Lore

taste and judgment in AI era

The rise of taste, human authenticity and judgment in an AI world | Adam Mosseri (Head of IG)

As AI makes building products and generating content radically easier, the scarce and decisive resource becomes human taste and judgment — knowing what to build, how to steer AI, and how to preserve authenticity — and Adam Mosseri traces this idea through Instagram's org design, hiring, product strategy, recommendation systems, AI-content policy, and his own leadership mistakes.

Lenny's Podcast · 2026-07-09 · English

Key ideas

  1. Mosseri opens with the episode's throughline: as AI lowers the cost of building, taste matters more, not less, because the bottleneck shifts to deciding what's worth building; the people who benefit most are those who are clear-eyed about AI's current strengths/weaknesses and have a 'nose' for where its capabilities are heading.

  2. He claims a widespread misconception is that Instagram's algorithm has detailed semantic understanding of users; in reality it has long relied on illegible embedding vectors ('a big ass number that correlates with surfing'), and only now are LLMs making those vectors describable in natural language.

  3. On AI-generated content, he argues it's simultaneously a tailwind (more content = more attention for an ad-driven business) and a challenge (Instagram isn't yet good at ranking it); his core prediction is that abundant synthetic content will make people seek out creativity, authenticity, and real people more, not less — so Instagram's long-standing bet on individual creators should benefit, provided the platform labels AI content and surfaces trust signals about posters rather than filtering AI out or judging content by the tool used.

  4. He argues Instagram's canonical big-company product team (~13 specialists) is being replaced by small 'pods' (~6-7 people) built around a generalist 'product staff' role that AI tools let absorb design, data-science, and research work, with senior specialists added only when truly needed — evidence, he says, of a broader 'age of the generalist' in which functional lines blur, though elite specialists still matter and companies must keep developing junior talent in every function rather than just retaining today's seniors.

  5. He reframes hiring and career advice for the AI era: durable baseline traits (grit, quick learning, self-awareness) still apply, but the emerging premium is on curiosity and willingness to try and fail publicly, while pure large-organization management skill is depreciating as teams shrink.

  6. He argues AI is 'resetting' individual impact — engineers at AI labs now spend most of their time planning and reviewing code rather than writing it, some previously low performers are newly thriving, and people with strong ideas but weak technical execution are newly unblocked — so success depends on whether one's strengths match what the tools and the business now reward.

  7. Contrary to his own expectation, he found AI is not naturally good at originating strategy — only useful for strategy when aggressively constrained with real inputs (technology state, personnel, competitive/regulatory/compliance landscape, brand identity) — because good strategy must be 'controversial'/disputable, and lazily-prompted AI strategy is predictable; this leads him to argue human judgment increasingly concentrates in vision and strategy as execution gets automated, with steering AI framed as its own management craft.

  8. He reframes the best product leadership not as personal visionary genius but as curation — of people, ideas, technologies, and strategies regardless of origin — combined with assembling leadership teams for both complementary skills and interpersonal chemistry/trust.

  9. Using Instagram's feed as a case study, he argues that naive stated user preferences (e.g., wanting a purely chronological feed) are empirically wrong once tested — chronological feeds reward high-frequency institutional posters over individuals and lower both usage and long-run satisfaction — and generalizes this into a broader stance that nearly every contentious product debate (including privacy vs. safety) involves real trade-offs rather than a clearly 'right' answer.

  10. He argues public debate about a platform Instagram's size (3 billion users) is inevitable, so leadership should proactively engage, explain trade-offs, and pre-plan communications strategy for any test that could leak or go viral, rather than staying silent — a stance he traces to his own history running Newsfeed at Facebook and becoming the company's public face.

  11. Revisiting the 2022 feed-redesign backlash, he argues it conflated four separate issues (a limited 4% iOS test, leaning into Reels, leaning into recommendations, and creators upset about falling organic reach), that the underlying investment was directionally correct and drove subsequent growth, and that the real lesson was inadequate explanation and pacing rather than the substance of the change.

  12. Reflecting on Reels versus TikTok, he admits building Reels on top of Stories (extending an existing primitive rather than creating a new surface) was a design mistake that weakened Reels' early distribution and let TikTok's pandemic-era explosion go largely unchallenged, calling it a major strategic fork in the road.

