What this covers
This is a conversation between Instagram's co-founders, Kevin and Mike, that moves from product craft to machine learning to management. The product thread starts with jobs to be done, which Kevin describes as customers hiring a product to meet social and emotional needs as well as functional ones. Instagram used it to make each product's problem explicit. The examples are concrete: uploading a photo in the background while the user types a caption, so that "done" feels instant, and tuning the feed for slow 3G phones, a practice the founders call Optimize for the Worst Device. They also note that Instagram had spent about $65,000 by launch, against roughly $500,000 raised.
The middle of the conversation turns to machine learning. Mike argues it pays off where data has reached the scale and quality that general-purpose prediction is valuable, not because it uncovers hidden patterns. Teams that chase accuracy drift into weak features like camera direction or weather and end up with overfitting. With libraries now free, the scarce skill is using them well. Smaller asides cover computer vision models that transfer to unrelated problems and cheap EEG headbands that classify calm versus stress.
The contested positions come last. Mike favors launching major changes to nearly everyone at once, as with Stories reaching about 99.5 percent of users, over hedged half-launches. He is skeptical of speculative R&D such as AR headsets, and he rejects a VC's claim that product doesn't matter. The two also discuss carrier zero-rating in India, global content rules on nudity, and management as a parallel track that needs mentorship and skip-level meetings.
Systrom and Krieger argue that great products come from repeatedly identifying the specific problem people hire a product to solve, staying community-first and humble, and shipping decisively, not from grand visions or technology for its own sake.
- Instagram's growth came from solving concrete, observed problems (photo quality, sharing, speed) and from usage data showing how specific groups actually used the product
- Jobs-to-be-done framing revealed that adjacent behaviors and friend-focused features drove usage, while speculative future features were deferred
- Decisive full launches and a 'we may not be right, but we're not confused' mindset beat hedged, incremental half-launches
Products must fit users' existing behavior, devices, and local context in each market.
- New products succeed when they build on behavior already happening on the platform, as with Indonesian sellers who posted items and took payment via WhatsApp; a commerce product built on that existing behavior is far more likely to succeed than one that requires entirely new behavior.
“I think you always need to build in these adjacencies.”
- In markets like India, users loaded as much content as they could at home over WiFi because mobile connectivity was rate-limited elsewhere, and Instagram's refresh behavior broke when the network failed, so the product had to account for a whole day's experience rather than a single session.
“there's things you can do around thinking about people's day long experience of Instagram versus sort of single experience of Instagram.”
- Carrier zero-rating deals, where a carrier makes Instagram free to users, were among Instagram's largest international growth drivers, with a cell provider in India producing rocket-ship growth, whereas a similar deal at AT&T in the US would have produced only a blip.
“One of our largest growth drivers ever internationally, was, there's a concept called Zero-rating”
- Sign-ups from a year-and-a-half-old Instagram version were not spam but phones sold preloaded with an old app in markets like Nigeria, so the product had to work for the version a user already had, not just the latest one.
“It was phones being sold, I think it was in Nigeria, and they came preloaded with Instagram, an old version”
- Instagram's community is a collection of many distinct communities, and product teams must go deep on one at a time, such as Brazilian entrepreneurs using Instagram as a storefront or French teens, by sending researchers and engineers to observe them firsthand.
“You get this sort of amalgamation of a bunch of communities”
Growth and monetization must be rebuilt from current conditions, not copied from past playbooks.
- When Instagram's growth fell off a cliff, the cause was usually an internal failure, such as a broken push notification system (99 times out of 100), rather than a competitor or a change in user behavior.
“99 out of a hundred times, it was the push notification system broke.”
- As Instagram aged, users became more selective about what they posted, creating a self-limiting feed that differed from the original multi-post-per-day product, so the company had gotten worse at the job users originally hired it for, and Stories restored it.
“people start being more selective about what they post.”
- Instagram aimed to build tools that let people feel they captured something new and creatively expressive, a principle that persisted and evolved as the product grew.
“We talked often about unlocking people's creativity as well.”
- Mike wrote Instagram's early business plan for a visa application as a hopes-and-dreams document based on a local discovery ads model for Burbn, which showed the founders had no formal monetization plan at the time.
“We'd raised on a hope and a dream and an app demo”
- Because Instagram is highly visual and the best ads have always been visual, the founders believed that a large enough audience would be monetizable, even though the exact ad format remained to be determined.
