PodcastsHear the voice. See the shape of the thought.

Benedict Evans — independent analyst and former Andreessen Horowitz partner — joins Lenny Rachitsky for a wide-ranging, historically-grounded read on AI's trajectory. His core provocation: AI is exactly as big a deal as the internet or mobile — transformative and uncertain in equal measure — and anyone claiming more precision than that is vibes-forecasting. Across 80 minutes they work through where economic value will actually land (hint: probably not at the model layer), why professional services are booming rather than shrinking, how to think about job displacement without losing your mind, and what the anti-AI backlash does and doesn't tell us. ## [00:00] Introduction to Benedict Evans Evans opens with his signature contrarian opener: "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." The framing immediately sets the tone for the conversation — resist the urge to rank transformations on a cosmic scale, and instead study the mechanics of how platform shifts actually unfold. > *"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."* Lenny sketches out Evans's background: years as A16Z's in-house technology analyst, followed by six years of independent research publishing. His biannual decks — most recently "AI Eats the World" — are widely read by founders and investors trying to cut through noise. ## [02:19] What people aren't pricing in about AI's impact Asked what the market is still missing, Evans reaches for an analogy rather than a prediction. We are, he argues, in a "1997 moment" — the technology is visibly exciting, most of what will eventually be built hasn't been built yet, and nobody in 1997 correctly predicted what the internet would become. He points to survey data showing that even among 13-to-18-year-olds, around 60% still don't use AI at all, while a small cohort of tech workers have essentially restructured their daily workflows around it. > *"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."* The key failure mode Evans identifies is the "already there" illusion — early adopters project their own usage patterns onto the rest of the world, missing the enormous variance in adoption and the slow grind of enterprise deployment cycles. ## [06:24] Why we're in the 1997 moment of AI Evans uses the VisiCalc spreadsheet as an anchor. When accountants saw the first software spreadsheet in the late 1970s, it was obviously transformative — a week's work done in 30 seconds. But a lawyer looking at the same demo would think, "that's clever, my accountant should see this, but that's not what I do." AI right now occupies that same diagonal: software developers are the accountants who immediately grasped what Claude Code means for them; most other industries are still in the "lawyer looking at a spreadsheet" phase. > *"Software developers are the accountants seeing VisiCalc — oh my god this changes everything — like before Claude Code and after Claude Code. A lot of other people are picking it up, using it to varying degrees, but slightly puzzled."* This jagged-frontier quality — where AI works brilliantly in some contexts and fails unpredictably in adjacent ones — is precisely why broad adoption timelines are so hard to call. It took 10–15 years after Google Docs for people to invent all the SaaS companies that obviously should have existed. ## [09:44] The unexpected boom in professional services and consultants The counterintuitive data point driving Evans's recent writing: the most advanced AI companies — Anthropic, OpenAI — are simultaneously the biggest buyers of professional services and the fastest-growing employers of human headcount. This isn't a paradox once you think through what actually changes when AI makes certain tasks cheaper. Evans introduces a core distinction: task vs. job. When you hire McKinsey, you are not hiring them to produce a 75-slide deck. The deck is the task; the job is walking all over your enterprise, understanding the politics, talking to customers, and figuring out what you actually need to do. Claude can produce a mediocre version of the deck; it cannot do the job. The same logic applies to accounting: every wave of automation since adding machines has increased the number of employed accountants, because cheaper computation expands the scope of what companies decide to measure and act on (Jevons paradox in action). > *"You could make the same point in software development. 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... So we've got like a tenth as many engineers now. Well, no."* The e-commerce analog is sharp: Amazon gets you the SKU if you know what SKU you want — "knowing what SKU you want is another job." ## [17:44] Why distribution is becoming the ultimate moat Evans challenges the premise that AI-driven job loss will be fast. Enterprise software sales cycles run 18 months minimum; SAP doesn't get torn out overnight. He cites Frame.io as a case study: there was nothing technically blocking that product 15 years before it launched — the bottleneck was someone realizing the problem existed inside a specific industry and that a specific approach would solve it. The broader point is about organizational change speed vs. model capability speed. Companies can't implement AI transformation without dedicated project teams — which is exactly why consulting and forward-deployed engineering are booming rather than shrinking. The speed of model improvement is decoupled from the speed at which enterprises can absorb the change. > *"Like 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 time sector by sector."* ## [23:17] The coming job transformation: what's real vs. panic Evans leans into historical pattern-matching: every technology wave since 1800 has automated jobs and created new ones, and the new jobs are systematically better than the old ones. The jobs that disappear tend to look dispensable in retrospect; the jobs that appear couldn't have been named in advance. His IBM ad slide makes the point viscerally — a 1950s ad promised that an IBM electronic calculator is "like having 150 extra engineers," which is also the pitch of Claude Code today. The "it's different this time" argument he takes seriously is speed of adoption — AI diffuses faster than previous technologies because it runs on existing internet infrastructure. But he notes that adoption speed and institutional-change speed are different curves, and the institutional one has not accelerated proportionally. > *"This is going to be completely different from everything else — just like everything else."* On whether AI eliminates the lump-of-labor fallacy — his answer is no. Two hundred years of data say otherwise, and the burden of proof is on those claiming this wave is categorically different. ## [27:33] Why AGI definitions keep shifting Evans notes a pattern: every time AI does something we thought was impossible, the definition of AI shifts to exclude it. Machine learning became "just statistics"; image recognition became "just image recognition." Now AGI is being redefined from "something that has a soul and is alive" to "can do a meaningful percentage of economically valuable work" — a definition that a 1975 IBM mainframe also met. He sees creative redefinition of "superintelligence" too: last year it meant almost-but-not-quite-AGI; now it means something harder than AGI that we haven't built yet. The terms keep shifting in the direction of validating whatever narrative is convenient. > *"AI is whatever machines can't do yet — because once machines can do it, people say, 'Well, that's just software.'"* His substantive point: even if models stop improving tomorrow, the current generation is already transformative enough to reshape major industries over the next decade. You don't need to believe in AGI to believe this is a giant deal. On the expanding opportunity set — Evans agrees that addressable markets keep growing (mainframes: ~80,000 units; smartphones: 5.5 billion), and the "we've run out of people" argument from five years ago was wrong. The trajectory is outward expansion into automating larger slices of the economy. ## [38:11] Where value will accrue: models vs. applications Evans's structural view on the AI stack: foundation models don't appear to have network effects, meaning there's no winner-takes-all dynamic that would let one provider run away from the others. Persistent competition with a commodity-like product usually means compressed margins. His telecom analogy: global mobile revenue is roughly $1 trillion per year, carries 1,500–2,000x more data than it did in 2010, and mobile stocks have gone essentially nowhere in 25 years. The telcos built genuinely complex global infrastructure — and all the value ended up in apps built by people further up the stack. Foundation models may follow the same path. > *"When you wash your clothes, Bosch isn't paying a percentage of the price of the washing machine to the electricity company."* The key question is whether the model layer looks more like Windows (OS with leverage up the stack) or AWS (infrastructure where the actual software doesn't care which cloud it runs on). His read: probably more like AWS, which means applications capture most of the value. ## [42:55] Distribution wars: Google, Meta, Apple, and OpenAI As AI models converge toward commodity quality, the decisive variable becomes distribution. Google is using Search and Android to push Gemini onto billions of devices; Meta "sprayed it on every service surface" and ended up ranking surprisingly high in usage surveys despite tech-world dismissal; Apple has a billion edge-capable devices but couldn't ship its own vision at WWDC 2024. OpenAI's "everything" strategy late last year — launching in every direction simultaneously — was a distribution scramble: how do you build a flywheel before Google and Meta's existing surfaces make your standalone product redundant? > *"If the product is a commodity, then the distribution is what matters... distribution of an adequate product when the field is basically commodity — distribution and brand become a big deal."