  13. He closes by extending the same steering/judgment philosophy to parenting: he sets boundaries and earned access for his kids' screen time and personally approves their apps, while also 'vibe coding' with his 10-year-old, framing AI literacy as something to be taught through guided practice rather than blanket restriction or unrestricted access — a small-scale echo of his argument that AI should be managed and steered, not banned or given free rein.

  14. Pods — A small (4-6 engineer) generalist team structure replacing the old ~13-person specialist-heavy product team. Apply: Staff a product effort with a handful of generalist engineers plus one generalist 'product staff' lead, adding senior specialists only when a specific need (e.g., pricing data science) demands it.

  15. Product staff (role) — An evolution of the PM role that uses AI tools to absorb parts of design, data-science, and research work previously done by separate specialists. Apply: Have the generalist lead run waterfall/funnel analyses, mock up design proposals, and shape research questions directly via AI tools rather than routing every task to a dedicated specialist.

  16. Three-trait hiring framework — Mosseri's baseline hiring filter across nearly any function: grit/drive, being a quick learner, and self-awareness. Apply: Screen candidates for these three traits first, on the premise that someone with all three can get good at almost anything, while missing any one usually causes problems.

  17. Curiosity + willingness to put yourself out there — A two-trait framework for predicting who adapts well over the next 5-10 years of rapid AI-driven change. Apply: Favor people who stay curious and are willing to try new tools/skills publicly and be corrected, likened to being willing to sound like an idiot while learning a new language.

  18. Vision/Strategy framework — Vision is the articulated end-state you want to reach; strategy is an opinionated, genuinely disputable path to that vision (not just 'be the best'). Apply: Test a proposed strategy by checking whether a reasonable person could disagree with it; if not, it's just an execution plan, not a strategy.

  19. Aggressive constraint-steering for AI strategy — A method for getting useful strategic output from AI by deliberately feeding it concrete real-world constraints rather than asking it to originate an answer directly. Apply: Prompt the model with specifics on technology state, team/talent motivations, competitive landscape, regulatory/compliance landscape, and brand identity, and tell it to be critical rather than agreeable.

  20. Centaur / reverse-centaur — Cory Doctorow's concept distinguishing a human-controlled human-AI pairing (centaur) from an AI-controlled one where the human just executes the AI's bidding (reverse centaur, likened to gig-economy app workers). Apply: Use it to check whether a given human-AI workflow keeps the human in control of strategy/direction (centaur) versus reducing the human to executing AI-set tasks (reverse centaur), and avoid designing the latter.

  21. Curator model of product leadership — A reframing of top product leadership as curating people, ideas, technologies, and strategies (regardless of who originates them) rather than personally generating every good idea. Apply: Build an environment where strong ideas from anywhere can surface and be selected/backed, instead of relying on the leader to originate the winning idea themselves.

  22. Leadership-team composition (skills + chemistry) — A framework for assembling leadership teams that weighs both complementary functional skills/weaknesses and interpersonal trust/chemistry, the latter described as 'more art than science.'. Apply: When staffing a leadership team (e.g., Trust & Safety's 5 functional leads), evaluate not just each person's individual strength but how their skills fill gaps in the group and whether the group has good rapport; split up individually strong people who simply can't work together.

  23. Embedding-space model with LLM description ('your algorithm') — A recommender-systems technique that places user interactions and concepts into a shared high-dimensional vector space, historically illegible to humans but now describable in natural language via an LLM. Apply: Surface a user-facing feature that shows an LLM-generated label for a region of the user's interaction embedding (e.g., 'deep pourover coffee snobbery') and lets them add/remove inferred interests.

  24. Exploitation-based vs. exploration-based ranking — Two recommender strategies: exploitation uses existing data on what a user already likes, while exploration tests whether they might like something outside their known preferences. Apply: Weight some ranking capacity toward exploration (e.g., showing a niche genre to a mainstream-artist fan) specifically to give small/niche creators a chance to reach new audiences, as TikTok does.

  25. Chronological-feed incentive analysis ('designing a system is like designing a city') — An approach to evaluating feed/ranking design by tracing the mechanics of a rule through to the incentives it creates, how people act on those incentives, and the resulting outcomes. Apply: Before defaulting to a stated user preference like chronological ordering, model out who benefits from post-frequency incentives it creates (e.g., publishers posting 50x/day) and measure resulting usage/satisfaction over time, not just immediate sentiment.

  26. Content-level AI labeling vs. account-level authenticity marking — Two distinct classification dimensions for AI-content policy: whether a specific piece of content was AI-made, versus whether an account belongs to a real person. Apply: Label AI-made content directly (with stated confidence: definitely/probably/not sure/definitely not) and separately expose account signals (profile-change history, account age) so users can judge authenticity for themselves.

