Lenny's Newsletter

▶ YouTubecompleted
Back to results

Lenny's Podcast · Jul 9, 2026

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

team structureAI-era product managementproduct staff roletaste and judgmentrecommendation systemsembedding modelsLLM explainabilitycreator economyAI content strategyexploration vs exploitationtoken spend governancehuman-AI collaborationcentaur modelproduct leadershipchronological feed incentivesAI transparencyorganizational designstrategy formulation

Adam Mosseri, head of Instagram, discusses how AI is reshaping product team structures toward smaller generalist "pods" with a new "product staff" role, while arguing that taste, judgment, and human authenticity become more—not less—valuable as AI automates execution. He also shares Instagram's strategic bet that AI-generated content abundance will drive users toward authentic human creators, making Instagram's creator-centric positioning a tailwind.

Watch on YouTubePartial match

Topics (12)

Product Staff: The AI-Era Generalist Role

3:25

Product Staff is Meta's answer to the AI-era PM: a generalist who can do meaningful design, data analysis, and research work without dedicated specialists, enabled by AI tools.

How it works

  • A Pod team is 4–6 engineers plus one Product Staff, replacing the ~13-person canonical team (iOS/Android/server engineers, PM, designer, DS, researcher).
  • Product Staff handles what used to require separate headcount: pulling waterfall analyses, making design calls, running lightweight research—using internal AI tooling.
  • Specialists (senior designer, senior data scientist) are brought in only when the problem genuinely requires deep craft, not as permanent team members.
  • The role is explicitly cross-functional by design: strong designers or data scientists converting into Product Staff bring deep craft *and* expand their influence beyond functional silos.
  • Why you should care

    This is the organizational model Meta is betting on *right now* in 2025–2026—not a future hypothesis. If you're a senior PM, this is your role's next evolution; if you're hiring, the question is no longer "which specialist?" but "how generalist can this person be?"

    Think of it like...

    It's the Swiss Army knife that replaced a full toolbox—not as sharp on any single blade, but you can carry it everywhere.

    Taste as the Last Non-Automatable Skill

    0:00

    Taste is the judgment about *what to build and why* that remains irreducibly human even as AI handles more of the *how*.

    How it works

  • As building gets cheaper and faster, the scarce resource shifts from execution capacity to decision quality about what to execute.
  • AI-generated output (code, design, content) converges on recognizable patterns—"you can tell this was Claude-designed"—meaning undifferentiated output floods the market.
  • Taste is what breaks that convergence: a clear aesthetic and strategic point of view that AI can assist but not originate.
  • Mosseri specifically argues designers—historically anxious about AI—are actually "long" bets precisely because they tend to have trained taste.
  • Why you should care

    As a PM, your competitive edge is no longer speed of delivery—it's the quality of your judgment about which problems matter. Frameworks and process are table stakes; the differentiator is the ability to make controversial, opinionated calls.

    Think of it like...

    In a world where any chef can now execute any recipe perfectly, the scarce skill is knowing which dish to put on the menu.

    AI Literacy About AI's Own Limits

    0:10

    The highest-leverage AI skill is not using AI fluently, but knowing precisely when *not* to trust it and where it will improve next.

    How it works

  • Most people hold a binary view: AI-positive or AI-skeptic. Mosseri argues neither is accurate because AI tools are "amazing at some things and remarkably bad at others."
  • The valuable skill is two-part: clear-eyed assessment of *current* limits AND a nose for *near-future* capability trajectory.
  • This applies directly to product strategy: knowing which workflows to outsource to AI now versus which still require human judgment.
  • Without this calibration, teams either miss AI leverage or trust AI outputs in areas where they will fail.
  • Why you should care

    In regulated, complex B2B environments, wrong trust in AI outputs creates real downstream risk—compliance, quality, stakeholder credibility. Calibrated skepticism is a professional asset, not a deficiency.

    Think of it like...

    It's like knowing the weather forecast's accuracy rate by geography before deciding whether to pack an umbrella—the tool is useful only if you know its error profile.

