Peter Sellis: Design Teams Like Extremist Organizations, and the Three Oxymorons of Great PMs in the AI Era

AI is making writing code, designing interfaces, and building prototypes faster. As the supply of features expands, what product teams genuinely lack has shifted: the core challenge is no longer merely "can we build it," but how to keep dozens or hundreds of people making decisions in the same direction without endless meetings; how to discern what is worth building and what should never ship despite being great; and who has the capability to take ownership of those trade-offs.

Peter Sellis offers an interconnected set of answers to these questions. As Snapchat's first product manager, he spent seven years there before serving as Head of Product at Discord. When interviewed by Lenny Rachitsky, he was between roles—freed from having to adhere to corporate PR lines—and shared insights that sitting executives rarely speak aloud in public.

His most eye-catching phrase was "design teams like extremist organizations." But this is neither an endorsement of violence nor a mere provocative soundbite for distribution. It points directly to the foundational premise of Peter's entire product philosophy: a team is, first and foremost, a decision system. A shared mission dictates where everyone is heading, while clear authorization determines who can act immediately; the value of a product manager lies in enabling this system to make better decisions than it would "without a PM."

Following this thread, Snap's advertising dilemma, Discord's return to its gaming core, Evan Spiegel's habit of rejecting good ideas, and the three sets of oxymorons embodied by great PMs are all answering the exact same question: when production becomes increasingly cheap, how does judgment become the true bottleneck?

1. Design Teams Like Extremist Organizations: Design the Decision System First

Collaboration Is Not Free

In most companies, "strengthening cross-functional collaboration" is treated as an unquestioned virtue; few managers openly ask teams to collaborate less. In Peter's view, this is precisely why the cost of collaboration has gone uncalculated for so long.

The more people involved in a decision, the greater the need to sync information, align calendars, explain context, and secure buy-in. As long as a single node fails to respond, the entire chain grinds to a halt. In Peter's words, coordination costs force a system to advance at the speed of its slowest node. Collaboration can certainly fill information gaps and uncover risks; the problem is that organizations frequently mistake "more people participating" for quality itself, without ever asking what uncertainty was actually eliminated by that participation.

In high-growth companies, this cost becomes especially glaring. When business scale surpasses the founder's capacity to personally call every shot, product leaders must make the choices that "the founders themselves would have decided just a year or two ago." If every decision requires re-convening every stakeholder, the larger the organization grows, the slower decisions become.

Two Conditions: Shared Mission and Clear Decision Rights

Peter extracts two specific characteristics from so-called "extremist organizations."

First, members understand with complete clarity why they are taking action. A vision cannot live solely in a founder's head or inside an annual strategy doc; it must be compressed into language that everyone can repeat and use to navigate daily trade-offs. Peter believes that product leaders at high-growth companies are, to a large degree, "vessels" for the founder's vision: they absorb the founder's judgment and distill the complete strategy into a handful of repeatedly invoked phrases. The value of these phrases is not to boost morale, but to serve as shared premises for decision-making when leadership is not in the room.

Second, who can decide what must be crystal clear. Peter highlights the DRI (Directly Responsible Individual) and single-threaded owner: a single, designated individual owns a given class of problems single-threadedly. The point here is not to find someone to take the blame when things go wrong, but to ensure that accountability and authority reside at the exact same node. If someone holds accountability without final decision rights, they remain reduced to begging for approvals everywhere; conversely, decision rights without a shared mission quickly devolve into fractured, siloed fiefdoms.

Shared mission answers "where do we head when leadership isn't in the room," while clear authorization answers "who actually makes the call." Only when both exist simultaneously can a team cut down on meetings and approval requests without losing its direction.

This design does not mean eliminating all collaboration. Decisions regarding security, legal, infrastructure, and the like may still require cross-functional input. What Peter actually demands is turning collaboration from an unthinking default into a deliberate, costly choice: only when another person's knowledge can substantively alter the decision is it worth adding more nodes to the chain.

2. Organizations Do Not Distribute Effort Equally: Empower the Strong, Accommodate the "Non-Standard"

Continually Raise the Stakes for Good Decision-Makers

Peter's second managerial conviction is even more jarring: when an individual consistently makes good decisions, he will continually load more responsibility onto them until the quality of their decision-making begins to degrade. This process is like feeding out more and more rope to see just how high they can climb.

This is not about dumping more busywork onto the busiest person. Peter observes decision quality and uses real-world problems to test the boundaries of a person's capability, rather than predetermining their ceiling via job levels, tenure, or job descriptions. The requisite counterpart to this is public air cover: when facing executive or external pressure, the manager never pushes the delegate out to absorb the political fallout. Internally, teams can conduct post-mortems, but externally, the delegator absorbs accountability first. Otherwise, so-called delegation will merely train a cohort of people who seek CYA confirmation at every single step.