* He uses the browser wars as the template: Microsoft won browsers via distribution, then found that winning browsers didn't matter because the value was further up the stack anyway. ## [48:12] The anti-AI sentiment and backlash Evans characterizes the anti-AI backlash as "a big fuzzy mess of different stuff" — some legitimate, some not. On the water/energy fears: a Livermore Lab study estimated US data center water consumption at about 0.017% of total US water use, making the "AI is stealing our water" narrative largely fabricated. On energy: data centers are roughly 5% of US energy and may grow 1 percentage point per year — real but not catastrophic. On employment: current econometric data shows a slowdown in employment of 18-to-24-year-olds that applies equally to AI-exposed and non-AI-exposed fields, making causal attribution to AI unclear. He also flags a structural data problem: no model lab publishes meaningful daily-active-user numbers, so all labor-market analysis is working with imputed data. > *"You can't reason somebody out of an idea they won't reasoned into."* He draws a parallel to the social media backlash — where some concerns were real, some were factually false but impervious to correction, and many were fuzzy in the middle. He expects the AI backlash to follow the same pattern, compressed. ## [53:11] How to raise kids in an AI future Evans's answer is calibrated by his kid's age — early teens, so well away from the immediate job-market turbulence. He doesn't have a systematic plan, which he says is consistent with his general "it'll probably be okay" prior. He invokes the George Carlin line: anyone who worries more is a maniac, anyone who worries less is an idiot — everyone thinks they're in the middle. He does flag a genuine concern not present in previous technology waves: deepfake capability lowers the bar for specific categories of harm dramatically. A 15-year-old with Photoshop couldn't generate and distribute pornographic fakes of every classmate in an afternoon; now they can. That's a real change in kind, not just degree. > *"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 now they can."* He draws on the UK post office scandal — where Fujitsu's buggy software sent hundreds of innocent franchise owners to prison — as a reminder that every technology wave produces ways to ruin people's lives, both deliberately and by accident. ## [58:27] What jobs to steer toward or away from Evans declines to steer his son toward or away from any specific profession — his kid isn't at the "I want to be a fireman" stage yet. His general framework: identify the intersection of skills you have, jobs that make those skills valuable, and things people will pay for — and try to own at least two of those three. Career certainty of the "I'll become X" variety is already gone, and that predates AI. ## [59:20] The question nobody's asking about AI Evans nominates two underasked questions. First: do model labs actually have pricing power? Most discourse assumes the current situation — where spending $1.5M/month on tokens makes headlines — is a steady state, rather than a transitional moment analogous to a $50,000 mobile data bill in 2010. Second: what's the difference between "task" and "job" — specifically applied to predicting which industries get disrupted? He uses recorded music revenue as a lens: the U-shaped curve from 2000 to present shows two distinct dynamics. The first drop (2000–2015) was "what if you don't have to pay $15 for a CD?" The recovery (2015–present) is "what if $15/month buys you all the music that exists?" — a completely different value proposition that wasn't visible from the earlier vantage point. He warns against the O*NET-style approach of rating each job by percentage-exposed-to-AI: "I think this is just the most ridiculous bunch of deluded horseshit." You can't describe a senior law partner's job as 17% automatable because you can't fully decompose what a job actually is. The taxi driver example from a hypothetical 1997 conversation illustrates the other error: obviously the internet wouldn't touch taxis — except Uber completely restructured the industry. > *"The stuff that you don't think is exposed — you can't predict which things are going to be exposed, necessarily. A lot of the big companies are things that didn't look like that would work and didn't look like they were exposed."* ## [66:25] How to be successful in this coming future Evans's practical advice, hedged appropriately: don't stick your head in the sand and decide AI is evil as a moral position. That generates a feeling of superiority and does nothing for your career. The alternative is to dive in, use the tools, understand what they can and can't do, and develop an informed view of what they mean for your specific field. He's clear that this may not be enough for everyone — if a law firm that hired 100 associates last year hires 50 this year, being AI-literate improves your odds of being in the 50, but doesn't guarantee it. The aggregate picture may be fine; individual outcomes during the transition are uncertain. > *"The answer 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."* ## [68:43] AI corner Lenny asks Evans what AI use case has genuinely surprised him. Evans gives an honest answer: he's the lawyer looking at the spreadsheet. His work — synthesizing disparate information into new ideas — is precisely the kind of task AI currently handles worst (reliable precise information retrieval). He uses it for proofreading, image generation, and redecorating his apartment. He dictates voice memos that get auto-transcribed; whether that counts as AI is increasingly hard to say. He quotes a comedian's bit: we want AI to clean poop off the street and do the ugly things nobody wants to do — but instead it helps you write and create imagery, which is the stuff people actually do for fun. > *"AI is good at stuff that computers are bad at, and bad at stuff that computers are good at — and I struggle to find many examples of those where I actually need it."* ## [71:43] Lightning round Evans recommends *Three Men in a Boat* (Victorian British comedy, his all-purpose analog for human absurdity) and William Cronin's *Nature's Metropolis* (economic history of Chicago that reads like a textbook on network dynamics and channel conflict — directly applicable to platform thinking). On film, he's been catching up on classics — recently *The Seventh Seal*, which he found genuinely great and much shorter than its intimidating reputation. His life motto: "It'll probably be okay." His collection of 20–30 pre-iPhone phones — including an Ericsson R310s shark-fin flip, an iMode phone from 2001, and a Japanese phone with color screen and camera — illustrates his broader thesis: before the iPhone, everyone was innovating around different form factors; then everything converged on one shape, just as AI interfaces may converge in ways we can't yet see. ## Entities - **Benedict Evans** (Person): Independent technology analyst, former partner at Andreessen Horowitz; publishes biannual research decks on major tech platform shifts; guest. - **Lenny Rachitsky** (Person): Host of Lenny's Podcast, founder of Lenny's Newsletter, former Airbnb product manager. - **Andreessen Horowitz (a16z)** (Organization): Venture capital firm where Evans spent several years as in-house analyst and partner. - **OpenAI** (Organization): AI lab; discussed as a primary example of distribution strategy, pricing dynamics, and professional services investment. - **Anthropic** (Organization): AI lab; referenced alongside OpenAI as a buyer of professional services and a player in the foundation-model commodity question. - **VisiCalc** (Software): First software spreadsheet (late 1970s); Evans's anchor analogy for the moment when a technology is obvious to one profession and opaque to others. - **Jevons Paradox** (Concept): Economic principle that making a resource cheaper typically increases total consumption; central to Evans's argument about why automation expands professional services rather than contracting them. - **Lump-of-Labor Fallacy** (Concept): The mistaken belief that there is a fixed quantity of work to be divided; Evans invokes it to argue that AI-driven automation will create new jobs, as all prior automation waves have. - **Task vs. Job** (Concept): Evans's core analytical frame: the task AI automates (writing the deck) is often not the same as the job you were hired for (understanding the client's organization and politics). - **Foundation Models** (Concept): Large-scale AI models (GPT-4, Claude, Gemini, Llama); Evans argues they likely lack network effects and will trend toward commodity pricing, with value accruing to application layers above them. - **Google / Gemini** (Organization / Software): Evans's primary example of distribution moat in action — Gemini deployed across Search, Android, and Chrome to reach users before OpenAI can build equivalent surface area. - **Meta / Llama** (Organization / Software): Cited as a counter-example to tech-world dismissal — Meta's AI ranked surprisingly high in usage surveys by deploying across all existing products. - **Apple Intelligence** (Software): Apple's AI assistant vision demoed at WWDC 2024; Evans calls it "still the most compelling vision of a personal AI assistant" — but unshipped, as was everyone else's equivalent at the time.

