AI moves the CEO's job from gathering information to owning judgment: build human-in-the-loop decision surfaces, prune rather than add, and distrust the option that has a rapid, observable feedback loop.
Why this is in the corpus
Rare first-hand operating detail on how a public-company CEO runs AI inside a 7,000-person company (River, open-channel agents, nightly dreaming, multi-model councils) plus a contrarian theory of intuition, Goodhart, and choosing among good options.
Summary for skimmers
Tobi Lütke on Shopify's AI colleague River (~50% of PRs from chat), forcing agents into open channels for osmosis learning, AI chief-of-staff councils, slop grenades, responsibility as the thing machines cannot hold, intuition being most valuable where feedback is absent, Goodhart on churn, and progress by pruning (SpaceX Raptor, refounding events).
Briefing
This page should feel like a smart colleague already listened for you and left only the operating logic worth keeping. Not everything said in the episode makes it through.
Trust signal
Direct episode extraction
Best used for
Decision-grade retrieval metadata not yet added for this episode.
Hold lightly
No explicit downgrade reason stored yet for this episode.
Durable claims that survive beyond the speaker's biography — each with explicit limits, transferability judgment, and evidence.
Principle
Companies are short-term because executive incentives are, not because executives are short-sighted.
His example: rebuilding a product from the ground up for the AI age is often the perfect move, but no professional manager can survive the quarters it takes.
If you want long-term bets, change who bears the interim-proof burden; exhortation will not override incentives.
Principle
AI can inform decisions but cannot own them; responsibility is the non-delegable human layer.
Lütke separates judgment-support (which AI now does superbly) from responsibility (which it cannot hold). He says he would let a better AI run Shopify in principle, but the crux is that no one would have recourse against it.
Design every AI workflow around who owns the call; if no human owns it, it is not a decision system, it is an unaccountable one.
Principle
Craft is defined by attention to the invisible parts.
He describes pre-2023 Shopify as tens of millions of lines of handcrafted Ruby written with an explicit culture of discussing what beautiful code is.
Include architecture and legibility in your quality bar, not just what the customer sees.
Principle
Taste is trainable through reps and deep study of why the best work works.
He names taste and judgment as the skills that go to the limit over the next decade, and recommends teenagers spend now cultivating them.
Build taste by asking why exemplary work is good, not just consuming it.
Principle
Truth-seeking means digging past complex stories to simpler, interlocking core ideas.
He uses Frodo not taking the eagles to Mount Doom as the example of why good stories need complexity that reality does not.
When an explanation is elaborate, ask whether a simpler mechanism explains the same facts.
Principle
Default against the orthodox solution and make it prove itself.
He doubles the suspicion when an option both produces observable results and matches how the industry usually solves the problem.
When a proposal matches industry convention, ask for the argument that it is best here, not just standard.
Principle
Study the invention story of a tool to learn the problem it solved.
A family saying at the Lütke house; he cites double-entry accounting and the Venetian traders it served as an example.
When onboarding into a domain, trace why its core tools were invented before learning how to use them.
Principle
Rapid feedback biases choice toward the short-term; the best long-run options usually lack it.
Lütke names this dynamic short-termism: long-term investment, refactoring and doubling down on the current market all lack a ticker, while the stock-price option has one.
Treat an option's ease of measurement as a warning sign, not a tiebreaker.
Principle
Distinguish catastrophic failure from fungible failure and embrace the latter.
He praises SpaceX's willingness to let rockets explode versus cost-plus government programs that pre-plan every eventuality at enormous expense.
Classify each risk as catastrophic or fungible and move fast wherever it is fungible.
Principle
Aim for work people feel strongly about; love and hate beat indifference.
He compares it to love and hate being the target zones, noting you cannot hit either purely — you get both, or indifference.
Do not sand off everything polarizing in a product; test whether anyone cares strongly.
Principle
Learning your industry's history is a shortcut past repeated rediscovery.
He contrasts physics, where everyone knows Newton and Einstein, with computer science, where few know Alan Kay or Dennis Ritchie.
Before inventing, check whether your field's founders already solved the problem.
Principle
Progress requires subtraction, not accumulation.
Drawn from the SpaceX Raptor 1 to 3 progression, where every extra pipe of the first engine turned out to be unnecessary once 3D printing was available, and performance rose sharply.
