Taste, Authenticity and Small Pods: Instagram's Adam Mosseri on Building Products in an AI World

Mosseri says the canonical product team at a company like Meta used to be a baker's dozen: a few Android, iOS and server engineers, plus a PM, designer, data scientist and, with luck, a researcher. That structure existed partly so that anyone writing code had a reviewer familiar with the codebase. This year, he says, Instagram has adopted "pods" of four to six more-generalist engineers working with one person in a new "product staff" role.

Product staff is an evolution of the PM who can also do some of what a designer, data scientist and researcher do, using the latest tools. Specialists are added based on the work: a pricing strategy needs a senior data scientist, and a novel experience needs a very senior product designer. The result is a core of roughly six or seven people who, with less to coordinate, can move faster and avoid design by committee.

He gives an example of why this is newly possible. A basic waterfall analysis of where people drop off when creating Reels used to need bespoke work from a data scientist, but internal tools can now pull it automatically. Mosseri says the functions are bleeding into each other, and some pods now have no dedicated data scientist, designer or researcher at all. He credits both AI productivity and the simple effectiveness of small teams.

Lenny relays that another leader he interviewed hires builders with great taste and people with deep domain expertise, and Mosseri agrees that taste matters a ton. With building easier, he argues, your time is better spent working out what to build. Although many designers are anxious as engineers and PMs start doing design work, he is "pretty long" on them because taste is much harder to imagine being automated.

Lenny observes that AI-built apps tend to share a recognizable look, and wonders why designer hiring has flatlined. Mosseri expects the answer to be that designers will not stay in traditional design roles. A great designer usually has strong opinions on product strategy and go-to-market, so some of the strongest product staff will be converts from design and data science. He adds that he started as a designer himself and notes his own bias.

Mosseri says anxiety is everywhere, but functions will persist in a different shape. Teams still need to hire and mentor people on the way to becoming senior, since you cannot have only super-senior data scientists with no one coming up behind them. A reasonable soft landing, he suggests, is for those not on that path to move into generalist roles. He also urges people not to be overconfident in predictions, given how much is in flux.

For hiring, Mosseri has long looked for three traits in almost any function: grit, being a quick learner, and being self-aware enough to take feedback. For the next five to ten years he adds staying curious and putting yourself out there. He compares it to learning a language, where a good predictor of success is whether you are willing to sound foolish and be corrected. With large organizations giving way to smaller teams, he expects less demand for people who manage thousands.

AI makes the job different rather than simply easier, he says. Engineers who once spent a large share of their time writing code now spend most of it planning and reviewing, and who succeeds depends on whose strengths fit that shift. He describes getting to program again for the first time in about ten years, calling himself a mediocre engineer who can now write code responsibly. He pushes back on the binary of being pro- or anti-AI: the tools are amazing at some things and remarkably bad at others, and the winners will be clear-eyed about which.

On cost, Mosseri calls a leaderboard of token spend a terrible idea and says it is easy to build a "token incinerator" that creates little value. Meta does not currently cap tokens, but he can imagine caps within a year or two if a strong engineer's burn rate approaches their salary, tied to the company's trust in their ability to use tokens with positive ROI. He treats tokens like any other resource, such as GPUs, labeling budgets and headcount, expects total usage to rise, and expects prices to fall as frontier models compete.

Asked where human brains stay most valuable as AI eats the development cycle, Mosseri names taste and judgment, especially around strategy. He compares the future to management: you define success, choose how prescriptive to be, and give feedback along the way. He defines vision as the state of the world or product you want to reach, and strategy as an opinionated path to it. A strategy should be something a reasonable person could disagree with, otherwise you are only competing on raw execution.

Lenny expected AI to be excellent at strategy, but Mosseri says it only becomes useful when you steer it aggressively using constraints such as technology, team motivation, competition, regulation and brand identity. A lazy request returns something predictable. He recommends telling the model to be critical and picking one that is willing to push back, since he does not want a pleaser. When Lenny raises the "reverse centaur" worry of AI directing humans, Mosseri cautions against outsourcing a workflow just because you can.

On product leadership, he says some great leaders are idea machines, but many of the best are curators of people, ideas, technologies and strategies. He does not mind whether a strategy comes from the leader as long as it is excellent and everyone has bought in. He also weighs how a leader fits the rest of their leadership team, including skills, weaknesses and a harder-to-measure chemistry, since trust and rapport let a team work through almost anything.