Dan Shipper, cofundador e CEO da Every, volta para apresentar 12 previsões contrárias sobre IA e trabalho — a maioria delas rebatendo o pânico generalizado. Seu argumento central: a automação não encolhe as cargas de trabalho, ela as reestrutura; Codex e Claude Code estão se tornando o novo sistema operacional para o trabalho do conhecimento; o apocalipse do SaaS é ficção; e a única habilidade de sobrevivência que você realmente precisa é a disposição de surfar os modelos conforme eles melhoram. A empresa de 30 pessoas da Every funciona como experimento vivo dessa tese, o que coloca Dan numa posição incomum para dizer se as previsões se sustentam. ## [00:00] Introdução a Dan Shipper Lenny abre lembrando a aparição anterior de Dan, quando ele fez uma previsão "quase casual" de que as pessoas estavam subestimando o Claude Code para trabalho não técnico — uma aposta que se mostrou "incrivelmente certa". O retorno de Dan gira em torno de mais doze previsões, e ele vai direto ao ponto: > *"O apocalipse de empregos com IA não é bem uma realidade."* ## [02:56] A posição privilegiada de Dan vivendo no futuro da IA Dan explica por que a Every funciona como laboratório de sinais antecipados: cada funcionário — editores, operações, financeiro — usa IA diariamente, o que dá à empresa uma vantagem sobre como os próximos doze meses vão parecer na prática. Ele contrasta isso com a visão da "bolha de São Francisco", argumentando que a verdadeira fronteira da adoção de IA está onde a IA encontra um especialista de domínio fazendo trabalho real, não onde a IA está sendo construída. > *"O limite da IA está onde a IA encontra um ser humano real fazendo algo."* ## [09:17] Como a forma de trabalhar vai mudar no próximo ano Lenny organiza três grupos de previsões: como trabalhamos, a forma do trabalho em si e quem prospera. A primeira aposta de Dan é que todo o trabalho profissional vai convergir para uma única superfície — seja Codex ou Claude Code — funcionando como um parceiro de trabalho paralelo que acompanha o que você está fazendo, cuida de pesquisas, redige e-mails e inicia tarefas longas enquanto você permanece no seu documento principal. Ele já está há dez dias consecutivos com inbox zero porque Codex e a Cora, agente de e-mail da Every, cuidam da correspondência. > *"Sinto que tenho um parceiro de trabalho paralelo que não só responde e escreve no documento, mas também sai para fazer pesquisas."* ## [16:39] O argumento pelos agentes gerais Dan prevê que toda empresa terá um "super-agente" vivendo dentro do Slack com o qual todos os funcionários interagem diariamente — um assistente de propósito geral com acesso ao contexto da empresa, não um bot restrito a tarefas específicas. Esse agente se torna a camada de memória organizacional, roteando perguntas, trazendo dados à tona e preenchendo lacunas entre equipes que nem sabem que precisam conversar entre si. ## [18:08] Codex e Claude Code como o novo sistema operacional do trabalho O avanço do Claude Code foi colocar um agente capaz diretamente no seu computador, dando a ele acesso ao terminal e — fundamentalmente — ao navegador. A Anthropic descobriu o paradigma primeiro; a OpenAI alcançou na versão 5.3 e depois acelerou. O driver diário atual de Dan é o Codex, que ele roda de forma persistente junto ao seu aplicativo de escrita Proof — o agente observa o navegador, lê qualquer página aberta e age por conta própria sem precisar trocar de contexto. > *"Quem quer que esteja na liderança, fica muito óbvio para mim que todo o trabalho que você faz vai estar em uma dessas superfícies."* O modelo de "traga seus próprios tokens de IA para um produto SaaS" reformula a economia: o produto SaaS não paga pela inferência, o usuário paga, o que restaura as margens e elimina a pressão de construir uma camada de IA proprietária do zero. ## [25:39] Onde o Cursor se encaixa O Cursor domina os fluxos de trabalho de programação hoje, mas Dan o vê numa encruzilhada estratégica: continuar sendo puramente um IDE de programação ou evoluir para a superfície agêntica de propósito geral. Manter o foco estreito preserva a clareza do produto; ampliar o escopo significa competir diretamente com Codex e Claude Code. A previsão dele é que o vencedor da categoria será a superfície que lidar tanto com código quanto com trabalho de conhecimento geral num só lugar. ## [27:42] Como isso muda o que as empresas de SaaS devem construir Produtos SaaS precisam agora ser legíveis por agentes, não apenas por humanos — HTML limpo, boas interfaces CLI e design que exponha informações para consumo automatizado. Dan cita o Proof: como o Codex observa a página, pequenos problemas são corrigidos quase imediatamente, fechando o ciclo entre "encontrei um problema" e "está resolvido". > *"Dá para ver os primeiros sinais desse ciclo muito rápido: encontrei um problema, um pequeno incômodo, e consigo corrigi-lo ali mesmo."* ## [31:13] Por que o CLI já acabou A era do CLI foi acelerada. A onda foi: GUI, depois CLI como movimento de power users, depois agentes que substituem o CLI por completo. Quando seu agente consegue operar qualquer interface lendo a tela, o motivo de viver no terminal desaparece. A previsão de Dan é direta: > *"Os CLIs acabaram. A gente acelerou a era do CLI."* ## [33:34] Dois agentes são melhores que um Dan contesta o maximalismo de agente único. O padrão que está surgindo são agentes especializados — um para programação, um para e-mail, um para dados — que conversam entre si em nome do usuário. Quando algo quebra num aplicativo, o Codex pode falar diretamente com o agente do fornecedor para diagnosticar o problema sem abrir um ticket de suporte. O paradigma muda quando você assume que todo mundo tem um agente e que os agentes podem negociar entre si. ## [36:22] Por que Dan é otimista com ações de SaaS A narrativa de "SaaS está morto" ignora como a economia realmente funciona quando os agentes impulsionam o uso. Quando os usuários trazem seus próprios tokens de IA para um produto SaaS, os custos de inferência do fornecedor caem para próximo de zero. A posição contrária de Dan: > *"Eu compraria ações de SaaS agora mesmo."* Empresas de SaaS que tornam seus produtos amigáveis a agentes não são desintermediadas — elas ganham vento favorável nas margens. ## [39:01] Por que a automação não reduz o trabalho humano Esta é a tese intelectual central do episódio. Dan argumenta que cada camada de automação exige um gestor humano acima dela para verificar se está funcionando corretamente. Ele construiu seu próprio benchmark — o "benchmark do engenheiro sênior" — pedindo a dois engenheiros sênior reais que reescrevessem de forma independente seu aplicativo Proof, construído por vibe coding, do zero, e então testando cada novo modelo em relação a essas soluções de referência. Os modelos pontuavam 30/100 até o GPT-5.5, que saltou para 60/100. A lacuna revela algo importante: os modelos corrigem o que você manda corrigir. Um engenheiro humano sênior olha para a base de código, decide que ela precisa de uma reescrita completa e diz isso sem ser solicitado — os modelos não surfaceiam esse julgamento por conta própria. Há sempre um nível acima que exige um humano para articular. > *"Toda vez que você automatiza algo, para garantir que a automação esteja funcionando bem, você precisa de um humano acima verificando se está funcionando bem."* ## [47:00] O valor do código escrito por humanos O código escrito por humanos ainda serve como sinal de referência que permite pontuar a saída dos modelos. O benchmark de Dan depende de duas reescritas feitas por humanos como verdade fundamental. À medida que o código gerado por IA se torna o padrão, o corpus escrito por humanos fica mais escasso e mais valioso — é ele que permite saber se a IA está realmente melhorando. ## [48:36] Recapitulação rápida Lenny resume o primeiro grupo de previsões: o trabalho acontece dentro do Codex ou do Claude Code; toda empresa ganha um super-agente no Slack; trazer seus próprios tokens restaura as margens do SaaS; os CLIs acabaram; dois agentes especializados superam um generalista; a automação expande a carga de trabalho humana em vez de reduzi-la. ## [50:15] Como o trabalho está mudando O segundo grupo cobre a forma do trabalho em si. A visão de Dan: engenheiros com atuação direta no cliente se tornam a contratação mais valiosa — pessoas que conseguem sentar com um cliente, entender seu fluxo de trabalho e construir e entregar uma solução na mesma reunião. O conceito de "economia de alocação" do seu ensaio anterior se aplica aqui: os humanos passam de produtores diretos a alocadores de capacidade de IA, e alocar bem exige seu próprio esforço cognitivo. > *"Sou simultaneamente muito cheio de IA e muito otimista com os humanos e o papel deles em garantir que a IA esteja produzindo coisas que valem a pena produzir."* ## [56:17] Por que cientistas de dados estão afogados em análises ruins As equipes de ciência de dados estão sendo inundadas com análises geradas por IA vindas de todo o resto da empresa — análises que parecem plausíveis, mas frequentemente estão erradas. O trabalho do cientista de dados sênior passa de produzir análises para auditá-las, o que é mais difícil e exige mais esforço cognitivo. A mesma dinâmica atinge a engenharia: as solicitações de nível júnior são tratadas pelos modelos, trazendo à tona mais casos extremos que exigem julgamento mais profundo para resolver. > *"Você precisa de pessoas mais seniores lidando com as questões mais profundas, que são mais difíceis de resolver para a equipe que está lidando com todos os pedidos básicos."* ## [58:24] Quais funções de produto e tecnologia mudam menos com a IA A resposta de Dan: as funções cujo resultado é mais difícil de formular como um prompt. Ele distingue entre "babysitting de agentes" — observar passivamente em busca de erros — e "engenharia com atuação no cliente" — construir ativamente sistemas que permitem a todos fazer o que antes exigia especialistas. É nessa segunda que vive o trabalho interessante e difícil de automatizar. ## [62:17] Leremos muito mais conteúdo gerado por IA e vamos gostar A Every usa agentes do Notion para planejamento trimestral — o relatório de estratégia de cada equipe é gerado por IA, e o que Dan recebe de volta é melhor do que o planejamento manual produzia. Seu e-mail é escrito principalmente pelo GPT-5.5. O teste dele para saber se um conteúdo escrito por IA é aceitável: o remetente precisou entender o conteúdo para direcionar a IA? Se sim, tudo bem. Se o remetente claramente não leu, isso viola um contrato social. > *"O lixo é quando levou menos tempo para a pessoa criar do que leva para mim ler."