For poor performers, Peter candidly admits that he lacks an effective mechanism for improvement. He tends to let them fail naturally rather than pouring his primary energy into plugging holes. This is his managerial choice and a self-admitted personal blind spot; it should not be packaged as a universally sound talent system. The exact same "hands-off" approach can feel like trust to a proven partner, yet feel like neglect to someone who lacks support. If a manager cannot tell the difference, so-called "empowering the strong" easily degenerates into an evasion of feedback and developmental responsibility.

Spiky Talent: Sharply Spiky People Shouldn't Be Shaved Down

At Discord, Peter managed growth product savant Nikita Bier. He compared Bier to the protagonist of Taken: virtually irreplaceable when facing critical missions that require specialized skills, but also prone to deploying that same extreme toolkit when handling mundane daily tasks. Peter even recounted that to book a 1-on-1, he had to message Bier across multiple chat apps, nearly resorting to booking a paid consultation slot on Bier's advisory platform.

This story is not meant to glamorize being difficult to work with; it illustrates Spiky Talent—talent whose capabilities are distributed in sharp spikes. They vastly outstrip normal people in certain dimensions, yet are noticeably incomplete in others. Peter views an organization's inability to accommodate such individuals as a failure of the organization itself: managers should deploy rare superpowers against the problems that genuinely require them, and use role boundaries and other team members' strengths to fill the deficits, rather than trimming everyone into standard 9-to-5 employees.

He admires NBA coach Phil Jackson precisely because Jackson could take players with radically different personalities, abilities, and temperaments and forge them into championship teams. Peter also compares top-tier product teams to rock bands: even the greatest band can put out a flop album. For creators who have repeatedly proven their judgment, a manager cannot discard their entire capability because of one misfire. Yet "tolerating a flop album" neither means skipping a post-mortem nor allowing one person's spikes to inflict limitless damage on the team.

Why Peter Views Median PM Quality as Low

Peter's assessment of the PM cohort is blunt, but his argument is not simply that "product people are bad." Instead, it is an explanatory model grounded in career incentives and talent mobility.

During the ZIRP (Zero Interest Rate Policy) era, capital was cheap, making it effortless for companies to "hire just one more person." Product management also became a popular on-ramp for non-technical individuals to break into tech and secure compensation close to engineering bands. At the same time, the outputs of product work are not as easily testable as whether code compiles and runs, allowing some individuals to survive long-term purely on process, meetings, and organizational rhetoric.

At the top end, once the strongest product thinkers master user intuition, commercial trade-offs, and organizational execution, they frequently leave the pure PM track: some start companies, while others step into executive leadership to take on end-to-end business accountability. With the strongest continuously departing, the role expanding rapidly, and weak performance proving difficult to measure, Peter argues that the median quality of PMs gets dragged down. He describes this distribution as a power law; a boundary should be maintained here: this is an analytical model based on career incentives and his personal experience, not an industry study proving that all PMs adhere to a specific statistical distribution.

Peter also observed a "horseshoe" dynamic: both the strongest and weakest engineers rarely complain about PMs, while the middle tier complains the most—for entirely different reasons.

This remains Peter's experiential observation rather than an empirical measurement of the engineering population. Its takeaway is clear: a quiet team does not signify an exceptional PM, nor do rampant complaints prove a PM is incompetent. The true baseline is Value Over Nobody—the product could already move forward with founders, engineers, and designers alone; a newly joined PM must prove they generate more value than "having nobody in that seat."

Are critical problems identified earlier? Are trade-offs made clearer? Do engineers and designers experience less rework and spend more time on high-leverage work? Do pivotal decisions turn out better because of this individual? If the answer is consistently negative, the PM has not only failed to create value, but may have actively introduced a new layer of communication overhead and a decision bottleneck.

3. Judgment Beats Output: Two Real-World Stress Tests from Snap and Discord

Snap's Challenge Was First and Foremost Business Math

Snap's advertising struggles cannot simply be chalked up to "an ad team that wasn't strong enough" or "missing a few monetization features." In Peter's view, it was bounded by three structural product constraints simultaneously.

The user base was too young. Young users, especially minors, face far tighter restrictions on ad tracking, targeting, and attribution. When advertisers struggle to verify whether an ad drove a purchase, they are far less willing to pay premium rates for impressions. CPM is the cost per mille (thousand impressions); Peter noted that CPMs shift dramatically as a user moves from age 17 to 18, and then to 21. The issue is not that young people lack spending influence, but that their influence is far harder for advertising systems to measure reliably.