    Strategy as a Human Residual in the AI Development Cycle

    24:05

    AI can assist strategy but cannot generate a genuinely differentiated one without heavy human steering, because real strategy requires integrating context—people, culture, regulation, brand identity—that isn't fully legible to a model.

    How it works

  • Vision = the state of the product you want to reach. Strategy = an *opinionated, controversial* path to get there. If a reasonable person can't disagree with your strategy, it's not a strategy—it's just competent execution.
  • AI given a lazy strategy prompt produces predictable output—"something the competition would expect you to do."
  • Effective AI-assisted strategy requires the human to first enumerate all the non-obvious constraints: team motivation, talent attraction, regulatory landscape, brand identity, competitive dynamics.
  • The interaction model looks more like management: define success, set constraints, give iterative feedback—not "generate strategy."
  • Why you should care

    If you outsource strategy to AI without this steering work, you converge on consensus answers. In competitive B2B markets, consensus strategy is a race to the middle. Your value as a senior PM is making the controversial call and defending it.

    Think of it like...

    Asking AI for strategy without constraints is like asking a GPS for the "best route" without telling it whether you're optimizing for speed, scenery, or avoiding tolls.

    Product Leader as Curator (Not Visionary)

    31:02

    The best product leaders are curators—of ideas, people, technologies, and strategies—not necessarily the originators of those ideas.

    How it works

  • A visionary model assumes the leader is the idea machine; a curator model assumes good ideas are distributed across the team and the leader's job is to create conditions where they surface.
  • Curation applies at multiple levels: which ideas to pursue, which people to hire, how teams *complement* each other (skill mix, personality chemistry, trust).
  • The curator still needs their own ideas—hard to evaluate what you've never generated—but isn't the primary source.
  • This reframes the hiring question: not "is this person brilliant?" but "how does this person fit the existing leadership chemistry?"
  • Why you should care

    In large orgs with complex stakeholder structures, the bottleneck is rarely idea generation—it's idea selection and alignment. Positioning yourself as a curator is also a more defensible and scalable leadership posture as AI generates more candidate ideas.

    Think of it like...

    A museum curator doesn't paint the art—they decide which art belongs together, in what order, and why it matters.

    Embedding Space and Algorithmic Illegibility

    35:04

    Recommendation systems don't know you like "surfing"—they know you as a vector of numbers that *happens to correlate* with surfing content, and only now are LLMs able to translate those numbers back into human language.

    How it works

  • Large embedding models map content and users into a shared high-dimensional space; similar items cluster together, but the dimensions have no human-readable labels.
  • Progress in recommenders over the past decade came from these embedding models, not from richer semantic understanding of user interests.
  • LLMs can now be pointed at a cluster in embedding space and asked to describe it in natural language—"deep pour-over coffee snobbery"—giving humans legibility into previously opaque systems.
  • Instagram's "see your algorithm" feature uses this: surface the embedding cluster → LLM describes it → user sees a topic label they can edit.
  • Why you should care

    For any AI-driven product, "explainability" and "user trust" are increasingly product problems, not just model problems. This technique—LLM-as-interpreter of opaque model outputs—is a practical pattern for building user-facing transparency into black-box systems.

    Think of it like...

    The algorithm had a perfect map of your taste written in a language only computers could read—LLMs just invented the translation dictionary.

    Authentic Human Scarcity as Platform Strategy

    0:48

    When synthetic content floods the market, authentic human perspective becomes the scarce premium good—and platforms built around creators are positioned to capture that demand.

    How it works

  • Power has been shifting from institutions (e.g., the New York Times) to individuals for years; AI accelerates the institution side by making mass content production trivial.
  • This increases the relative value of individual perspective, point of view, and the *person behind the content*—which is Instagram's core value proposition.
  • Instagram's strategic response: invest in individual creators broadly (not just influencers), improve ranking for authentic content, and provide AI-origin transparency so users can make informed trust decisions.
  • Content moderation logic stays the same (unsafe content removed, relevance via ranking), but the platform explicitly avoids value-judging content based on the tool used to create it.
  • Why you should care

    This is a live product strategy thesis from a 3-billion-user platform navigating a real market shift. The underlying logic—scarcity inversion when a previously scarce resource becomes abundant—applies to any product category where AI commoditizes supply.