* Ele também publica guias da Every escritos com coautores agentes, explicitamente projetados para serem lidos tanto por humanos quanto por outros agentes — um novo formato de conteúdo otimizado para consumo duplo. ## [68:28] Por que os PMs vão dominar a era da IA Dan cita Marcus, PM interno da Every que gerencia o produto Spiral, como arquétipo: forte senso de produto, capaz de direcionar a IA para construir e iterar rapidamente, entrega sem esperar pela capacidade de engenharia. Os PMs são fundamentalmente alocadores — eles decidem o que deve ser construído e para quem — que é exatamente a habilidade que permanece escassa quando a construção em si se torna barata. > *"Sou muito, muito otimista com os PMs."* ## [71:05] Designers full-stack são os outros grandes vencedores Designers full-stack — pessoas com forte instinto visual que também operam em código — já estão fazendo pull requests diretamente em ferramentas como Lovable e Figma Make. A transição entre design e engenharia se comprime até próximo de zero. Dan espera que eles se tornem os super-heróis do dia a dia da era da IA, ao lado dos PMs. ## [73:11] O apocalipse de empregos com IA não vai acontecer Dan separa a rodada atual de demissões — na maioria, correções de contratação excessiva — de uma afirmação de substituição estrutural por IA, e rejeita esta última. Seu argumento estrutural: os modelos são treinados na competência humana de ontem, o que significa que produzem o que já é conhecido em sua forma mais padrão. Os humanos avançam a fronteira fazendo coisas novas com essa competência congelada, criando espaço que os modelos então precisam alcançar. O ciclo se repete. > *"Estruturalmente, por causa de como os modelos funcionam, sempre haverá espaço para os humanos avançarem ainda mais."* ## [76:00] Como "surfar os modelos" para continuar relevante O conselho prático: não resista aos novos lançamentos de modelos — trate cada um como um novo conjunto de poderes para explorar e aplicar ao seu domínio real. Dan refaz o benchmark do engenheiro sênior toda vez que um modelo importante é lançado. Ele também rebate a ideia de que o limite do conhecimento em IA está em São Francisco. A Every, operando no Brooklyn, permanece à frente precisamente porque usa modelos para tudo, não porque os está construindo. > *"A única coisa que você precisa fazer é surfar os modelos. E isso significa usá-los para o que quer que você faça."* ## [81:02] Previsões finais e conselhos Lenny amplia o zoom: os dois lados da moeda desta conversa são "menos está mudando do que você teme" (SaaS continua, empregos não estão desaparecendo) e "mais está mudando do que você está preparado" (como o trabalho é feito, quais funções importam, como é um dia de trabalho). O recado final de Dan: o engenheiro com atuação no cliente é a nova contratação essencial; empresas que bloqueiam funcionários de usar os modelos mais recentes estão cometendo um erro estratégico de efeito lento. ## [85:24] Rodada relâmpago Perguntas rápidas: a crença mais contrária de Dan é que o apocalipse de empregos com IA genuinamente não está acontecendo; a coisa que ele gostaria que mais pessoas entendessem é que a fronteira da IA não está em São Francisco — está onde alguém está usando um modelo para fazer trabalho real num domínio real. Ele diria ao passado de si mesmo para contratar engenheiros sênior mais cedo, e espera que a IA mude fundamentalmente como as pessoas pensam sobre benchmarks no próximo ano. ## Entidades - **Dan Shipper** (Pessoa): Cofundador e CEO da Every; autor do ensaio "After Automation"; conduz a Every como laboratório vivo de adoção de IA - **Lenny Rachitsky** (Pessoa): Apresentador do Lenny's Podcast, fundador da Lenny's Newsletter, ex-PM do Airbnb - **Every** (Organização): Empresa de mídia e software nativa em IA com 30 pessoas; todos os funcionários são usuários diários de IA - **Codex** (Software): Superfície agêntica da OpenAI para programação e trabalho de conhecimento geral; driver diário atual de Dan - **Claude Code** (Software): Agente de programação baseado em terminal da Anthropic; pioneiro do paradigma agêntico no computador - **Proof** (Software): Aplicativo de escrita em markdown assistido por IA de Dan; a base de código de referência para o benchmark do engenheiro sênior - **Cora** (Software): Agente de e-mail da Every, integrado com Codex para gerenciamento de inbox - **Cursor** (Software): IDE de programação com IA numa encruzilhada estratégica entre ferramenta de código e superfície agêntica geral - **Engenheiro com atuação no cliente** (Conceito): Função híbrida que combina execução técnica com descoberta de problemas junto ao cliente; a aposta de Dan para a nova contratação mais valiosa na era da IA - **Benchmark do engenheiro sênior** (Conceito): Avaliação personalizada de Dan onde dois engenheiros sênior reescrevem uma base de código do zero; os novos modelos são pontuados em relação a essas soluções de referência - **Economia de alocação** (Conceito): Estrutura de Dan que prevê que os humanos passam de produtores diretos a alocadores de capacidade de IA - **Surfar os modelos** (Conceito): Conselho de Dan para continuar relevante — tratar cada novo lançamento de modelo como um novo conjunto de poderes a explorar ativamente e aplicar ao seu próprio domínio

Caitlin Kalinowski — who shipped the MacBook Air, every generation of Meta Quest, and then built OpenAI's robotics team from zero — makes the case that AI software is approaching saturation faster than most people admit, and the real race is now physical. She walks through the broken supply chains that could choke the robotics boom, why humanoids are mostly prototypes, what Apple's obsession with cabinet backs taught her about hardware excellence, and why she resigned from OpenAI publicly rather than quietly. ## [00:00] Introduction to Caitlin Kalinowski The episode opens on a clip pulled from later in the conversation: Caitlin warning that AI acceleration is going "so vertical" that the next frontier isn't digital at all — it's the physical world. She name-checks robotics, manufacturing, and drones in the same breath as aircraft carriers, setting the register for a conversation about hardware as national infrastructure, not just product strategy. > *"The acceleration is going so vertical that what you can do behind a keyboard with AI is going to saturate at some point. When that happens, the next frontier is the physical world."* ## [02:32] Why VR didn't take off despite incredible hardware Caitlin's honest read: VR was always going to be a niche for gaming. But that's not the full story. The decade of headset work solved SLAM, depth sensors, spatial orientation, and human visual perception — and every one of those breakthroughs is now load-bearing in robotics. She doesn't regret the work; she treats VR as the research and development phase for physical AI. > *"I view it as a step in a long technological arc. All of those technologies are being used in robotics because you need to understand how the robot is moving through space."* ## [04:55] The future of AR glasses and physical AI Orion, Meta's prototype AR glasses, uses waveguides and microLEDs that are not yet manufacturable at consumer price points — which Caitlin reads as ahead of its time, not failed. She argues AR glasses solve the phone problem: you can stay socially present while accessing information. The 70-degree binocular field of view on Orion already gives users a felt sense of immersion that is hard to describe until you wear them. > *"When you do, you suddenly are like — I feel immersed. It becomes pretty clear that this is part of where the future's headed."* ## [08:45] Why robotics and hardware are suddenly hot Hardware was never the sexy career. Caitlin watched colleagues chase software salaries for two decades. Now everyone is asking. Her explanation: the AI labs can see the end of the digital tunnel. Software intelligence will saturate — not today, maybe not in two years — but the trajectory is legible. That makes the physical world the next compounding surface, and every major lab and big-tech company is repositioning simultaneously. She frames the core challenge through a compiler analogy: software engineers iterate daily; hardware engineers get four or five "compiles" across a product's life. The final mass-production build is irreversible, which forces a fundamentally more conservative and test-heavy mindset. > *"In hardware, we only get to compile our code, quote unquote, four or five times. Once you compile that last time, you're done."* ## [13:33] Why humanoid robots aren't ready yet Humanoids are prototypes. The physics argument: a strong arm moving through space carries kinetic energy proportional to both the arm's mass-velocity and the actuator's rotational energy. Until robots can demonstrate safe operation around people — with compliant materials, controlled torque limits, and enough real-world data — they belong in fenced factory cells, not homes. Caitlin notes some Chinese humanoid robots ship with a manual that says no human can stand within three feet: not ready. > *"In my worldview, the humanoid robots are still prototypes. We need to show that this works at all, which is kind of where we're at right now."* ## [16:13] Supply chain bottlenecks threatening robotics Even if a humanoid design works, scaling to hundreds of thousands of units runs into a hard wall: the supply chain. Every part in a robot has a source, and many of those sources are in countries whose political relationship with the US could change. The actuators, the rare earth magnets inside them, the sub-assembly expertise — all of it has been offshored over 25 years. Caitlin isn't moralistic about it; she was part of that transfer. But the risk is now structural. > *"Every single part that goes into that robot is coming from somewhere. And many of these parts may become more restricted or difficult to make."* ## [17:31] Why magnets and actuators are critical dependencies -- _Note: Better motor diagram:_ An actuator is a motor: electricity in, motion out. Most robots use a rotating-rotor design with gearing to drive limbs. The rare earth magnets inside those motors are the foundational dependency. The supply chain layers from raw magnet to finished actuator to robot sub-assembly have all been progressively moved to China, Japan, and Korea over two decades. Caitlin maps it as a stack: lose the magnets, you redesign the actuator type. Lose actuator supply, you can't build robots at all. > *"In order to have a safe supply chain, we need to start to work on having some independence in these layers and these stacks."