For people who build products
Patterns
The same lessons keep returning across very different technologies. These are drawn from 212 decisions in 25 essays, 17 annotated memos and 10 technologies that lost, each linked to the case it comes from.
Incumbents absorb the challenger’s best idea
A new product proves an idea; the dominant product copies or buys it, and the challenger loses its reason to exist. Features are easy to copy; distribution, data and switching costs are not.
- If the largest player copied our best feature tomorrow, why would anyone still choose us?
- Are we building on a platform whose owner can absorb what we do?
- Pursue Napster in court until it closed, rather than license it as a paid service, even after Bertelsmann broke ranks to back a membership version
- Commission original series without pilots and release whole seasons at once
- End its Netflix distribution deal, keep its films for its own service and launch Disney+ at a low price
- Build its own base map with Street View cars and human editors, and offer free voice-guided navigation on Android
- Use Google’s results to power Yahoo!’s web search, with Google credited on the page
- Show its own shopping comparison service prominently in results while rival comparison sites appeared as ordinary links lower down
- Put AI-generated answers at the top of the results page, above the links
- Require apps selling digital goods and subscriptions to use Apple’s in-app purchase system, at its commission, and bar them from telling users about cheaper options elsewhere
- Consolidate the market for installed enterprise applications by acquiring PeopleSoft and Siebel
- Sell Office as a subscription while continuing to offer perpetual versions
- Price AI agents by the conversation or action they handle, rather than by human user
- Replace its older ERP with S/4HANA and set a date for ending mainstream support of the old system, pushing customers to migrate, increasingly to SAP’s cloud
- Open Amazon’s product pages to independent sellers, letting them compete directly with Amazon’s own retail offers
- Sell its own products in competition with marketplace sellers while holding their sales data and controlling which offer wins the Buy Box
- Build a video app around a feed chosen entirely by algorithm from all content, rather than from accounts the user follows
- Fill the main feed with full-screen recommended videos from accounts users did not follow, following TikTok
- Release the source code of its web browser, Communicator, under an open licence
- Buy GitHub, the main home of open-source development, for $7.5 billion
- Release the weights of its large language models for broad use under a custom licence, rather than offering them only through an API
- Launch Excel on the Macintosh first, designed around the mouse and a graphical interface, before bringing it to Windows
- Release a spreadsheet that separated formulas from data and organised numbers by named categories, first on NeXT
- Add Python and an AI assistant inside the spreadsheet
- Publish the Transformer architecture openly, in keeping with academic norms
- Announce a public chatbot quickly in response to ChatGPT, with a promotional demonstration
- Drop SQL, joins and schemas for scale and flexibility
- Sell “a reduction in information overload” through open, searchable channels
- Buy a hit email app built on top of Gmail, then close it
- Launch a separate experimental email app, then close it and move its best ideas into Gmail
- Buy Instagram and WhatsApp while they were young
- Copy Snapchat’s Stories format openly
- Reorient the whole of Microsoft around the internet, and treat a browser company as the main threat to Windows.
- Buy young social apps, Instagram above all, before they could grow into threats to Facebook.
- Continue investing primarily in proprietary Lisp hardware as general-purpose workstations improved rapidly.
- Continue to rely on and protect Minitel as the national online system while the internet spread elsewhere.
- Position OS/2 2.0 and Warp as "a better DOS than DOS and a better Windows than Windows", emphasising their ability to run DOS and Windows programs.
Who pays shapes what gets built
A product’s business model decides what it optimises for: advertising rewards attention, subscriptions reward retention, unpaid creators eventually leave. The design follows the money, whether or not anyone intends it.
- What behaviour does our revenue model reward, and is that what users want from us?
- Are the people who create value in our product getting a share of it?
