Adam Mosseri: Building Instagram for an AI world

Why does it matter? Because the algorithm everyone feared was just a big number that happened to correlate with surfing.

Adam Mosseri has run Instagram for eight years and overseen three billion monthly users — and the central revelation of this conversation is that the surveillance-level intelligence most people assumed was profiling their feeds was never really there. The mind-reading algorithm, the strategic AI advantage, the canonical 13-person product team: all of it is being rebuilt from scratch in 2026, and Mosseri explains exactly how.

• Instagram's recommendation engine was collaborative filtering on opaque vectors — semantic understanding of your actual interests is only arriving now, via LLMs. • An AI content flood makes human provenance scarcer and more valuable, not less — a structural tailwind for creator platforms. • The canonical cross-functional product team is dead; pods of 6–7 built around a new "product staff" generalist are replacing it. • A strategy that no reasonable person can argue against isn't a strategy — it's a goal dressed up with slide deck confidence.

Instagram's algorithm never actually knew you — it just had a big number that happened to correlate with surfing

"It just has this big ass number that happens to correlate with surfing." That's Mosseri describing how Instagram's recommender has actually worked for years — not semantic profiling, not intent-reading, but collaborative filtering on giant, illegible vector embeddings. The system didn't know you liked surfing or deep-pour coffee or anything else in legible terms. It knew which numbered vectors clustered together. "Until recently, we don't really know as much about you as you think. We were just like, oh, you liked these photos, these people also like those same photos and they like these other photos, so you might like those."

What's changing: LLMs can now translate those illegible embedding spaces into plain English. Instagram's "Your Algorithm" feature — where users can see and adjust what topics the platform thinks they care about — runs on exactly this. An LLM maps a user's engagement history onto an embedding space and surfaces something like "deep pourover coffee snobbery" as a descriptor. "Now only now are we actually getting as sophisticated as I think people have assumed we've been for many years."

The practical upshot: most criticism of algorithmic surveillance has been aimed at a capability that didn't exist until this year. The sophisticated individualized profiling was always more imagined than real. The privacy debate just shifted ground.

The AI content flood is a tailwind for Instagram — because human provenance is about to become the scarce asset

Most platform analysts frame the AI content wave as a threat to creator-driven apps. Mosseri calls it a tailwind — specifically for Instagram, specifically because of how it's positioned.

His logic: "In a world where there's an abundance of synthetic content, I actually think people are going to seek out creativity and authenticity and people more, not less." As AI commoditizes content production, what becomes scarce is verifiable human origin and genuine point of view. A journalist, an artist, someone selling hand-sewn scarves — the person behind the post becomes the differentiator, not the technical polish of the output. "Instagram was never just about the content. It was always about to a certain degree the person behind the content, the point of view, the reason they're sharing it, their perspective."

He explicitly rejects filtering AI content by default — "I don't think we should judge content based on the tool that made it" — while arguing for infrastructure that makes provenance legible. His long-term instinct: labeling camera-captured content as non-AI may eventually be more practical than labeling AI content, as detection becomes unreliable.

The competitive framing: Instagram hosts more creators, by most definitions, than any other platform. If the value of individual human provenance rises, that's a structural advantage no one can easily replicate.

Instagram killed the 13-person cross-functional team — the 6-person pod around a 'product staff' generalist is what replaces it

"For the longest time, the canonical team was something like two or three Android engineers, two or three iOS engineers, two or three server engineers, maybe a generalist, a PM, a designer, a data scientist, a researcher if you were lucky." Mosseri counts the old model out: a baker's dozen, give or take, for every product area. In 2026, Instagram is replacing it with pods of four to six engineers plus one "product staff" — a PM who absorbs design, data science, and research tasks via AI tools, pulling in senior specialists only when the work genuinely requires it.

The unlock that makes this viable: routine data analysis — the kind of funnel waterfall that used to require a data scientist's bespoke queries — can now be generated automatically with internal tools. Product staff can pull it in minutes. What remains for actual specialists is work that's genuinely creative or technically demanding: pricing strategy requiring serious statistical design, novel user experiences requiring world-class interaction design.

The payoff isn't just cost. Smaller teams decide faster and make sharper calls. "By virtue of having less people to coordinate, they can often move faster and make better decisions — a little bit less design by committee." The specialist roles don't disappear — they become senior-only, high-leverage positions rather than required seats on every team.

