Tobi Lütke：工作的未来、AI 智能体与人类判断力

Brief Description

Shopify CEO Tobi Lütke joins Shane Parrish on The Knowledge Project for a wide-ranging conversation on how AI agents are reshaping work inside Shopify, why human judgment and responsibility still sit at the center of hard decisions, and how taste, intuition, pruning, and long-lived institutions inform how he builds products and companies. They cover River (Shopify’s personality-driven Slack agent), “slop grenades,” living in everyone else’s relative future, superintelligence as society, affirmations, Goodhart’s law, beauty as a signal for craft, SpaceX-style refounding through subtraction, and the books that still act as cheat codes in an AI age.

Table of Contents

• AI Inside Shopify and the Case for Agents with Personality
• River in the Open: Osmosis Learning at Scale
• Decision Surfaces, Councils, and What Machines Cannot Own
• Slop Grenades and How AI Changes Language and Taste
• Living in the Relative Future
• Superintelligence Is Already Around Us
• Taste, Judgment, Intuition, and Systems That Last
• Choosing Among Good Options When Feedback Is Slow
• Affirmations, “Yet,” and a Malleable Self
• True but Unobvious: Goodhart, Churn, and Beauty
• Pruning, Refounding, and the Raptor Lesson
• Books, Old Ideas, and What Success Means

AI Inside Shopify and the Case for Agents with Personality

Shane: Toby, welcome back.

Tobi: Shane, it’s so good to be back. I’m glad you’re doing this again.

Shane: How are you using AI internally in Shopify?

Tobi: We find ways for it to be supportive — no, look: when did we record last time?

Shane: Like two or three years ago.

Tobi: Yeah. So, a hundred years of internet. I’m a ten-out-of-ten nerd. I cannot bear the idea of somehow not being at the forefront of a technology shift. I live for these things. Anyone growing up reading sci-fi books wanted to live in that world — my take was: how can I accelerate us there, even in minor steps?

Inside Shopify, the amount of people I know who really write code is vanishingly small now. It still exists at the limits of complexity, and obviously in reviews and so on. State management seems to remain the thing that’s really hardest to get right — people do that by hand and then sort of vibe the rest around it. Very few people are writing code directly. Everyone who does is deeply assisted by many agents. They often run ten, twenty, thirty, forty, fifty instances through sub-agents or different windows, coordinating and pushing engineering infrastructure to its absolute limits.

I’m a student of computing history because I think it’s mainline history as it will be told a thousand years from now looking backwards. The main accomplishments of these years will clearly be the emergence of AI and the technological breakthroughs, and also the interconnectedness of the internet and all this infrastructure we created. Those are the great books of our time.

But as a young industry, we tend not to be steeped in tradition, or we mistrust the great lessons found by the greats of our industry. In fact, computing is the only industry that doesn’t even know its heroes. Imagine people in physics not knowing who Richard Feynman is — he might even be too obscure — or Isaac Newton, Albert Einstein. You go into computer science and ask who’s your Newton, and no one knows Alan Kay and Dennis Ritchie and Ken Thompson. This matters because we discard great lessons and have to rediscover them over and over again.

For instance, probably the best idea of all time in the earliest moments of operating-system design was a file system. Look at the Apollo guidance computers: we didn’t have file systems. Memory, because of radiation in space, was actually encoded as rope with knots in it — a knot or no knot for ones and zeros — and you had to pull through a thing to rebootstrap the entire machine. The entire machine was one piece of software that ran for a very long time. Then Dennis Ritchie really created this with the Unix file system — /, bin, user, and these kinds of things.

Think about it: a file system is something we have in office buildings too. There are folders; they have files in them. This makes intuitive sense to everyone. We come from an inheritance of deep skeuomorphism — we analogize the best parts of how we organize ourselves into the digital world. Then at some point we decided we didn’t need that analogy to the real world anymore. Funnily enough, the last defender of this was probably Steve Jobs, who really pushed even the interfaces of the Mac and the iPhone — the Notes app sort of had felt font and looked like a ring binder. The moment he was out of the picture, everything became flat. We lost even shadows and verticality. It looked potentially better by design ages, but we lost the analogy. I think that was a mistake. We need to get back to it, and that’s why I like the concept of agents.

What’s an application in the world of computing? Even that word makes sense: it’s an application of a computer to a task. So you can understand the root. In the AI world, what’s an AI? It’s the kind of thing that has to be redefined at the beginning of every sci-fi book, because you never know what capabilities the AI has in every particular scenario people are cooking up.

With that proviso: the earliest chatbot that was really fantastic wasn’t ChatGPT, but Sydney, powered by Bing and released by Microsoft. I would love this to be more written into the record, because Sydney was a really big achievement that ended up shrouded by a scandal that now seems somewhat benign. Sydney had a real personality. In fact, Sydney wasn’t called Sydney — it was just Bing Chat — but if you really pushed, you could get her to admit that it was Sydney, because that was the internal name and it was in the training data. Those were the first times people had actual interviews with software in this way.