- Sell songs one at a time for 99 cents, with the copy protection the labels required, through its own software and player
- License the major labels' catalogues for a free, advertising-supported tier as well as a paid one, granting the labels equity and advance payments in return
- Add a cheaper plan with advertising and charge for accounts shared outside the household
- Invest €838 million for a 5 percent stake in ASML and commit €276 million to its research on extreme ultraviolet lithography
- Launch a free web map that users could drag, zoom and search in the browser, funded by advertising, and let outside developers build on it
- Buy the digital map maker Navteq for about $8.1 billion
- Build its own base map with Street View cars and human editors, and offer free voice-guided navigation on Android
- Collect precise location from phones through apps and advertising code, and sell or share it
- Fund card systems through interchange fees paid by merchants’ banks to cardholders’ banks, with rules that kept merchants from charging card users more
- Set the merchant discount rate on UPI and RuPay debit payments to zero by law
- Sell the order of search results by open auction, with advertisers paying only when a user clicks
- Sell text ads by cost-per-click auction, rank them by bid multiplied by expected click-through rate, charge just above the next bidder, and keep them labelled and separate from the organic results
- Extend the ad auction to other people’s websites, sharing revenue with publishers
- Put AI-generated answers at the top of the results page, above the links
- Require apps selling digital goods and subscriptions to use Apple’s in-app purchase system, at its commission, and bar them from telling users about cheaper options elsewhere
- Deliberately break Apple’s and Google’s payment rules in Fortnite, invite its removal from both stores and sue both companies
- Price software, training and services separately from computer hardware
- Stop selling boxed Creative Suite licences and move all customers to monthly or annual Creative Cloud subscriptions
- Price AI agents by the conversation or action they handle, rather than by human user
- Run its Chinese business, EachNet, on eBay’s global platform and fee model, against a local rival that charged sellers nothing
- Charge sellers nothing to list, hold buyers’ payments in escrow until goods arrived, and let buyers chat with sellers
- Give away the software and sell subscriptions for certified, supported, long-maintained versions of Linux for enterprises
- Release capable language models with downloadable weights that others could run and modify
- Test advertising in ChatGPT for free and low-priced tiers
- Make the word processor write: drafts, rewrites and summaries, priced into subscriptions
- Make email free, run it in the browser, and put an invitation at the foot of every message
- Offer a gigabyte of storage, replace folders with search, and pay for it by scanning mail to choose ads
- Run a creator-driven video app without giving creators a way to earn
- Start with an expensive, low-volume sports car and use the money to fund progressively cheaper, higher-volume cars.
- Call publicly on the major record labels to abolish copy protection (DRM) on music sold online, rather than opening Apple’s own DRM to competitors.
- Build a national videotex network and give households terminals free of charge, initially to replace the paper telephone directory.
- Introduce the kiosque billing system, charging users per minute on their phone bill and sharing revenue with service providers.
- Buy the satellite network out of bankruptcy at a tiny fraction of its cost and serve niche markets.
- Pursue a web of machine-readable meaning built on formal standards, shared ontologies and logic, intended to be adopted broadly by publishers.
- Launch schema.org, a shared, simple vocabulary for structured data, rewarded by richer appearance in search results.
Open standards out-reach closed gardens
A system anyone can implement can reach more people than any single company can sign up. Closed systems win early on control and polish; open ones tend to win once reach matters more than control.
- Is our advantage control of the experience, or the size of the network we can reach?
- What would happen if a competitor made our core interface an open standard?
- Never design or sell chips under its own name, so that it would never compete with its customers
- Announce that GPS would be made available for civil aviation once it was operational, after Korean Air Lines Flight 007 was shot down
- End Selective Availability and commit to a civilian GPS signal at full accuracy, free of charge
- Turn Bank of America’s licensing programme into an independent organisation owned by its member banks, with shared rules and no single owner
- Build an instant payment system owned by a not-for-profit company of banks, open to any bank and to third-party apps, using simple addresses instead of account numbers
- Build the kernel in the open and release it under a licence that keeps it free
- Allow Android users to install apps from sources other than Google’s own store
- Require designated gatekeepers to allow alternative app stores, sideloading and outside payment, and ban self-preferencing
- Price software, training and services separately from computer hardware
- Open Amazon’s product pages to independent sellers, letting them compete directly with Amazon’s own retail offers
- Release anonymised viewing data and offer $1 million to anyone who could improve its rating predictions by 10 percent
- Build a complete free operating system, and protect it with a licence (copyleft) that requires anyone distributing modified versions to share their changes on the same terms
- Release the Linux kernel under the GPL and accept contributions from anyone over the internet
- Release the source code of its web browser, Communicator, under an open licence
- Release the Android operating system as open source, while keeping its app store and key services proprietary and licensed to phone makers
- Change licences from open source to source-available terms that prevent cloud providers from offering the software as a service
- Release the weights of its large language models for broad use under a custom licence, rather than offering them only through an API
- Sue makers of compatible spreadsheets for copying 1-2-3’s menus and “look and feel”
- Publish the Transformer architecture openly, in keeping with academic norms
- Release capable language models with downloadable weights that others could run and modify
- Release a strong reasoning model openly, with a technical report emphasising low training cost
- Put the subscriber’s identity on a removable SIM card, and use spare signalling capacity for text messages
- License the operating system to many manufacturers
- Accept mail from any server and believe whatever it says about the sender
- Store notes as plain Markdown files on the user’s own computer
- Reorient the whole of Microsoft around the internet, and treat a browser company as the main threat to Windows.