Real strategy must be controversial — and prompting AI lazily gives you exactly the same strategy your competitors will also generate

Any strategy that no reasonable person can argue against isn't really a strategy — it's a goal with better formatting. Mosseri's definition is exact: "Strategy is an opinionated path to achieve a vision. It can't be 'be the best' or 'be amazing.' It has to be controversial — a reasonable person should be able to disagree with it, because otherwise you're probably just trying to compete on raw execution."

That test becomes more important as AI makes it easy to generate the appearance of strategy. Ask a model for one without loading the real constraints, Mosseri warns, and "you're going to get something pretty predictable that probably the competition would expect you to do." The inputs that make strategy genuinely differentiated don't live in public market data — they live in a leader's head: the specific personnel dynamics, the regulatory exposure at Instagram's scale, the brand identity, the kinds of talent a bold move might attract. "You need to think long and hard about what are all the different inputs that need to be considered."

His prescription: treat AI strategy sessions as a back-and-forth, not a one-shot query. Tell it to be critical. Load every real constraint before asking for a path. The humans who understand all the constraints still own the hardest part.

Engineering used to be 40–60% writing code — now it's mostly planning and reviewing, which flips who the winners are

"Engineering used to be maybe not majority but a large percentage — 40, 50, 60% — writing code. You know, it's not now." The job changed faster than anyone's performance record could adapt, and Mosseri is blunt about what that means: prior success at the mechanical act of coding doesn't predict future success at directing AI to do it.

Two camps are splitting. People whose edge was raw execution speed — fluent, fast code writers — face the sharpest displacement. People who had strong opinions about adjacent functions but lacked the technical skill to act on them are newly unlocked. Mosseri puts himself in the second group: "I now get to program again for the first time in maybe 10 years, and I am not a good engineer — I'm a mediocre engineer on a good day — but now I can write code responsibly, which is just an amazing thing."

The people who thrive, he says, are "the ones who are cleareyed about what AI is good at and what it's not good at, and also have an instinct or a nose for what it will be good at." Judgment about the tool's trajectory matters as much as current fluency. Sensing where the models are heading — not just knowing where they are — is what compounds.

The best product leaders are curators, not idea machines — and AI generating ideas on demand makes that distinction urgent

Most people joining product roles want to do vision and strategy. Mosseri's observation about the leaders he actually admires cuts against that: "A lot of the best end up sort of being curators — curators of people, curators of ideas, curators of technologies, curators of strategies."

The distinction matters more as AI can generate ideas cheaply. If ideation is no longer the constraint, the scarce capability is recognizing which idea is right, committing to it regardless of its origin, and building the team environment where the best ones surface. "I don't really care if I'm hiring a strong lead for an area if the strategy comes from them or comes from somebody else. I just care that there is an amazing strategy and everyone has bought into it and that we're executing against that strategy well."

Great curators, Mosseri adds, also think about team chemistry with the same rigor as strategy — not just individual competence but how a leadership team fits together, where the trust is, where the friction will emerge. "A leadership team with strong trust and rapport can work through most anything." Getting the people configuration right is the curator's second job — and it's what lets the ideas flow.

Token burn per engineer could match salary within a year or two — treat it as headcount budget now, not a free utility

"You can imagine, at least in a year or two coming, that the burn rate of a strong engineer might be the same as their salary or their cost of employment." Mosseri frames AI compute the way he'd frame any other finite resource — GPUs, storage, payroll — something to allocate deliberately, not leave running as ambient infrastructure.

Instagram isn't at hard caps yet. They've controlled costs mainly by shutting down obvious waste: "It's not that hard to build a token incinerator, and that doesn't create a lot of value. And as soon as you actually look at the dollars in and value out, you might just be like, 'Oh, that's just a bad idea.'" No leaderboards for token spend, either — that just gamifies volume with no connection to output value.

The discipline to build now: track token spend per team and per project, connect it to value produced, and establish the mental model that AI compute is a budgeted resource before external pressure forces the accounting.

Everything Mosseri describes is the same underlying shift: the gap between apparent and actual intelligence is finally closing

What unifies every thread in this conversation is a compression of latency — the distance between what a system seems to be doing and what it actually does is shrinking toward zero. The algorithm is finally becoming what users imagined. AI is finally becoming a budgeted constraint. The fake sophistication is becoming real sophistication. That convergence doesn't shrink the human role — it raises the ceiling for judgment, taste, and conviction. Tools that can generate anything make the people who know what to build, and why, more indispensable than ever.

---

Topics: product management, Instagram, AI and work, team structure, recommendation algorithms, content strategy, product leadership, AI content, org design, strategy frameworks