Tobi Lütke: Empowering a World of Rebels | Knowledge Project Podcast

The key to fixing insomnia is understanding that sleep comes when you are sleepy (not tired), and the way to build sleep hygiene is to only go to bed when sleepy, create a bedside reading chair for when you wake up and can't immediately fall back asleep, and stick to a consistent sleep schedule.

“you cannot sleep when you're just fatigued and not sleepy but you can always sleep when you're sleepy um so what to do from it on an effective [basis] is uh you you do create this routine a lot of these tools are useful um you uh but and this is the hardest part you do not do basically that's some minor allowances anything in your bed better specific it's you see you'll be your best for sleeping you have a chair close by uh that's you take your book there um you would read there um you do not bring your phone to bed ever”

Specialists develop extraordinarily deep expertise in narrow domains, while generalists can develop 80% competence across many fields much faster than becoming 100% expert in one; the Pareto Principle suggests most fields can be learned to 80% in 20% of the time required for 100%.

“Specialists will like everyone who knows that there should be a specialist no space should you know like it's like everyone should be in specialist now they should be like this is this is a very particular set of character traits um which wants to dedicate itself to a [very narrow kind of thing] what Specialists don't understand is like how how incredibly good journalists can be at at 50 things right like they think like a bit many people figure journalists is sort of like a jack of [all trades master of non-legates] this is one of the dumbest sentences I've ever heard that's like not like that's just usually someone who hasn't made a choice and is not good at developing any any kind of like a very fairly rare people um um uh you you like Pareto Principle and no one tells you but first 80 of every field can be done 20 of time”

Risk in business is most often not about taking bold action but about inaction or following best practices; the real risk is strategic obsolescence that comes from doing the same thing as everyone else, which only becomes apparent when you fast-forward time.

“it is actually risky to not take uh if I try higher but like I a lot more people have it's not risky in the moment though right to to follow best practices it's only risky as you fast forward time yeah and I mean we should acknowledge why if you make like if you do what everyone else does you are like say like well that's what industry does um if you make an edit to it you are you will have a responsibility go to bed for for that change um at least that's it from if you risk avoidant this is something you don't you don't want no if of course if it works it's great”

Stable Diffusion being released as open-source created an acceleration in the field because it enabled developers and practitioners to contribute improvements, not just academic researchers, and it has been getting faster and more efficient through these contributions.

“what what is so profound I think stable diffusion is that it's gotten released as open source and now what's happening with it is just a wants to it's like all acceleration it has but all like this is already the property maybe minus crypto the fastest progressing field in the world and um has just found two or three more gears there because if you track like I mean we started like [I] think first versions on top of a line graphics cards on your own machine you can execute this by this open source um you could make an image in 15 seconds I think they're down to three yeah and and like with no Hardware changes and just like well because um the tooling behind machine learning was again it's a very academic field”

Technology IS progress, and if people adopt a pessimistic view of technology, they are adopting a pessimistic view of human progress itself, which is both factually incorrect based on data and also corrosive to optimism and agency.

“if you have the people who do like like um on the airwaves uh you know like every person on the planet ends up becoming oh well everyone becomes sort of average of five people they spend the most time with Fred like that's just sort of a really well documented or oh if it's fundamentally true but it's super useful as if as a as a mental [model] and I think that uh you know one of those people is a media for most people and um um so you're adopting a set of beliefs which are I think extremely colored um and in this particular case I think somewhat unfortunate because you just recast the applied field of human progress as something that is causing detrimental outcomes”

The Turing machine, and by extension all computing, represents something more fundamental to life and human civilization than is commonly appreciated; it is as fundamental as DNA and deserves to be integrated into how we understand human history.

“I think what's the one of the greatest bit bits of like if technology is [progress] the turing machine is the engine of progress and uh that that almost all like net human purpose is going to be built on uh because everything is going to round to zero um and it was probably mainly needed for bootstrapping the turing machine”

Programming synthesizes key properties of mathematics (theoretical purity and abstraction through symbols and equivalence) with practical linguistics, creating a new field that can reason about computable, resource-constrained systems that mathematics alone cannot address.