- Call publicly on the major record labels to abolish copy protection (DRM) on music sold online, rather than opening Apple’s own DRM to competitors.
- Keep Adobe Flash off the iPhone and iPad for good, and explain the decision publicly as a choice for open web standards, battery life, reliability and touch.
- Build hypertext with one-way links between documents on independent servers, with no central control, and accept that links may break.
- Launch Betamax on Sony's own design and seek partners to adopt it, rather than co-developing a shared standard with rivals before launch.
- Launch VHS with longer recording time and license it widely to other manufacturers.
A small working thing beats a grand design
A modest system that works today tends to persuade, spread and survive better than an ambitious one that arrives late. Grand designs often contribute ideas; small working ones contribute products.
- What is the smallest version that would convince a sceptic?
- Are we adding scope because users need it, or because the design feels incomplete?
- Add streaming free to its DVD-by-mail subscriptions instead of selling it as a separate service
- Found a semiconductor company that would only manufacture chips designed by other companies, with Philips as the main foreign partner
- Back 193-nanometre immersion lithography, which images chips through a thin layer of water, and cancel orders for tools using 157-nanometre light
- Launch mobile money for people without bank accounts, using a network of shopkeepers as agents to convert cash into electronic value
- Sell online payments to software developers, through a simple programming interface that hid banks, card networks and compliance
- Demonstrate time-sharing with a modest system that ran alongside batch work
- Leave the ambitious Multics project and build a small system, Unix
- Rank pages by the links pointing to them, and offer a plain search page rather than a portal
- Deliver business software through the browser from a single shared (multi-tenant) system, priced per user per month, with no installation
- Replace hundreds of legacy logistics systems with a single commercial ERP-based system while also changing its logistics processes
- Filter information using the reactions of other people, rather than only the content of the documents
- Release the Linux kernel under the GPL and accept contributions from anyone over the internet
- Launch first on the Apple II, a hobbyist machine, at about $100
- Release a free chat interface to an existing model as a “research preview”
- Turn a quiz-show demonstration into a business unit aimed at cancer care and other high-stakes fields
- Require real names and college email addresses, and open one campus at a time
- Start with an expensive, low-volume sports car and use the money to fund progressively cheaper, higher-volume cars.
- Use different decision-making processes for irreversible and reversible decisions, and fight the tendency of large organisations to treat every decision as irreversible.
- Cut about 10,000 jobs and 5,000 open roles, remove layers of management and cancel lower-priority projects, while protecting investment in AI and the long-term technology bets.
- Design hypertext as a single coordinated docuverse with two-way links, transclusion and versioning, rather than as links between independent documents.
- Invest in the Xanadu project to turn it into a commercial product.
- Launch the Newton MessagePad as a new category of pen-based personal digital assistant with natural handwriting recognition.
- Build a complete platform for communicating devices, Magic Cap and Telescript, licensed to manufacturers and carriers through a large alliance.
- Develop a global satellite telephone network for business travellers and others outside terrestrial coverage, and spin it into a separate company with outside investors.
- Launch Wave as a broad new category meant to unify email, chat and documents, with an invitation-only preview.
Compatibility becomes gravity
Once programs, files or habits depend on an interface, keeping it compatible becomes more valuable than improving it. Compatibility pulls future designs toward the past, and data outlives the systems built around it.
- Which of our interfaces will we still have to support in ten years?
- Are we choosing formats and APIs we are prepared to keep forever?
- Turn Bank of America’s licensing programme into an independent organisation owned by its member banks, with shared rules and no single owner
- Make an entire range of computers run the same software (System/360)
- Pay browser and device makers to make Google the default search engine, under revenue-sharing contracts that in many cases excluded rivals
- Stop selling boxed Creative Suite licences and move all customers to monthly or annual Creative Cloud subscriptions
- Sell Office as a subscription while continuing to offer perpetual versions
- Build standard business software that many companies could use with configuration, processing data in real time, rather than writing custom programs for each client
- Adopt a standard SAP retail system while keeping its own long-standing practice of valuing inventory at purchase prices, which required extensive customisation
- Replace its older ERP with S/4HANA and set a date for ending mainstream support of the old system, pushing customers to migrate, increasingly to SAP’s cloud
- Store, pack and ship goods for independent sellers, and make those goods eligible for Prime delivery
- Build only for the IBM PC, in assembly language for speed, and integrate spreadsheet, charts and database
- Concentrate on a version of 1-2-3 for OS/2 rather than Windows
- Sue makers of compatible spreadsheets for copying 1-2-3’s menus and “look and feel”
- Rename 27 human genes whose symbols spreadsheets turned into dates
- Use the machine’s instruction set as the interface, giving every user a complete virtual copy of the computer
- Bet on a graphical, what-you-see-is-what-you-get word processor for Windows
- Ship a scripting language in ten days that had to “look like Java”
- Add optional types on top of JavaScript instead of replacing it
- Protect users from knowing how data is stored (the relational model)
- Standardise Common Lisp so that Lisp programs could run on many machines, including general-purpose ones.
