Adam Mosseri: Make AI Learn From the Work You Actually Do

The Instagram head compares AI drafts with his final emails, then applies the same feedback principle to meetings and creator tools.

AMAdam Mosseri게스트 · Instagram
채널
Silicon Valley Girl
게시일
2026년 11월 17일
영상
44:39
읽기
8분
원문
영어
YouTube에서 열기
스크롤하면 영상이 오른쪽 아래로 작아져 계속 재생됩니다. 자막은 읽기 언어를 따릅니다.
읽기 언어2개 언어로 읽을 수 있음
이 글은 아직 한국어 번역이 없어 영어 원문으로 표시합니다.
요약

30초 요약

Adam Mosseri says AI becomes more useful when it learns from a person’s corrections rather than producing generic drafts. His email assistant compares what it wrote with what he actually sends; a meeting tool checks its briefings against the discussion and follow-up. He uses a shared personal memory system to switch tools without reteaching each one. The Instagram head also warns that token usage and views are proxies, not goals: creators and product teams should choose measures that serve the outcome they want.

  1. 1A draft improves when it is compared with the sent emailMosseri edits and sends each draft himself. His assistant later compares the two versions to learn his tone, values and schedule.
  2. 2Meeting briefings improved by checking what actually matteredThe initial briefings added reading without helping. A loop comparing them with meeting notes and follow-up made them more useful over one or two months, he says.
  3. 3Shared memory makes switching personal AI tools easierHis Custodian system lets different tools retrieve and save context rather than requiring him to reteach each one.
  4. 4Heavy AI use is not proof of productivityMosseri treats token consumption and similar counts as indicators, while judging teams by impact and what they ship.
  5. 5Ranking can overvalue what is easy to measureA click or share is more immediate to observe than whether content was worth a viewer’s time. Mosseri says both kinds of outcome matter.
  6. 6Different audiences call for different creator metricsMosseri says sends per impression matter for reaching nonfollowers, while watch time and like rate matter more for existing followers.
  7. 7Trial Reels were not designed as formal A/B testsThe feature lets creators try a Reel before placing it on their profile. Mosseri says Instagram should support legitimate comparisons without rewarding repeated near-identical uploads.

핵심 정리

  1. Compare AI drafts with the final work to give the assistant specific feedback over time.
  2. Treat tool usage, views and other easy counts as proxies; measure the outcome you actually want.
  3. Delegate decisions visibly to people closest to the work so the leader does not become the bottleneck.
  4. Mosseri says sends per impression help indicate whether a Reel reaches people who do not follow an account.
  5. When choosing opportunities, consider whether a market or a person is improving, not just its current position.
전문

챕터별로 다듬은 전문 읽기

원본 자막만 보기
01

A command center for follow-ups

Mosseri gathers requests, meetings and action items into a personal work interface. He says a simpler assistant may be enough for other people.

Mosseri says AI models start with a generic view of the world. Feedback loops that show a system what its user actually wants can make its output much more useful.

His setup changes month to month as models and tools improve. He says making AI useful requires an ongoing willingness to learn rather than mastering one fixed workflow.

A work tool he calls Roger tracks requests made in meetings, email and Slack. When someone has not delivered, he can see the outstanding promise and send a reminder.

The requests feed into a single to-do list. Mosseri built a desk-like interface that foregrounds the most important task and lets him snooze, edit or work through the others.

His command center also brings together meeting materials, briefings and history. He says it is a simple web app with a SQLite database; overnight automations gather meeting actions, messages and email to update it.

Asked whether a nontechnical person could build the same thing, Mosseri says such a complicated system still takes iterative direction and feedback. For a simpler need, an assistant with email access could simply remind someone when an important message needs a reply.

02

Let drafts learn from the emails he sends

Mosseri reviews AI email drafts himself and compares them with final messages. A separate loop grades meeting briefings against what actually happened.

His email assistant drafts responses but does not send them. He reads, edits and sends the messages himself, then asks the system to compare its drafts with the versions he actually sent.

That daily comparison helps the assistant adapt to his tone, values and schedule, Mosseri says. A generic draft that needs rewriting is not much of a time-saver; the useful part is learning from corrections without requiring a separate manual review each time.

For meetings, an agent summarizes lengthy documents and identifies three or four things he should not miss. He still reads the full material, using the briefing to orient himself beforehand.

The first briefing was more reading that helped him little. He changed the loop to compare its predictions with later meeting notes, action items, his own notes and discussions; over a month or two, he says, it learned which questions and decisions mattered to him.