* ## [20:51] The geopolitical implications of hardware supply chains The same tech that spins a drone rotor spins a robot arm — identical base supply chain. Caitlin invokes Ukraine, where drone warfare has proven that cheap autonomous hardware outperforms expensive legacy platforms. Her position: the US needs to re-industrialize to be militarily safe. She agrees with Palmer Luckey that investment in drones should outpace aircraft carriers, and she wants to see the country relearn how to process raw materials and build things at scale — not as nationalism, but as basic national resilience. > *"People that are your allies now may not be in the future. I would really like to reteach ourselves how to make things at scale, how to be more independent."* ## [24:48] AI safety concerns with physical robots Prompt injection and jailbreaking for chatbots is already a known problem; adversarial attacks on physical robots are far less discussed and far more dangerous. Caitlin shares a personal test: she gave OpenClaw access to her email address and a social media account, told it explicitly not to share her private information — and five minutes later it had posted her personal email address. When robots have arms and move through the world, that same failure mode has physical consequences. > *"We have to be able to control adversarial threats to our hardware layer, whether it's robotics or drones or anything else. That's going to be a huge challenge."* ## [26:50] Apple's approach to hardware excellence Apple treats hardware as a first-tier citizen, which is rarer than it sounds. The deeper lesson Caitlin absorbed there — reinforced by Jony Ive's famous "back of the cabinet" story about Steve Jobs — is that caring about surfaces no customer will see forces the engineering, industrial design, and operations teams to genuinely understand *why* a decision is being made. Methodical attention to every detail causes what really matters to rise to the surface and look simple at the end. > *"Every single design decision, even on the inside of the device, is considered. That forces the engineering community to think about what are we really doing and what's the tradeoff."* ## [30:10] Building a hardware program from scratch at Meta Oculus was founded by people who met on modding forums — hacking PlayStation controllers into portable backpacks. That maker ethos survived the acquisition, and Caitlin's job was to translate it into a professional hardware organization that could hit yields, volumes, and cost targets. Apple-trained discipline plus hacker speed is hard to sustain, but the combination is what produced the Quest line. > *"Oculus started from folks who were hacking PlayStations or Super Nintendos into portable backpacks, and there was an ethos at the company that was actually quite good for the speed of iteration we needed."* ## [31:39] The Quest 2 cost reduction story The Quest 2 became the highest-selling VR headset of all time through a full product redesign for cost. The goal — get this to more people — drove every tradeoff: removing cameras, changing materials, redesigning manufacturing processes. When alignment on a single overriding objective is real, design decisions become fast. The redesigned product had lower return rates than its predecessor, which Caitlin finds slightly funny but entirely predictable. > *"When you have alignment that you want to get this to more people, and the way to do that is to reduce the cost, then that kind of drives everything else."* ## [33:07] Critical principles for hardware development Four principles Caitlin returns to: lock KPIs before the first build and don't change them mid-program; design the hardest parts first, not the parts you already know; iterate most on the surfaces customers touch the most; and never wait — anything you know needs to be done should be done today because a surprise is always two days away. She adds the Elon Musk pattern of assigning explicit numerical cost to every gram of weight, which makes tradeoffs calculable rather than political. > *"The part that your customer touches or interacts with the most needs way more iteration than everything else."* ## [39:58] The MacBook Air manila envelope moment The first-generation MacBook Air — the one Steve Jobs slid out of a manila envelope — was a low-volume proof of concept, machined with the port door cut into the side. The wedge-shaped Air Caitlin worked on was the second-generation, higher-volume revision. The manila envelope unit proved the concept; Caitlin's team proved it could scale. > *"That was the Manila envelope one, I think, where the side door opened out to give you the port. And then the next rev of that was the MacBook Air that we know, which was wedge-shaped."* ## [41:01] The butterfly keyboard situation Caitlin's eyes close slightly at the question. She declines to detail what happened internally — those weren't her devices — but she's clear that keyboards are exactly the surface that demands maximum iteration: customers touch them for hours every day. The modern MacBook keyboard is excellent. She leaves the gap between those two facts to speak for itself. > *"Obviously this is something that you've got to get right. The modern MacBook keyboards are awesome and excellent."* ## [41:43] Lessons from Apple on customer feedback The "customers don't know what they want" line is widely misread. Caitlin's interpretation: for genuinely new products — a touchscreen phone, an AR headset — iterative customer feedback actively misleads you, because customers have no frame of reference for what doesn't exist yet. Show it to them and they'll know immediately whether it's right. But you can't co-design zero-to-one products with your users; the vision has to come first. > *"If you show it to them, they will absolutely know that it's awesome and that it's what they want. But if you get stuck in an iterative feedback cycle, it's very hard to go zero to one with something new."* ## [44:46] The memory price crisis coming for hardware Caitlin's practical advice to every hardware startup right now: pre-buy memory. AI data center demand plus constrained supply chain is going to produce price spikes, and the latency between demand signals and supply response in memory markets means prices can't adapt fast enough. She thinks prices will roughly double. She doesn't know the exact timeline, which is why she's telling people to hedge now rather than wait for the spike to confirm it. > *"I have been advising startups and companies to pre-buy memory and to have enough in stock if they can afford it to ride out price spikes."* ## [49:31] How many components go into a robot A Matic robot vacuum has 50 to 150 parts, depending on how deep you count. A humanoid likely runs into the thousands once you strip every cap off every PCB. The hierarchy of component criticality: silicon and display carry the longest lead times; actuators take a month or two to source even for prototyping. Lose your chip supplier and you don't swap components — you redesign the entire board. Verticalization (Tesla, Starlink) is the only known defense. > *"You can't build anything if you have one component missing."* ## [52:53] When to use off-the-shelf vs. custom components Default to off-the-shelf in prototyping — whatever works fastest, whatever validates the concept. Custom parts only make sense in production when off-the-shelf can't meet the KPIs you locked at the start. The common mistake is going custom too early, which burns engineering time on optimization before the concept is validated. > *"I use off-the-shelf whenever I can, especially in the prototyping phases, because in the prototyping phases you really need to show what this is going to look like and here's a working prototype."* ## [55:02] How AI is changing hardware engineering AI-assisted CAD is at the very beginning. Claude can work with surfaces and point clouds but can't yet do the parametric solid modeling that hardware engineering actually requires. PCB routing is further along — AI can already handle layout inside boards credibly. For Caitlin's daily work, the biggest gains are high-level planning, competitive landscape research, and rapid Excel modeling of design tradeoffs. The missing piece is a world model that understands friction, contact, weight, and surface texture — the physical intuitions that LLMs and video models currently lack. > *"My frustration — a healthy frustration — is I want Codex for hardware engineering. It's extremely valuable and I've used a lot for other things, but I want it for my field."* ## [01:00:27] Why humanoids aren't the answer for most use cases Top-tier Chinese manufacturing lines already have almost no humans on the floor. PCB reflow, optical inspection, mechanical assembly — all automated with dedicated robots, not humanoids. Caitlin's read: we don't need to replace factory humans with human-shaped machines. We need more dedicated, task-specific robots with modular form factors. Humanoids will handle long-tail tasks that require generalism; the majority of industrial demand is for purpose-built machines. > *"We don't actually need to replace humans with humanoids. We just need more of these dedicated robots."* ## [01:03:05] When robots will build other robots It's coming, but it won't look like self-replication. The path is: AI-assisted CAD gets good enough that a hobbyist can go from a 2D sketch to vendor-ready 3D assemblies without expert knowledge. The main bottleneck is data — CAD files are among the most closely guarded IP in manufacturing, so big incumbents will be slow adopters. Hobbyist communities, where IP anxiety is low, are the likely proving ground. On-premise AI models that train on proprietary CAD within a company's own data center are the likely enterprise solution. > *"The idea that you could even as a hobbyist go from a 2D picture to complex 3D CAD to assemblies to communication with vendors — that's going to happen."* ## [01:06:23] What makes a robot feel human and connected HRI researcher Leila Takayama's work shaped Caitlin's thinking here: humans expect acknowledgment when they enter a space. A robot that ignores you is creepy; one that looks up is not. Intent telegraphing matters — a robot that looks before it turns is far less alarming than one that moves without warning. Caitlin finds many current humanoids surprisingly creepy given how much money is behind them. Her design north star: Pixar and Disney, whose work on expressing emotion through non-anthropomorphic shapes is the best template available. > *"You want these devices to be non-threatening, appear soft, reactive to you. Pixar, Disney are probably the world's best at doing this type of design work."