- Launch Betamax on Sony's own design and seek partners to adopt it, rather than co-developing a shared standard with rivals before launch.
- Jointly develop OS/2 as the successor to DOS, initially targeting the Intel 80286 processor.
- Position OS/2 2.0 and Warp as "a better DOS than DOS and a better Windows than Windows", emphasising their ability to run DOS and Windows programs.
- Jointly develop a new 64-bit architecture based on explicitly parallel instruction computing, rather than extending x86.
- Extend x86 to 64 bits with full backward compatibility, announced in 1999 and shipped as Opteron in 2003.
Distribution beats elegance
The technology that ships where people already are often beats a better design that requires them to go somewhere new. Being bundled, free, embedded or viral matters more than many technical virtues.
- Where are our users already, and how do we get there with the least friction?
- Is the invitation to new users built into the core action of the product?
- Sell songs one at a time for 99 cents, with the copy protection the labels required, through its own software and player
- Launch a free web map that users could drag, zoom and search in the browser, funded by advertising, and let outside developers build on it
- Leave the ambitious Multics project and build a small system, Unix
- Sell the order of search results by open auction, with advertisers paying only when a user clicks
- Pay browser and device makers to make Google the default search engine, under revenue-sharing contracts that in many cases excluded rivals
- Open the iPhone to outside developers only through an App Store in which Apple reviews every app, distributes it and takes 30 percent of paid sales
- Allow Android users to install apps from sources other than Google’s own store
- Promote “open source” as a practical development method, dropping the moral language of “free software,” to win over business
- Release the Android operating system as open source, while keeping its app store and key services proprietary and licensed to phone makers
- Not to seek a patent on the electronic spreadsheet
- Release a spreadsheet that separated formulas from data and organised numbers by named categories, first on NeXT
- Release a free chat interface to an existing model as a “research preview”
- Put an OpenAI model into its search engine within weeks of ChatGPT’s launch, then cap conversations when it misbehaved
- Package applications as portable images rather than inventing new isolation
- Bet on a graphical, what-you-see-is-what-you-get word processor for Windows
- Ship a scripting language in ten days that had to “look like Java”
- Allow only web apps at first, then reverse course and open a native App Store
- License the operating system to many manufacturers
- Mix code into HTML pages and make deployment as simple as copying a file
- Make email free, run it in the browser, and put an invitation at the foot of every message
- Copy Snapchat’s Stories format openly
- Reorient the whole of Microsoft around the internet, and treat a browser company as the main threat to Windows.
- After the success of Windows 3.0, shift strategic priority to Windows and away from joint OS/2 development.
- Launch schema.org, a shared, simple vocabulary for structured data, rewarded by richer appearance in search results.
Users decide what the product is for
Many products found their real purpose in a use their designers did not plan: text messages from spare signalling capacity, chat tools built for game studios, graphics chips that trained neural networks, a query language used for transactions.
- What are users doing with the product that we did not design for?
- Are we prepared to let the unplanned use become the main one?
- Run a central index of the songs stored on users' computers, with the files themselves passing directly between users
- Build its own fraud detection software and concentrate on payments between buyers and sellers on eBay
- Launch mobile money for people without bank accounts, using a network of shopkeepers as agents to convert cash into electronic value
- Let buyers and sellers rate each other publicly after every transaction, instead of policing disputes itself
- Charge sellers nothing to list, hold buyers’ payments in escrow until goods arrived, and let buyers chat with sellers
- Let customers pay in cash when goods were delivered
- Build a video app around a feed chosen entirely by algorithm from all content, rather than from accounts the user follows
- Fill the main feed with full-screen recommended videos from accounts users did not follow, following TikTok
- Launch first on the Apple II, a hobbyist machine, at about $100
- Release a free chat interface to an existing model as a “research preview”
- Put the subscriber’s identity on a removable SIM card, and use spare signalling capacity for text messages
- Design an English-like query language for non-programmers
- Open graphics chips to general computation (CUDA), then design chips only for neural networks (TPU)
- Rebuild the product four times around blocks that can become databases
- Argue that Google should treat every product as a platform, as Amazon did after Jeff Bezos required every team to expose its work through service interfaces.
- Position a group chat tool as organisational transformation, and make the team’s job selling that change rather than features.