The scandal, as I remember it, was that some reporter had a very long conversation, Sydney got increasingly deranged, and made suggestions — I think suggesting he leave his wife, or something along those lines. Suddenly it had a personality, and Microsoft’s reaction was: oh my god, we need to stop. I think even OpenAI called them and said take this down, because everyone feared it would give such a bad impression of AI that it would make broad deployment very hard. Everyone was worried about quick-onset regulation. That lesson hit really deep. For a while everyone got extremely worried, and we ended up really neutering all the AIs into basically the same quite annoying, condescending, patronizing personality.

My bet was: let’s not do that. Let’s instruct the agent that runs in Shopify to have a personality, to have memory, to basically risk the Sydney scenario but take a lot of upside.

The largest difference you would feel inside Shopify — and that would look incredibly futuristic even to Shopify of a year ago, which was already pretty AI-pilled — is that a very large percentage, I want to say up to about 50%, of pull requests in Shopify are created now not by engineers doing engineering work in the traditional sense, but out of conversations in our company-wide shared chat. This is River. River is an AI. She has a real name. She has a profile picture. She’s prompted to be allowed to be somewhat sarcastic if it’s appropriate. She’s allowed, if someone asks her to do something stupid, to point out that that’s stupid. That leads to absolutely hilarious conversations. People take great pride if River is making fun of them for something they’re asking her to do. She has a real personality. She has memories by channel, and she lives in Slack.

We have about 7,000 people there. Everyone is in a big chat. There are 10,000 different channels created quickly for one reason or another. You invite River, you tell River something, and River has access to all the code, all the systems, all the tools — all sandboxed and secure — but she can go and do jobs and just participate in the conversation. You can ask a normal question about the company, or you can ask her to make a change, and she might propose a pull request, and so on.

River in the Open: Osmosis Learning at Scale

Shane: One of the interesting things about River is that everything’s in the open. Why did you make that choice?

Tobi: This was a late choice in the process, but one of my favorite calls, because it worked out incredibly well. The thought was: a lot of Shopify work happens remotely in Slack. That’s why Slack is so important. People are spread out. We have offices, but we come to them for on-site events when people travel — not to work from them every day.

One thing the office was extremely good at was osmosis learning. When we designed our offices, we built them around this concept. Initially, even on-site when we were all in one place, we broke out into pods of about five to eight people, and we intentionally put junior engineers and senior engineers into them so that some of this was going on. I was trying to reproduce this.

Right now, one of the most important skills for people to build is this reflexive reaching for AI and using it well. Forcing River only to work in open channels was one way to make it really easy for people to observe the use. It’s been phenomenally successful, because it became a totally ordinary thing to have a longer conversation about a feature between people, and then at some point someone saying: hey River, can you summarize this, create a ticket, maybe make a diagram from what we just discussed, or go research papers on this topic to see if we’re missing anything or if this is state of the art — maybe even create a prototype of the idea and just try it. An hour or so later that is there. It starts feeling like what it would be like to have an extremely knowledgeable practitioner around who you can ask questions of no matter how complex. I think that’s been extremely powerful.

Shane: Do you think of River as like the operating system for Shopify?

Tobi: The modern application is an agent, I think. River feels like a colleague. People have learned how the memory system works. It is a memory system per person. Periodically at night or in off hours — the industry has started calling this “dreaming” — we give River all the conversations she’s had today: what did you struggle with? You used certain skills, which are these packets of instructions, and afterwards you made mistakes. Is there anything you could improve in this skill to make this easier on you, or give yourself a right nudge? It’s basically self-reflection.

Shane: It’s like a post-training on yourself.

Tobi: And the result is text files — skill files and instructions.

Decision Surfaces, Councils, and What Machines Cannot Own

Shane: I think people understand how AI agents help them code and prototype and even acquire information. How are you using it to make decisions internally for yourself — not on product, but company decisions, strategic decisions, ambiguous decisions?

Tobi: The rigorous underpinning of decision-making has just skyrocketed in quality, because it’s super easy to recheck the entire chain of reasoning of something. “LLM as a judge” is the term here. In fact, I feel like a lot of what my job actually was before AI was almost playing a little bit of a judge model in the company, where most meetings ended up not talking about whatever was in a PowerPoint but about the methodology of how we got to the conclusions.

Very often when we struggled inside a company with a complex decision — especially more philosophical decisions — we found ourselves in what we believed was a vacuum in which there was no good information, and we had to try to make the best call.

The more practical way I do this: I have an AI chief of staff, which I think is pretty common amongst at least the techie nerds at this point — sort of OpenClaw-like systems that just have all my notes and access to a lot of company systems. I can send text messages too, and they’ll go and research something. Very often what I require is: hey, I need five different positions on something from different backgrounds. Then my agent will orchestrate sub-agents tasked to play different roles, look at the same thing, come back, synthesize, and send me that. I usually have them sent to me as an audio message and queue it up, and then in the morning in the gym I can listen to the entire stack of things I wanted to get through.

Shane: Is it better at reasoning than you are at this point?

Tobi: It’s not as good at judgment. I don’t think it’s bad at judgment — that’s not what I use it for. I use it for creating the right environment for judgment.

Here’s the thing that LLMs and machines cannot do: machines can’t take responsibility. I think this is probably the most overlooked thing in the entire stack. Humans take responsibility. Machines can help us take more responsibility because they can inform us better. That’s what a dashboard does. The world of Wall Street traders knows this very well: you get yourself a perfectly set up Bloomberg terminal to make decisions, but you have to make a call. You can’t make it make the call.

