By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Head of Instagram Reveals How the Top 1% Actually Work With AI
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
Great conversations run long. Their best ideas don't have to.
uListen is free on iPhone, iPad, Android, Chrome, and Claude.
See howCHAPTERS
Jump to publisher timestampsChapter summaries
Why AI needs feedback loops to become truly useful
Adam Mosseri explains that large language models are "average of the internet" by default, and become far more valuable when you build loops that help them learn from real outcomes. Marina frames his setup as unusually advanced for running a large organization, setting up the discussion of how he actually uses multiple AI tools day-to-day.
- •Models can be capable but generic out of the box, so loops make them more personalized and effective
- •A learner’s mindset is required because tools and models change rapidly
- •His AI systems evolved through iteration, and early versions were not helpful
- •He describes learning every day in a way he hasn’t had to in years
Roger: an AI follow-up system that tracks promises and accountability
Marina asks about Roger, the tool Adam built to track things people agreed to do across meetings, email, and Slack. Adam describes using it to surface overdue follow-ups and keep accountability high across a large team.
- •Roger captures requests and commitments from meetings, email, Slack, and more
- •It nudges Adam to follow up when someone hasn’t delivered after agreeing
- •Adam uses it to send friendly reminder messages and prevent items from slipping
- •The goal is maintaining accountability without relying on memory alone
The “desk” command center: one place for tasks, meetings, briefings, and history
Adam explains the visual model behind his command center: one primary focus item plus a long list of other tasks just out of view. He built a “desk” interface that consolidates Roger, Caddy, meeting briefings, and meeting history into a single view.
- •The desk mirrors how he mentally visualizes work: one main focus plus a backlog
- •It consolidates multiple tools and information streams into one experience
- •He can snooze, edit, and blast through tasks like feedback, unblocks, and nudges
- •It’s not perfect, but he says it makes him more productive
How he built the system (and why non-technical users aren’t fully there yet)
Adam details the implementation: an internal harness, internal APIs, a simple web app, and nighttime automations that extract meeting and communication context. He also explains that fully non-technical users still need to guide and iterate to get something good today, though simpler agent setups can work for most people.
- •Built as a web app with a simple SQLite database plus nightly automations
- •Automations gather meeting outcomes, action items, requests, and messages
- •He says no-code prompting alone usually isn’t enough for a complex command center yet
- •For simpler needs, an agent with email access can remind you about important replies
AI-drafted emails that actually sound like him: training via draft-vs-sent comparisons
Marina asks about AI drafting most of his emails; Adam explains he still edits before sending. He describes a loop where the system compares the AI’s draft to the final email he actually sent and learns from the differences over time.
- •AI drafts certain email responses, but Adam reviews and edits them
- •A loop saves the AI draft and compares it with the final sent email
- •Over time, it learned his tone, values, and schedule-related preferences
- •He emphasizes loops reduce the need for manual feedback every time
Meeting briefings that improved from “garbage” to genuinely helpful
Adam walks through a meeting-briefing agent that summarizes long documents and highlights what matters most. The first version produced extra reading with no benefit, so he added a loop that evaluates briefing quality using post-meeting artifacts and signals what Adam actually cares about.
- •An agent summarizes long pre-meeting materials into key “don’t miss” items
- •Adam still reads full docs, but uses the briefing to anchor attention
- •The loop pulls post-meeting notes, action items, Slacks, and Adam’s notes to self-grade
- •Over a month or two, briefings became helpful and kept improving weekly
A kid-friendly podcast generator app for his children (topics, age, length, voice)
After the sponsor break, Marina asks about an app Adam built that generates podcasts for his kids. Adam describes grounding sports content with timely score APIs, selecting a kid-friendly voice, and letting users set listener age and episode length.
- •Built APIs for up-to-date sports scores and scaffolding around time awareness
- •Different prompts for different interests (scores, history, players, rankings)
- •Users set listener age, episode length, and topic before generating audio
- •He tuned voice and engagement style specifically for his kids
Custodians: a shared memory and skills layer so he can switch AI tools easily
Adam explains a personal system called Custodians that prevents “starting from zero” whenever he tries a new model or assistant. It stores text-based memories and skills, maps them into an embedding space for conceptual search, and provides read/write guidelines so multiple tools can share context.
- •He frequently switches among different AI tools and models, which makes re-teaching annoying
- •Custodians provides shared activity, memory, and skills so tools benefit from each other
- •Memory is stored as text and searchable by concept via embeddings
- •Tools are instructed to save useful memories under guidelines, enabling plug-and-play experimentation
Using AI to stress-test decisions—and delegating decisions publicly to avoid bottlenecks
Marina asks how Adam uses AI for decision-making; he says he often uses it to stress-test choices. He also explains his philosophy of delegating decisions to capable people and making delegation explicit so teams don’t default back to him.
- •AI is used to stress-test decisions rather than replace judgment
- •He tries not to make every decision himself to avoid becoming a bottleneck
- •Delegation is made public so everyone knows whose call it is
- •Clear accountability changes team behavior and can be empowering
Hiring in a fast-changing world: grit, fast learning, and self-awareness
Adam outlines three traits he looks for in hires across functions: grit/hustle, ability to learn quickly, and self-awareness. He explains how he tries to test learning speed in interviews and notes that job boundaries are blurring as designers, data scientists, and engineers increasingly overlap in tasks.
