CHAPTERS
- 0:00 – 2:15
AI raises the bar: taste, judgment, and knowing AI’s limits
Adam frames AI as making building cheaper and faster, which increases the premium on deciding what to build—and building it with taste. He argues the winners will be clear-eyed about what AI is good at today, what it’s bad at, and what it’s likely to become good at soon.
- •AI lowers execution costs, so choosing the right problems matters more
- •“Taste” becomes a durable differentiator as tools commoditize output
- •Success comes from realism about AI strengths/weaknesses (not hype or dismissal)
- •Instinct for where AI is heading becomes a career advantage
- 2:15 – 8:08
Meta’s new “pods”: smaller, generalist teams and the rise of “product staff”
Adam describes how Meta is shifting from larger, specialized “canonical” product teams to smaller pods. These pods rely on more generalist engineers and a new “product staff” role that blends PM, design, research, and data work—pulling specialists in only when needed.
- •Old model: many specialists (iOS/Android/server, PM, design, DS, research)
- •New model: 4–6 generalist engineers + a “product staff” generalist + occasional specialist
- •Smaller teams coordinate less, move faster, and reduce committee-driven decisions
- •AI tooling enables generalists to do work that required specialists a year ago
- 8:08 – 14:00
Role blurring and career anxiety: why design may stay resilient
They discuss how AI is pushing functions to bleed into each other, creating anxiety—especially among designers and other specialist roles. Adam is optimistic about design because strong designers often bring taste and broader product instincts that translate well in a more generalist world.
- •Specialist tasks (e.g., basic DS analysis) are becoming more “toolable”
- •Designers may migrate into broader product roles rather than disappear
- •Best “product staff” may come from design or data science backgrounds
- •AI-made work has recognizable “vibes,” making distinct taste more valuable
- 14:00 – 16:48
Hiring for the AI era: grit, learning speed, self-awareness—and curiosity
Adam outlines the core traits he’s always hired for, then adds what’s becoming more important as AI accelerates change. He emphasizes curiosity and a willingness to try things (and look foolish) as the main predictors of adapting well.
- •Baseline traits: grit/drive, quick learning, and self-awareness
- •New premiums: curiosity and willingness to experiment publicly
- •Analogy: language learning requires being willing to “sound like an idiot”
- •Fewer roles will require managing huge orgs as teams trend smaller
- 16:48 – 21:13
AI is reshaping who succeeds: planning/reviewing replaces “pure doing”
AI is changing jobs fundamentally, not just boosting productivity. Adam notes engineering time is shifting from writing code to planning and reviewing it, and AI lowers barriers for cross-functional contribution (designers coding, engineers doing analysis, etc.).
- •Engineering is moving toward planning/reviewing more than writing code
- •AI enables people to contribute outside their formal function boundaries
- •Tools are uneven: excellent at some tasks and “remarkably bad” at others
- •Competitive advantage comes from disciplined, non-binary thinking about AI
- 21:13 – 23:23
Token spend, ROI, and why “leaderboards” are dangerous
They dig into the real operational costs of AI usage—tokens as a resource like headcount or storage. Adam argues that token-spend leaderboards incentivize waste, and predicts future budgets/caps may be necessary as usage grows (even if unit costs fall).
- •“Token incinerators” are easy to build and rarely worth it
- •Tokens should be managed like any constrained resource (GPUs, labeling, payroll)
- •Future state: AI burn rate per engineer could rival salary costs
- •Caps may become healthy, tied to trust and ROI expectations
- 23:23 – 25:56
Where humans stay essential: vision, strategy, and managing autonomy (for teams and agents)
Adam argues humans will increasingly focus on vision and strategy while AI handles more execution. He frames leadership as defining success, setting constraints, and calibrating autonomy—skills that may translate to supervising AI agents too.
- •Human value concentrates in taste, judgment, and strategy
- •Vision = desired future state; strategy = opinionated path to get there
- •Good strategy must be debatable/controversial, not generic aspiration
- •Leadership requires balancing prescriptiveness vs. wasted exploration
- 25:56 – 30:24
Why AI isn’t automatically great at strategy (unless you feed it constraints)
Lenny challenges the assumption that AI should excel at strategy; Adam explains why it often produces generic answers. Effective strategic use requires enumerating constraints (people, incentives, regulation, brand) and running an iterative, critical back-and-forth with a model that pushes back.
