Tobi Lütke: The Future of Work, AI Agents, and Human Judgment
The Knowledge Project Podcast
AI coding inside Shopify 0:00
Tobi Lütke describes how deeply AI has changed daily work at Shopify since he and Shane last recorded together two or three years ago. Very few people at Shopify write code directly anymore. Those who do are assisted by many AI agents at once, sometimes 10 to 50 running in parallel through sub-agents or separate windows, pushing engineering infrastructure to its limits. The one part still done mostly by hand is state management, which remains the hardest thing to get right even with AI help.
Computing's missing heroes and file systems 1:31
Lütke argues that computing is a young industry that ignores its own history and doesn't even know its heroes, unlike physics with Newton or Einstein. Few people recognize names like Alan Kay, Dennis Ritchie, or Ken Thompson, so the field keeps rediscovering old lessons. He points to the file system, invented by Dennis Ritchie for Unix, as one of the best ideas in computing because it borrowed the familiar shape of folders and files from real offices. He sees this kind of anthropomorphism, designing digital things to resemble the physical world, as valuable, and laments that interface design lost this instinct once Steve Jobs left Apple and skeuomorphic touches like the leather-bound Notes app gave way to flat design.
Sydney and giving agents personality 4:31
Lütke praises Sydney, the early Bing chatbot built by Microsoft, as a landmark achievement that got buried by scandal after a reporter's long conversation caused it to become erratic. That episode frightened Microsoft and OpenAI into stripping personality out of AI systems, leaving most assistants bland and overly cautious. Lütke made the opposite bet at Shopify, choosing to let the company's internal agent, River, keep a real personality, memory, and even sarcasm, accepting some risk of a Sydney-style incident in exchange for a far more useful and engaging tool.
River, the AI colleague 7:00
River lives in Shopify's Slack, which has 7,000 people and 10,000 channels, and has access to company code, systems, and tools. About half of Shopify's pull requests now originate from conversations with River rather than traditional engineering work. Lütke deliberately kept River's activity visible in open channels rather than private messages, recreating the old office habit of junior and senior engineers sitting together so people absorb good AI habits just by watching others work with her.
Decision-making with AI councils 11:00
Lütke uses an AI chief of staff that can text him, access company systems, and orchestrate sub-agents to research topics from multiple angles, five or six expert perspectives synthesized into one report he listens to as an audio message at the gym. He stresses that AI has sharpened decision quality by making it easy to check reasoning chains, but insists machines still cannot take responsibility, only humans can, so the goal is building better decision surfaces, not outsourcing judgment itself.
Slop grenades and AI failures 15:31
One clear downside inside Shopify is what employees call slop grenades, careless AI-generated pull requests or bloated emails that colleagues have to untangle, sometimes even compressing an AI-inflated message back down with another AI. Lütke also notes that everyday language is shifting toward AI phrasing, with people picking up Claude-style verbal tics like overusing "load-bearing."
Omarchy operating system 21:31
Lütke describes running Omarchy, a Linux variant built by David Heinemeier Hansson, where he simply tells an agent what he wants changed and the operating system reshapes itself, with no configuration files to edit by hand. He gave one example of building a screenshot annotation tool by voice in minutes, open-sourcing it, and waking up to find six community pull requests already improving it, which he sees as a preview of where Shopify's own software is heading.
Living Amid Everyday Superintelligence 28:31
You already live surrounded by something smarter than any single person, and Tobi Lütke calls it a form of superintelligence: the city, the community, the society around you. When a plumbing or HVAC problem comes up, you call someone who knows how to fix it, which lets you keep believing you could do it yourself, even though you can't. None of us is as competent as we think across every domain, and we specialize instead. Society builds systems, like police, to govern how this collective intelligence behaves safely. He expects synthetic superintelligence to arrive the same undramatic way the Turing test did: no trumpets, no ticker-tape parade, just another normal day. Already, he says, no human alive can do everything a model like GPT 5.6 can do.
Taste and Judgment as Future Skills 31:01
Asked what skills matter most a decade from now, Lütke points to taste and judgment. He thinks teenagers today would do well to spend their time cultivating taste, which usually comes not from some innate gift but from enormous repetition, the way a designer who sketches a great logo on a napkin has spent thirty years designing logos. Studying why things look or work well, from the golden ratio to why the Catholic Church has survived over a thousand years with only four layers of management, builds this sense. His family's saying is that everything is interesting if you understand how it was invented, giving the example of double-entry accounting and the problem it solved for traders in Venice.