“it's it's called a sort of um it's theoretical cohesion a very equivalency sign is like those two sides are exactly the same thing not just reasonably similar um it's it's it's sort of a language but it's actually it doesn't have a concept of abstraction it's like everything is just these formulas picture and um um so you end up with this new word of a turing machine which is actually the midpoint between human language and Mathematics”

Transformer models and large language models represent a breakthrough in making machine learning practical and accessible, allowing millions of people to contribute improvements rather than just academic researchers.

“machine learning has been unreasonably effective already um but why do Transformer models change everything like it's another one of us they've Transformer models feel a little bit more of a vocational than academic and like they're probably not as good an analogy to the brain as maybe some of her other like approaches or at least it's up as puplay as some of them um bad boy back it's crazy what's happening like it's crazy”

The model of 'hire smart people and get out of their way' is not effective; instead, founders should maintain deep involvement in decision-making, maintain what Tobi calls a 'trust battery' with key leaders, and collaborate on important decisions to benefit from diverse perspectives.

“I fell a little bit into this work through org chart treated as a trust for you know like uh hire smart people and then get out of their way I think that's a waste of everyone's time it's not even good for the like it's a nice story but it's like dumb [we have] a um idea of a trust battery that we talked about in previous times you talked about and today about um the trust battery builds we work together we work alongside we make key [decisions] together as a team everyone brings perspective we listen to absolutely everyone but the choice at some point has to be someone's”

The convergence of Transformer models, large language models, and diffusion models is creating possibilities that will define the next era of human capability; this is the most exciting time in technology and requires both theoretical breakthroughs and practical engineering to implement.

“everything changes every time there's like a significant breakthrough in terms of speed and and and computation because uh like it is extraordinary how many things we can do [with] language large language models and uh diffusion models and Transformers and I I just like I wish I think in a more inspired world yesterday's breakthroughs in the world of stable diffusion would be every single day of the last couple months at the front page news”

All of business should be rederived from programming and systems design principles because programming uniquely solves the problem of building resilient systems that must operate in chaotic, non-deterministic environments with incomplete information.

“I think all of business has to be rederived from programming principles I I had this intuition I think we even had this situation on previous episodes like we talked about this wherever I tried it it worked better and all other approaches”

Legitimacy—the accumulated social credit from an organization's founding story—is a resource that can only be spent by the founder; when a founder leaves, the account remains but cannot be drawn upon, making it impossible for non-founder successors to execute the same kinds of changes.

“if the founders are still there they can still draw on as a count of legitimacy and so my role I see um very often is that um I have the like pretty humbling uh ability to like like use the resources deposited into this uh backup card of legitimacy in the company [you make changes and because I can speak for the entire story of a place yeah now if I would leave Shopify this bank account would only increase but no one could draw on it and therefore it can't be an active ingredient in a company”

The human brain's compute capacity and intuition are more powerful than artificial computers combined, especially for making decisions in the moment with incomplete information; people should be more confident in their intuitive pattern-matching abilities.

“the super computer in our brain is absolutely utterly in support not unsurpassed by auditory machines combined um coming to work to produce everything you've ever seen to make the best decision in the moment”

Tobi is optimistic about the future and believes 'the best is yet to come'; in 20 years, these AI developments will be studied as the formative years of human civilization, when humanity had to ask existential questions about organization and prioritization.

“The Best is Yet to go it is it's and it's gonna be freaking amazing there is no chance like I'll take a long bet for a lot of money on this if anyone wants to take it [it'll] have to be a real long bet but like if I'm right I'll have to pay out um but for rest of human history but 20 years before this point I mean 20 years after I'm going to be more studied and will be treated as the coming of eight of our species um they are going to be treated as the formative years where we went through the questions we had to ask ourselves”

The quality of a decision is independent of the outcome; a good decision with incomplete information that leads to a bad outcome is still a good decision, while a lucky decision with the right outcome is still a bad decision if it wasn't based on available information.

“accuracy very often is um picking the right result but that's actually independent of making good decisions uh like that's just that's I think resulting is what poker players call this and I think it's not really the right thing to do”

The Turing machine is not just an abstract mathematical object but represents something analogous to DNA in biology: it is a mechanism that reads and writes to itself, creating more of itself, which suggests computing is more fundamental to understanding life and systems than is commonly appreciated.