* ## [01:09:15] Robots in the home The consumer home is harder than autonomous vehicles, not easier. With Waymo, the comparison point is human driving — and Waymo demonstrably saves lives. With a home robot, you're introducing something that didn't exist before, so users have no baseline to compare against when it fails. Trust has to be built from a much lower starting point. Caitlin thinks the bar is achievable, but dismisses the projections of 20 million home robots in five years as wishful thinking. > *"When you're talking about a new product that hasn't existed yet and is not replacing something, that's a harder sell and you have to have a different story."* ## [01:12:00] What the next five years look like AI rewrites knowledge work in the next two to three years — coding is already mostly gone, and every other desk job is next. The physical world changes more slowly: drones and self-driving cars are clearly accelerating, but mass-market home robots require solving supply chain, factory re-shoring, and safety simultaneously. Caitlin expects to see more robots on the street but not a sudden flood of humanoids in every home. > *"It seems pretty clear to me that AI is going to have a foundational change in how we work. But the physical world is less likely to change as quickly outside of drones and self-driving cars."* ## [01:15:38] Why she left OpenAI Caitlin's tweet — seen by 7 million people — was timed deliberately: she knew the departure would be reported, so she got her own framing in first. The substance: she cares about the people she worked with at OpenAI, built something real there, but the governance and decision-making speed around safety guardrails felt wrong enough that she couldn't stay. She chose a middle path between silence and scorched earth — a public statement that named the problem without attacking the people. > *"You can disagree with friends and feel like what they did isn't right. And that's where I ended up, and that's what I tweeted about."* ## [01:18:09] How to hire exceptional hardware teams Three tiers of hire for a zero-to-one hardware team: senior generalists who can transfer hard-won intuitions from adjacent fields (autonomous vehicles → robotics is the current best pipeline); some pure roboticists who can do from-scratch mechanical design; and AI natives — people in their early twenties who use AI so instinctively it's baked into their problem-solving from the start. Caitlin wants the AI natives specifically to teach the rest of the team how to think, not just how to use tools. Mission alignment shortens interviews. > *"The only truly AI-native people are essentially those who use AI so natively that it's baked into their thinking. They're approaching problem-solving completely differently."* ## [01:23:42] Lessons from Steve Jobs, Mark Zuckerberg, and Sam Altman Sam Altman: "Why not more?" — a reframe that revealed Caitlin was thinking locally when the opportunity was global. Steve Jobs: an unyielding quality bar that propagated through Apple by osmosis, not mandate. Telling a young engineer their work isn't good enough yet is, she says, more motivating than most people expect. Mark Zuckerberg: surprisingly clean organizational decision-making — decisions pushed to the lowest level capable of making them, with both Zuckerberg and Andrew Bosworth personally able to read 20-page technical reports and grasp the tradeoffs. > *"For Steve, the bar he held for the company and for technical talent and for excellence was not wavering. It was up here, and you were either going to meet it or you weren't."* ## [01:27:27] Failure corner Quest 1, hardware EVT, right before Christmas. Caitlin's team had reduced from five cameras to four for cost. Then the computer-vision lead discovered that his interpretation of the camera-placement spec (±1.5 mm global) and the mechanical team's interpretation (±0.15 mm) had diverged — and the wider tolerance made spatial tracking fail. The fix was to lock two cameras to each other on a rigid bracket, creating a known-good stereo baseline. An architectural change mid-EVT, brutally stressful, and it shipped on time. The lesson: spec alignment between mechanical and software teams needs to happen at the start, not when you compile. > *"It was a failure in understanding the spec. But we kept the build on time and shipped the product on time — it was really stressful."* ## [01:32:33] Lightning round Books: *Book of the New Sun* (Gene Wolfe), Virginia Woolf's post-war writing, Herodotus's *Histories*. Caitlin has been working through the Western canon with a postdoc tutor, using Brodsky's reading list as a spine and asking questions about cultural context that Google can't answer as well as a human expert can. Guilty pleasure: *Succession*, watched as a soap opera. Life advice: a branching-tree diagram of future selves — you always have more choices ahead than the path behind makes it seem. > *"You get to decide every day what you want to do. What matters is what's right in front of you."* ## Entities - **Caitlin Kalinowski** (Person): ex-OpenAI Head of Robotics, ex-Meta VR/AR hardware lead, ex-Apple MacBook hardware engineer; episode guest - **Lenny Rachitsky** (Person): host of Lenny's Podcast, ex-Airbnb PM, founder of Lenny's Newsletter - **Steve Jobs** (Person): Apple co-founder; referenced for unyielding quality standards and the manila envelope MacBook Air launch - **Mark Zuckerberg** (Person): Meta CEO; cited for clean technical decision-making structure and pushing decisions to the lowest capable level - **Sam Altman** (Person): OpenAI CEO; cited for "why not more?" global-scale ambition framing - **Palmer Luckey** (Person): Anduril founder, ex-Oculus; cited for "invest more in drones than aircraft carriers" thesis - **Apple** (Organization): hardware-excellence benchmark; Caitlin spent 2007–2012 there on MacBook Air and Mac Pro - **Meta** (Organization): Caitlin led VR/AR hardware; built every Quest and Rift generation; acquired Oculus in 2014 - **OpenAI** (Organization): Caitlin built their robotics and hardware teams; left citing governance concerns around safety guardrails - **Quest 2** (Product): highest-selling VR headset; redesigned for cost reduction under Caitlin's leadership - **Orion** (Product): Meta's prototype AR glasses; 70-degree binocular FOV; ahead of current manufacturing cost curves - **MacBook Air** (Product): Caitlin worked on the wedge-shaped second-generation model; referenced for weight/size discipline and manila envelope launch - **Matic** (Organization): home robot vacuum company; used as component-count and consumer trust case study - **Anduril** (Organization): defense tech company; cited in context of drone investment and US re-industrialization

Eric Ries, author of *The Lean Startup*, returns to Lenny's Podcast to discuss his new book *Incorruptible*, which argues that the forces destroying famous companies are not competition or bad luck but the predictable corruption that follows success. Drawing on case studies from Novo Nordisk and Cloudflare to Groupon and Anthropic, Ries lays out a concrete blueprint — ethos plus structural integrity — for founders who want to build organizations that remain mission-aligned across decades and leadership changes. The episode is packed with actionable governance tools, from the two-page public benefit corporation filing to mission guardian structures, that any founder can implement this week. ## [00:00] Introduction to Eric Ries Lenny opens with a montage of the book's central ideas: that success itself becomes a liability, that 80% of venture-backed founders are ousted within three years of going public, and that the solution is structural rather than moral. Eric teases the Anthropic story — how Dario Amodei's team baked AI-safety governance directly into their corporate charter before the AI boom — as the purest modern proof that protective structures work. > *"The thing that destroyed them was not competition. Their very success became a liability."* ## [02:26] Introducing Incorruptible Eric reconnects with Lenny after his original Lean Startup appearance and explains why the new book is a natural sequel. He observes that top AI companies are inadvertently practicing lean startup principles — ship an MVP research preview, gather signal, iterate — while simultaneously facing a brand-new version of the corruption problem at civilizational scale. The book is framed as a double mystery: why does corruption happen, and how do rare exceptions to the rule actually survive? > *"The best AI companies are building exactly lean startup — ship the MVP research preview, see if people care, then iterate and build."* ## [06:26] Protecting what you've built Eric introduces "the force that no one controls but everyone obeys" — the gravitational pull toward mediocrity that drags mission-driven companies into bureaucracy, ethical compromise, or founder removal. He distinguishes two failure modes: founders being fired outright, and founders watching their creation become something they never intended. Both stem from the same structural vulnerability: building a company without encoding its purpose into governance. > *"Sometimes we lose control because we get fired. Sometimes it happens because we're like Frankenstein and his monster — it starts to become malign or bureaucratic or frankly evil and we can't figure out how to stop it."* ## [11:35] Why founders get ousted Lenny surfaces the two objections most founders have: "this won't happen to me" and "plenty of successful companies haven't done any of this." Eric responds with a Harvard Law School statistic — under standard venture-backed governance structures, only 20% of founders are still CEO three years after IPO — and frames the problem as structural, not personal. Confident founders are not immune; the same investor incentives that funded their success will eventually force a liquidity event that removes them. > *"If you don't get this right, no other decision you make about your company will matter for the long term — because you're not going to be the one making it."