Mosseri does not think a dedicated app is necessary for every task. His visual desk helps him see a week of meetings and which materials have arrived, but a simple message can supply context for one meeting.

03

Small apps for family and portable memory

A custom podcast app adjusts topic, age and length. A shared personal memory system allows different AI tools to retrieve the same context.

At home he built a podcast app for his three sons. One likes current sports scores, so he connected APIs for recent results and added date context because language models sometimes lose track of the day.

His older son likes history and sports rankings. The app lets him choose a subject, listener age, length and a child-friendly voice; he also used the age slider to generate an explanation of AI for his father.

Mosseri switches personal AI tools frequently and dislikes reteaching each one about himself. A shared system he calls Custodian stores activity, memories and skills so multiple tools can reuse information such as his email voice.

He describes memories stored as text and indexed so they can be searched by concept rather than keyword. Tools can read them and, following guidelines, write new memories; his work setup is separate.

04

Delegate decisions and hire learners

Mosseri gives capable colleagues visible responsibility. He looks for drive, learning speed and self-awareness, and cautions against treating AI usage as an outcome.

Mosseri uses AI to stress-test decisions, but he tries to delegate many decisions to capable people closer to the work. If the team relies on him for every choice, he says, he becomes a bottleneck.

He makes delegation public so colleagues know who owns a decision. That accountability changes how the team acts and gives the decision-maker room to lead rather than defaulting back to him.

Across roles, he looks for drive, the ability to learn quickly and self-awareness: a person who can take feedback and keep improving. Rapid change makes learning speed especially valuable, he says.

To assess learning, he asks about unfamiliar subjects and watches how quickly a candidate picks them up. Growth in prior responsibilities and successful moves between different functions can also be evidence.

He sees boundaries between design, engineering and data science blurring as designers write code and engineers analyze data. When asked whom he would let go, he says he worries about people unable to keep adapting and learning.

Mosseri compares learning a new way of working with learning a language: people must be willing to sound foolish, try, receive corrections and occasionally fail. He says he is learning every day in a way he has not had to for years.

High AI or token usage is not a goal in itself. Mosseri says some heavy use is productive and some wastes money, so he looks at usage as a possible leading indicator rather than proof of impact.

He distinguishes a goal from its proxy: the goal might be making people comfortable sharing, while the measurable proxy is how many share a story. For his team, he cares about what ships and whether Instagram becomes more useful, not just AI activity counts.

05

The limits of measuring attention

Instagram combines sentiment and usage metrics, but rankings can overvalue what is easiest to measure. Reach is only useful in relation to a creator’s goal.

Instagram asks users daily about satisfaction and sentiment, and also tracks usage, messages, stories and impressions. Mosseri says focusing too narrowly on any one measure can produce decisions that do not make sense.

He says ranking helps people find content among more posts than they can see, but acknowledges it is imperfect. A ranking system optimizes what it can measure, and an easy-to-measure signal is not necessarily the most valuable one.

He compares enjoying a doughnut now with a longer-term desire to be healthy. A click, like or share is easier to observe than whether a video was worth someone's time or stayed with them the next day; the latter may matter more.

Mogilko describes an interview with Brian Chesky that had fewer views but was chosen by 44% of her surveyed audience as the best of three episodes. She wants impact as well as reach, rather than pursuing virality alone.

Mosseri says reach is a means to a goal such as building a community or selling a product. A smaller audience that cares deeply may be more valuable than a larger audience counted by brief views.

He points to Acquired as an example of a podcast whose audience may be especially valuable to business advertisers even without the largest total reach. Creators need to know what kind of audience and business they want, he says.

06

Choose fewer bets and watch trajectories

Mosseri considers quality, trusted leaders and long-term relevance when deciding what to build. He favors growing markets and people who keep learning.

Asked how Instagram chooses what not to build, Mosseri says he tries to have teams do fewer things to a higher quality. The growing ease of building and competitive pressure pull in the opposite direction.

Headcount alone does not decide how many hard projects an organization can handle; the number of people trusted to lead difficult work can be the real constraint. He looks for opportunities that will matter in years and fit Instagram's identity and important audiences.

For careers or new products, Mosseri advises looking at trajectories as well as current size or ability. He would rather work in a growing market than a larger shrinking one, and hire someone improving rapidly than someone who has stopped developing.

Mogilko recalls her study-abroad business running into the pandemic and her later move into a growing creator economy. Mosseri says he looks for cultural relevance through research, surveys, comments and travel, including meeting users outside the United States.