“I think what terrain like invented um is a uh his EP um um is not as separate as people make it out to be if you look at the cells that make all life this is kind of what DNA is like it's a it's a read head and the right and the right head and it like it creates more DNA it creates more cells”

Russia's invasion of Ukraine forced Europe to reconsider its energy policies, creating a secondary effect where climate change may actually be solved faster due to the need for energy independence and rejection of fossil fuel dependence—a perverse silver lining to the crisis.

“Russia invites Ukraine which seems to be a recurring topic here everything in primary is bad secondary effects many many many of them are bad like there's a secondary effect of everyone having to look at the energy policies and realizing holy health should be probably defeat nuclear power plants back so you know we actually um you know in a really odd way um Europe's dependence on Russia for natural gas pipelines Plus the act of aggression by Russia may actually cause climate change to be solved which is like an unbelievable [story]”

Computer science and its practitioners were historically sidelined by academia and industry as a cost center, but the 'revenge of the nerds' is that companies founded on computing principles are now out-performing traditionally managed firms, though the broader public has not yet integrated this into their understanding of why the world is the way it is.

“the Revenge of the nodes is actually that we now build all these companies which have gone out performing everyone else and and and frankly everyone has to kind of understand but that isn't because we can technical skills but actually because we and I'm including myself here in sort of this engineering discipline”

Business systems should be thought of like software systems with explicit models, configurations, and auditable functions that can be tested, versioned, and iterated; this allows organizations to make changes systematically rather than ad-hoc, making them more adaptive to changing conditions.

“I made all of this executable I now go and um just edit one file and then it rebuilds a model of a business I think and then if you like it in a engineering that's sort of engineering thinking company the bat becomes the new desired model and the goal of uh the um HR discipline is actually not to make all these arbitrary decisions but actually to look at what is the model of a company what is the actual state of a [company] whatever moves that gets the actual State as close to the model as possible”

Understanding second and third-order effects can lead to cynicism if one focuses on how perverse incentives and unintended consequences are inevitable; the antidote is to focus on facts and systems thinking while maintaining agency to influence outcomes.

“the danger with secondary advert but tertiary effect thinking his it really lends itself to cynicism um because because um sometimes you see these connections which are like Felix trying out the inconvenient for for for for the main story if you tell ourselves um but I I also think facts are friendly uh I mean I mean and also those are not Factor facts they are like compounding systems”

Applied computational philosophy, combining principles from computer science, systems design, engineering, and other fields, should probably be a core academic discipline with a dedicated faculty, not relegated to 'computer science' departments.

“I think that's completely unappreciated right now as as a source of um business thinking value uh and um uh I find I find it's very very interesting this is really what I mean with applied computational thinking in a applied computational philosophy being almost the higher order [academic department] that should exist on the campus and by way it should probably have the nicest building”

The book 'Why We Sleep' is anxiety-inducing and misleading because it makes unsupported claims about universal sleep requirements and the catastrophic consequences of sleep deprivation, whereas actual data (from modern sleep trackers) shows significant individual variation.

“there's a book which initially I thought was good called why we sleep which is actually very bad um from everything I know now um I I mean it's it's not bad for Coke as incorrect in the macro sense but it gives you it it I think the book itself is anxiety-inducing um because it makes claims on which I don't think prove out on the um importance of sleep and how everyone's goals should be to sleep eight hours and there is no variance”

Professionally managed (non-founder-led) companies tend to excel at operational excellence and quantifiable metrics through careful stakeholder management and planning, but they struggle with subtraction and cannot draw on founder legitimacy to make disruptive changes.

“professionally managed companies are often like the mental picture comes to my mind it's like a river stone right something that's like really smooth around all the edges it's like um there's like it's a very similar uh skill level on everything but usually they're like if they have significant competencies they're often about just operational excellence and uh you're very very quantifiable things um everything that can be measured they tend to be very good at”

The rapid technological change and pace of new infrastructure emergence (internet, mobile, SaaS, broadband all converging in the 2000s) created a Goldilocks zone for startup creation, and this environment has become less favorable now; future entrepreneurship will more often involve extracting great ideas from large innovative companies and applying them to other fields rather than creating entirely new categories.

