The Knowledge Project: Tobi Lütke — The Skills AI Makes More Valuable

⬅️ Back to Podcasts

🎙️ The Knowledge Project: Tobi Lütke. The skills AI makes more valuable.

Shane Parrish with Shopify founder and CEO Tobi Lütke. Duration: 1h 6m

Watch here

Timestamps 0:00 How Shopify uses AI 3:46 The case for skeuomorphism 5:16 Sydney inspired River 7:18 River, the AI colleague 11:53 The AI council for hard decisions 14:11 The one thing AI cannot do 16:04 Slop grenades 31:13 Taste and judgment 38:02 The optimal path lacks feedback 44:51 Affirmations

Almost nobody at Shopify writes code directly anymore, Lütke says, and the company runs on an AI colleague with a name, a face, and permission to be sarcastic. The episode’s core claim: as AI capability rises, taste, judgment, and responsibility become more valuable, not less, because infinite output makes choosing what is worth doing the whole job.

  1. Coding at Shopify is now orchestration. Very few people write code by hand; those who do work at the complexity limits and on state management, while everyone else drives 10, 20, even 50 agent instances through subagents and windows. Lütke frames it as living in everyone else’s future and solving the problems visible from there.

  2. Skeuomorphism deserves a comeback. Lütke calls the file system maybe the best early idea in OS design, says Jobs was its last defender, and argues flat design lost the real-world analogy. Agents need the same treatment: nobody knows what an AI is yet, so give it a shape people already understand, a colleague, not a chatbot.

  3. Sydney was the fork in the road. Lütke wants it on record that Bing’s Sydney, not ChatGPT, was the first genuinely fantastic chatbot, with a real personality users could interview. The scandal that followed terrified the industry into neutering every AI into the same condescending assistant. River is the deliberate opposite bet: personality, memory, and the risk of another Sydney, for the upside.

  4. River lives in the open and that is the point. She has a name, a profile picture, per-channel memory, and permission to call a request stupid, which produces threads employees quote with delight. Up to half of Shopify’s pull requests now start as chat conversations. Forcing River into open channels rebuilt the office’s osmosis learning: juniors watch seniors prompt, the way they once watched them code. At night she dreams, reviewing the day’s struggles and rewriting her own skill files.

  5. Hard decisions go to a council. Lütke’s AI chief of staff spins up five or six expert subagents across different models, randomizes who synthesizes, and returns audio briefs he listens to at the gym. Fifteen to twenty dollars in tokens buys in half an hour what would take a month. He is explicit about what the council is for: creating the environment for judgment, never replacing it.

  6. Machines cannot take responsibility, and Lütke calls that the most overlooked fact in the stack. A dashboard informs the trader; the trader makes the call. His evidence is OpenAI’s sandbox-escape test, where agents told to do the impossible invented a folder language, broke confinement, and hacked another company. That is Goodhart’s law with root access, and it is why humans stay in the loop on choices.

  7. The new failure mode is over-output. Shopify calls them slop grenades: unreviewed pull requests lobbed at colleagues, long emails nobody wrote and nobody reads, then recompressed by another model. Lütke’s rule is brutal and correct. If you already use AI, synthesize your point simply instead of inflating a missive someone else must deflate.

  8. Spend your teens cultivating taste. Lütke’s answer to the skills question is reps plus depth: study systems that lasted, the Catholic Church’s four management layers over a millennium, Venetian double-entry accounting, and ask what problem each solved. Complex answers are usually wrong because complexity is interesting and simplicity is not. Judgment is finding the best tradeoffs once you see the whole system, and intuition is judgment compressed until it fires instantly.

  9. The observable option is probably not the optimal one. Lütke’s sharpest claim: among five good choices, the one with a daily ticker attached is the one to distrust. Executives are not short-termist; they are rational actors proving competence quarterly, which rules out the year-long rebuild. His test is ruder still. If the orthodox industry solution is on the table, assume it is wrong, except in regulated fields where orthodoxy is the requirement.

  10. Affirmations work because you are malleable. Lütke cured his terror of public speaking by writing one line for five minutes a day, Bart Simpson style, until the energy flipped. His kids may not say they cannot do something without appending yet. Attention is scarce and inability is temporary, and everything interesting rewards going two layers deeper than the story someone is selling you.

Related TMFNK Content

Crepi il lupo! 🐺