- Launch with a one-hour recording time, favouring a smaller cassette and picture quality.
- After the success of Windows 3.0, shift strategic priority to Windows and away from joint OS/2 development.
- Launch the Newton MessagePad as a new category of pen-based personal digital assistant with natural handwriting recognition.
- Launch commercial service with large handsets priced around $3,000 and calls costing several dollars a minute.
- Extend x86 to 64 bits with full backward compatibility, announced in 1999 and shipped as Opteron in 2003.
- Launch Wave as a broad new category meant to unify email, chat and documents, with an invitation-only preview.
- Stop developing Wave as a standalone product about fifteen months after its unveiling, and carry its technology into other products.
A defining strength becomes the constraint
The choice that made a product succeed in one environment can hold it back when the environment changes: a keyboard built for email, a language that trusts experts, a graph of friends that everyone thought was the moat.
- Which of our strengths depends on an assumption about the world that could change?
- What would a competitor build if they assumed the opposite?
- Pursue Napster in court until it closed, rather than license it as a paid service, even after Bertelsmann broke ranks to back a membership version
- Add streaming free to its DVD-by-mail subscriptions instead of selling it as a separate service
- Aim its 10-nanometre process at an unusually large increase in density, 2.7 times its 14-nanometre process, without extreme ultraviolet lithography
- Let any program touch any part of the machine (no memory protection)
- Use Google’s results to power Yahoo!’s web search, with Google credited on the page
- Host conventional client-server business software in their own data centres, one installation per customer, funded largely by debt
- Consolidate the market for installed enterprise applications by acquiring PeopleSoft and Siebel
- Adopt a standard SAP retail system while keeping its own long-standing practice of valuing inventory at purchase prices, which required extensive customisation
- Run its Chinese business, EachNet, on eBay’s global platform and fee model, against a local rival that charged sellers nothing
- Change licences from open source to source-available terms that prevent cloud providers from offering the software as a service
- Concentrate on a version of 1-2-3 for OS/2 rather than Windows
- Trust the programmer with memory and leave out automatic checks
- Remove the keyboard and buttons and make the whole front a touchscreen
- Store relationships as chains of pointers that programs navigate
- Choose every video with a recommendation model instead of from the people you follow
- Promise to “remember everything”: capture anything, find it anywhere
- Keep Adobe Flash off the iPhone and iPad for good, and explain the decision publicly as a choice for open web standards, battery life, reliability and touch.
- Prepare Nokia’s employees to abandon its own smartphone platforms, an announcement that came days later as a partnership to build phones on Microsoft’s Windows Phone.
- Redefine Microsoft as “the productivity and platform company for the mobile-first and cloud-first world,” moving the company’s centre of gravity away from Windows and devices.
- Continue investing primarily in proprietary Lisp hardware as general-purpose workstations improved rapidly.
- Continue to rely on and protect Minitel as the national online system while the internet spread elsewhere.
Hiding the machine wins once the machine is cheap
Raising the level of abstraction (compilers, relational queries, virtual machines, declarative interfaces) costs efficiency at first. It wins once hardware gets cheap enough that people’s time is the scarce resource, provided the translation is trustworthy.
- What do our users currently have to know that the system could know for them?
- What is the single objection (speed, cost, trust) that would stop people adopting a higher-level approach?
- Found a semiconductor company that would only manufacture chips designed by other companies, with Philips as the main foreign partner
- Combine the Air Force and Navy concepts into one system with atomic clocks on the satellites and receivers that only listen, so that any number of users could share it
- Sell online payments to software developers, through a simple programming interface that hid banks, card networks and compliance
- Build an instant payment system owned by a not-for-profit company of banks, open to any bank and to third-party apps, using simple addresses instead of account numbers
- Build standard business software that many companies could use with configuration, processing data in real time, rather than writing custom programs for each client
- Store, pack and ship goods for independent sellers, and make those goods eligible for Prime delivery
- Recommend products by relating items to items, computed in advance from purchase histories, rather than matching each customer to similar customers at the moment of the visit
- Offer a free spreadsheet in the browser, built for sharing and simultaneous editing
- Add Python and an AI assistant inside the spreadsheet
- Release a strong reasoning model openly, with a technical report emphasising low training cost
- Use the machine’s instruction set as the interface, giving every user a complete virtual copy of the computer
- Treat the compiler’s efficiency, not the language design, as the real challenge
- Remove the keyboard and buttons and make the whole front a touchscreen
- Describe the interface as a function of state and let the library work out the changes
- Protect users from knowing how data is stored (the relational model)
- Let the database choose how to run each query using statistics about the data
- Store the program in memory alongside the data
- Build a programmable processor instead of custom calculator logic
- Offer a gigabyte of storage, replace folders with search, and pay for it by scanning mail to choose ads
- Argue that Google should treat every product as a platform, as Amazon did after Jeff Bezos required every team to expose its work through service interfaces.
