算法也需要人：Adriel Frederick 讲判断力、多元团队和增长

Frederick describes his career as moving from consulting in oil, gas and electric power, to a startup, to four years on Facebook's growth team working on user acquisition, then Lyft's marketplace, covering rider pricing, driver incentives, matching and operational tools. He now runs Reddit X, a team exploring new modes of interaction on Reddit, such as faster communication, video and audio, and different levels of permanence. A recent project is an avatar marketplace where creators sell art, using NFT technology so creators have a public way to claim rights to their work.

Lenny asks how to set up long-horizon, R&D-style teams inside an established company. Having worked on the other side, in growth and marketplace, Frederick has seen "organ rejection" of teams that look very different from the rest of the organization. His first lesson is that the rest of the company must see the team as core and critical to the mission, or others who fight for resources will resent it.

His second lesson is that it must be everyone's success, not just the R&D team's win. The third is to avoid giving the impression that all innovation happens on that team, since other teams will still innovate and the new team simply takes on what the organization needs but others lack capacity for. The goal, he says, is for the team to be seen as part of the organization.

Frederick was the first Black product manager at Facebook, and he says his Trinidadian background shapes how he builds product. Trinidad is ethnically, religiously and economically diverse, and its schools are melting pots, so people learn each other's holidays, beliefs and habits. He cites its mix of roughly 35% Indian, 35% African and 25% mixed heritage, and says that environment gave him a different perspective at work.

He gives a concrete example. Many tech colleagues assumed one phone number and one device equals one person, but growing up in Trinidad he knew prepaid numbers change, phones can have two SIM cards and expensive devices are often shared. That knowledge shaped his design of registration at Facebook, and he believes it helped Facebook's growth even after he left.

He recalls first feeling like an imposter in a room of about 30 PMs, then realizing colleagues valued his perspective because Facebook wanted to grow globally. His team included people from Trinidad, Israel, Brazil, Russia, China and elsewhere. They argued about how their friends and cousins would use a product, which let them settle in 15 minutes what might otherwise take two weeks to resolve by recruiting a panel of users, although they still had to go and talk to real users.

On fostering diversity, Frederick says it becomes easy when a company recognizes concrete business value, not social pressure. Diverse hires then stay, retention is the first step to growth, and recruiting gets easier because people see it is not lip service. Lenny adds that building for the world from a single homogeneous location can feel colonial, and Frederick agrees it doesn't feel right.

Facebook, Lyft and Reddit have all been in the news for controversy. Frederick says a PM must act as a damper on the team in both directions: when something is hailed as the best ever, remind people there's more to do, and when the press is brutal, remind them they're not as bad as portrayed and keep charging toward the mission. He thinks some controversy is unavoidable, since changing a pattern of behavior creates conflict with whoever benefits from it.

The hard part is separating valid criticism from complaints about a shifting source of power. He had little sympathy for New York taxi medallion owners, because cabs often drove past him when he hailed them. But he saw real merit in complaints about driver pay, such as payment for time spent driving to a pickup or for the empty trip back after a long ride, which he says pushed Lyft to fix it.

To judge criticism, he stayed close to users, including by driving for Lyft himself. He tells of a pickup 15 minutes away for an elderly passenger, Rick, whose destination turned out to be two minutes away because he had been drinking and worried about breaking his hip. The ride had real value, but neither driver pay nor pricing reflected the 15-minute pickup, which clarified why drivers complained.

During the Prop 22 press cycle he kept designing a product to pay drivers for that time, keep prices reasonable and avoid riders going unpicked, and was ready for either vote outcome. His advice to PMs is to plan the work, work the plan, and solve for customers in the middle of the noise. Lenny adds that talking to customers also reveals that not everyone shares the loud complaint.

Asked for his most stressful memory at Lyft, Frederick describes having to unwind a pricing algorithm change that had taken major effort. Revenue-management experts, including PhDs and authors of a textbook, advised the team. It worked well in three cities, but rolling it to more cities was painful, and when someone later wanted to change prices it took months to implement.

After a bad week he concluded the team had to rebuild it, since the company needed some ability to influence prices to run an effective marketplace. They had gotten caught up in the algorithmic complexity and ignored operational flexibility. His takeaway is that operational requirements and control must be treated as first-order requirements, rather than just handing an algorithm an objective and some data.

He explains operational control with Lyft's roughly 300 cities, each with slightly different pricing that may need adjusting for seasonal traffic, fuel prices, taxes or competitor moves. An algorithm has no visibility into such things, so a person needs to be in the loop. In a Chicago snowstorm, for example, someone must judge how to raise driver pay, and every snowstorm is different.

Frederick generalizes the Lyft lesson: algorithms need people to make judgment calls. He says it was there at Facebook too, such as deciding how often to show ads versus friends' posts or recommended content, which varies by market. He criticizes people who believe there is an algorithmic solution to everything, and says PMs on machine-learning-heavy products should treat the question of where a human decides and where the machine runs as a product design problem with a human-computer interface.