“these examples are not coming along at the pace anymore as the early 2000s right like um or you know a lot of a multi-billion dollar companies um uh in in Technologies uh like come from is it a three four five year period of startup creation right which was this extreme Goldilocks zone of um uh everything was changing the mobile SaaS internet uh Broadband um member Broadband folder um all of these things happen at the same time”

The key advantage of the engineering/computing mindset in building large organizations is experience designing systems that must be resilient to chaos and non-determinism, which trains the mind to think in terms of abstraction and model-building in ways that historical business training does not.

“we spent our formative years designing systems that are have to be resilient to like weird non-deterministic like the kinds of things that are even that you can't even really put into a market formula right like the hacket loss and like um and you know our word actually is one of Chaos in which we are trying to pretend that it isn't and then we have to put abstract systems almost in the heritage of linguistics”

Most books and media narratives about history in the last 50 years have not integrated computing and the information revolution into their main narrative; computing remains relegated to a sideshow while the 'real' history is told through political and economic events, but this is an extraordinarily lossy model of what is actually happening.

“I think 120 years from now when we talk about human history um like these books felt like hacker by Stephen Levy and a dream machine and um look code by pets old and you know office kind of these books are actually um going to be referenced books that show the actual Mainline history of our species right and currently we like it's sort of like relegated like the newspapers talk about I mean current affairs which we showed like they are also important but it feels like we are living on this in in this weird word but actually has not integrated its own Mainline history track uh into the public Consciousness”

The widespread pessimism and anxiety about the future and technology, despite objective improvements in life expectancy, poverty, disease, and opportunity, is driven by media consumption patterns; people think they are doing OK but believe everyone else is doing poorly, creating a sentiment-reality gap.

“everyone in the world is like I'm doing great everyone's anxiety is based on their understanding of everyone else which everyone else literally disagrees with like so we all think everyone has to do bad at everything that's getting worse but it isn't like it it's it the numbers don't bear down there's countries I think South Korea was like 9010 like 90 people are doing fine and they're only fake like ten percent think anyone else is okay”

Good decision-makers are not necessarily those who make correct decisions in hindsight, but those who can identify the relevant inputs for a decision, properly weight them, and explain their reasoning in a way that makes the decision obvious in retrospect.

“everyone who looks at the info kind of come to the same I I've actually never really seen this situation where people are truly making different choices like there's always if there are like people on different sides you just have to like okay let's take it from the top and like talk or click State or the unstated assumptions that are actually inputs into decision making and once was uh flushed out it becomes pretty obvious”

Decisions should be revisited periodically as new information emerges and as your understanding of the inputs evolves; maintaining a decision record allows you to assess whether the decision would change based on new evidence.

“and then after you make this decision you don't just if it's an important decision like strategy or something you don't like throw away this entire book you actually capture it and you revisit it because again the world keeps changing and by the way your appreciation of inputs are that you considered end up evolving as well and so you should actually at any moment we run a function over the inputs”

The transition from building companies on old tools (process, story, rules) to building them with computational thinking will make incentive system design more complex but more auditable, allowing companies to move faster while remaining aligned.

“now with computation impactual Computing I think we can Implement more complex systems that are more auditable be able to provide some of the analysis if things are going in the right direction and so in in a macro sense um and and will actually make it much easier to implement uh in a more complex incentive systems”

The transition from financial optimization of compensation systems (using stock options for tax benefits) to explicit, auditable systems (clear salary + choice of how to receive it) represents applying engineering thinking to HR, making it more transparent and adaptable.

“so I mean we switched to a very simple system where everyone gets a number this is what you make every year and um so everyone agrees but like the ground truth is of compensation and then you can say we want this in cash do you want visit stop options um uh and there's going to be other things like uh so like we will hopefully get to a place where you can use uh like like we have currently I think two or three sliders but you can uh choose how you want your conversation and then like they're hopefully will be 10 or so to get as we build more capacity for different options”

Founder-led companies are often described as immature, but this is incorrect; they are instead 'spiky objects' with very strong aptitudes in specific areas (often matching the founder's strengths) and sometimes surprising weaknesses, but organized around solving a particular set of problems.