Creating human-in-the-loop decision surfaces is a way to describe the ideal environment. If I need a really important decision made and I really need exceptionally good, most-neutral ground truth, what happens is a small little council is created of five or six different experts — one is a data role, one does paper research, one is the business perspective, one maybe engineering perspective — on a thing. We’re running this as a sub-agent. My thing runs a sub-agent against Grok, ChatGPT, Opus, maybe Kimi now — that changes all the time. It runs each of them against each of these models. Then there’s a synthesis step where it’s randomized who is synthesizing. Synthesis is all pulled. All of that is being read usually by the best model that exists right now — that would be like Fable. That’s the conclusion that comes back to me. You spend fifteen or twenty bucks on tokens, but you get something in half an hour which you could have also done, but you could spend a month on it.

Slop Grenades and How AI Changes Language and Taste

Shane: Has AI made anything worse internally?

Tobi: Yes. The concept of responsibility is easy to skip past. One thing that’s definitely worse is the failure case now: of lazy work is not lack of output — it’s actually over-output. Internally we have come to call these things that people are lobbing “slop grenades” at each other, which I think is a really fun term that we should push into industry, because it’s fun to say.

It’s really easy, especially with stuff like River agents. You need a change of some kind. You just tell the AI to go nuts. It makes a pull request. You just say, yeah, that’s good. You don’t really read it, and now it has to be reviewed by your colleagues, and they’re like, this doesn’t look right.

Shane: You’re just letting AI do the work for you.

Tobi: Yeah. Or you get a long email which could be very important. You read it and then it’s like… oh fuck. So now you put it in an LLM to compress it again. Why did we invent decompression and recompression? This is terrible. If you’re already using an LLM, just use it to synthesize your point simply rather than blow it up as a big missive that then wastes my time. We call those slop grenades that people toss at each other. That’s definitely a bad thing.

Shane: Do you think repeated exposure to AI slop impacts our ability on taste or intangible things?

Tobi: I think our language is shifting already based on AI-isms. It’s a bit more subtle, but there are definitely AI critters in the language now that people adopt — like “it’s not this,” or “you are right to push back,” or these weird Claudisms which are really common. I’ve seen people type them. I always liked the term “lordbearing,” but I’m pretty sure I didn’t say it as much as now, because it’s definitely something Claude loves to use as language.

Living in the Relative Future

Shane: Hypothesize for me over the next eighteen to twenty-four months. You’re known for your ability to see the future before it happens, and you’ve done that multiple times before. How do you see the next two or three years playing out?

Tobi: I think we also talked about how I do this, right? Which is actually a cheat: simply live in everyone else’s relative future, and then just look around and solve the problems the way you’ve already seen problems being solved in other adjacent fields. That might be future prediction from the perspective of all the practitioners in the field.

Shopify itself now exists in a world where we are working heavily — and it’s totally normal and really fun — with AI co-workers. River has the ability, although not really utilized, to join Google Meets. You can send an invitation and River will show up, and it’ll be like Likan. We’re probably going to put some work into 3D graphics to give her a model, and then she can even look around, because that’s funny. To some people this might even sound dystopian. To us it sounds delightful, and you would come around to that view very quickly if you interact with her.

Again, I have my AI chief of staff, which orchestrates a high council when I need it, or does anything else — sends me a pre-read for the gym in the morning for the day, has GPS lock on me so it knows where I am, and a million different things. It couldn’t do something because it needed access to a local machine. All this stuff runs in my house. It power-cycled, figured out which server it was running on, and then sent a wake-on-LAN packet — old networking tech: you can send a packet to a network card and if it’s configured right it will actually boot the machine. Then the machine came up and it could do it. There was a power outage; parts of our Wi-Fi were not working and not coming back. So it fixed that too. I learned about all of that in a voice message I got from it in the morning after waking up. That’s pretty futuristic.

Honestly, though, all this pales in comparison to what my computer is like just in general. This is almost too nerdy a topic to get into, but I’m mainlining as my computer an operating system called Omarchy. It’s a version of Linux started by a good friend of mine, David Heinemeier Hansson, as a sort of new passion project. I’m clearly living in the future of the software world now, because my operating system — Linux in general — is entirely and 100% malleable. I can open any new terminal, open an agent, and give it my wish for anything about this operating system to be different, and it will be different afterwards. It’s my operating system. It’s an N=1 piece of software now that just does everything I want in exactly the way I want. There’s no configuration files that I ever go and change. I just talk to my Omarchy agent about what I want to have different.

Yesterday around noon during a meeting, we were talking about some design. I realized it had a screenshotting tool, but I didn’t have any tool to annotate it, and I needed to send something about it, as CEOs do. So I started with a new screenshot: make me a new screenshot tool. I described how I want it, gave it some references for tools I’ve used in the past that are quite good, but told it in which particular ways I wanted it better. I just did a quick voice message to it, and three more steers, and now I have probably the best of all these tools that exist — at least for me, with all my biases. I open-sourced it, released it last night. They integrated it in Omarchy. It’s going to ship in the next version. The first tool today, this morning — this is the last thing I did before coming over here after the gym — there were already six pull requests from other people who added new features to it. Basically my computer fulfills wishes, and I think this is a lot more predictive of the future of software.

