提前排好棋路：Peter Deng 如何把产品做到数十亿人

Asked whether AGI is coming and will solve everything, Peter Deng says nobody really knows what that world looks like, but he doubts it solves problems on its own. In his view it is a new source of energy that still needs a lot of hustle and elbow grease from builders to be channeled into something humans want to use.

He draws an analogy with databases. When they arrived, it seemed amazing that you could store and query data quickly, and entrepreneurs built great products on top of them. Today that looks obvious, and Deng suggests that in 10 or 15 years a super-intelligent thinking machine may seem equally obvious.

On fears about the technology, he points out that every new technology triggers them, even the bicycle. Humans co-evolve with technology, moving from fearful to familiar to masterful, and he notes how much the startup landscape has shifted in just 18 months since ChatGPT launched. That track record is the basis for his optimism.

Deng thinks education is the area where AI will change things most in ways people underestimate. His nine-year-old son, who used ChatGPT before it was public, started asking questions in new ways and even built a custom GPT that writes a sentence containing every letter of the alphabet on any topic. That is a task that would be very hard to write as a traditional Python function.

He doesn't expect his son to need to code for a living, but sees programming as valuable training in structured thinking. The skill he expects to matter most is knowing how to ask the right questions, since everyone will have access to the same AI. Being inquisitive and turning that into good prompts becomes the differentiator.

He compares this to the calculator, which didn't end math but let people do higher-level math, and to Google making memorization less necessary. The next step, he says, is that code appears when summoned, so people should develop skills at the next level of abstraction that draw on creativity and curiosity. Lenny adds that teachers are using ChatGPT for curricula while students use it for essays, an arms race that suggests the whole system will have to change.

Deng credits a college course called Language and Thought, taught by Herbert Clark, with a thesis that stuck with him: language affects how you think. Having grown up speaking Chinese, he noticed his own thinking shift when he learned English. He cites a recollection of studies on Russian speakers distinguishing shades of blue, while admitting he would need to check the details.

In practice, he obsesses over wording. A recent class presentation of his contained about 20 words in total, and he spent hours on them. In vision docs and PRDs, he argues, careless words get misinterpreted and carry the wrong connotations, with downstream effects that multiply.

He finds it fitting that the AI breakthrough came from large language models, since so much knowledge is condensed in words. He adds that ChatGPT often works by writing and interpreting Python, itself a language, and recalls Ilya Sutskever's argument that predicting the next word well requires understanding the world that produced the text.

The most counterintuitive lesson Deng learned, at Uber, is that sometimes the product, meaning the pixels in the app, doesn't matter as much as you think. Price and ETA were the real product, and fixing a bug had less impact than improving something people cared about more. He says it pained him as a design-minded PM, and that the same applies to B2B products that are loved by users but don't make business sense.

A related insight is that many of today's most valuable tech companies didn't begin with a technological breakthrough. Facebook took a database of human connections and kept polishing it, and Uber combined GPS in pockets, cars and a human need, then built heavy technology later. Deng calls Uber "largely an operations company" and praises the colleagues who ran it that way.

Instagram is his example of craft winning: the idea of visual sharing was simple, and Kevin Systrom and Mike Krieger's product taste made it take off. Lenny frames the tension as either the product doesn't matter because operations dominate, or no breakthrough means the product matters a lot. Deng agrees both hold, and adds that even a company with a real technical edge will soon find that the experience matters once the advantage fades.

For companies building on top of LLMs, Deng says the defensibility comes from data and workflow. Proprietary data is a start, but the flywheel is how you keep generating it, and the workflow is how the product integrates into people's lives. Models are malleable and do what they are trained on, so being deliberate about the data you can access matters. He and Lenny cite Windsurf's use of accepted and rejected code suggestions as an example.

Distribution advantages can be overcome if the product is far better. Lenny points to Cursor, Windsurf, Lovable and Bolt outpacing Microsoft's Copilot despite Microsoft's talent and head start, and Deng adds Granola, which competes against Google Meet, Teams and Zoom through small, delightful touches. He says the question used to be who has the best model, and is now who has the best workflow and product.

On what product does at model labs, Deng says being in the same building as researchers creates a symbiotic link between post-training and product. He expects the work to be less about raw intelligence and more about fine-tuning what the model does. Lenny relays that Mike Krieger found his biggest wins came from putting PMs on research teams, and Deng says OpenAI independently reached the same conclusion with its post-training team.

Deng says the one-to-100 phase is different from finding product-market fit. Having quadrupled Instagram usage in two years, he argues you must think ahead and build systems that let you go sustainably faster, comparing it to feeling the G-forces of a rocket takeoff. The Facebook Newsfeed has barely changed in about 12 years, he suggests, because the team thought carefully about the whole sharing loop and its information architecture.

The same applied elsewhere. At Uber, the team re-architected a spaghetti-code rider app and built a dedicated pickup and drop-off effort, finding the right abstraction for venues so airports and other places could be configured at scale. On Messenger, thoughtful push-notification infrastructure helped the app grow from zero to 4.7 billion messages sent per day in about two and a half years.

Asked when to make this shift, he says it isn't a binary switch but a ramp, and recommends a portfolio approach, citing Google's 70-20-10 and suggesting perhaps 50/50 for a startup. His second tip is to measure everything, since you wouldn't fly a plane without instruments. He starts almost every team by building a growth team, because hiring a growth leader forces you to log what you haven't and ask the right questions, and because growth leaders, unlike an analytics team, are tied to outcomes and so get listened to.