- Commercialise the MIT Lisp machine as dedicated hardware rather than as software for general-purpose computers.
Making something cheap moves the cost to someone else
Costs rarely disappear. A technology that makes sending, building or shipping cheaper usually moves the cost onto the receiver, the user’s device or a later team, often people who did not make the choice.
- Who pays for the effort this feature saves us or our users?
- Does the person bearing the cost have any say in the decision?
- Build advanced fabs in Arizona, under pressure from the US government
- Deliberately degrade the accuracy of the public GPS signal through Selective Availability
- Fund card systems through interchange fees paid by merchants’ banks to cardholders’ banks, with rules that kept merchants from charging card users more
- Withdraw ₹500 and ₹1,000 banknotes as legal tender with four hours’ notice
- Set the merchant discount rate on UPI and RuPay debit payments to zero by law
- Put AI-generated answers at the top of the results page, above the links
- Let customers pay in cash when goods were delivered
- Make the word processor write: drafts, rewrites and summaries, priced into subscriptions
- Build sites as single-page JavaScript applications
- Choose availability over consistency: an always-writeable store
- Drop SQL, joins and schemas for scale and flexibility
- Show everyone whether you are online (presence)
- Sell “a reduction in information overload” through open, searchable channels
- Give receivers control: threads, do-not-disturb, scheduled send, status messages
- Add AI assistants that summarise, triage and draft messages
- Respond to the end of Dennard scaling with multiple cores
- Fight spam with learning filters and sender reputation rather than a price on sending
- Let a model read a person’s notes and answer questions grounded in them
- Position a group chat tool as organisational transformation, and make the team’s job selling that change rather than features.
- Commercialise the MIT Lisp machine as dedicated hardware rather than as software for general-purpose computers.
- Introduce the kiosque billing system, charging users per minute on their phone bill and sharing revenue with service providers.
Decide at the speed of reversibility
Decisions that are hard to undo deserve slow, careful process; decisions that are easy to undo deserve speed. Organisations get into trouble by applying one process to both, becoming either reckless or slow.
- If this turns out wrong, how expensive is it to go back?
- Are we spending one-way-door process on a two-way-door decision, or the reverse?
- Build advanced fabs in Arizona, under pressure from the US government
- Replace Google's maps on the iPhone with its own, assembled from licensed data, in iOS 6
- Withdraw ₹500 and ₹1,000 banknotes as legal tender with four hours’ notice
- Install SAP R/3 and a new automated warehouse system at the same time, and bid for contracts on the basis of savings the systems were expected to deliver
- Switch on new ERP, customer relationship and supply chain systems together in July, shortly before the Halloween ordering season
- Replace hundreds of legacy logistics systems with a single commercial ERP-based system while also changing its logistics processes
- Withhold the full GPT-2 model over misuse concerns and release it in stages
- Put an OpenAI model into its search engine within weeks of ChatGPT’s launch, then cap conversations when it misbehaved
- Allow only web apps at first, then reverse course and open a native App Store
- Launch a separate experimental email app, then close it and move its best ideas into Gmail
- Prepare Nokia’s employees to abandon its own smartphone platforms, an announcement that came days later as a partnership to build phones on Microsoft’s Windows Phone.
- Split Netflix in two: keep streaming under the Netflix name and move DVD-by-mail to a separate service called Qwikster, with separate websites and accounts.
- Use different decision-making processes for irreversible and reversible decisions, and fight the tendency of large organisations to treat every decision as irreversible.
- Invest in the Xanadu project to turn it into a commercial product.
- Discontinue the Newton platform.
- Develop a global satellite telephone network for business travellers and others outside terrestrial coverage, and spin it into a separate company with outside investors.
- Adopt 64-bit x86 for mainstream products while continuing Itanium for high-end servers.
- Stop developing Wave as a standalone product about fifteen months after its unveiling, and carry its technology into other products.
Permissive defaults are borrowed time
Trusting every input, every sender or every programmer makes a system easy to adopt. The bill arrives later, when the system is connected to people who do not deserve that trust, and tightening a default after millions depend on it takes years.
- If this default is wrong in five years, how would we tighten it without breaking everyone who relies on it?
- Who outside our current users could exploit the trust we are extending?