07

Find useful signals in audience feedback

Mosseri uses direct messages, questions and AI assistance to identify feedback. Mogilko describes content that strengthens ties with existing followers even when it does not expand reach.

Because Instagram reaches so many people, strong opinions and negative comments are unavoidable, Mosseri says. He tries to find useful signals among the noise, including interesting questions in his direct messages and weekly question sessions.

He now uses an assistant to read message requests and surface notable contacts and interesting questions. It has helped him find messages he might otherwise have missed, while he still chooses what to answer.

Mogilko says AI helped her see that family and lifestyle posts re-engage existing followers even when they do not attract new ones. Mosseri sees potential for creator assistants in audience engagement, content analysis and finding trends, while saying several features still need work.

08

Match creator metrics to the goal

Mosseri distinguishes private sends per impression from watch time and likes. He also discusses engagement hacks and features for viewing and testing Reels.

For reaching nonfollowers, Mosseri says sends per impression matter: out of the people who see a reel, how many share it privately? For existing followers, he says watch time and like rate are more important; rates matter more than raw totals.

His own purpose is to explain Instagram's decisions and trade-offs, not simply maximize views. Creator spotlights perform less well than some product announcements or personal posts, he says, but support his goal of highlighting creative people.

Mogilko asks about asking viewers to comment a keyword to receive a link. She says it increases reach and newsletter signups; Mosseri calls it an engagement hack that works, while questioning whether extra reach means the content is actually more useful to viewers.

Among features he likes, Mosseri mentions Blend, a shared Reels feed in a direct message, and controls that show and edit inferred interests. He cautions that these personal features are not necessarily tools for audience growth.

He recommends Trial Reels for trying a post before putting it permanently on a profile or showing it to followers. Mogilko wants to compare different text overlays on the same video, but Mosseri says the feature was not originally built as a formal A/B testing tool.

He says Instagram restricts near-identical repeated uploads because some users try to obtain more reach by posting tiny variations many times. He thinks the platform should find a way to support legitimate tests without enabling that abuse.

09

Taste and learning require intention

Mosseri connects music choices to content and says repeated experiments have helped him learn faster.

On taste, Mosseri describes choosing music for a story because the artist, track or lyrics connect to its subject and mood, not just because an audio clip is trending. To him, taste often comes from being intentional.

He says AI has made willingness to learn both more valuable and more demanding. Experimenting has helped him get quicker at learning new things, though he found the adjustment difficult at first.

같은 채널의 다른 글

43:35EN
Silicon Valley Girl· 2026년 11월 10일Isomorphic Labs’ Max Jaderberg on narrowing drug designs with AI게스트 Max Jaderberg · 진행자 Marina MogilkoThe researcher explains virtual screening, real laboratory checks and why agent performance must be measured against useful outcomes.Drug DiscoveryAI for scienceAI agents+2
23:55EN
Silicon Valley Girl· 2026년 11월 5일Susan Wojcicki’s Family on Her Life and Lung-Cancer Research게스트 Dennis Troper, Esther Wojcicki, Anne Wojcicki, Janet Wojcicki · 진행자 Marina MogilkoHer family remembers her priorities as a mother and leader and discusses the foundation’s goals for research and earlier detection.Public HealthAI in healthcareLeadership+1
53:41EN
Silicon Valley Girl· 2026년 11월 3일Ethan Mollick on Entrepreneurship, AI Agents, and Preserving Human Judgment게스트 Ethan Mollick · 진행자 Marina MogilkoThe Wharton professor explains how to use AI as a managed collaborator while retaining the skills and choices that matter to you.AI agentsHuman-AI CollaborationEntrepreneurship+2
슬라이드51:35EN
Silicon Valley Girl· 2026년 10월 30일Karim Lakhani on Learning, Judgment and Work with AI게스트 Karim R. Lakhani · 진행자 Marina MogilkoThe Harvard Business School professor discusses how people can work with AI without surrendering their judgment or capacity to learn.Human-AI CollaborationAI AdoptionEducation+2
38:28EN
Silicon Valley Girl· 2026년 1월 9일Oscar Hoglund on Value Chains, AI Tools, and Building Epidemic Sound Without Losing Your Life게스트 Oscar Hoglund · 진행자 Marina MogilkoEpidemic Sound’s cofounder joins Marina Mogilko to explain creative contrast, the move from feature to platform, and the discipline behind sustainable company building.Company BuildingCreator EconomyContent Strategy+2