- •“Lazy prompting” yields predictable strategies competitors expect
- •Real strategy must include constraints: talent, motivation, regulation, compliance, brand
- •AI can clarify thinking if instructed to be critical and engaged iteratively
- •Model personalities differ; pick one that will challenge you
- 30:24 – 34:23
Great product leaders as curators: ideas, people, and team chemistry
Adam reframes strong product leadership as curation more than pure vision. The job is to create the environment where great ideas surface, choose the right approach for the broader context, and build leadership teams with complementary skills and strong trust.
- •Best leaders aren’t always “idea machines”; they curate ideas/people/tech/strategy
- •Success = great strategy + buy-in + execution, regardless of who originated it
- •Hiring includes fit with the leadership team, not just individual excellence
- •Team chemistry and trust determine whether small issues become big issues
- 34:23 – 38:06
What Instagram’s algorithm really “knows”: embeddings, illegibility, and “See your algorithm”
Adam dispels the myth that Instagram has a semantic, human-readable model of your interests. Historically, recommenders relied on embedding vectors that correlate with interests but aren’t interpretable; LLMs now help translate those spaces into understandable topics and give users more control.
- •Algorithms often don’t “know you like surfing” in words—just numeric correlations
- •Embedding spaces act like maps where similar content clusters together
- •LLMs can label/describe regions of the embedding map in plain language
- •“See your algorithm” aims to give users agency to view and adjust inferred topics
- 38:06 – 40:57
Why chronological feeds disappoint: incentives, relevance, and user satisfaction over time
They unpack why purely chronological feeds often feel desirable but underperform in practice. Chronological ordering rewards high-volume publishers, buries meaningful posts due to time zones and recency, and tends to reduce long-term satisfaction even if it feels good initially.
- •Chronological feeds incentivize posting volume, favoring institutions and publishers
- •Recency isn’t the same as importance; meaningful posts can be missed
- •System design creates emergent incentives—like “designing a city”
- •Meta’s experiments show default chronology reduces usage and sentiment over time
- 40:57 – 46:19
AI content as a tailwind: abundance makes authenticity and creators more valuable
Adam argues AI-generated content will increase supply, but that may push users toward creativity and authentic human perspective. Instagram’s advantage is its creator ecosystem; rather than banning AI content, he favors transparency and giving users the context needed to judge trust.
- •More content can mean more attention, but ranking AI content well is hard
- •In synthetic abundance, people may seek creators, authenticity, and point of view
- •Don’t judge content solely by the tool used; focus on quality and intent
- •Prefer labeling/metadata so users can make informed trust decisions
- 46:19 – 48:01
Labeling AI: content vs account identity, detection limits, and new spam vectors
They distinguish labeling individual pieces of content from labeling accounts, and why both matter. Adam notes detection will get harder as models improve; long-term, it may be easier to label “camera-captured” content. He also highlights AI-enabled impersonation and scam accounts as a major abuse risk.
- •Two labeling targets: AI-created content and AI/inauthentic accounts
- •Detection confidence will vary; platforms must communicate uncertainty honestly
- •Possible future: label verified “non-AI / camera-captured” rather than “AI”
- •AI opens new spam/impersonation vectors (e.g., synthetic personas selling scams)
- 48:01 – 1:03:03
Learning from competitors and handling backlash: TikTok exploration, public criticism, and iPad regret
Adam praises TikTok’s exploration-based ranking for breaking new creators and explains Instagram’s push toward originality, recency, and breakout distribution. He then discusses why he engages publicly, how redesign backlash got conflated, and lessons from failures like Facebook Home and early Reels—plus his regret about not building Instagram for iPad sooner.
- •TikTok excels at exploration-based ranking that helps niche creators break through
- •Instagram is investing in originality, recency, and more breakout opportunities
- •Public communication: debate happens anyway, so Meta should participate with transparency
- •Backlash lesson: tests can leak; controversial experiments require comms strategy
- •Failure learnings: Facebook Home, building early Reels on Stories, and missing iPad timing
- 1:03:03 – 1:08:28
Parenting with boundaries: kids, screens, and AI literacy (vibe coding with a 10-year-old)
Adam shares how he sets rules and earns-based screen time for his kids, emphasizing boundaries and parent-approved apps. He also describes a pragmatic approach to AI literacy—teaching creation (like building a game via Claude Code) while still protecting learning and attention.
- •Boundaries and earned screen time (not on-demand access)
- •Parents should approve apps; supports policy requiring app-store permissions for kids
- •Balance concern: AI can harm learning if it replaces thinking, but literacy is essential
- •Hands-on creation: co-building a multi-level game via vibe coding with Claude Code