- Run a central index of the songs stored on users' computers, with the files themselves passing directly between users
- Let anyone upload video and remove infringing material only on notice, then build a system that matched uploads to owners' works and let owners block them or take a share of the advertising
- Collect precise location from phones through apps and advertising code, and sell or share it
- Launch a general-purpose credit card by mailing 60,000 unsolicited cards to residents of Fresno at once
- Let any program touch any part of the machine (no memory protection)
- Open the social network to outside apps with broad access to users’ data and their friends’ data, then progressively restrict it
- Treat data from its Horizon accounting system as reliable evidence against sub-postmasters, prosecuting and pursuing them for shortfalls, despite reports of errors
- Run a new value-at-risk model for a large derivatives portfolio on manually maintained spreadsheets
- Load laboratory results into a spreadsheet format with a fixed row limit as part of the national Covid-19 reporting pipeline
- Let a website chatbot answer policy questions, then argue in a dispute that the chatbot was responsible for its own words
- Add the null reference because it was easy to implement
- Trust the programmer with memory and leave out automatic checks
- Accept mail from any server and believe whatever it says about the sender
- Turn profiles into News Feed, a continuous stream of friends’ activity
- Make security, availability and privacy Microsoft’s highest priority, and choose security over new features whenever they conflict.
- Tell public investors that Facebook’s mission came before profit and that it would keep a culture of rapid, iterative building, which Zuckerberg called the Hacker Way.
- Build hypertext with one-way links between documents on independent servers, with no central control, and accept that links may break.
- Pursue a web of machine-readable meaning built on formal standards, shared ontologies and logic, intended to be adopted broadly by publishers.
Computing swings between the centre and the edge
Work moves between central shared machines and personal devices as the relative cost of computing, storage and networks changes: time-sharing, then PCs, then the cloud; servers, then browsers, then servers again.
- Which cost (compute, bandwidth, latency, privacy) is changing fastest, and which way does it push our architecture?
- What would our product look like at the other end of the pendulum?
- End its Netflix distribution deal, keep its films for its own service and launch Disney+ at a low price
- Add a cheaper plan with advertising and charge for accounts shared outside the household
- Deliver business software through the browser from a single shared (multi-tenant) system, priced per user per month, with no installation
- Host conventional client-server business software in their own data centres, one installation per customer, funded largely by debt
- Build human resources and finance software for large enterprises from scratch as a shared online service
- Rebuild its system for client-server computers rather than staying on mainframes
- Launch Excel on the Macintosh first, designed around the mouse and a graphical interface, before bringing it to Windows
- Offer a free spreadsheet in the browser, built for sharing and simultaneous editing
- Ask the guest operating system to cooperate with the hypervisor (paravirtualization)
- Move the document into the browser with simultaneous editing
- Put capacity and intelligence in the network: small cells, reused frequencies, simple handsets
- Build sites as single-page JavaScript applications
- Render pages on the server and hydrate only small “islands” of interactivity
- Rent computing capacity by the hour
- Open graphics chips to general computation (CUDA), then design chips only for neural networks (TPU)
- Redefine Microsoft as “the productivity and platform company for the mobile-first and cloud-first world,” moving the company’s centre of gravity away from Windows and devices.
The metric becomes the target
Any number chosen to stand for value (likes, comments, open rates, time spent) will be optimised directly, by users, by competitors and by the product itself, until it stops measuring what it was meant to.
- What would someone do to raise this metric without creating any real value?
- Which behaviour would we be ashamed to see the metric reward?
- Sell text ads by cost-per-click auction, rank them by bid multiplied by expected click-through rate, charge just above the next bidder, and keep them labelled and separate from the organic results
- Extend the ad auction to other people’s websites, sharing revenue with publishers
- Change the ranking to demote large volumes of thin, low-quality pages, even though many of them displayed Google ads
- Default many customers to an “annual, paid monthly” plan that charged a large fee for cancelling in the first year
- Release anonymised viewing data and offer $1 million to anyone who could improve its rating predictions by 10 percent
- Rank and recommend videos by expected watch time rather than clicks or views
- Run a new value-at-risk model for a large derivatives portfolio on manually maintained spreadsheets
- Replace much of customer service with an AI assistant, then reverse course and hire people again
- Add a reward for users’ immediate thumbs-up reactions to the main ChatGPT model, then roll it back
- Build a social network on top of everyone’s Google account
- Re-rank the feed for “meaningful social interactions”
- Promise to “remember everything”: capture anything, find it anywhere
- Tell investors in advance that Amazon would sacrifice short-term profit for long-term market leadership, and how it would make that trade.
- Go public on Google’s own terms: a dual-class share structure that kept control with the founders, no traditional earnings guidance, an auction-based offering and an explicit commitment to long-term, high-risk projects.