  27. Pre-planned comms/crisis strategy for controversial tests — A practice of deciding in advance what the public message will be (proactive or reactive) for any product test that could leak or provoke backlash, before the test even launches. Apply: For any design change or test that could be controversial, agree on the talking points ahead of time so the team is ready to explain it publicly even before deciding whether to fully launch it.

  28. 'Execute to find fit' — The practice of shipping/running an idea specifically to determine whether it has market fit, accepting it may need to be killed. Apply: Greenlight a genuinely uncertain idea (as with Facebook Home) not to guarantee success but to definitively learn whether the idea itself was good, then be willing to shut it down if it isn't.

  29. Extending existing primitives (design philosophy) — A design approach that favors building new capabilities on top of existing product surfaces rather than creating entirely new ones. Apply: Weigh this against the alternative of a standalone surface — Mosseri notes that applying it to Reels (built inside Stories) was, in retrospect, the wrong call given Stories' low read-through and content overload.

  30. Earned screen-time system — A household rule where kids trade completed tasks (e.g., homework sessions) for capped device time rather than receiving screen time on request. Apply: Set a fixed exchange rate (e.g., three 30-minute homework sessions = 90 minutes of weekend iPad time) so time is earned, not just given.

  31. Parental app-approval model — A practice (and a policy position Meta advocates) of requiring parental approval before a child can download any app. Apply: Personally review and approve each app before a child installs it, rather than allowing unrestricted downloads.

  32. Vibe coding with kids — Using an AI coding assistant (e.g., Claude Code) together with a child to build a real project as a hands-on AI-literacy and creativity exercise. Apply: Sit down with a child to co-build something concrete (Mosseri and his son built a 19-level 8-bit platformer) rather than leaving AI tool use unsupervised or off-limits.

  33. Token-budget resource-allocation analogy — A framing that treats AI token spend like other finite resources (GPU/CPU/storage/RAM capacity, opex, headcount payroll) that need eventual budgeting. Apply: Instead of leaderboarding raw token usage, plan for future per-person or per-team token caps sized to the trust placed in that person's ability to spend tokens ROI-positively.

  34. Waterfall analysis — A data-science technique for mapping each step of a funnel (e.g., the reels-creation flow) to identify where users drop off. Apply: Run this drop-off analysis using now-automatable internal tools, letting a generalist product staff member do work that used to require a dedicated data scientist.

Insights

Meta reportedly ran (and Mosseri calls 'a terrible idea') a leaderboard of engineers' token spend; there are currently no token limits at all, but he expects a strong engineer's AI 'burn rate' could rival their salary within a year or two, eventually requiring trust-based caps modeled on how companies allocate GPUs, CPUs, storage, and headcount budgets.

Contrary to his own prior assumption that AI would excel at strategy given its access to market/competitor/metrics data, Mosseri found it does not — useful strategic output requires deliberately constraining the model with concrete inputs, otherwise it produces predictable answers competitors would already expect.

Instagram's new 'your algorithm' feature uses an LLM to translate previously illegible embedding vectors into natural-language interest labels (his example: 'deep pourover coffee snobbery'), letting users view and edit what the algorithm thinks they're interested in.

He proposes it may eventually be more practical to label camera-captured (non-AI) content rather than trying to label AI content, since confidence in AI-detection is inherently limited and likely to degrade further as models improve.

He names a concrete abuse pattern to crack down on: non-disclosed 'AI monk'-style personas selling bogus supplements — his stated line isn't being an AI creator, but using non-disclosure to deceive.

The old 'canonical' Instagram/Meta product team was 'on the order of a baker's dozen' (2-3 Android, 2-3 iOS, 2-3 server engineers plus generalist PM/design/data-science/research); the new 'pod' model often has just 4 engineers plus a single 'product staff' generalist, with no dedicated designer, data scientist, researcher, or content designer.

Facebook once required designers to pass a technical/programming interview loop, later dropped it because it made hiring too hard, and Mosseri — after roughly 10 years away from coding — says he's personally programming again thanks to AI tools, admitting he was never a great engineer.

His first PM project, Facebook Home (an Android-OS-level fork built with HTC hardware), was, in his words, a 'spectacular failure,' yet he calls it the most educational year-and-a-half of his career and says he's glad it was ultimately killed.