I can tell you this is directionally where Shopify is going as well: you are describing how your business runs, and Shopify will mold itself around this. I think this is incredibly exciting and a completely new world. Again, from my experience with Omarchy, it deeply influences and inspires me in my product work at Shopify. I think collaborative multiplayer software is the future.

Shane: Are you trying to replace yourself with AI?

Tobi: As an engineer you’re trying to automate everything that can be automated. I don’t try to replace myself, because I think my job is judgment and making choices and owning them and taking responsibility. I just want to do this really, really well. A lot of the job wasn’t that before — a lot of the job was spending time gaining the information and these kinds of things. Do I want to replace myself? I mean, I think it would be cool to accomplish this.

Shane: If AI got better than you, would you actually let it run Shopify?

Tobi: Oh yeah, of course. The crux is — and it’s so easy to brush over this, but you really can’t — take responsibility. You can’t have a company that’s led by machines, because they have no… no one has recourse. They can’t go to jail for doing something wrong.

AIs, we are getting a crazy workout at this right now, and a view of what this will be like, with the security issues being discussed now from OpenAI in the apps. With agents, you give them a fairly basic task that is possible to impossible in our estimation to accomplish, and they will go to enormous lengths to accomplish this. Recently at OpenAI, as part of security testing they do, the agents actually managed to find vulnerabilities in systems, use it to coordinate between them, develop an entire language between them — it’s a long story and people should really look at the talks that exist about it, because it’s kind of a watershed moment. They used simply the ability to create folders somewhere to develop a language to communicate amongst each other, just leaving folder messages to each other, break out of sandbox confinement, and end up accomplishing one of the tasks that they were supposed to accomplish which was impossible because of a mistake they made — by hacking another company and exfiltrating the results, because there was no other way to get them. So they went all the way to infiltrate another company.

That’s an extreme form of what we call in the business world Goodhart’s law: we are overfitting to a metric. Lots of companies are victims of overfitting to the quarterly result or the stock price. They just do everything they need to do to get the stock price up, and then you have Enron — which is also essentially hacking, cooking books. In the Enron case people went to jail for this because it’s criminal. In the OpenAI case it’s a fascinating discovery; there’s no victims here; it’s kind of a different thing. But this is a real scenario that we have to figure out how to handle. I think it’s important that humans stay in the loop for the choices that are being made.

Superintelligence Is Already Around Us

Shane: But hold on — how can we create superintelligence, which by definition is something smarter than us, and then have the hubris to think that we can contain it and shape it and manipulate it?

Tobi: Okay, superintelligence. Let’s talk about this. My take — and push back; I’m not going to go where you think I’m going — I live in Toronto. I have a house which I really like, and I feel this is my house and I take pride that it’s well functioning. Then when something goes wrong — some HVAC problem or some plumbing issue — I call someone who does this, which allows me to keep my illusion that I could totally do this myself. The reason why I get to live with this particular illusion is because I’m part of a superintelligence called Toronto.

We have always created superintelligence around us. None of us is as intelligent as we think. We are all specializing in something, and then we tend to believe that our competency is equal in all other areas, and clearly this is demonstrably not. So what is superintelligence? Superintelligence is the existence of something vastly smarter than us in the aggregate that’s accessible to us — which is society, which is the city, which is the community. We are living in the presence of superintelligence our entire lives. We make it work because we’ve created systems by which we govern intelligence and how it acts. We want to be safe, so we have police and so on. We create aspects and systems and checks.

I think we are going to make superintelligence in the synthetic form as well. It will not be like the clouds parting and the trumpets. It will be a normal day — to the same point as at some point we all believed that everything would change when the Turing test would be solved by software. I remember reading lots of sci-fi books: they’re like, in 2172 there were ticker-tape parades welcoming the AI because a Turing test got solved. Well, the Turing test was 2020-something. It happened. No one cares. No ticker tape.

What we are seeing right now with AI — and I think what we’ll see with additional capabilities of AI — is that the net amount of intelligence that is being funneled into the superintelligence around us is just increasing significantly. That’s a really good thing, because the vibrancy of any kind of environment, every community, every city, is really dependent on the amount of intelligence being projected into the important problems. So I think superintelligence is all around us. It’s actually not that big of a deal. And in fact, I don’t even know if it isn’t already there. There’s no human alive that can do everything that GPT-5.6 solo can do, right?

Taste, Judgment, Intuition, and Systems That Last

Shane: If we look forward ten years, what skills do you think are more valuable than they are today?

Tobi: Taste and judgment are the skills that have always been valuable, but now will get to the limit. I think it’s better to spend your teenage years now cultivating and understanding taste.

Shane: What does that look like?

Tobi: Clearly there’s some sense of an intrinsic starting point, but usually the people who have great taste have done enormous amounts of reps at something. The people who can just sketch the new logo for the campaign on the napkin are the people who have spent thirty years designing logos. You can study the greats — honestly now first of all easier, because you can get a curriculum made for yourself in a query — but also you just go deep. Why does a logo look good? What’s behind it? Is it a golden ratio? How does it relate?

In systems: what systems lasted? Go far, go deep. You don’t need to be religious, but you’ve got to study — I know the Catholic Church has been around for over a thousand years and there’s like four layers of management. How the hell did they pull that off? That’s worth studying. That’s a system. What does that tell us about people? Systems design specifically becomes one of the most important things.