Mosseri says the biggest misconception is the reverse of what people think: they assume the algorithm has a detailed semantic understanding of their interests. In practice, recommender progress over the last five to ten years has come from large embedding models that produce illegible vectors, so the system does not know you like surfing; it has a number that happens to correlate with surfing. Much of the logic is that people who liked these photos also liked those.

LLMs are changing this because they can describe in words what a region of the embedding space represents, such as "deep pour-over coffee snobbery". Instagram's "your algorithm" feature lets people see the topics it thinks they like and add or remove them. Mosseri hopes to extend it to non-topical requests, such as seeing friends more, or not seeing seven photos in a row. He says only now is the system becoming as sophisticated as people assumed it was for years.

On why a purely chronological feed is not the answer, he points to incentives. If everyone's post lands at the top as soon as it is published, the feed fills with publishers who can post fifty times a day while your best friend posts once a week. Recency is one input to relevance but not the only one, and Instagram's surveys show that when chronological is the default, usage and overall sentiment fall. He says the goal is to give people agency in a way that works for the whole community.

Mosseri calls the rise of AI content a tailwind but also a challenge. More content can mean more attention for an advertising business, but he admits Instagram is not yet good at ranking AI content, which ranges from great to crap. He says power has been shifting from institutions to individuals, with sports as an easy example, so Instagram has invested in creators broadly, including journalists, artists and anyone sharing as themselves. Amid abundant synthetic content, he expects people to seek out creativity, authenticity and people.

He does not think Instagram should judge content by the tool that made it. Rather than filtering AI content, he would label it, give more information about who posted something, such as how old their profile is, and let people decide whom to trust. He praises AI creators with a clear point of view and aesthetic, including an account of singing doll-like figures and a Paris-based artist who chains several models together, comparing the latter to a painter with a new tool.

Detection is hard and may get harder as models improve, so Instagram should be honest about its confidence. Mosseri thinks that, long term, it may be more practical to label camera-captured content than AI content. He also separates labeling content from labeling accounts, noting a new spam vector of fake accounts, such as an AI monk selling bogus supplements, that do not make clear they are synthetic.

Asked what he envies in other platforms, Mosseri names TikTok's ability to break small talent. He contrasts exploitation-based ranking, which uses the data you have, with exploration-based ranking, which tests content to find what someone might like without knowing it yet. The latter helps niche and small creators find new audiences, and Instagram has invested in ranking for originality, breakout content and recency. By his own account Instagram is catching up, ahead in a couple of areas, and has line of sight to being best in class at recommendations for the first time in his tenure.

On the comment-section criticism he receives, Mosseri says he is not trying to win over people whose minds are made up, but to lay out the trade-offs for everyone watching. Never seeing anything uninteresting means a diet of lowest-common-denominator content, while discovery means the occasional miss, and privacy and safety are likewise in tension. He started speaking publicly while running News Feed at Facebook, reasoning that the debate would happen with or without the company, and recalls the harsh first comment on his 2009 News Feed redesign. His lesson was that rearranging someone's digital desk without warning would reasonably annoy them.

On the backlash to the video-style feed, Mosseri says several things got conflated: a test to 4% of iOS users that was never going to launch, the push into Reels, more recommendations from accounts you do not follow, and creators upset about reach. Press coverage and creator criticism then amplified each other. His takeaway is that Instagram pushed a little too fast, but he would rather face occasional backlash than lack video, DMs, Stories or ranking. He now has a communications plan ready for any controversial test before it leaks, since you cannot test at three billion users without being noticed.

Mosseri's first project as a PM was Facebook Home, a fork of Android with an HTC phone, which he calls a spectacular failure. He had declared himself PM after the PM quit and learned about carriers, OEMs, certification and operating systems, which he says taught him more than any other year. He is glad he ended it, adding that sometimes the best thing is to execute an idea without market fit well enough to find out whether it was a good idea.

His second failure was building the first version of Reels on top of Stories around 2019. Stories had momentum, but their read-through rate is low, so most Reels were never seen and then disappeared. He believes TikTok took off during the pandemic in 2020 and Instagram was out of position, a big fork in the road for the last eight years. As a designer he wanted to extend existing primitives, and he says he was wrong.

On his three children, aged ten, eight and six, Mosseri emphasizes boundaries: they earn iPad time through homework, and he approves their apps. Planes are an exception. He supports no phones in classrooms but also worries about kids who do not learn to use AI, so he has been vibe coding with his oldest, who built a 19-level platformer game. He wants his kids to be digitally and AI literate without a free-for-all, and he asks listeners to remember that contentious tech debates involve real trade-offs.