Opening the app goes straight to the camera. Feeds naturally possess continuous inventory slots where ads can be inserted as users scroll. To fulfill the promise of "the fastest way to share a moment in front of you," Snap opens straight to the camera. Snapping, picking friends, and sending forms a very tight creation and communication loop; forcibly injecting ads here destroys the core user experience. Peter considers failing to find an ideal monetization path for the camera one of his major failures at Snap.

Messaging tools suffer low monetization efficiency per unit of time. Messaging can be ultra-high frequency and represent an indispensable daily habit, but people open it to reach specific individuals as quickly as possible, not to continuously consume ad-insertable content. High usage frequency does not automatically translate into high ad inventory.

When these three constraints compound, Snap could boast a massive, highly engaged user base and still find it nearly impossible to replicate the ad economics of Meta's feeds. Yet this does not mean Snap failed to build an advertising business. Peter recalled that while Snap started later than Twitter, Reddit, and Pinterest, its ad revenue reached roughly $4 billion faster. This figure is drawn from his interview recollections; the point is that the team built a substantial business despite disadvantageous structural constraints, not that the structural constraints ceased to exist.

Peter also entertained a counterfactual: if the Snapchat+ subscription had launched four years earlier, would the advertising business have faced less growth pressure? Subscriptions for communication products typically require users to first build daily habits and perceive long-term value before paid penetration gradually compounds; time allows paid habits to accumulate, but also introduces churn. Thus, this observation reflects that subscriptions are equally bound by the mathematics of time, rather than proving that launching four years earlier would have guaranteed success.

Growth Is Often Hidden in the Days Core Users Don't Show Up

Following the backlash to its 2018 redesign, Snap's overall DAU briefly plateaued quarter-over-quarter. Rather than pinning their primary hopes on another brand-new narrative, the team returned to the most fundamental experience of their existing users, particularly overhauling performance on Android. Peter largely attributes the multi-year growth rebound that followed to this rigorous investment in core performance.

Discord's experience illustrates the gap between external buzz and the primary business engine even more sharply. Midjourney was built on top of Discord and generated immense media coverage; yet Peter noted that if you look at Discord's overall usage curve, you can barely detect any trace of Midjourney's arrival. What truly underpinned the platform remained friends chatting, partying up, and hanging out before, during, and after playing games.

That is why Discord refocused on multiplayer gaming in 2024. Outside observers might assume that PC multiplayer gamers already have Discord installed and that the core market is saturated. Yet real behavior still reveals plenty of friction: some players don't join voice channels, others use in-game voice chat, and some resort directly to FaceTime. Growth does not necessarily mean searching for an entirely different audience; it can mean removing the reasons why core users drop off at pivotal moments.

Peter offered a concrete tactical path: identify the cohort using the product roughly 10 days a month, and nudge them to 11, 12, or 13 days. For a communication product, having a user return two or three more days each month increases the likelihood that they and their friends are online simultaneously, spark interactions, and continue to retain. When this shift happens across a massive base of existing users, the compounding impact can vastly exceed a new marketing campaign that yields floods of low-retention signups.

This does not imply that every product should pursue daily active use. Tax filing, flight booking, and monthly invoicing are inherently low-frequency; artificially forcing usage days upward will only create annoyance. This methodology is valid only when the team has verified that the core need naturally recurs on a daily cycle.

CPV: Bringing the Mission Directly into the Roadmap

At both Snap and Discord, Peter utilized Core Product Value (CPV) to translate a shared mission into actionable criteria for daily trade-offs.

Snap's phrasing was: "The fastest way to share a moment with people you care about." "Fastest" invites scrutiny over how long loading takes; "share a moment" demands that the app open instantly to the camera and complete the share; "people you care about" anchors the focus squarely on closest friends rather than generic broadcast traffic.

Discord's phrasing was: "The best way to chat and hang out with friends before, during, and after playing games." A new use case that drives media buzz, if it fails to improve how friends connect around games, may well be far less important than fixing a single voice-channel join failure.

A single sentence is not enough. Peter stressed that teams must identify actionable metrics that share a causal link with user retention and can be directly influenced by engineers and data scientists. Because he did not disclose the proprietary internal metrics of either company, we cannot fabricate an exact formula on his behalf. CPV must retain two distinct layers: a memorable user promise, alongside behavioral signals capable of verifying whether that promise was delivered. Slogans alone cannot guide trade-offs, while numbers alone easily lead teams to mistake metric-gaming for genuine value creation.