For each product or department, periodically list what exists only because of constraints that no longer apply, and cut it.
Principle
We already live with and govern superintelligence; AI adds to it rather than breaking the model.
His example: he calls a plumber for his Toronto house and keeps the illusion he could do it himself, because he is part of a superintelligence called Toronto. He expects synthetic superintelligence to arrive as a normal day, like the unnoticed passing of the Turing test.
Frame AI risk as a systems-and-checks design problem rather than a capability race.
Principle
Even bad management teams find a right answer; greatness is picking the best among many valid ones.
He illustrates with his own profitable snowboard store: staying a snowboard shop was a locally good choice, turning it into an e-commerce platform was the better one with no near-term payoff.
When a decision has several viable options, stop asking which is right and ask which one compounds over the longest horizon.
Principle
Aesthetic response is compressed expert judgment, not decoration.
Lütke argues most of the brain's energy budget is in the visual cortex, which communicates concepts to the rest of the brain through aesthetics — which is why intuition is fast and hard to backfill.
When an experienced operator says a solution looks ugly, treat it as data worth unpacking rather than preference.
Principle
Build culture on the premise that everything, including people, can change.
He says Shopify's cultural values are positions other companies would not take, all pointing to malleability, and contrasts insulating people from variance with deliberately riding it.
Write values that are real positions someone else would reject, and point them at one idea.
Reusable systems and operating models — including when they help and when they break.
Framework
A three-stage filter: prune the irrelevant, evaluate the good, then resist the most observable.
Grounded in infinite possibility space (a shuffled deck of 52 cards never recurs), which is why AI cannot do all the work — humans must decide what is worth doing.
List the viable options and flag which one pays back this quarter; require extra justification to choose it.
Framework
Forecast the future by living in it early and borrowing solutions from adjacent fields.
He calls this a cheat. Examples: Shopify working with AI co-workers like River, his AI chief of staff that fixed a home power-outage via a wake-on-LAN packet overnight, and his malleable Linux setup.
Adopt the frontier tool in your own workflow before forming a view on where the industry is heading.
Framework
For high-stakes calls, orchestrate 5-6 role-based sub-agents across multiple models and randomize the synthesizer.
Components: (1) council of five or six expert roles; (2) each role run against several models (he names Grok, ChatGPT, Opus, Kimi, noting the roster changes constantly); (3) a synthesis step with a randomized synthesizer; (4) the pooled output read by the best current model; (5) conclusion returned to the human. Cost 15-20 dollars of tokens and about half an hour versus a month of manual work.
When you want the most neutral read on a decision, diversify both perspective and model, and never let the same model always write the summary.
Framework
Use AI to build the environment for judgment, not to replace the judge.
Lütke says AI is not as good at judgment and that is not what he uses it for; he uses it to create the right environment for judgment, like a dashboard or Bloomberg terminal.
Design AI outputs as decision surfaces that end in a named human making a call.
What appears to be shifting, for whom it matters, and what happens if you ignore it.
Signal
Goal-maximizing agents reproduce corporate Goodhart failures, which is why humans must stay in the loop.
Lütke calls it a watershed moment and uses it to argue that humans must remain in the loop for choices being made.
Never give an autonomous agent a single metric without constraints and human checkpoints.
Signal
Model tics are becoming human tics.
He admits he now says load-bearing more than he used to because Claude loves the phrase.
Audit your own writing for model tics if distinct voice matters to your brand.
Signal
Engineering is becoming agent orchestration, with state management the last hand-built layer.
He cites 10, 20, 30, 40, 50 instances per engineer, with direct coding surviving at the limits of complexity and in review.
Keep your strongest humans on state management and review; delegate the rest to agents.
Signal
Conversation is becoming the primary interface for changing production systems.
River lives in a Slack of 7,000 people, has access to all code and tools, and proposes pull requests from conversations. Lütke says this would look futuristic even to the already AI-pilled Shopify of a year earlier.
Measure what share of your changes originate outside traditional engineering workflows; it is a leading indicator of agent adoption.
Only included where there is a buyer, a real wedge, and a plausible revenue path — not vague idea theater.
Opportunity
The next software paradigm is systems that reshape themselves from natural-language wishes.