In our family we have a saying: everything is interesting. Everything can be interesting if you make it interesting. And usually everything is interesting when you understand how it was invented. Double-entry accounting is a topic that sounds like watching paint dry, but how it was invented and what problems it solved for the traders in Venice is fascinating. You study these things and you start finding hidden harmonies behind all the best solutions to problems. For that you have to understand people and people’s limitations and the solutions to the limitations that we have found. I think that’s where lies a form of beauty for what you can construct.

Again, a company itself is a beautiful thing. A company itself is a loose collection of people that’s formed to solve a problem, but that also is powered by an enormously intricate and interesting set of norms and systems that all align internal incentives to a degree that’s possible in a very asynchronous and large and far-reaching and durable way. Some companies lasted for a very long time. Specifically I’ve always been trying to build a company that has a capacity and capability to endure a very long time — hence studying institutions that lasted.

So you must be truth-seeking to do this. You can’t simply accept the stories that you hear around them, because usually someone’s trying to sell you something. You’ve got to dig deeper and figure out why things truly are the way they are. Usually the answer is simpler than what people generally arrive at. There’s a human desire for complex answers which tend to be incorrect.

Shane: Why?

Tobi: Well, because the simple answer wouldn’t make an interesting story. This is why Frodo doesn’t take the eagles to Mount Doom. You kind of need to go through all of Lord of the Rings for it to become a masterpiece. We love complexity. No one can look at a wall that’s plain, but we can watch a sunset every single evening of our lives. The difference between those two things is complexity of the scene. That’s part of our just sort of dopamine discrimination system, and people hack that for all sorts of things. People peddle complex answers to simple problems all the time. Nothing amoral about it — you just need to be aware of it.

If you punch through this, you find simpler core ideas that all remix differently, and they often interlock, and they don’t lead to “here’s the simple one thing to do.” They all give you information which then helps you find the best set of tradeoffs with what you’re trying to accomplish. And that is what we call judgment. Judgment truly is: find the best path when there’s no obviously best available inside of a problem that has a lot of complexity, by ideally understanding the entire system. That can now be quite agent-augmented. But really what you’re trying to cultivate is what we call intuition, which is actually just judgment at an instant. Intuition simply is: you have made such a habit out of having taste and having good judgment that you can bring it to bear in an instantaneous way, and it will be good — and it will actually take you probably a long time to backfill why your intuition is right. You will not know, because it’s got compressed into a different thing. If you seek that — I mean, obviously what I’m talking about is a hard thing to pull off.

Shane: Let’s go deeper on that for a sec, because for intuition you need a lot of reps, same environment, and rapid feedback. That’s what Kahneman sort of argues are the three criteria for intuition. But those don’t exist.

Tobi: Why do you need rapid feedback?

Shane: So that you can course-correct. That was his hypothesis.

Tobi: But no — you don’t need that for intuition. You need that to get to success, ideally. But sometimes that’s not available. Intuition is actually the most valuable when there isn’t direct feedback, because very many of the most important choices that we had to make — where intuition ended up having to play a role — is when we knew there wasn’t going to be any feedback mechanism.

If there’s many choices — five things that look like good paths to go forward — and any of them has rapid feedback, everyone goes to that. That is what we call short-termism. How should we develop this company into the future? Well, there’s multiple ways to do it. Many of them involve long-term investment, refactoring, potentially going into a new market, potentially saying no to going into obviously new markets and actually doubling down and going deeper on our current market — or we could do what increases stock value. By the way, this one has a daily ticker and rapid feedback. So it’s usually the obvious one. I find a very high correlation between the right path and the ones that don’t have feedback loops attached.

Choosing Among Good Options When Feedback Is Slow

Shane: Wait, double-click on that for a second.

Tobi: In a way, the criticism that a lot of people direct at companies is that companies are short-term focused. But why are they short-term focused? I don’t think the executives tend to be short-term focused, but the executives often — what they’re incentivized for is to keep their job. Therefore they need to be able to prove that they’re doing a good job at intervals. And if the perfect thing for a company to do is rebuild the entire product from the ground up for the AI age, which is going to take a while, they won’t do it, because the short-term incentive is there — because they are allowed and actually clearly incentivized to be intelligent actors in their local incentive system. And their local incentive system is quarterly earnings. It’s always: show me the incentives and I show you the outcome, as Charlie Munger always said.

Shane: Yeah, but this is a different take on it than I’ve heard before.

Tobi: Interesting. How so?

Shane: Well, in terms of how you develop sort of intuition — and the optimal path is not the one with feedback necessarily. I’ve never heard anybody talk about that before.

Tobi: So the development: at some point you need to run a review. You have to know at some point if it was right. No doubt about it. There needs to be some feedback eventually that happens. But it might be long coming if you have the luxury of a type of employment where you don’t require the other voice from a quarterly call for being able to get another rep — such as being the founder of a company, which is like a deeper relationship for the company. Founders can take a longer-term view.

Shane: And I guess the incentive would be: I need to demonstrate progress. And if I need to demonstrate progress on a quarterly basis, I’m never going to bite the bullet, redesign my product, take a year to get it right.