This brings the discussion back to the opening organizational challenge. A shared mission should not be a corporate value pasted on the wall; like CPV, it must provide real constraints when confronting the question: "Should we build this feature?" Otherwise, so-called high autonomy is merely abdicating decisions downward in the absence of standards.

4. The Scarce Value of PMs in the AI Era: Subtraction, Models, and Credible Ambition

When Features Are Easily Copied, Taste Is the Ability to Reject Good Ideas

In its early days, Snap faced aggressive copying from Instagram. Peter recalls that Instagram not only copied Stories, but also duplicated several features that had performed poorly on Snap itself; later, Instagram refined Stories around its own product ecosystem, producing a version that Peter himself conceded was superior. This example illustrates that features themselves rarely serve as durable moats: being able to copy an interface does not mean understanding why a feature fits the product; but once understood and rebuilt, first-movers do not maintain a permanent lead.

AI has lowered the barriers to producing code, design, and content even further. Current tools excel at answering "what else can we add," but rarely volunteer "what should we strip away." Consequently, the scarce value of a product leader has shifted from pushing more features out the door to discerning which features reinforce core value and which were built simply because building them was easy.

The product taste Peter learned from Evan Spiegel is reflected precisely in this restraint. He noted that he has seen Spiegel reject good ideas that were far better than features many consumer startups actually ship.

He uses the metaphor of an art museum curator to explain this: the vault holds countless precious works, but wall space is strictly limited. If a curator hangs every piece in the collection, they have failed to curate. Value emerges from deciding which works should be seen right now, in what relationship, and which great works must remain in the vault.

That is why Peter tests a product person's taste with one specific question: What have you built that was already in your hands and that you personally loved, but ultimately decided not to ship? Proposals that failed user testing do not count. What truly demonstrates restraint are those high-quality initiatives that the team eagerly wanted to launch, but deliberately abandoned due to timing, boundaries, or overarching coherence.

Systems Thinking Must Be Expressible as a Model

"Having judgment" cannot rely on personal vibe alone. Peter's standards for systems thinking are exacting: asking "why" five times in a row does not equate to understanding a system. Instead, one must be able to describe stocks, flows, probability distributions, inputs, outputs, and the levers that can shift outcomes using mathematical language.

A model can simply be a spreadsheet that ultimately lives in a shared folder. Its value does not lie in guaranteed predictive accuracy, but in forcing the team to articulate: Where do users enter? Where do they drop off? If a variable shifts, along what path does its impact propagate? Are we debating the same set of facts, or is everyone imagining a completely different system?

This is precisely the value a PM can create over "having no PM." No matter how smoothly a meeting is facilitated, it pales in comparison to converting an ambiguous debate into testable relationships between variables. A model will not make decisions for the team, but it will expose the assumptions on which those decisions depend.

Long-Term Greedy: When You Extend the Time Horizon, What Should You Do Today?

In the advertising business, Peter uses the phrase "Long-Term Greedy." This is not a moral platitude, but a commercial chain of causality: how much revenue a platform can extract from advertisers in the future is intimately tied to how much return on investment it delivers to those advertisers today.

ROI (Return on Investment) weighs the return an advertiser generates against the cost incurred. A platform can boost returns through better matching, or reduce spend through lower pricing. The latter leaves money on the table for the platform in the current period, but may lead advertisers to realize that their budget yielded superior returns, encouraging them to stay, continue spending, and steadily ramp up their budgets over time. Long-term revenue comes from customers repeatedly capturing value, not from charging them as much as possible in a single transaction.

The hardest moments to uphold this principle come when a public company misses quarterly revenue expectations, potentially triggering layoffs. The short-term pain is immediate and visible, while the advertiser's reinvestment months down the road has yet to materialize. Peter gave his team a shared vocabulary with "Long-Term Greedy": taking less revenue today is not charity, but an intentional choice for a longer accounting cycle. Yet this bet must ultimately be validated by advertiser retention, budgets, and realized ROI.

He extends this exact logic to conversational AI advertising. Traditional feed auctions account for an ad's relevance to the user; in a conversational setting, Peter envisions having to calculate the "exchange of trust." When a user asks, "Which microphone should I buy?", recommending a suboptimal product to pocket a higher bid generates revenue in the moment, but erodes user trust in future answers. He speculates that long-term trust is correlated with user retention and advertiser value, and should ultimately factor into the ranking mechanism. This remains Peter's unverified hypothesis: the specific format, calculation methodology, and whether trust plays a direct causal role were not empirically validated during the interview.