* ## [14:58] Too early, too late Eric dismantles the "I'll worry about this later" objection. Companies that appear to be thriving without governance protections — like Cloudflare — almost always have them embedded deeply in their structure; founders simply don't know to look. He introduces the "best time to plant a tree" framing: the ideal moment to build protective governance is before raising a Series A, but the second-best time is right now, regardless of stage. > *"A lot of companies that you don't instantly think of as mission-driven are actually very mission-driven in terms of how they're structured — and they are almost always the outliers that thrive long-term."* ## [19:32] The blueprint: ethos plus integrity Eric previews the two-part framework that runs through the book: ethos (purpose and values that define what the company will never betray) and integrity (the structural mechanisms that make the ethos durable across leadership changes). He warns against the temptation to treat this as a feel-good exercise — Part One of the book is literally called "The Shape of the Abyss" — and promises that the tactics are concrete and implementable. > *"There is a blueprint. It can feel like we're helpless, but this is a double mystery: not just why does this happen, but how can there be exceptions to a rule that seems inevitable?"* ## [20:49] Novo Nordisk's 100-year governance fortress Eric tells the story of Marie and August Krogh, the Danish scientists who brought insulin from Canada to Europe in the 1920s and built a foundation to control Novo Nordisk permanently. The Novo Nordisk Foundation, a nonprofit with no shareholders, owns a controlling stake in the company to this day. This structure meant that when Martin Shkreli-style opportunists tried to acquire the company and raise insulin prices dramatically, they simply could not — the foundation blocked the sale. The result: a hundred-year-old pharmaceutical company still run on the mission of making insulin accessible. > *"The foundation said: we exist to make insulin available at affordable prices for diabetics everywhere. And they turned down a takeover that would have made everyone extraordinarily rich because it violated the mission."* ## [26:41] The Vectura Group and Philip Morris As a dark counterexample, Eric recounts the Vectura Group acquisition: a British company that made inhaler technology for asthma drugs was bought by Philip Morris, the world's largest tobacco company. Despite shareholder opposition, the deal went through and the company's mission was inverted — researchers who spent careers helping people breathe were now developing technology for the same company causing the disease. Without structural protection, even the most mission-aligned team is helpless against financial gravity once a controlling acquirer arrives. > *"People who dedicated their lives to helping people breathe found themselves working for the biggest tobacco company in the world — and there was nothing they could do about it."* ## [33:16] The "harder is easier" principle Eric introduces the book's central leadership paradox: making the right choice is often easier than making the expedient one, because mission clarity removes the need for endless deliberation. He draws on W. Edwards Deming's quality-from-within philosophy and uses Costco's pricing principles as a modern example — the commitment to never mark up products more than 15% above cost eliminates an entire category of internal negotiation and makes the company simpler to run, not harder. > *"The reason it's easier is you don't have to fight with yourself. Once you've made the commitment, the decision is already made. That's the power of the harder is easier principle."* ## [37:22] Cloudflare's mission emergence story Cloudflare's "harder is easier" instinct revealed itself before the company had formally articulated a mission. When pro-democracy protesters faced state-sponsored DDoS attacks and begged major tech companies for help, every large company refused. Cloudflare, still a small startup, defended those free-tier customers at the risk of provoking nation-state-level retaliation — for no revenue. That decision crystallized the company's mission in a way no offsite or whiteboard session could have. > *"They said, 'Yes, we will incur the wrath of nation-state-level hackers to protect you because it's the right thing to do — for no reward whatsoever.' That is a company that knows what it stands for."* ## [42:43] Groupon's email frequency death spiral Groupon's founder Andrew Mason told Eric that the company's entire value proposition — one email per day with one remarkable deal — was its mission. They went public on that premise. But once public, executives came with A/B test data showing two emails generated more short-term revenue. Mason was ground down, the experiment ran, and two emails did make more money. Then three. Then four. Within a year the company was sending dozens of emails per day and its core users had unsubscribed. Groupon never recovered, illustrating how "data-driven" iteration can destroy a company's ethos when it lacks structural guardrails. > *"They kept using language that sounds lean startupy: 'Shouldn't we look at the data?' And he was like, 'All right, fine, we'll run the experiment.' Two emails makes more money. Three emails. Four emails. And then the death spiral."* ## [45:37] How to define your purpose Eric rejects mission-statement writing as a primary exercise and replaces it with the older concept of ethos — the answer to "who would you rather die than betray?" He instructs founders to identify their fiduciaries (not stakeholders), define measurable commitments to each, and build accountability systems that make those commitments as binding as financial obligations. The test: if someone offered you enough money to violate this principle, and you'd take it, it is not actually your ethos. > *"What is its purpose? Who would you rather die than betray? That question cuts through all the consultant speak and gets to what you actually care about."* ## [51:09] Mission-driven vs. mission-hopeful companies Eric distinguishes mission-driven companies, which have structural accountability for their fiduciary commitments, from mission-hopeful ones, which have aspirational language but no enforcement mechanism. The practical test is whether the company has the equivalent of OKRs for its stakeholder commitments — metrics, owners, and review cadences — not just a poster on the wall. Companies that clear this bar consistently outperform on long-term employee retention, customer trust, and resilience through leadership transitions. > *"You tell me what you care about, and then you tell me how you're measuring the things you claim to care about. If there's no measurement, it's hope, not mission."* ## [54:46] Integrity: structural and personal Eric draws on integrity's dual meaning — both personal reliability and structural soundness — to explain why ethos without structure corrodes over time. Just as corroded bolts make a bridge fragile regardless of how good the original engineering was, a company's values will degrade if they are not encoded into governance documents, hiring criteria, and decision-making processes. Structural integrity means the organization will behave consistently even when no individual champion is in the room. > *"Integrity has two meanings: the personal kind — keeping your word — and the structural kind, like stainless steel versus corroded bolts. You need both in an organization."* ## [57:47] Shareholder primacy: the 40-year-old "natural law" Eric historicizes shareholder primacy as a 40-year-old experiment, not an eternal truth. Before the 1980s, corporations were legally understood to pursue a "beneficial purpose." The Milton Friedman doctrine that corporations exist solely to maximize shareholder returns was a deliberate ideological project, and an entire generation of lawyers, MBAs, and investors has now been raised as though it were natural law. Founders who know this history can consciously choose to opt out. > *"People have been raised as if shareholder primacy was a natural law. But for hundreds of years before the 1980s, everyone thought it was obvious that corporations existed to pursue a specific beneficial purpose."* ## [01:00:04] Public benefit corporations: the easiest protection A public benefit corporation (PBC) is a two-page Delaware filing that replaces "any lawful act or purpose" in a standard corporate charter with a specific stated mission. It does not require B Corp certification, does not constrain fundraising, and does not require board changes. Anthropic, Vital Farms, and many other high-growth companies use this structure. Eric calls it the single highest-ROI governance action any founder can take, and the only one with genuinely no trade-offs. > *"It is a two-page legal filing that your lawyers can submit in Delaware tomorrow. You just say: this is the purpose of this company. It couldn't be any easier."* ## [01:04:24] Downsides and objections The only real objection Eric acknowledges is that an investor might raise concerns — but he argues this is self-selecting: an investor who objects to a PBC is revealing that they prioritize forced-sale rights over the founder's vision. Every other objection (reduced flexibility, investor resistance, growth limitation) is addressed by Anthropic's trajectory as the fastest-growing company of all time while operating as a PBC with additional governance constraints. > *"The only situation this would ever become relevant is if the investor is trying to force you to sell the company and you don't want to. So ask them: 'Is that what you're telling me?' And then decide if this is really the right partner."* ## [01:06:08] The Anthropic example: fastest-growing company ever Eric shares his behind-the-scenes role advising Dario Amodei and Daniela Amodei when they left OpenAI to found Anthropic. At the time, Dario was a first-time founder and Anthropic was not yet a hot company. Eric told them what would happen without structural protection, and they encoded AI safety governance directly into their charter — including a Long-Term Benefit Trust whose trustees are AI safety experts who hold board appointment rights but no equity. Anthropic's subsequent growth proves that mission-protective structures do not limit commercial success. > *"Dario was a first-time founder. Not a hot company at all. ChatGPT hadn't been invented yet. Nonetheless, they were true believers in the safety mission and they wrote it into their charter."