“the best ads have always been visual. So there was clearly a plausible path to that working”
- Instagram's first ad product ran on a whiteboard calendar with Sharpie-written advertiser slots, a deliberately simple approach that tested whether advertisers would pay before investing in a sophisticated auction system.
“View one of our auction system was literally a whiteboard with a calendar that was drawn in Sharpie.”
- The viral growth channels that Instagram used, such as cross-posting to Twitter and Facebook social graphs, have vanished, so a playbook from 2010 cannot be copied and growth must be figured out iteratively with current distribution mechanisms.
“Time travel back to 2010 and we can have a conversation.”
Major product changes should launch decisively to nearly everyone rather than being tiptoed in.
- Instagram Direct's first version, which let people add others to a private photo for conversation, did not meet expectations and was replaced by persistent group threads, which broke a niche live-action roleplay community that used ephemeral group creation.
“We fired that product and we evolved it to eventually to be like threads that were more permanent”
- Major product changes should be launched decisively to nearly everyone at once rather than tiptoed in, because half-launches create social confusion and make users doubt the company's conviction; Instagram Stories launched to about 99.5 percent of users in 2016 with only a brief holdout.
“for stories, we launched it to basically everybody in the world at the exact same time.”
- Every major change that caused significant pain was either the best or the worst decision the company made, and leaders must accept that disruptive changes will sometimes break the product for some group while avoiding paralysis from assuming they will fail.
“You can't treat every major disruptive change as being necessarily the right thing for the product, but if you automatically assume it will be the wrong thing for the product, you'll just be paralyzed”
- If a team automatically assumes a major disruptive change will fail, it will be paralyzed and never make changes; instead it should launch with conviction, since half-launches create confusing social dynamics among users.
“for stories, we launched it to basically everybody in the world at the exact same time.”
- A product team must sometimes change the product in ways that break some use cases; Instagram Direct's redesign broke a niche LARP storytelling use case, and while the affected group was small, a tiny percentage of a billion users is still a lot of people.
“You have to be occasionally okay, changing the product in a way that might actually sort of break the product in some way for some group of people.”
Teams should be run like product companies, with management kept as a distinct, mentored track.
- Instagram ran its infrastructure teams like product companies by requiring each team to write the job it performed for other teams, which cut gold-plated projects nobody adopted and focused work on what mattered.
“I had all of our infrastructure teams actually write up what job they were doing for the other teams”
- Management should be a parallel career track to individual contribution rather than a promotion, because people should not self-select into management for status; people should be allowed to transition between tracks, and new managers need real mentorship and skip-level check-ins.
“make management a parallel career track to individual contribution.”
- Reid Hoffman's approach of finding great partners to make projects happen taught Mike the value of leverage; Mike notes he tends to do too much himself, and that lesson matters more now that the partnership has no team to delegate to.
“for Reid, I think it's leverage.”
- Employee survey metrics such as optimism about Instagram's future and expected tenure were most strongly correlated with whether the team was doing great product work, not with management initiatives, which were needed only to fix blocking problems.
“almost every metric in the end was correlated with, are we doing great work on the product?”
- Founders should not self-select out of opportunities because they feel they do not know enough yet; Instagram's founders had no computer science background and no management experience, but learned by remaining open-minded and hiring people who could teach them.
“Never self-selecting out of things because you don't feel like you know them yet.”
- Promoting someone to manager without real mentorship fails; new managers need coaching in their first year and skip-level meetings to check on their direct reports.
“Actually that's a real transition and requires real mentorship.”
Machine learning needs quality data and a reasonable hypothesis, not hunts for hidden patterns.
- Data has become a first-class citizen in programming: early Instagram analytics pipelines were ad-hoc nightly scripts that often failed, whereas modern data pipeline tools are declarative, well-structured, and built for whole teams rather than one analyst.
“data has become, going to say like, a first-class citizen in programming”
- Before throwing data into a prediction system, you need a reasonable hypothesis about why the data should yield a useful prediction; otherwise teams end up adding ever more features from a long tail of weak signals that overfit the data.
“you need to have a reasonable hypothesis about why it is that throwing a bunch of data into some kind of prediction system is going to yield a useful prediction.”