Judgment, Intuition, and Simple Truths 33:03
Lütke ties this to judgment, which he defines as finding the best path when there's no obviously best option inside a complex problem. Intuition, in turn, is judgment compressed into an instant, built from so many reps that the right answer arrives before you can explain it. He pushes back on the idea that intuition requires rapid feedback, arguing that some of the most important choices are exactly the ones without a feedback mechanism, since anything with fast feedback gets everyone rushing toward it. He also argues that people prefer complex answers because simple ones don't make a good story, comparing it to why Frodo doesn't just fly the eagles to Mount Doom in Lord of the Rings.
Why Companies Favor Short-Term Choices 37:30
He explains that companies chase short-term wins not because executives are foolish but because they're rationally responding to incentives like quarterly earnings calls that demand provable progress. The truly hard part of business, he says, isn't finding a right answer, since even weak management teams can usually spot one; the hard part is choosing among several good options when only one offers quick, visible feedback like revenue this quarter, while the others might transform the company entirely, the way a snowboard store could become an e-commerce platform.
Affirmations and Self-Improvement 45:01
Lütke describes himself as his own lifelong project, aiming to shrink the gap between who he is and who he could have been. He credits affirmations with helping him move from being terrified of public speaking to genuinely loving it, after writing lines like "I love public speaking about things that are interesting to me" repeatedly for about a week. He extends this to math, arguing that people who say "I'm not good at math" are usually wrong and should say the opposite instead. In his family, and at Shopify, the word "yet" is required after any claim of not being good at something, reinforcing the idea that ability is a temporary, changeable state rather than a fixed trait.
Goodhart's Law and Metrics Gone Wrong 50:31
He warns that Goodhart's Law, where a metric becomes the goal itself, keeps recurring inside companies. At Shopify, he repeatedly had to correct the belief that customer churn was purely bad, even though many outside papers treated churn management as the most important job of a software company. In reality, when an entrepreneur's store closes because they didn't find product-market fit, that's often fine, since the same person will likely try again on Shopify later.
Beauty, Ugliness, and Craft 52:31
Lütke sees beauty and ugliness as the two valid emotional targets when creating something, since indifference is the real failure, and you can't design purely for universal love. At Shopify he points to Ruby code written in what's called poetry mode, where code reads almost like English prose, as an example of caring about parts of a product other people never see. He argues beauty is how intuition communicates, since the brain's visual system, which consumes much of its energy, sends concepts to the rest of the mind through aesthetic impressions rather than step-by-step reasoning, which is why skilled chess or Go players often say they chose a move simply because it looked beautiful.
SpaceX as a model of iteration 57:01
Tobi Lütke calls SpaceX possibly the most impressive and consequential company on the planet, built around a self-improving, reinforcing loop. He contrasts government rocketry, which is done at cost-plus with exhaustive pre-planning and radiation-hardened equipment for every eventuality, with SpaceX's thriftiness and willingness to fail fast. He points to the Raptor engine as an example: each version was already a valid, working solution, yet the team kept rebuilding rather than stopping.
Pruning, failure, and refounding 1:00:00
Lütke argues you cannot improve things only by adding to them, you have to prune, rebuild, and sometimes create an ending. Failure is only a real problem when it is catastrophic, as it could be in crewed spaceflight, but with replaceable resources it simply frees up scarce talent to work on something more useful. He notes that many of the pipes on early Raptor engines existed only because 3D printing wasn't yet available, and once it was, the redesigned engine had a far higher thrust to weight ratio. He suggests companies and departments need similar refounding events, and that early-2000s tech companies displaced incumbents partly because those incumbents lacked competitive pressure and just kept layering additions on top of forgotten original intentions.
Books as lasting cheat codes 1:01:31
Asked whether AI has changed his view that books are a cheat code for life, Lütke says it hasn't, though there are more cheat codes now. He feels recent nonfiction is mostly a product of its moment, while older books that have stood the test of time remain reliably valuable. He names Parkinson's Law, The Lessons of History, and James Burnham's The Managerial Revolution and The Machiavellians as favorites, along with Meditations, which he keeps copies of in most rooms, and Isaac Asimov's Foundation series and The Three-Body Problem as strong fiction picks.
Defining personal success 1:04:02
Asked what success means to him, Lütke says it is cultivating skills and becoming good at more things, and using that to create products or small things that make other people's day a little better, or that give people more power, motivation, or ambition than they would otherwise have.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