Lenny cites Mark Zuckerberg's remark that the angry emoji reaction is not used in the algorithm. Frederick calls the opposite view techno-utopian: feeding all data to an algorithm with an objective, which fails because algorithms don't understand long-term effects, human response or product intent. His pricing example is an objective to win market share, where the algorithm would drop prices to the floor; adding a profit constraint fails when a rival does the same, so people must choose where to win.

He proudly built products that help people see that game and decide where to play. In his view, software including ML is a tool like a screwdriver: machine learning excels at optimizing for a given objective, but people must set constraints and strategic choices. Algorithms can then amplify that intent across millions of individual decisions, making the product an extension of the person.

On AI, he is not yet worried about systems taking over, since text and image generators still lack decision-making and judgment and can't choose a paper's topics for you. When Lenny raises Sam Harris's argument that progress could accelerate far beyond human ability, Frederick accepts our linear-thinking limits and says even a low-probability outcome with a high cost deserves attention.

On growth hacking, Frederick says finding small changes with outsized impact is valuable, but people go wrong by assuming hacks alone will carry a product. He finds that disrespectful to users, who catch on eventually, and without fundamental value you become a one-hit wonder. The products that moved the needle at Facebook focused on the marginal user and made the product easier, through hard work on a few things: easy to find, easy to get into, and very easy to find friends.

He gives an example in friend recommendations. The People You May Know approach found mutual friends and spiraled a new user into a single friend group, so his team devised a way to surface friends from multiple groups, a game changer for new users. He also credits Tom Allison and a small team who quietly rebuilt an algorithm over a few months and crushed it. Hacks like button copy are table stakes, he says, but the core actions need grinding on.

He endorses the line "no silver bullets, just many lead bullets, and a few cannonballs". Cannonballs he names are signup with phone numbers, getting SMS delivered worldwide, and good friend recommendations: big, fundamental product changes that required major investment.

On Facebook's well-known "10 friends in 14 days" milestone, which predated him, he says its brilliance was in making a goal memorable and galvanizing, with a number, a time frame and a graph everyone could watch. He heard that Zuckerberg ended the debate by picking 10 in 14 days. There was nothing magic about it and there were quirks, like a dip in retention at 11 or 12 friends from code that stopped help at 10, but the organizational momentum was worth it.

Frederick defines the marginal user as someone on the cusp of taking the action you want. For registration he looked for countries with lots of traffic but terrible conversion, then found the worst-case person, likely on a feature phone on a slow network far from a data center. Studying that experience exposed a list of problems: wrong language, poor country detection and phone number formatting, and slow connections.

He cautions against using data alone to find the marginal user. Funnels give clues, but problems are often orthogonal to the funnel, and you have to watch people. He saw someone in India about to enter their full legal name, though nobody called them that, which meant friend requests would not be accepted and friends would not recognize them, showing up as a problem deeper in the funnel.

He built Lyft's experimentation platform and loves data, but says it is a complement, not the whole answer, and that companies with lots of data can think they are swimming in answers. He likes to go to the worst case to see everything that's wrong, then use marginal-user thinking to prioritize which barriers are closest to being resolved.

Asked how many experiments succeed, Frederick estimates around 60% succeed and 40% should be turned off. He warns of a hidden cost: time spent on many small tests that could have gone to something bigger. A laziness can creep in, because small ideas are easier to come up with, design, build and report on, which leads to incremental thinking.

He recommends a healthy portfolio, for example 80% of energy on a few cannonballs and 20% on small bets, though he says these are just illustrative numbers. Very early, when the big things are known, it might be nearly all cannonballs, since experimentation costs time. As a product matures there are fewer big opportunities, experimentation gets cheaper at scale, and refinement takes a larger share.

Asked which skills he has grown most, Frederick names organization design and empathy. He once believed the smartest, most technically competent individual contributors rise, but learned the job is building a great team, setting direction, unblocking and guiding people. Nothing meaningful gets done by one person, he says, despite "hustle porn", so he reflected on when he was productive and tried to create that for others: clear goals, a sense of safety, easy processes.

On empathy, he says it's one thing for users and different for colleagues. Getting in someone's shoes is not the hard part; the hard part is taking your own shoes off, clearing your agenda and understanding their goals, motivations and fears. Then he can find a middle ground, sometimes deciding to leave them alone, and he notes it helps that he looks and sounds different, since showing shared objectives dissolves problems.

In the lightning round, Frederick recommends *The Prize* and *The New Map* by Daniel Yergin, on the geopolitics of oil. He enjoys the podcast Revisionist History, and the Netflix show *Mo*. His favorite interview prompt is "teach me something you don't think I know", which tests empathy, knowledge and communication. As a thought leader he names the team behind Radiant Nuclear, and he asks listeners to talk for five minutes with someone different from them.