Adam Mosseri, head of Instagram (over 3 billion monthly users), discusses how AI is reshaping product development at Meta, the evolution of Instagram’s algorithm and creator ecosystem, and his approach to leadership, criticism, and parenting in the AI era.
Product teams at Meta are shifting to small, generalist “pods”
The canonical team of ~13 specialists (separate iOS, Android, server engineers, PM, designer, data scientist, researcher) is being replaced by pods of 6–7 people.
Pods center on 4–6 generalist engineers plus a “product staff” role — an evolution of the PM who can also do design, data analysis, and research using AI tools.
Specialists (senior designers, data scientists, researchers) are brought in only when the work demands deep expertise.
Smaller teams coordinate less, move faster, and avoid design-by-committee; AI productivity gains amplify this effect.
Functional boundaries are blurring: data scientists’ mechanical work (e.g., waterfall analyses) is now automated, letting product staff handle it; designers and engineers are crossing into each other’s crafts.
Mosseri expects functional lines to keep blurring but believes senior specialists will remain essential — both for their craft and to mentor the next generation.
Hiring for large-organization leadership will decrease as teams stay small; the “age of the generalist” favors range, but deep expertise still has a place.
Hiring traits: curiosity and willingness to try things now outweigh specialization
Two rising premiums: curiosity and willingness to put yourself out there (try tools, make mistakes, sound foolish) — analogous to language learning.
Declining need: pure people-management-at-scale roles, since orgs are flatter.
AI is resetting who succeeds: engineers now spend most time planning and reviewing AI-generated code, not writing it; people who couldn’t previously execute cross-functional ideas can now do so with AI assistance.
The most effective people are clear-eyed about what AI is good at today, what it’s bad at, and have an instinct for where it’s heading.
Meta manages AI token spend like any other constrained resource
No hard token caps yet; early “token incinerators” (low-ROI experiments) were shut down once costs were visible.
Token budget is allocated like GPU, labeling, or headcount budgets — based on expected ROI and trust in the team.
Mosseri expects engineer burn rate (tokens + salary) to converge in 1–2 years, at which point proportional caps will make sense.
Costs may rise from volume before model prices fall due to frontier-model competition.
Human judgment remains central to vision, strategy, and taste
As AI eats execution (coding, analysis, design comps), human cycles shift to defining success, setting constraints, and giving feedback — essentially management of agents.
Vision (articulating the desired future state) and strategy (an opinionated, controversial path to get there) are where brains add the most value.
AI is not automatically great at strategy: it needs heavy steering with context (team dynamics, talent attraction, regulatory landscape, brand identity) and iterative back-and-forth; lazy prompts yield predictable, me-too strategies.
Models differ in willingness to push back; Mosseri prefers ones that challenge him (e.g., Mythos, Fable) over sycophantic ones.
Great product leaders act as curators — of people, ideas, technologies, strategies — creating environments where the best ideas surface and teams have trust/rapport.
Team chemistry (complementary skills, good vibe) is as important as individual talent; sometimes two great people simply can’t work together.
Instagram’s algorithm knows less about you semantically than people assume
Historically, recommenders relied on opaque embedding vectors (giant number arrays), not human-readable interest profiles; the system knew “people who liked X also liked Y” without understanding “surfing.”
LLMs now let Instagram translate those vectors into legible topics (e.g., “deep pourover coffee snobbery”) — the “Your Algorithm” feature shows users inferred topics and lets them add/remove.
Future: users will express non-topical preferences (“more fun,” “see friends more,” “no 7 photos in a row”) — giving agency back in a recommendation-dominated feed.
Chronological feeds disappoint at scale because of incentive distortions
Pure chronological feeds incentivize high-volume posting (publishers, brands), drowning out low-frequency friends/family.
Recency is one relevance signal, not the only one; a sister’s engagement announcement matters more than a brother’s sandwich photo.
Surveys show chronological-by-default lowers both usage and overall satisfaction over months, even if the switcher feels happy initially.
Trade-offs are inherent: zero unwanted content means lowest-common-denominator feed; discovery means occasional misses.
AI-generated content is a tailwind for Instagram, not a threat
More content = more potential attention, but ranking AI content well is still a work in progress (great vs. crap).
Long-term trend: power shifts from institutions to individuals (e.g., athletes > teams); Instagram’s bet on creators (broadly defined: journalists, artists, small businesses) aligns with this.
In a flood of synthetic content, people will seek creativity, authenticity, and human point-of-view more, not less — Instagram’s scale as a creator platform becomes a moat.
Mosseri opposes filtering out AI content; instead, label it (content-level and account-level) so users can decide.
Detection will get harder as models improve; labeling camera-captured (non-AI) content may be more practical long-term.
Account-level signals (profile age, changes, verification) help users judge trustworthiness; spammy AI personas (fake monks selling supplements) are a new vector to combat.
Examples of great AI creators: “Plastic Dream Sequence” (aesthetic doll animations), “If Only AI” (dreamscapes with clear artistic voice) — the tool doesn’t matter; the point of view does.
TikTok’s exploration-based ranking inspired Instagram’s push for originality and breakout creators
Exploitation ranking (show what’s already popular) is easy; exploration ranking (test niche content to find hidden interests) helps small creators find audiences.
Instagram has invested heavily in originality, breakout rate, and recency to stay culturally relevant — catching up to TikTok, with line of sight to best-in-class recommendations.
Mosseri acknowledges the “disappointed dad” vibe: never satisfied, always pushing.
Public criticism comes with the territory; perspective and boundaries sustain him
Started engaging on Twitter (journalists’ turf) during News Feed days: debate happens with or without you, so participate with humility.
2009 News Feed redesign backlash taught him: rearranging millions of people’s “desks” without warning reasonably angers them.
Coping: put criticism in perspective (users’ daily habits disrupted), step away (family, outdoors), accept cycles of intensity.
2022 feed redesign test (4% iOS) conflated with Reels push, recommendation increase, and creator reach complaints — creator/press echo chamber amplified it.
Lesson: at 3B users, any test can leak; need a comms strategy before launching experiments (not if, when).
Pricing experiments are especially volatile; leaders should share lessons on managing leak-driven cycles.
Biggest failures: Facebook Home and building Reels on Stories
Facebook Home (Android fork + HTC hardware): spectacular failure, but taught him more in 18 months than any other period — carriers, OEMs, OS, certification; also learned to kill a product cleanly when market fit is absent.
Reels v1 built inside Stories (2019): wrong foundation — low read-through, content disappeared unseen. Delayed proper Reels launch to summer 2020, ceding critical pandemic-era growth to TikTok.
Designer instinct to extend primitives (Stories) backfired; sometimes a new primitive is necessary.
Screen time with kids: boundaries, earned time, curated apps, and co-creation with AI
Kids (10, 8, 6) too young for social media; each has an iPad with earned weekend time (homework = screen minutes).
Parents should approve every app; Meta advocates for this at policy level.
Exceptions: long flights (survival), and — experimentally — “vibe coding” with 10-year-old using Claude Code to build a 19-level platformer game.
Goal: digital/AI literacy without free-for-all; co-create, don’t just consume.
Schools banning phones in classrooms likely good for education; balance needed between critical thinking and AI fluency.
Closing message: remember the trade-offs and the people making them
Every contentious debate (ranking, privacy, safety, AI) involves real trade-offs; they’re rarely as simple as public discourse makes them seem.
Mosseri invites criticism but asks observers to remember: decisions are made by people trying their best in a complex space.