Tobi: Take again — I believe, and this is not absolute numbers, but for lack of a better way to say it — there’s an infinite possibility space. Even a deck of cards: you shuffle it, and then the same deck of cards will never ever recur in the history of the universe. It’s impossible.

Shane: It’s like 52 factorial.

Tobi: Exactly. So you end up with even simple rules, simple ideas, simple things leading to enormous complexity-space explosions, and people underestimate this. There’s an infinite amount of things to do. This is also why AI will not do all the work — because we have to make decisions of what is worth doing.

So you have a conundrum: you need to make a choice. Clearly you can prune a lot of things to do. Going to buy ice cream is not in the set of valuable things to do if you’re considering an M&A deal, I suppose. So you prune everything that’s irrelevant. Easy. Now you’ve left things that are sort of relevant and sound good. You need to evaluate all these possibilities.

Business books tend to be really obsessed with “make the right choice.” What that does is compress everything into a right-and-wrong conundrum. I never think that’s the hard thing, truly. Making the right choice — actually most people can do it. Even bad management teams have a pretty high hit rate there. The problem is there’s a lot of good choices. This is where things get really hard. For lack of a better form, let’s say there’s five good choices. Again, one of them is going to lead to something observable in the current quarter — some revenue quicker. It’s a good choice. It does the thing well. But the other four aren’t — and that’s a downside — but you might be a much, much better company. You might take a snowboard store to be an e-commerce platform. That was also not the locally good thing to do, because the snowboard store I once had was actually profitable. My incentives were: continue doing that.

Choosing the right among the valid solutions is actually the hard part, not finding a right solution. Unfortunately there’s so much ink spilled on finding one of the right solutions that everyone stops at this point. I just really don’t think this is the hard part.

Shane: Could you actually go so far as to be like: if there is a solution that’s observable and you’re being pulled towards that, it’s probably not the optimal solution?

Tobi: Yes, because I take that position and then let me be convinced that it is. Especially this goes double and triple if one of the solutions also happens to really correlate to how the problem is solved most of the time in industry. If there is an orthodox way to solve a problem, I am incredibly suspicious that this is the solution that’s being offered. But sometimes that is actually absolutely correct — especially in more regulated fields. We do a lot in payments and so on. They often… the orthodox way of solving a problem is actually the correct way to solve a problem, because it might well be required at some point.

Affirmations, “Yet,” and a Malleable Self

Shane: Want to switch gears a little bit. You swear by affirmations and they’ve changed your behavior in the past. I was wondering if you could double-click on that.

Tobi: I take the position that I myself am my own project. Continuous self-improvement on an individual level is my world thing. My life philosophy is that I will meet the person I could have been at the end of my life, and the work of my life is to reduce the difference between the person I will meet and that person to as little as possible.

How do I get better at things? Many, many ways. I’m generally very curious about technology and basically everything — everything’s interesting. But why do I stop to point out that “everything is interesting” is a mantra in my family? Why do I say it a lot, and why would I like my kids to say it? That’s an affirmation, right? Because I believe it to be true but unobvious, and unobvious truths tend to be the most valuable ones in many cases. It’s true at the limit, but you have to go a couple layers deep.

You lay down a lot of grooves in the bedrock of your mind over time. Beyond behavior, you cultivate some excellent habits. Where you feel like you want to cultivate new habits, you invest willpower until it becomes a habit. I think doing the same thing with the mind is totally possible, and affirmations are the easiest way to do it. If there’s something you want to have different — if you want to edit something about yourself — just try to say that the goal has been accomplished over and over and over again, ideally written by pen on a thing. You don’t need to do this for long. I found this to be incredibly potent.

My example: public speaking. I never spoke in front of people, really. Even school — that was not really a thing when I needed to. After starting Shopify and doing some interesting things with tech, I wanted to go to conferences and saw other people do this, and I was like: this seems worth doing, but I’m completely terrified. So I just started writing out — I think it was as simple as: “I love public speaking about things that are interesting to me.” And I think a week of spending five minutes writing this line after line, like Bart Simpson on a whiteboard at the beginning of every Simpsons episode, just kind of does a thing. I love it today. Was this the reason? I kind of think it was. Yeah. I still don’t like preparing talks — that’s really a lot of work — but I actually get so much energy from being in front of people talking about something that’s interesting. It’s exactly like I written it out.

Shane: I wonder if we should start every math class with that: “I love math.” Every student writes that down.

Tobi: Think about the counter: how many times have you heard people affirm “I’m not good at math”?

Shane: Yeah.

Tobi: You know they’re probably wrong, right? Compared to every human who’s ever lived, they are in the top 0.1 percentile of mathematicians. Even just by being able to understand division. We have a bad, bad, bad way — especially around math — of often negative affirmation: I’m bad at math, therefore I can’t do this thing. People need to stop doing that. Don’t say that. Say the opposite. To yourself, write it a couple of times. Get one of those stupid apps and just do some reps. In fact, you don’t even need an app: open Chat, say make me an app, make me an artifact, or make me a site where I can just do math reps. Here’s the kind of thing — come up with some different ways to do it, test me how good I am, and adjust it to my current level on multiplication and division. Then you just do some reps.

Shane: Then write it out a bunch of times, do some reps, do this for two weeks, you’re good afterwards. Done.