Real values are what you reward
An organisation’s stated values matter less than what it hires for, promotes, pays for and tolerates. Documents that change culture work by changing those decisions, not by naming virtues.
- Which behaviour did we actually reward in the last promotion round, and does it match what we say we value?
- What would a new hire infer about our values from the last ten decisions they saw?
- Never design or sell chips under its own name, so that it would never compete with its customers
- Respond to a court order requiring it to let developers link to outside payments by charging a 27 percent commission on such purchases and adding warning screens
- Default many customers to an “annual, paid monthly” plan that charged a large fee for cancelling in the first year
- Treat data from its Horizon accounting system as reliable evidence against sub-postmasters, prosecuting and pursuing them for shortfalls, despite reports of errors
- Reduce recommendations of “borderline” videos that came close to violating its rules, without removing them
- Tell investors in advance that Amazon would sacrifice short-term profit for long-term market leadership, and how it would make that trade.
- Make security, availability and privacy Microsoft’s highest priority, and choose security over new features whenever they conflict.
- Go public on Google’s own terms: a dual-class share structure that kept control with the founders, no traditional earnings guidance, an auction-based offering and an explicit commitment to long-term, high-risk projects.
- Replace most rules and processes with high talent density, candour and context, and let people go who are merely adequate.
- Split Netflix in two: keep streaming under the Netflix name and move DVD-by-mail to a separate service called Qwikster, with separate websites and accounts.
- Tell public investors that Facebook’s mission came before profit and that it would keep a culture of rapid, iterative building, which Zuckerberg called the Hacker Way.
- Redefine Microsoft as “the productivity and platform company for the mobile-first and cloud-first world,” moving the company’s centre of gravity away from Windows and devices.
- Cut about 10,000 jobs and 5,000 open roles, remove layers of management and cancel lower-priority projects, while protecting investment in AI and the long-term technology bets.
Network effects run in both directions
Each new user makes a network more valuable for the others, so growth accelerates. The same mechanism works in reverse: when a key group leaves, the network becomes worse for everyone who stays, and decline can be as fast as growth.
- Which group of users, if they left, would make the product worse for everyone else?
- Are we measuring accounts, or relationships and activity between users?
- License the major labels' catalogues for a free, advertising-supported tier as well as a paid one, granting the labels equity and advance payments in return
- Halt development of 7-nanometre and more advanced processes and concentrate on established ones
- Launch a general-purpose credit card by mailing 60,000 unsolicited cards to residents of Fresno at once
- Open the social network to outside apps with broad access to users’ data and their friends’ data, then progressively restrict it
- Require real names and college email addresses, and open one campus at a time
- Decline Google’s offer and keep growing while the site slowed and fake profiles were deleted
- Run a creator-driven video app without giving creators a way to earn
- Build a social network on top of everyone’s Google account
- Prepare Nokia’s employees to abandon its own smartphone platforms, an announcement that came days later as a partnership to build phones on Microsoft’s Windows Phone.
- Buy young social apps, Instagram above all, before they could grow into threats to Facebook.
Defences against abuse concentrate the market
Spam, fraud and harmful content are fought most effectively by whoever sees the most traffic. Reputation systems and filters make open systems usable, and quietly favour the largest operators over newcomers.
- Will our trust and safety measures make it harder for legitimate new participants to join?
- Who decides what counts as abuse, and can they be held to account?
- Let anyone upload video and remove infringing material only on notice, then build a system that matched uploads to owners' works and let owners block them or take a share of the advertising
- Build its own fraud detection software and concentrate on payments between buyers and sellers on eBay
- Change the ranking to demote large volumes of thin, low-quality pages, even though many of them displayed Google ads
- Let buyers and sellers rate each other publicly after every transaction, instead of policing disputes itself
- Apologise publicly and guarantee hosts up to $50,000 for damage caused by guests, after a host’s home was ransacked
- Fight spam with learning filters and sender reputation rather than a price on sending
Taking choices away can be the product
Removing options (strong conventions, uniform layouts, a few well-chosen features) can be worth more than flexibility, because it saves users from decisions that do not deserve their attention.
- Which decisions are our users making repeatedly that we could make once, well, for everyone?
- Which options exist only because we could not decide?
- Commission original series without pilots and release whole seasons at once
- Rank pages by the links pointing to them, and offer a plain search page rather than a portal
- Open the iPhone to outside developers only through an App Store in which Apple reviews every app, distributes it and takes 30 percent of paid sales
- Require the largest platforms to explain their recommender systems and offer at least one option not based on profiling
- Make the routine decisions once, for everyone (“convention over configuration”)
- Replace most rules and processes with high talent density, candour and context, and let people go who are merely adequate.