“in a in a way where you have Riverstone on one side you really have sort of a sort of volcanic ash kind of spiky object thing um of a rock on the other side like they like some have very very strong aptitudes spiky object sometimes surprising weaknesses it's it's not a well-rounded thing in Rand but it's usually perfectly appropriate in terms of well an ideal case it's appropriate in its aptitudes for the um a particular set of problems that the company a company serves”

The ability to subtract—to remove products, initiatives, and people—is one of the most valuable and rare capabilities a founder can provide to a company, but it requires bravery and legitimacy and can be done perhaps 10-100 times more easily by founder-led CEOs than by professionally managed ones.

“what does office mean well the best thing Founders can do with subtraction this is surprising but like it's it's the the reason like it's much much much easier to add things um than it is to uh to to to to remove um things um adding things is a lot more expensive than removing things actually um uh however it requires some measure of bravery”

Innovation should happen everywhere in a company, not just in product R&D; every function—from HR compensation systems to sales processes—should be reimagined from first principles, and the best acquisitions are those that bring in ideas that can be adapted and improved rather than replacing internal innovation.

“I think every group in a company should be able to go to a conference and talk about how we are doing like how our group is better than sort of a general implementation of this discipline or idea um and by us doing things differently and how you're thinking about this first principle thinking applied to uh any area can can yield huge upside and um I mean ideally that I think the best uh like like a way I would like that to see this is actually that for our own purposes to build the best company that supports it being extremely Innovative for entrepreneurs in the retail space um we are innovating across all the disciplines”

The proper framework for evaluating a decision is to ask: given the information available at the time, was this a good decision? If information was available but not considered, that is a failure of due diligence. If information was unavoidable and the decision was still wrong, that was acceptable risk-taking.

“my my test interpreter is always like wherever any 20 minutes I could have spent to pull this piece of information I didn't consider right into the set of things I uh used as inputs to make the decision if so I wasn't equal to the task of making the choice at the time and I need to um understand this better in the future and approve for”

Tobi's role now and in the future will be to understand possibilities created by computing, test them, and contribute to the historical record by participating in exploring what becomes possible; being a 'custodian and explorer' of computing's possibilities is his fundamental contribution.

“the like there's no chance you would want to live through any different times if a future is fantastic and will probably become [immortal] relatively soon and all these kind of things um and um maybe not us but our children they think well uh and and so um a lot of what people will be preoccupied with in the future is clearly what's what's next and what's possible and many of the biggest questions will remain to be studied and and to guide our curiosity but like it's here where it all was like pop on the table and it was being wounded the world thing and I just [it] seems it seems so incredibly great to be around to be able to study this stuff it's so fantastic to go and um actually be a public like play a tiny little role like make the title Dent into it”

Media institutions have a grudge against technology because they have been negatively affected by it, but this is a mistake; new media forms (podcasts, founder-operated video, newsletters) are thriving and this is actually a Golden Age of media, not a decline.

“because media I mean a media has a crutch against technology for frankly understandable reasons like um everyone in media works for an institution that probably was negatively affected by Tech now I think media [itself] like if you if you trip some out it's actually doing great like podcasts are doing great like um a builder operated founder operated videos killing it”

Excessive business travel pre-pandemic was treated as a marker of commitment and sincerity despite being destructive to family relationships and personal health; this was a form of unnecessary 'penance' that should not return.

“I think um I think it's hugely Pro um spending time with people and I think travel is a um instead of a text on doing this which is totally like easily worth it um but I think there was a lot of trouble before that um pandemic I think which was almost like I don't know feels like it was almost Penance for meetings like it's just like hey there like it's like your sincerity had to be proven at some kind of altar of sincerity by you sacrificing uh days of travel and airports and cross-continent travel um and compromising your time with your children”

Shopper's new compensation system gives employees a transparent base salary and then lets them choose how to receive additional value (cash, stock options, and potentially other instruments), treating them as adults who can make informed choices rather than using complex financialized derivatives that obscure the actual value exchange.

“we switched to a very simple system where everyone gets a number this is what you make every year and um so everyone agrees but like the ground truth is of compensation and then you can say we want this in cash do you want visit stop options um uh and there's going to be other things like uh so like we will hopefully get to a place where you can use uh like like we have currently I think two or three sliders but you can uh choose how you want your conversation”

Bureaucracy is not inherently bad; it is often a set of checks and balances that exist for good reasons (preventing rodents, in the wall metaphor), but the problem is that bureaucracy often lacks perspective hierarchy—a hole in a wall might need to be papered over locally, but at a higher level it might need structural repair, and at an even higher level the wall might be irrelevant.