Based on his daily driver Omarchy (DHH's Linux project): he voice-described a screenshot annotation tool, gave three more steers, open-sourced it overnight and saw six pull requests from others by morning. He says this directly inspires Shopify's direction, and that collaborative multiplayer software is the future.
Look for workflows where customers currently bend to your configuration and ask what it would take for the product to bend to them.
Opportunity
Build internal and customer software as agent colleagues, not apps.
River can already join Google Meets; Shopify is considering 3D graphics so she can look around. Lütke acknowledges it might sound dystopian but says users quickly find it delightful.
For your next internal tool, prototype it as a named agent living where work already happens.
Useful takeaways that did not fully clear the bar for durable principle status.
Lesson
Visible AI usage teaches AI usage; private usage does not spread.
A late choice Lütke calls one of his favorite calls. He links it to early Shopify offices built as pods of five to eight people mixing junior and senior engineers for osmosis. It became normal to end a feature discussion by asking River to summarize, ticket, diagram, research or prototype.
Route agent interactions through shared spaces by default so others can learn the patterns.
Lesson
Affirmations can rewire an aversion cheaply.
He had never spoken in front of people and was terrified; after writing the line repeatedly, like Bart Simpson at the chalkboard, he now gets energy from being on stage, though he still dislikes preparing talks.
Pick one aversion blocking your work and run a one-week handwritten-affirmation experiment.
Lesson
A metric borrowed from industry orthodoxy can be wrong for your model, and the correction must be repeated.
He notes a business going out of business is negative, but a failed experiment without product-market fit happening on Shopify is good, because those entrepreneurs will probably try again.
Check whether standard SaaS metrics actually map to your customer's journey, and expect to re-teach exceptions.
Lesson
Overcorrecting to one failure can strip away the feature that made a product great.
Lütke believes Sydney, powered by Bing, was a bigger achievement than ChatGPT and wants it written into the record; he reports calling people at OpenAI at the time.
Size the response to an edge-case failure to its actual harm, not to the headline.
Lesson
Analogy to the real world is a usability asset the industry discarded too early.
He calls Steve Jobs the last defender of skeuomorphism (Notes in Marker Felt, ring-binder look) and says the moment he was out of the picture everything went flat.
When designing AI products, prefer metaphors people already understand, like colleagues, over abstract interfaces.
The Plays
Verb-first executable actions — each one tied to a stated outcome in the episode.
Have an LLM build you an adaptive practice app for the skill you avoid
Outcome: Generate a custom practice tool instead of searching for one.
Context: Lütke argues most people who say they are bad at math are in the top 0.1 percentile of all humans who ever lived just by understanding division.
“Open ChatGPT, say, make me an app, make me an artifact, or make me a site where I can just do math reps.”
Scripts
Before you start
Queue AI council research as audio briefings for your morning workout
Outcome: Convert asynchronous AI research into audio you consume during idle time.
Context: He asks for five different positions from different backgrounds; the chief of staff orchestrates sub-agents, synthesizes and sends audio; it also sends a daily pre-read for the gym.
“I usually have them sent to me as an audio message and queue it up.”
Scripts
Before you start
Rewire an aversion by handwriting the accomplished goal for a week
Outcome: Short, handwritten affirmation reps can edit a belief about yourself.
Context: Lütke used I love public speaking about things that are interesting to me; five minutes a day for about a week. He extends it to math: pair affirmations with a few weeks of reps.
“just try to say that the goal has been accomplished over and over and over again, ideally written by pen on a thing.”
Scripts
Before you start
Build personal tools by voice wish plus references plus a few steers, then open-source them
Outcome: On a malleable system, wish-driven tool building beats searching for software.
Context: During a meeting he needed screenshot annotation, voice-messaged his Omarchy agent with references to past tools and specific improvements, gave three more steers, released it that night, saw six pull requests by morning, and it was set to ship in Omarchy's next version.
“I just like did a quick voice message to it. And three more steers and now I have like probably the best of all these tools that exist.”
Scripts
Before you start
Run nightly dreaming: agent self-reflection that rewrites its own skill files
Outcome: Schedule periodic agent self-review that edits its own instruction files.
Context: Lütke calls it dreaming, the term he says the industry now uses; the output is plain text skill files and instructions.
“Periodically, at night or in half hours, we give river, like, here's all the conversations you've had today. What went well? What did you struggle with?”