    Think of it like...

    When printed photos became free and infinite, handwritten letters became more emotionally valuable, not less.

    Exploration vs. Exploitation in Recommendation Ranking

    48:35

    Recommendation systems must choose between playing it safe with known preferences (exploitation) and taking calculated risks to discover unknown ones (exploration)—and the balance shapes the entire creator ecosystem.

    How it works

  • Exploitation-based ranking: show content similar to what the user has already liked. High engagement, low risk, low discovery.
  • Exploration-based ranking: show content the user hasn't seen or requested, test whether they respond. Lower average engagement, but enables new creators to find audiences.
  • Exploration is harder to execute (you need to absorb engagement dips) but is better for niche creators, originality, and long-term platform diversity.
  • Instagram explicitly invested in exploration to increase "the number of pieces of content that break out"—directly inspired by TikTok's strength here.
  • Why you should care

    If you're building any recommendation or discovery surface—internal tool, marketplace, content feed—this trade-off applies. Pure exploitation converges toward incumbents and popular content; only deliberate exploration investment surfaces new value.

    Think of it like...

    Exploitation is a restaurant recommending your usual order; exploration is a sommelier saying "try this—I think you'll be surprised."

    Chronological Feed Incentive Collapse

    38:53

    A pure chronological feed is a design that sounds user-friendly but structurally rewards institutional publishers who post constantly, causing individual and personal content to drown.

    How it works

  • In a chronological system, recency = visibility. The incentive for any account is to maximize posting frequency.
  • Institutions (news outlets, brands) can post 50 times a day; individual friends post once a week.
  • The resulting feed is dominated by institutional content—the opposite of what users say they want.
  • Algorithmic ranking can surface the sister's engagement post over the brand's filler content, but requires accepting that recency is only one input into relevance.
  • Why you should care

    This is a clean example of how system incentives diverge from user intent—a core product design failure mode. Anytime you design a feed, marketplace, or ranking surface, ask: what behavior does this incentive structure reward at scale?

    Think of it like...

    Designing a party where whoever talks loudest gets the most attention—technically fair, but the quietest guests with the best stories get drowned out.

    Token Spend ROI and AI Budget Governance

    22:10

    As AI token costs approach the scale of engineer salaries, organizations need the same budget discipline for AI spend that they apply to headcount and infrastructure.

    How it works

  • A "token incinerator" is easy to build: an AI workflow with high spend and unclear value. Without ROI scrutiny, teams default to building them.
  • Meta's initial approach: look at dollars-in vs. value-out; shut down obvious waste. No blanket caps yet, but Mosseri anticipates caps becoming necessary.
  • The analogy: GPU/CPU/RAM/storage are already managed resources with allocation decisions. Token spend is heading toward the same governance model.
  • Caps should be proportional to team trust and demonstrated ROI—not uniform.
  • Why you should care

    In enterprise contexts with procurement cycles and budget ownership, AI spend governance will land on PMs and product leads to justify. Getting ahead of this with a clear value-tracking model is easier before finance demands it than after.

    Think of it like...

    It's the same problem as cloud cost governance in 2015—everyone was spinning up instances freely until the AWS bill arrived and someone had to own it.

    Centaur vs. Reverse Centaur in Human-AI Collaboration

    28:15

    The centaur model keeps humans in strategic control with AI as a powerful tool; the reverse centaur inverts this, with AI directing and humans executing—a dynamic to actively resist.