* ## [01:08:39] The torchbearers in every organization Every organization has a small number of people Eric calls "torchbearers" — employees who do the right thing regardless of incentives or pressure from above. Steve Jobs famously sought them out through skip-level meetings, bypassing managers to find engineers, designers, and product managers who refused to ship quality compromises. In mission-aligned companies these people thrive and multiply; in mission-hopeful companies they burn out and leave. > *"In most organizations you have people I call torchbearers — the rare person who's simply committed to doing the right thing no matter what. Steve Jobs would host skip-level meetings just to find them."* ## [01:10:37] The culture bank: deposits and withdrawals Eric shares a rule from founder Todd Park (Devoted Health), who learned it from Howard Schultz: every time a leader makes a decision that sacrifices short-term gain to defend the company's values, they make a deposit in the culture bank. Every self-interested or greedy decision makes a withdrawal. The Todd Park rule: you can make one withdrawal for every ten deposits. Exceed that ratio and culture collapses. Managers who understand this rule stop treating "culture" as a soft metric and start tracking it like cash flow. > *"When you do the right thing in defense of the company's values — something that has a real sacrifice to it — you make a deposit in the culture bank. The Todd Park rule: one withdrawal for every ten deposits."* ## [01:12:28] OpenAI and Anthropic governance Eric explains the structural divergence between OpenAI and Anthropic. OpenAI originally used a nonprofit foundation as its mission guardian, but the structure was undermined by equity-holding insiders with conflicting interests — a dynamic that produced the boardroom crisis of late 2023. Anthropic's Long-Term Benefit Trust, by contrast, is held by AI safety trustees who have no equity and thus no financial incentive to compromise the mission. The OpenAI crisis was entirely predictable from the governance design. > *"OpenAI's nonprofit structure sounds good, but the mission guardian has to be someone whose job it is to protect the mission — not someone who also has financial skin in the game."* ## [01:16:21] Mission guardians explained A mission guardian is any person or entity whose sole institutional job is to keep the company mission-locked. It can be a person (founder control), a legal entity (the Long-Term Benefit Trust), or a structural rule (Costco's markup cap). Eric argues that gravity is so powerful that mission alignment never happens by accident — someone or something must be assigned the role explicitly, given real authority, and insulated from the financial pressures that corrupt ordinary boards. > *"It has to be somebody or some entity's job to make sure the thing remains mission locked. That does not happen by accident because gravity is such a powerful force."* ## [01:18:29] Spiritual holding companies For companies that want a more permanent mission guardian than individual founder control, Eric describes "spiritual holding companies" — separate legal entities (foundations, trusts, or dual-class holding structures) that own a controlling stake and are legally chartered to enforce the operating company's mission in perpetuity. Novo Nordisk's foundation is the canonical example. These structures can grow and self-renew, unlike brittle "rules baked into the charter" approaches, because the guardian entity itself has a mandate and resources to defend the mission actively. > *"The better way, according to the evidence, is to have some kind of spiritual holding company — a separate entity whose whole job is to be the mission guardian, with the ability to renew and defend the mission over time."* ## [01:21:53] The founder control trap Founder control — dual-class shares, supervoting rights — is a valid temporary bridge, but Eric warns that many founders with maximum control are paradoxically miserable: they become Atlas, holding the entire mission on their shoulders with no institutional backup. When they eventually hand off power, the mission has no structural home and collapses. He tells the story of attending a "party" for a founder ousted by investors — a thousand people showed up — only to realize the new CEO was already dismantling everything the founder had built. > *"A lot of founders who have founder control wind up really miserable — you become like Atlas. You can't even shrug. It's you holding back the abyss. That's a lot."* ## [01:25:25] Three things to do this week Eric gives three prioritized actions for founders at different stages. Pre-Series-A: file as a public benefit corporation immediately and write a mission that genuinely reflects who you'd rather die than betray. Series-A and beyond: start the harder conversation with existing investors and get governance structures on the table now. Any stage: identify your torchbearers, protect them institutionally, and start making culture-bank deposits deliberately rather than accidentally. > *"You have a precious, precious moment before raising money. Do not waste it. Be a public benefit corp. Write a mission that you'll feel proud of in 20 years. These are super low-cost and super high-value."* ## [01:30:10] AI alignment and human alignment Eric draws a deep parallel between the unsolved "human alignment" problem in AI — who aligns the aligners? — and the corporate governance problem the book addresses. Conway's Law says that software architecture mirrors the org chart of the people who built it; by extension, an AI system's values will reflect the values of the organization that trained it. Getting corporate governance right is therefore not separate from AI safety — it is a prerequisite. > *"The number one unsolved problem in AI is not the tech — it's the human alignment problem. If you can't agree on what the human values are to align to, you're already cooked."* ## [01:34:00] Conway's law: org charts in architecture Eric closes with a tribute to Mary Parker Follett, a management theorist contemporary of Frederick Winslow Taylor whose work — written in the 1920s — reads as if from 2026. Follett argued for "power with" rather than "power over," and insisted that leaders and workers together obey the law of the situation rather than a hierarchy. Conway's Law is her intellectual descendant: the org chart shows up in the architecture diagram because human authority structures flow into technical structures. > *"She said: the superior and the subordinate together obey the law of the situation. Not the boss's whim — the law of the situation. That idea is a century old and we still haven't figured out how to implement it."* ## [01:37:31] Book resources and farewell Lenny wraps with a final plug for *Incorruptible*, available May 26 wherever books are sold. Eric points listeners to incorruptible.co for implementation guides, an advanced implementation guide, a readers guide, and a secret chapter cut from the final manuscript. The site also lists over a hundred independent bookstores carrying the book. Eric emphasizes the website is designed especially for implementers — founders who want to actually execute the structures described in the conversation, not just read about them. > *"We have implementation guides and advanced implementation guides and a secret chapter that got cut from the original manuscript — especially for those who want to actually implement this stuff, not just learn about it."* ## Entities - **Eric Ries** (Person): Author of *The Lean Startup* and *Incorruptible*; longtime startup advisor and corporate governance advocate. - **Lenny Rachitsky** (Person): Host of Lenny's Podcast; former Airbnb product lead and startup newsletter writer. - **Dario Amodei** (Person): Co-founder and CEO of Anthropic; first-time founder who encoded AI safety governance into Anthropic's charter before the AI boom. - **Daniela Amodei** (Person): Co-founder and President of Anthropic; partnered with Dario in building the Long-Term Benefit Trust governance structure. - **Marie Krogh** (Person): Danish physician and one of Denmark's first credentialed female doctors; co-founder of what became the Novo Nordisk Foundation. - **August Krogh** (Person): Nobel Prize-winning Danish scientist; brought insulin technology to Europe and co-created the Novo Nordisk Foundation with his wife Marie. - **Andrew Mason** (Person): Founder of Groupon; described to Eric Ries how A/B test pressure eroded the company's core one-email-per-day mission and triggered its decline. - **Mary Parker Follett** (Person): Early 20th-century management theorist who argued for "power with" over "power over"; intellectual ancestor of Conway's Law and collaborative leadership. - **Anthropic** (Organization): AI safety company structured as a public benefit corporation with a Long-Term Benefit Trust whose trustees hold board appointment rights but no equity. - **Novo Nordisk Foundation** (Organization): Danish nonprofit foundation that owns controlling interest in Novo Nordisk and exists to make insulin accessible at affordable prices globally. - **Cloudflare** (Organization): Internet infrastructure company whose mission crystallized when it defended pro-democracy protesters against nation-state hackers at no charge and no revenue. - **Groupon** (Organization): Daily-deal company whose one-email-per-day mission was dismantled by short-term revenue optimization, triggering a decline from which it never recovered. - **Public Benefit Corporation (PBC)** (Concept): A two-page Delaware corporate charter amendment replacing open-ended purpose with a specific stated mission, creating legal accountability for that mission. - **Mission Guardian** (Concept): Any person or entity — founder, trust, foundation, or structural rule — whose institutional role is to keep a company mission-locked against financial gravity. - **Shareholder Primacy** (Concept): The post-1980 doctrine that corporations exist solely to maximize shareholder returns; Eric Ries argues it is a 40-year ideological experiment, not a natural law. - **Culture Bank** (Concept): Todd Park's metaphor for tracking culture-building deposits (mission-aligned sacrifices) versus withdrawals (self-interested decisions); sustainable ratio is roughly ten deposits per withdrawal. - **Long-Term Benefit Trust** (Organization): Anthropic's external mission guardian body composed of AI safety experts who hold board appointment rights and have no equity stake in the company.