- •Core hiring criteria: drive, quick learning, and self-awareness (ability to take feedback)
- •He tests learning by discussing topics candidates wouldn’t already know
- •Role changes can signal learning if they’re meaningful and sustained (not constant churn)
- •Technology shifts are blurring lines between design, engineering, and data science
Learning like a second language: embracing failure, avoiding “token maxing,” and focusing on impact
Adam says people who struggle to evolve and keep learning are hardest to retain in a shifting environment. He compares learning new ways of working to learning a language—requiring comfort with sounding wrong—then discusses why high AI usage metrics can be misleading and why he focuses on real outcomes instead.
- •He emphasizes growth over time: getting stronger month over month in capability
- •Learning requires willingness to “sound like an idiot” while you build fluency
- •He critiques treating token usage as a goal because it can be productive or wasteful
- •He warns companies conflate proxy metrics with real goals, and he prioritizes shipped impact
What Instagram measures (and what it can’t): sentiment, ranking trade-offs, and deeper value than easy metrics
Adam describes how Instagram measures sentiment using in-product questions about satisfaction and fun, alongside usage metrics. He explains why ranking exists, then highlights a core limitation: ranking tends to optimize for what’s measurable, even though harder-to-measure outcomes can matter more.
- •Sentiment is measured by surveying large numbers of people on-platform about satisfaction-related questions
- •They also track health indicators like daily use, messaging, stories, and impressions
- •Ranking is used because there’s too much content, and without ranking usage and sentiment drop
- •He distinguishes easy-to-measure first-order preferences from harder-to-measure second-order value (what was worth your time)
Views aren’t the whole story: reach as a means, audience quality, and deciding what not to build
Building on Marina’s example of an episode with fewer views but strong audience impact, Adam argues reach is a means to an end and that not all views are equal. The conversation then shifts to how he decides what not to build at Instagram, emphasizing focus, quality, talent density, and aligning bets with future opportunity and product identity.
- •Reach is framed as a tool to achieve goals like community, brand, evangelism, or sales
- •Views are described as a shallow measure (including very short views) and not all views are equal
- •He argues for focus: doing fewer things at higher quality amid incentives to do more
- •He prioritizes opportunities likely to matter in a few years, aligned with Instagram’s identity and key audiences
Spotting opportunities and extracting signal from feedback with help from AI
Adam explains that trajectories matter more than snapshots when choosing markets, products, or people. He describes seeking culturally relevant signals via research, comments, DMs, travel, and user conversations, then discusses filtering noisy feedback and using AI to surface worthwhile DMs and questions he should answer.
- •He prefers trends/trajectories over current size: growing markets and rapidly improving people
- •Signals come from surveys, research, ethnography, and direct user feedback like DMs and comments
- •He tries to find signal in feedback while accepting negativity as part of working on a large platform
- •He uses an AI system to surface notable messages and a shortlist of interesting questions to respond to
Creator tactics and features: what drives unconnected reach, comment-to-DM “hacks,” underused tools, and Trial Reels limits
Adam shares what metrics correlate with reaching non-followers and how he balances performance with mission-aligned posts that do worse. Marina asks about “comment for a link” tactics; Adam calls them an engagement hack that works but questions whether it improves consumer experience. He also highlights lesser-known features and clarifies Trial Reels were designed to preview content, not to run full A/B tests.
- •For reaching non-followers, he highlights sends per impression and the importance of rates over totals
- •He distinguishes metrics that matter for followers (watch time and like rate) versus non-followers
- •He says “comment for a link” tactics work but frames them as a hack and questions their value to viewers
- •He calls out Blend and a tool to view/edit interest topics as favorites, and explains Trial Reels aren’t meant for A/B testing variants
Taste, music choice, and how AI raises the stakes on learning new things
Marina and Adam discuss intentionality as a component of taste, using music selection as an example where the track should connect to the content through artist name, title, or lyrics. Adam closes by describing how AI increases the demand and value of learning, and how experimenting more has helped him get faster at picking up new things.
- •They emphasize intentionality: music and creative choices should connect to the content and match the vibe
- •Adam describes a preference for a “through line” via lyrics, track name, or artist name
- •He says AI makes continuous learning more necessary and more valuable
- •He frames learning as a muscle, and experimenting has made him quicker at learning over time
Publisher timestamps
- 0:00Inside Adam Mosseri’s AI Setup
- 1:13Roger: Keeping Track of Promises and Follow-Ups
- 3:26How Adam Built His AI Command Center
- 5:33How He Gets AI to Write Emails That Sound Like Him
- 7:18Turning AI Meeting Briefings Into Something Useful
- 9:52The Podcast App He Built for His Kids
- 12:17Switching AI Tools Without Starting From Zero
- 13:53Using AI to Stress-Test Decisions
- 15:10The Three Qualities Adam Looks for When Hiring
- 17:14Why You Have to Be Willing to Sound Like an Idiot
- 18:25Why More AI Usage Doesn’t Always Mean More Productivity
- 21:00What Instagram’s Algorithm Can and Can’t Measure
- 23:02Why Views Don’t Tell the Whole Story
- 25:59How Adam Decides What Not to Build
- 27:47How to Spot Your Next Opportunity
- 30:15Finding Useful Feedback in Comments and DMs
- 32:39What Creators Can Ask an AI Assistant
- 33:56The Metrics That Matter for Reaching New Audiences
- 36:02Do “Comment for a Link” Tactics Work?
- 36:59Adam’s Favorite Underused Instagram Features
- 37:59What Trial Reels Were Actually Built For
- 39:03What Taste Has to Do With Choosing Music
- 40:20How to Get Better at Learning New Things