To protect taste and craft while pursuing numbers, Deng gives two people different charges: one to grow the product and one to maintain its design quality. He sees that tension as healthy and says he created it at Instagram and saw it at Facebook, Airtable and ChatGPT. His approach is to think of the team as a product, assembling people who spike in different ways, with the leader adjudicating disagreements.

He describes five PM archetypes that came out of a jam session at Uber to help with hiring. The consumer PM is half designer and obsessed with details; the growth PM is half data scientist and wants to test everything. The GM or business PM is half MBA and thinks in margins and incentives; the platform PM loves building tools that make others faster; and the research PM, previously called the algorithms PM, mixes researcher, engineer and product taste.

Most people, he says, lead with one type and are balanced by a secondary one. He counts himself as primarily a consumer PM with a growth-PM secondary. As an investor, he looks for founder-market fit, since a consumer PM dropped into a dull regulated industry will likely get frustrated, and he hopes the framework gives people room to be a different kind of PM.

Deng says he spent as much time recruiting and building the OpenAI team as on the product, and he is prouder of that team than of the products. He insists leaders should look at where the team needs a spike and find the best person for each, rather than treating people as warm bodies. Lenny notes that this reframes hiring from a checklist for an individual to the composition of the whole team.

His personal razor is "in six months, if I'm telling you what to do, I've hired the wrong person." It keeps his bar high, tells new hires what success looks like, and shifts the conversation from hitting an OKR to whether they are calibrating well enough that the hire will soon be leading him. Mistakes become learning opportunities, and he thinks it applies to any manager, including in interviews.

His second criterion is growth mindset, which he learned to prize after a hiring mistake at Facebook. In his org he does the final interview and tests only for that, trusting others to assess product sense, design and execution. He asks candidates to describe a painful mistake, what they learned, and how they work differently now, and he shares his own mistakes in return, which he believes sets a tone of "it ain't a loss, it's a lesson."

Joanne Jang, who worked with Deng at OpenAI, credits him with a managing-up phrase: say you'll do the thing, do the thing, say you did the thing. Deng traces the underlying idea to a former Uber colleague's saying that "repetition doesn't spoil the prayer." Stating what you'll do lets you calibrate with your manager, and closing the loop makes sure your work is seen, which matters especially for introverts who avoid attention.

He also helped Jang shape a new role. He praises her as the rare person with as much technical depth as product taste, asked her to write up what she was excited about, and then created and put her in charge of the model designer function. The first two model designers were hired after a search, and he credits that technical-plus-taste role for much of ChatGPT's feel.

His wider view is that fit is a two-way street and that people should lean into what they love rather than force-fit an archetype. He sees today's wave of new companies as a chance to find a place that resonates, much as podcasting didn't exist 20 years ago but now lets people express themselves. Lenny draws out the leadership lesson: notice what excites people and ask them to describe it clearly in a document.

Asked what makes a successful product person, Deng says to obsess over the details of craft while having the wisdom to know which details don't matter. He admits he wasted time on trivial details when young. His example is Uber Reserve, built around the peace of mind of a 6 AM flight: you say when your flight is or when you want pickup, the product warns if you're cutting it close, and driver incentives are designed in too. He says it is now approaching a $5 billion a year business.

He and Lenny connect this to dogfooding. Deng leased a white VW, put an Uber sticker on it and drove for two weeks before joining Uber, echoing what Uber's CPO Sachin said about driving with a phone a few feet away. Deng also recalls the IDEO five-stage design-thinking process of empathize, define, ideate, prototype and test, and values the first two for forcing you to feel your customers' pain and to state the problem precisely.

He warns that shortcuts miss the point. Having flown to LA with Kevin Systrom to watch Instagram users through one-way glass, he says asking ChatGPT to summarize user interviews means you can't empathize with the summary. You have to be in the room, hearing the words and intonation, a view Lenny backs with Bezos's advice to trust the anecdote when it disagrees with the data.

Deng explains why he left Google for early Facebook: he was drawn to how humans are wired, and he saw a team that understood people's fundamental desire to connect. Their mission, to make the world more open and connected, felt more human than the vague mission statements of rivals such as Friendster or MySpace. He also says he has always optimized for learning, and moved roughly every two and a half years during his nine and a half at Facebook.

The takeaways, as Lenny sums up, are to ask whether you'll learn, whether the company's insight matches human behavior, and whether you share its view of the world. Deng looks for a unique insight and strong conviction in founders, quoting a saying from Instagram's Mike and Kevin: "We may not be right, but at least we're not confused."

For his failure story, he describes Bolt, an Instagram camera-first app built by a top-notch design and engineering team and tested in a small market. It failed, and retention curves that never flattened told them so. He says the lesson is that even the best team can't predict hits, and that some of the technology was reused in the main app, because "it ain't a loss, it's a lesson."

In the lightning round, Deng recommends Sapiens, The Design of Everyday Things by Don Norman and The Silk Roads by Peter Frankopan, and names The Wire as a favorite show and Granola as a recent product he loves. His life motto, taught by his father, says that if you move a tree it dies but if you move a person he thrives, which fits his preference for new experiences.

At Felicis he invests at seed, seed-plus and Series A, where he can help founders see the next stage and scale from one to 10 and 10 to 100. He asks whether a startup has unique data and a data flywheel, a crafted workflow, and insight into which product details matter. He prefers LinkedIn for outreach, and says AI is a new element humanity has discovered and he wants to be inspired by what creative people do with it.