    How it works

  • Centaur: human upper body (judgment, direction) on horse body (AI capability/speed). Human decides what to do; AI enables doing it faster or better.
  • Reverse centaur: AI directs the work, humans are the execution layer—like gig economy workers following app instructions with no agency.
  • The risk in product development: if AI generates strategy, roadmap, and priorities, and humans just execute on those outputs, you've created a reverse centaur organization.
  • Mosseri's strategy response: keep vision and strategy explicitly human-owned; use AI to assist within human-set constraints.
  • Why you should care

    This is a decision-rights question, not just a productivity question. In regulated industries and complex organizations, surrendering strategic judgment to AI output creates accountability gaps that surface in audits, failures, and stakeholder conflicts.

    Think of it like...

    The difference between a pilot using autopilot (centaur) and a pilot who just follows whatever the autopilot says without understanding or overriding it (reverse centaur).

    AI Content Labeling vs. Creator Account Verification

    46:32

    Knowing that a *piece of content* is AI-generated and knowing that an *account* is AI-operated are two separate trust signals that serve different user needs and face different technical feasibility challenges.

    How it works

  • Content labeling: mark individual posts as AI-generated. Hard technically—current detection works but degrades as models improve. May invert: label camera-captured content as "not AI" rather than labeling AI content.
  • Account labeling: mark accounts as AI personas vs. real humans. Addresses spam vectors where AI personas promote products without disclosure.
  • The key judgment Mosseri makes: don't filter content based on AI origin—filter based on safety and relevance. But do provide transparency so users can make informed trust decisions.
  • These are product design decisions, not just moderation decisions: they shape what signals users have to calibrate trust.
  • Why you should care

    For any platform handling user-generated content or third-party AI agents, this two-layer transparency model (content origin + account identity) is a concrete design pattern for trust and provenance.

    Think of it like...

    Labeling a photo as "edited" tells you about the artifact; labeling the photographer as "stock photo service" tells you about the source—both matter, but for different reasons.

    Why this matters to you

    Your team structure is about to shrink—and that's a feature, not a bug.

    Meta is actively moving to pods of 6-7 people where a 'product staff' generalist replaces dedicated designers, researchers, and data scientists. If you're a senior PM in a large org, this is the organizational direction; understanding the role's shape and the specialist-on-demand model matters for how you hire and structure work now.

    AI is eating execution—your defensible value is judgment and taste.

    Mosseri is explicit: vision and strategy are where human cycles will concentrate as AI handles more delivery. For PMs in regulated, complex B2B environments, this means doubling down on opinionated, controversial strategic calls—not process compliance or delivery throughput.

    Recommendation system design has concrete lessons for any AI-driven product.

    The exploration/exploitation trade-off, embedding illegibility, and LLM-as-interpreter pattern are directly applicable to any product with personalization, discovery, or content ranking surfaces—not just consumer social.

    Relevant transcript passages (5)
    6:19
    a lot of what a data scientist does at a big company, for instance, is relatively mechanical. And so, you know, there's stuff that they do that is really more like, you know, art and science. And the stuff that's really more like just pulling data, data management, you know, so some of the tools that we're building internally to understand, for instance, a traditional data science question would be a waterfall.

    Concrete example of which data science tasks AI tooling is absorbing first—relevant for any PM thinking about team composition or which specialist functions to retain.

    26:57
    I think if you ask an AI just for a strategy lazily, you're not going to get something great. You're going to get something pretty predictable that probably the competition would expect you to do. I think if you want a really more effective one, you need to think long and hard about what are all the different inputs that need to be considered.

    Directly actionable for any PM using AI in strategy work—the failure mode is precisely the lazy prompt, and the fix is explicit constraint enumeration.

    10:13
    I actually think some of our strongest product staff are going to be converts from design and from data science who are just looking to expand their reach. And they were influential across functional boundaries before, but this world where those functional boundaries are just wildly blurred just allow them just to jump in.

    Reframes the career risk for designers and data scientists—the path is expansion, not replacement. Relevant for cross-functional leadership and talent retention conversations.

    59:31
    You can't launch something to three billion people and not test it first, but you can't test something at our scale and not expect people to cover it and not and be so you have to be ready to talk about it before you even know you want to launch it.