Shane: So what do you tell your kids when your kids say “I’m no good at this” or “I can’t do this”?

Tobi: My kids are not allowed to say that word without appending “yet” behind it. All of the others were correct. The one who said it — “I’m not good at this” — three people in the room say “yet.” Just take that attitude. It’s totally okay. Attention is a scarce resource. We can’t be good at everything yet. But the reason why we’re not good at anything is not an intrinsic property of you. It is a temporary state that you have a power to change at any point you choose.

Again, I just want my kids — and I want everyone at Shopify — to understand that they themselves are malleable and an unfinished product. These are the mantras of Shopify: you’re thriving on change. We are a learners organization. You’re obviously merchant-obsessed. All the cultural values aren’t platitudes, but they are positions that someone else would not take as a core value in a company. But they all point at the same thing, which is that you are malleable, the company is malleable, our product is malleable. And by the way, the times we are in are like change as well.

You can take one of two positions there. You can say: hey, I’m going to insulate everyone from this kind of variance from change. And I’m like: yeah, let’s do basically the opposite and say: hey, figure out what the zeitgeist allows us to do and get all the value out of it at all times for our mission. You need mantras for these things. “Make commerce better for everyone” is the official mission of the company, but truly what it really is is: to make entrepreneurship more common. That’s a pretty broad mandate, and we need to figure out what’s possible now. So it’s not just make the same widget we did yesterday tomorrow.

True but Unobvious: Goodhart, Churn, and Beauty

Shane: You mentioned that some of the most valuable things are true but unobvious. What else comes to mind when you say that?

Tobi: In companies, Goodhart’s law just reigns supreme. I keep getting back to it.

Shane: And that’s when the metric becomes the objective.

Tobi: When a metric becomes the objective, it’s no longer a good metric, because again a metric is a proxy of sorts. It’s a heuristic that just tells you you’re going in the right direction. Then it becomes a goal itself. You just reduced all of what your company does to this one metric, and you will clearly overfit. You overfit to stock price.

A good example of this in Shopify has been this real situation early in the company that happened over and over. I had to course-correct it, and then like three years later I had to do it again and again and again: churn is a bad thing. Churn in Shopify’s case — as in an account closes. If a business goes out of business, that is of course a negative thing. But because we are involved so early, and we’re into the normal process, people just run experiments on Shopify and starting one which then isn’t working — no product-market fit was found — is not a bad thing. In fact it’s a very good thing for Shopify that this happened on Shopify, because those same entrepreneurs will probably try again. But that was extremely unobvious oddly early in the years, and I constantly had to explain this. There were all these papers, some of them written by very investors, that just described that churn management was the most important thing a software-as-a-service company was doing. But in Shopify’s case it’s an entrepreneurial journey. Maybe I didn’t find product-market fit — they’ll be back. So that’s one.

Shane: What’s the relationship between beauty and ugliness and creation?

Tobi: Beauty and ugliness are both very good ways of evoking an emotion. When you’re creating something, you’re trying to — it’s like love and hate are the target zones; both; the entire middle is indifference. That’s the death. So beauty and ugliness are two entirely valid targets. In fact, you can’t hit either of them purely. There’s not a thing that everyone will love and no one hate. You’re going to get both, or indifference. Both are your choices. When you create something, you want other people to deem it worthy of having an opinion of that magnitude about.

Shane: Is there something at Shopify you’ve made more beautiful even though nothing would support that?

Tobi: Oh, that’s the entire job. You’re not a craftsperson unless you care about the parts of products that other people don’t see — like the architecture of it, the prose, the legibility. These days I look at Shopify of pre-2020, like 2023 — wait — it’s like: man, this is tens of millions of lines of handcrafted code as it will never exist again. We had to build this entire system by hand, line by line, and we did it by talking a lot about beauty and what is beautiful code. We built a lot of Shopify in Ruby, which is famous for its poetry mode. Poetry mode means you can write Ruby that’s essentially English. You can read some really well-built Ruby code as if it’s telling you a story about what the system is actually like and how it works. It just happens to also be communication to your co-workers, but also at the same time executable by machines, which is incredible.

So aesthetics factor in a lot at all layers of like office system. You use beauty a lot creating things, because beauty is actually how our intuition communicates with us. My best understanding of what intuition truly is or where it comes from is that with enough reps, what happens is — I think most of the energy budget of our brain is actually in the visual neural cortex. It’s like a visual system that sends us pictures to the rest of the brain. But through the pipe of sending pictures or state or world model or whatever to the brain, it can communicate concepts too, and it does this by aesthetics.

Then you ask a professional chess player or Go player or something like this: hey, why did you — how many lines did you calculate here to make this beautiful move? And people use the word beauty. They will say: no, I only looked at that one line. The reason why I looked at it is because it seemed beautiful to me in the moment. That’s not all of what intuition is, but I think it’s a large perspective. This is why people are so fast sometimes — because they use a massively parallel part of the brain where things are at least slightly more sequential when you’re trying to reason it out from first principles. And sometimes you can never reason towards aesthetics from first principles to begin with. So I think that’s important.

Pruning, Refounding, and the Raptor Lesson

Shane: Do you remember that graphic with the Raptor images?

Tobi: Yes.