- Machine learning is useful where data quantity and quality have reached the point that a general-purpose prediction is valuable, not because it reveals hidden patterns nobody knew existed; those patterns probably were not there to begin with.
“It's not, I think, useful in the magic sense of, 'Oh, all these patterns that we never knew were there.' They probably weren't there to begin with.”
- The best founders using ML get interested in what their underlying data actually looks like, since problems such as images arriving upside down are only discovered by diving into examples, and careless users miss them.
“I think the founders who get interested in what their data actually looks like underneath.”
- When ML teams chase accuracy by adding features, they drift into a long tail of weak features such as camera direction or weather, which may show short-term predictive power but overfit the data and produce downstream distortions.
“you find your team is mostly doing feature engineering on this long, long tail of stuff... you're probably over-fitting the data.”
Early products should favor simple, focused work over complex, speculative, or distracting efforts.
- Early-stage founders should balance unscalable hands-on work, such as handling support tickets personally, with building for scale, and should not become so enamored with scalability that they lose a deep understanding of the problem.
“bridge the things that were really unscalable, handling support tickets ourselves at the beginning”
- Instagram's core operating principle was to do the simple thing first, such as using Django rather than building a custom web framework, because complex early solutions cost months and were often unnecessary.
“We had this phrase at Instagram, which we used religiously, which is, 'Do the simple thing first.'”
- A case-study lesson the founders adopted is that a team may not be right but should not be confused; a two-person team should focus on one thing while staying humble about being wrong, rather than diversifying efforts.
“We may not be right, but we're not confused.”
- Instagram's early no-external-meetings policy for four months before launch was adopted because each minute away from the desk was time the product was not being built, and it drove investors crazy.
“We had a no external meeting policy for four months before launch”
- Companies often confuse product-heavy with R&D-driven, and pursuing speculative technology like AR headsets wastes effort that could solve problems for a billion users today; R&D is needed somewhere, but most product teams should focus on what can be solved now.
“I don't think enough teams think about the opportunity costs of being super pie in the sky.”
Differentiation comes from data and problem choice, and trust comes from authentic, human communication.
- Since open-source and cloud providers supply most core technology, companies should differentiate on the data they can ingest, provide, or the problem they solve, rather than reinventing the same technology as hundreds of other companies.
“Now, it's what you do with it, and now you differentiate on what data you can get in, what data you can provide or what problem you're solving for people”
- Users can tell a company's underlying intentions and ethos, so being authentic and speaking to users as real people, such as posting announcements from the founders, builds the trust that community-first requires.
“people can tell who's behind the company, so be authentic and talk to them as real people rather than as an abstract user base.”
- A VC claimed that product does not matter and go-to-market is everything, which Mike rejects as applicable only within certain categories, since his investing edge lies in which problems are solved and how well.
“I don't think I'd enjoy investing because if there is any amount of edge or unique view we have on the world, I think it's what problems are solved”
- Tech people are drawn to hobbies that are precise and process-driven, such as coffee, photography (ISO and exposure), flying, and cooking, because each rewards technical control.
“There's probably four things that I think people in tech are naturally drawn to.”
Machine learning on brain and vision data converges with how humans perceive and generalize.
- Neuroscience and computer science diverged around the 1970s and have only recently begun to reconverge, and in the next five to ten years ML may increasingly reflect parallels to how human vision works.
“the neurological side or the neuroscience and the computer science, I'd probably diverge in the 70s and haven't really reconverged”
- Computer vision algorithms trained on one set of images often work surprisingly well on other problems, sometimes not even vision-related, suggesting a convergence with how the brain generalizes its learned seeing skills.
“you train a bunch of computer vision algorithms on a set of images for one particular problem. And weirdly, it works really well for other sets of problems, sometimes not even vision-related.”
- Cheap consumer EEG headbands such as Muse, which cost about $200 to $250 versus $10,000 to $20,000 for lab equipment, are useful mainly because machine learning models trained on brainwaves can classify states such as calm versus stress.
“the only reason that the Muse, which basically helps you meditate, is useful is that they've trained a bunch of ML models on brainwaves.”
- Wearables such as the Oura ring can reveal how stress shows up in heart rate variability before bed, which adds insight beyond confirming what one already feels, such as poor sleep.
“what does it look like when I am more stressed out before bed in terms of heart rate variability.”