    Practical governance point about experimentation communication strategy at scale—the leak is not 'if' but 'when', requiring pre-prepared narratives even for unreleased tests.

    37:30
    I think we can come up with I don't want to see seven photos in a row, but I'm happy to see six photos in a whatever your heart can, you know, whatever your mind can come up with. So, we have a lot of work to do. And so, I'm excited about that.

    Signals the product direction for algorithmic transparency and user agency—LLM-interpreted embedding topics that users can edit—a concrete pattern for trust-building in AI-driven personalization products.

    Key Insights (10)
    • The canonical large-company product team (13 people with dedicated specialists) is being replaced at Meta by pods of 6-7 with a generalist 'product staff' at the core—specialists brought in only when craft depth is genuinely required.
    • Taste and strategic judgment are Mosseri's top candidates for the last non-automatable human skills in product development—more durable than technical execution, domain knowledge, or even management.
    • Recommendation algorithms have never had the semantic understanding of user interests that users assume—they operated on opaque embedding vectors. Only now, via LLMs interpreting those embeddings, is human-legible explainability becoming possible.
    • A pure chronological feed is a perverse incentive system that rewards posting volume, causing institutional content to crowd out personal content—the opposite of user preference.
    • AI-generated content abundance is a potential *tailwind* for Instagram specifically because it increases the scarcity value of authentic human perspective—the platform's core positioning.
    • Strategy generated lazily from AI prompts produces consensus, predictable output—genuinely differentiated strategy requires the human to first enumerate all non-obvious constraints (team dynamics, regulatory context, brand identity, talent market) before steering the model.
    • Token spend governance is an emerging organizational problem: treat AI inference costs like headcount and compute, with ROI scrutiny—not as an unlimited productivity enabler.
    • The best product leaders are curators, not visionaries—creating conditions for good ideas to surface and selecting among them, rather than being the primary idea source.
    • Exploration-based recommendation ranking (discovering unknown interests) is harder than exploitation (reinforcing known preferences) but is the mechanism that enables new creators to break through and platforms to stay culturally relevant.
    • Hiring signals trending up: curiosity, willingness to try things publicly and be wrong, self-awareness. Trending down: large organizational leadership scale (fewer roles managing thousands of people).
    Action Items (8)
    • Audit your current team structure against the pod model: identify which specialist functions are doing primarily mechanical work that AI tooling could absorb, and which are doing genuine craft work that warrants dedicated headcount.
    • If you use AI for strategy work, build a constraint-enumeration step before prompting: explicitly list team dynamics, regulatory landscape, brand identity, competitive context, and talent considerations before asking for strategic recommendations.
    • Implement a 'leak readiness' protocol for any experiment that could be controversial: define the reactive and proactive communication narrative before the test launches, not after it's spotted.
    • Evaluate your AI spend with a 'dollars in, value out' lens—identify any workflows that are effectively token incinerators with unclear ROI and shut them down before finance demands justification.
    • Map your product's discovery surface against the exploitation/exploration trade-off: if you're purely exploitation-based (showing users what they already like), model the cost to originality and new-entrant discovery.
    • If you run a recommendation or personalization system, assess whether LLM-as-interpreter of embedding clusters is a viable path to user-facing explainability—Instagram's 'see your algorithm' feature is a live reference implementation.
    • Apply the centaur test to any AI workflow you're deploying: is the human setting constraints and reviewing output (centaur), or is the human executing AI-directed tasks without strategic input (reverse centaur)? Restructure the latter.
    • Consider a two-layer transparency model for any platform handling third-party or AI-generated content: content-origin labeling (is this AI-generated?) separate from account-identity labeling (is this a real person?)—and be honest with users about your detection confidence levels.
    Skip if: Skip if you already have a fully formed view on AI-era team restructuring and are not interested in consumer social platform mechanics—the recommendation system and creator economy sections are Instagram-specific and may not transfer directly to B2B or regulated industry contexts.