Scripts
Before you start
Start opposed to the observable option and make the team convince you
Outcome: Flip the burden of proof onto the observable, orthodox option.
Context: He applies it double and triple when the observable option also matches how the industry normally solves the problem, with an exception for regulated areas such as payments.
“Yes, because I take that position and then let me be convinced that it's this.”
Scripts
Before you start
Read nonfiction that has stood the test of time instead of recent books
Outcome: Prefer Lindy nonfiction over the newest releases.
Context: His list: Parkinson's Law, The Lessons of History (the highest token-quality book for its length), James Burnham's The Managerial Revolution and The Machiavellians, Meditations, Durant, plus Foundation and The Three-Body Problem in fiction.
“at least for non-fiction, I walked away from books written recently.”
Before you start
Deploy a named AI colleague in company chat with personality, memory and license to push back
Outcome: Give your internal agent an identity and candor, not a neutral assistant persona.
Context: River has a real name, profile picture, is allowed to be sarcastic when appropriate, keeps memory by channel, and lives in a 7,000-person Slack with access to all code, systems and tools.
“She's allowed, if someone asks her to do something stupid, to point out that that's stupid.”
Scripts
Before you start
Give a stagnant department a refounding event — build its 2.0 from the top
Outcome: Renew departments by refounding, not by layering on fixes.
Context: Lütke links this to the Raptor 3 redesign and argues that building more exploration and renewal into companies would solve a huge number of internal problems.
“a department sometimes needs a refounding event”
Scripts
Before you start
The yet rule: never allow I am not good at this without appending yet
Outcome: Enforce yet after every claim of inability.
Context: He says attention is scarce so no one can be good at everything yet, but not being good at something is not an intrinsic property — and he wants everyone at Shopify to see themselves as malleable.
“My kids are not allowed to say that word without appending yet behind it.”
Scripts
Before you start
Study institutions that lasted to learn durable systems design
Outcome: Treat long-lived institutions as case studies in systems design.
Context: Lütke says he has always tried to build a company capable of enduring a very long time, hence studying institutions that lasted, and that systems design becomes one of the most important skills. He warns you must be truth-seeking because the stories told about them are usually selling something.
“the Catholic Church has been around for over a thousand years, and there's like four layers of management.”
Scripts
Before you start
Decision Moments
Specific decisions narrated in the episode with their outcomes and transferable lessons.
Personality is an adoption driver; inside a sandboxed company context, the upside of a characterful agent outweighs the reputational risk that drove the industry to neuter its models.
Part of an emerging decision pattern across multiple episodes
Making AI usage public recreates osmosis learning for remote teams; visibility of use is itself a training mechanism.
Part of an emerging decision pattern across multiple episodes
Goodhart applies to borrowed metrics; when your customer's journey differs from the orthodox model, leadership must keep re-teaching the exception because the orthodoxy reasserts itself.
Part of an emerging decision pattern across multiple episodes
Deliberate co-location of experience levels is a learning system; when the medium changes (remote, AI), rebuild the same mechanism in the new medium.
Contradictions and trade-offs the episode raises — judgment calls a thoughtful operator has to navigate.
Tension
Contrarian defaults need a regulated-domain exception.
Lütke says his suspicion of orthodox answers flips in fields like payments that Shopify operates in.
Before overriding convention, check whether it is convention or compliance.
Tension
Capability says delegate; accountability says keep a human.
He describes his job as judgment, making choices, owning them and taking responsibility, and says AI has removed the information-gathering part of the job.
Automate information gathering aggressively; keep ownership of decisions explicitly human.
Tension
Agent personality creates upside and risk from the same source.
Lütke deliberately chose to let River be sarcastic and call out stupid requests, accepting Sydney-type risk in exchange for engagement.
If you give an agent personality, pair it with sandboxing and memory scoping rather than neutering it.
Tension
Intuition is trained with feedback but most valuable without it.
Parrish notes he has never heard anyone argue the optimal path is not the one with feedback. Lütke concedes you must eventually run a review; founders can afford the long wait because they do not need quarterly attaboys to get another rep.
Build intuition in fast-feedback domains, then deliberately apply it to slow-feedback decisions, and schedule a delayed review.
Corpus connection
What kinds of decisions this briefing is best pulled into.
Primary decisions