AI Makes Ambition More Critical, and Hollow Demands for "Faster" Cheaper

By dramatically lowering build costs, AI enables teams to pursue goals they once never dared imagine. Peter observes that people and organizations possess an innate tendency to maintain the status quo; genuinely shifting a trajectory requires overcoming this "static friction." Consequently, leaders capable of continuously raising a team's ambition will become far scarcer.

Yet raising ambition does not mean walking into a conference room and asking, "Why not make this 1,000x bigger?" or "Can we move a little faster?" Peter ridiculed this kind of hollow prodding. Credible ambition must be accompanied by concrete judgment: What conditions have changed? Why do the previous constraints no longer apply? What is the next building block we can lay down? What step must we drive today? Otherwise, "thinking bigger" merely offloads the difficulty back onto the execution team.

5. The Three Oxymorons of Great PMs

In the end, Peter does not provide a one-directional checklist of skills. Instead, he presents three sets of paradoxical traits that must coexist. He describes them as "three pairs of wolves" living inside great PMs. The challenge is not pushing one extreme to the limit, but knowing how to let both sides constrain each other depending on the context.

Confidence × Humility: Own the Direction, Return the Credit to the Front Line

Great PMs require an almost borderline-arrogant confidence in their direction. Without this conviction, it is difficult to make calls when evidence is incomplete, or to shield an empowered team when executives push back. Yet Peter believes this confidence must coexist with genuine humility.

PMs are naturally positioned to receive disproportionate credit because they articulate the product narrative, report outcomes, and connect cross-functional threads. A person with sufficient inner security does not need to claim every shred of credit to validate their worth. They can clearly point out which designer caught and corrected a critical detail at the final hour, or which engineer pulled an all-nighter to deliver for a partner launch.

With only confidence, a DRI becomes a dictator who elevates personal judgment above all evidence; with only humility, a leader easily turns "respecting everyone" into an inability to make the hard call. The healthy combination is: own the decision externally, invite rigorous debate internally; take accountability first when outcomes falter, and distribute credit accurately when things succeed.

Process × Chaos: Have a Checklist, but Tear It Up When Premises Change

Peter values people who are exceptionally organized and understand the power of great process. He expects PMs to build checklists, define metrics, map system dynamics, and turn recurring challenges into repeatable playbooks. Yet he places equal weight on an individual's tolerance for ambiguity and upheaval, even admiring a certain healthy disrespect for authority.

Because process is valuable only as long as its underlying premises hold true. Once user behavior, technical conditions, or market constraints shift, continuing to faithfully execute an outdated roadmap is not discipline—it is a refusal to face reality. Great PMs ask: What failure mode is this process designed to prevent? Is that failure mode still relevant today? Has new evidence invalidated our original hypothesis?

Those who embody only process treat Jira tickets, PRDs, and review cadences as actual output; those who embody only chaos tear down yesterday's plans every single day, leaving teams unable to ship reliably. Combining both sides enables an organization to operate with operational rigor, while retaining the agility to reconfigure instantly when real disruption hits.

Long-Term Patience × Daily Action: Know What to Wait For, Know What to Do Today

The exceptional PMs Peter has worked with typically carry a mental roadmap looking six months, a year, or even two years out. They know which building blocks must be laid in sequence, and accept that complex systems do not produce instantaneous breakthroughs. Snap's Android performance turnaround, Discord nudging core users from 10 to 13 days a month, and compounding reinvestment driven by advertiser ROI are not shifts that happen on a single launch day.

Yet these individuals simultaneously possess a bias for daily action that borders on "annoying." They never use long-term thinking as an excuse for inaction today; instead, they relentlessly eliminate the next point of friction and knock down the next testable milestone. Peter noted that when such people wake up in the morning, their entire team wakes up with them—not because they fabricate busywork, but because every team member understands how today's incremental step connects to the distant horizon.

Long-term patience alone easily morphs into an eternally sound strategy that never actually lands; daily action alone produces a mountain of checked boxes that fail to explain where they are leading the product. The cadence of a great PM uses the long-term destination to constrain daily actions, while using daily empirical evidence to adjust the long-term direction.

These three pairs of oxymorons ultimately converge on the exact same standard of value. A shared mission demands confidence, yet equally requires hearing front-line ground truths; autonomous teams need structured process, yet must navigate situations where no playbook exists; core growth and compounding revenue require patience, but depend on eliminating one tangible point of friction every single day.

AI can help teams generate options faster, but it will not make these choices on behalf of an organization. What Peter actually raised is not the status of the PM, but the barrier to entry for their existence: if an individual merely adds documentation, meetings, and feature volume, the team would function perfectly well without them. Only when they empower the right people under a clear mission to make better decisions—and know which good things must be stopped before the release line—do they deliver net value above "having nobody at all."