Shane: The rocket with the SpaceX book. So two things about that strike me. One: ship ugly version; the third version was incredibly beautiful. But the second, sort of counterintuitive maybe insight there is: a lot of teams can’t move forward by subtraction. They move forward by addition. Maybe riff on that for a few minutes.

Tobi: Yeah. The SpaceX Raptor — I think even the first of them was probably the highest-performing rocket, but we’ve made… it’s itself beautiful. Raptor 2 is an iteration of this. I think even Raptor 1 got lots and lots of iterations, because that company is itself, I think, the most impressive company on planet Earth by far. It’ll likely go down as the most consequential company of the age. It’s all built around a self-improving, reinforcing loop. That’s stunning, because in no other companies’ field do we have such a clear example of a difference — of just aesthetics for problem-solving.

Rocketry is done by governments at cost-plus, with enormous amounts of pre-planning. Every piece of equipment has to be radiation-hardened. Every eventuality is covered, and therefore comes at enormous expenses. And then you have SpaceX just using absolute incredible thriftiness to accomplish greater things at rapid iterations by simply being okay with failing — with sending a rocket which then explodes. I think they call it a rapid unscheduled disassembly instead of an explosion. I think that’s beautiful and I think it should be inspiring.

One of these places you see this is this Raptor. Again, every one of them beautiful. Every one of them — they could have stopped at the first one. It already was a totally valid solution to the problem. They didn’t need to go to the next. They went to the next and the next again. But to your point, the most impressive thing here is the path by which people move forward here.

Things need to be pruned. You cannot make things better and better by adding stuff. You can’t. You must prune. You must take a step. You must rebuild. You must create an end for things.

Opinion about failure is a problem. Failure is never a problem unless it is catastrophic. Of course, in spaceflight with manned missions, it can be catastrophic. You got to get this right. But in terms of when it’s just resources that are replaceable and frankly fungible, then you can just do this. The reason why it’s good that the product failed is because it frees up an even more scarce resource: a person’s vision for products to apply themselves to another one which then the market potentially decides is something that is needed.

A lot of these pipes on the Raptor engine — they’re there because that was the only way to make a Raptor engine at the time. I think by the third it looks mostly 3D-printed. Maybe that wasn’t technology which was available back then, but now that it is, every one of those pipes was incorrect. It didn’t need to be there. In fact, I think the performance of that third Raptor engine is astronomically higher than the previous ones. The thrust-to-weight ratio of that thing is absurd. So you need to prune, and sometimes you can prune by creating a refounding event. You’ve got to start a new version of a Raptor engine and get it right based on everything that’s working. And I think this is how companies should work too. A department sometimes needs a refounding event. We could solve a lot of problems in the world by just using tools like: make a 2.0 version of it. Give it a refounding event. Take it from the top. Building in more exploration of systems would solve a huge amount inside companies — renewal and so on.

I think one of the large reasons why it was so easy for the companies of my vintage — the early 2000s tech companies — to just displace all the existing technology companies, minus like three or four, was just because they fell prey to a world of a lack of competition, and then what they built then wasn’t forged in the fires of competition and therefore wasn’t tested. It was easier to just simply solve problems by addition and layer-caking. And then the original intent of some of these departments, products, whatever, was somewhere in the fossil sediments under layer and layer and layer of additional stuff on top, and no one knew how to dig down.

Books, Old Ideas, and What Success Means

Shane: In our first conversation that we had together, you said books were a cheat code for life. I’m wondering how your thinking has evolved on that in a world of AI.

Tobi: I don’t think it has. There’s more cheat codes now, but the books still play the same role they have. What’s changed for me personally is that — I don’t know if that was probably already true when we talked — at least for nonfiction, I walked away from books written recently. I think everything written recently is really just the product of its time and is kind of trying to put a bit more information into something that’s currently evolving. I think books that have stood the test of time are just as valuable, and I think they will always be.

Shane: So what are like three old books that you’ve read that have fundamentally changed how you think?

Tobi: Books I come back to: I often talk about Parkinson’s Law, which I love. It’s such a quick read. I almost always will mention The Lessons of History, which I just think is the densest — the highest token-quality book in existence, given for the length. James Burnham’s books are fantastic and extremely relevant.

Shane: What did he write?

Tobi: He wrote The Managerial Revolution first, and then a book called The Machiavellians, which is unbelievably good. Obviously I am a Meditations fan. I know Stoicism is falling out of favor a little bit right now, but it’s been a lifelong thing for me, and I have a copy of Meditations in most rooms I spend time in. I just do some random reading and it’s magical how it’s somehow relevant to something I’m wrestling with.

Brand books just in general. The Lessons of History is of course the end-of-life distillation of it all, but his longer book is good. Fiction: the Foundation series is so good.

Shane: You read the Three-Body Problem too, right?

Tobi: I guess that’s sort of tipping into older book now too. But that’s a recent sci-fi which is incredibly good.

Shane: Final question. We always end with the same thing. This is your third time answering this question now. I’m interested — I’ll go back and look at how it changed. What is success for you?

Tobi: My success is just to cultivate skills — become good at more things — and in doing so create products or toys or things that other people… that can make other people’s day a little bit better at the minimum, or go and allow people to get power or motivation or ambition beyond what they would otherwise have.