He explicitly admits building Reels on top of Stories rather than as a standalone surface was a design mistake stemming from his own philosophy of extending existing primitives rather than adding new ones — a choice he now says let TikTok's pandemic-driven growth largely go unchallenged.

The 2022 redesign backlash he's known for stemmed from a test limited to just 4% of iOS users that was never intended to fully launch, amplified by a feedback loop between vocal creators and press coverage.

He and his 10-year-old built a 19-level, 8-bit-style platformer (styled after Super Mario, complete with an in-game skins/weapons store) using Claude Code, after an attempt six months earlier 'totally didn't work' with older models.

Instagram believes that, for the first time in his tenure, it has 'line of sight' to matching or beating TikTok on some recommendation dimensions, crediting TikTok/ByteDance's exploration-based ranking (surfacing content users don't yet know they'd like, e.g. showing Afrobeats to a Bieber fan) as the inspiration for its own investment in originality and breakout-content ranking.

«No, I think taste matters a ton. In a world where it's easier to build things, it's more important to make sure that your time is spent figuring out what you should be building in the first place.»

— 00:00

«I think people assume that there's a much more detailed semantic understanding of everybody's interests and preferences in the algorithm than there is.»

— 00:30

«I actually think people are going to seek out creativity and authenticity and people. I don't think we should filter out AI content. I think we should let you know if content is AI content or not. That's hard. By the way.»

— 00:52

«So you know on the order of a baker's dozen.»

— 02:54

«I mean in some ways it's probably the age of the generalist.»

— 11:26

«We don't know anything about what's going on.»

— 13:54

«Are you willing to sound like an idiot? Are you willing just to say it and be corrected and not be offended and then just get better and better?»

— 15:37

«Especially if you talk to anybody at these labs, they're spending most of their time planning and reviewing code. That is a very different job.»

— 17:26

«They're amazing at some things and remarkably bad at others.»

— 19:16

«I think you'll still be able to write code for free. It'll just be with a smaller model.»

— 19:52

«This is the worst the model will ever be.»

— 21:10

«It's a terrible idea. No leaderboards for token spend.»

— 21:25

«I want the real real sort of intelligence. I don't want the pleaser.»

— 27:57

«It's a horse head on a human body.»

— 28:29

«It just has this big ass number that happens to correlate with surfing.»

— 35:20

«That is like deep pourover coffee snobbery.»

— 36:22

«No, you actually engage a lot more when it's this algorithmic feed of things we think you will love.»

— 38:24

«I think it's going to be a tailwind, but I think it's going to be a challenge.»

— 41:17

«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.»

— 42:39

«I think we should judge it based on the content, the point of view, the person behind the content.»

— 44:20

«They call me disappointed dad.»

— 50:09

«The debate was going to happen with or without us. So we might as well participate.»

— 52:44

«I just came and I rearranged your desk and I didn't tell you, I didn't warn you, I didn't even explain why. Like you would be pissed and that would be reasonable.»

— 55:52

«Now, I would much rather have backlashes like that every couple years or continue to evolve and continue to stay relevant than the alternative, which would have been like we didn't have video, we didn't have DMs, we didn't have stories, we didn't have ranking, and we wouldn't be on having this podcast right now.»

— 58:31

«The cost of leaning in is that you're going to occasionally like make a mistake and you're going to definitely um pay for it.»

— 58:48

«On one hand, you know, I'm a designer. I'm trying to not add new things to the product. I'm trying to extend existing primitives and that was the idea. On the other hand, I was wrong.»

— 62:39

«It's unbelievable what a 10-year-old who still types with three fingers can do.»

— 65:44

«I just think you don't even have to tell this to other people, but just remember that this world and technology is complicated and there are almost always tradeoffs.»

— 67:23

Reception

Reception is mixed: several viewers praise the conversation's insight and authenticity while a vocal minority distrusts Meta's ethics and dismisses the CEO's credibility on 'taste,' amid a lot of unrelated spam and off-topic comments.

A coherent, first-person executive account that threads one real argument — taste and judgment as the scarce resource in an AI-abundant world — through org design, hiring, AI strategy limits, recommendation systems, and candid retrospectives on Instagram's own missteps (Reels-on-Stories, the 2022 redesign backlash); its strongest content is the concrete internal detail (pod structures, token-budget thinking, the 'your algorithm' feature) rather than any independently verified claim about AI's broader societal effects.

68:29

↳ Lenny's Podcast · YouTube

Watch original