labs, or does it move up the stack to the applications built on top? Benedict is skeptical that OpenAI, Anthropic, and peers will retain durable margins, arguing that with no winner-take-all dynamic among several competing frontier labs, current loss-making token pricing is a temporary disequilibrium that should compress toward commodity, utility-like economics, similar to the global mobile industry's flat stock performance despite exploding data usage. He extends this into a distribution thesis: as products commoditize, incumbents with existing reach (Google, Meta, Apple) can win with a merely-adequate AI product purely on distribution, drawing a direct parallel to Internet Explorer beating the better Netscape through Windows bundling.

On the current anti-AI backlash, Benedict is more skeptical of some claims than others. He cites a Lawrence Livermore Lab study putting US data-center water consumption at roughly 0.017 percent of total national water use to argue the water-crisis narrative is largely fabricated in aggregate, while conceding real localized harm exists. He's more careful on jobs, noting that current economic data shows no clear consensus that AI is driving the youth-employment slowdown, since the same pattern holds across degree-holders and non-exposed fields, but stresses that AI labs release essentially no usage data, leaving the entire debate to academic economists reverse-engineering government statistics and paid consultancy surveys. He also notes that terms like "AI" and "AGI" keep getting redefined downward the moment a capability starts working, quoting AI scientist Larry Tesler's line that "AI is whatever machines can't do yet."

Asked what practical advice to give listeners worried about their careers, Benedict's answer is deliberately unglamorous: don't retreat into moral objection, because it feels good but changes nothing; the people who will do well are the ones who dive in and understand what the tools can and cannot do, since that is now a real hiring differentiator as roles like law-firm associate positions get scarcer. He closes with a lightning round covering his favorite books (Jerome K. Jerome's "Three Men in a Boat" as his personal "I Ching," William Cronon's economic history of Chicago), a recent rewatch of Bergman's "The Seventh Seal," and his collection of roughly 20-30 vintage phones from his years as a telecoms analyst.

Notable Quotes

"AI is whatever machines can't do yet. Once machines can do it, people say, well, that's just software." - Benedict Evans, quoting AI scientist Larry Tesler

"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." - Benedict Evans

"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 Blue Sky and shout at everybody about how evil AI is. 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 it." - Benedict Evans

"The model is just like the dumb thing underneath that powers the feature." - Benedict Evans

"It's probably going to be okay, guys." - Benedict Evans