“see a hole in a wall and what you would try to do is paper over it which might make perfect sense in the in in the sort of local context that you have but actually if you the next person with a larger perspective says actually we we use driver patches um because that doesn't cause problems of rodents and um the person like if they talk to the next person they say you know what that wall that this hole is actually bad for it shouldn't be there let's actually get rid of entire walk”

Tobi has applied engineering skills and systems thinking to business strategy and organizational design, which he previously thought would not work, but has found they are surprisingly effective and represent a convergence of his engineering background with business leadership.

“now I have a um idea of a trust battery that we talked about in previous times and I'm doing [it] I'm not putting engineering skills on speed dial anymore they are now the core of how I make decisions and I I I I just that's been just a fascinating like if I [could] travel back in time like and just give this as a hey this is this will happen I wouldn't but I would be extraordinarily surprised by this I I really was almost in a like hey I this skills took me here they will not take me there I actually need to go for a full change of identity and just like embrace me looking for strategy and like a bit more of a sort of old school things a bit [um] I thought this is gonna work it's blown up in my like I think it can work in peace time uh I think it cannot”

Shopify shifted from all-office to permanently remote ('digital by design') because the input variable changed (people could not go to offices) and when you re-ran the decision function, the new logic led to a new policy, which was communicated early to reduce ambiguity for employees.

“I did that with I decided we built a lot of offices we did them extremely well I think they got very good at building offices but then a change happened in a world and and one of the variables was like well can are people allowed to go to it like like uh for officers I don't think I had that that was an unstated input that wasn't a stated input but it was one we won't get flipped we decided like okay no offices anymore”

Tobi now sleeps exactly 6.5 hours per night (his personal average) and can fall asleep in ~5 minutes with 95% sleep efficiency, having gone from chronic insomnia to excellent sleep quality in 3 days using CBT-I.

“by day three except for a night I have [been] sleeping perfectly ever since I mean once a month I maybe like if we walk walk me through this this is like what what are the steps to to do that um I sleep at like I'm going I'm going reading at 10 every evening basically that's sort of like maybe sometimes a little bit earlier depending on how good my book is um and then I'm the go to my bed and I'm sleepy and I I sleep immediately I fall asleep like my sleep Effectiveness rating on my sleep Checkers is basically 95”

The previous model of acquiring all innovation from outside companies (acquire-to-innovate) is fundamentally opposed to Tobi's philosophy; instead, he wants Shopify to innovate everywhere internally because that requires founders to demonstrate that they can solve the full spectrum of problems.

“a sort of an ergonomic for a lot of companies investor in the end state is to become companies that maintain what they've got um and and then acquire all Innovation I I think that's a I mean I would not be proud of [overseeing a company like this] I I I I if if Shopify would turn under my watch into that I think someone else could do my job better um uh like I think companies are like it's Innovation is a Lifeline I think you want to innovate everywhere”

Tying employee lifetime compensation to stock price movements that occur weeks before grant dates—unrelated to company performance or individual contribution—violates basic fairness and creates unnecessary randomness; a employee's compensation should not depend on whether Russia invaded Ukraine the week before their grant date.

“Shopify is the stock price is like this crazy function over time based on again sentiment market value so tying everyone's lifetime earnings really put like to this other thing just like why are they swinging compared to the thing that we are working on feels like just such a pattern violation I think it's just like a steel man wipers stealing me argument but it should affect what uh uh uh you [in a like a Shopify employee uh makes us lifetime earnings whether or not Russia invaded the Ukraine the week before for Grant”

Machine learning progress is accelerating so rapidly that Tobi is having difficulty keeping up with all developments despite paying attention, suggesting that tracking and understanding all progress requires dedicated effort and even well-informed people are losing sight of the pace.

“I can watch three four or five different fields right now or November progressing at Breakneck Pace where I mean especially machine learning is the first time I think I found my [total] capacity like I actually have to say this progresses so fast but I think my model of what's possible is for him right even though I'm trying to make that not true I think sometimes I just don't notice that something is accelerating and far behind but then I can catch up there I'm actually having trouble staying staying uh on top of it”