多种子，少浇混凝土：Alex Komoroske 谈 LLM 时代的园丁式产品开发

Komoroske, who spent 13 years at Google and later led corporate strategy at Stripe, keeps a roughly 600-page Google Doc called Bits and Bobs. It is a weekly reflection he writes for himself, deliberately a little illegible so readers have to work with it. He takes notes constantly in meetings, and tells people that if he is on his phone it means they said something interesting.

Behind it is a personal open-source tool, The Compendium, which holds 17,248 unpublished working notes. During the week he cleans up notes every day or two, and the tool uses embeddings to surface similar older ideas. On Fridays he ticks the ones that still resonate, distills them over the weekend and publishes on Monday mornings.

He says the mundane "will take every square inch you give it," so you must deliberately make room for reflection. Early in his career he worked from home on Fridays, skipping meetings to ask what he could have done before the week started to save time. If he had explained the same strategy in ten one-on-ones, for example, he would write it as a document in 30 minutes.

Komoroske calls LLMs a truly disruptive technology and says the industry is trying to apply late-stage playbooks to an era they do not fit. LLMs sit between humans and plain computing in cost, and they undermine the assumption that software is expensive to write and cheap to run. In his view this means advertising-funded consumer products cannot cover inference costs, and a prototype that once took months no longer differentiates anyone.

He compares the shift to a room that tilts five degrees: everything looks the same but gravity pulls in a different direction, so old intuitions fail. LLMs are a squishy computer that does roughly what you meant rather than exactly what you said. A product that punches the user in the face 5% of the time is not viable even if it improves to 1%, so he advises taking them for granted as duct tape rather than expecting a flawless oracle.

For product builders, he says we are in the community-gardening phase, not factory farming, so curiosity and play matter most. He likes WebSim, where you type a URL and the model invents the site, because it could only exist in an LLM world. When output is cheap and mostly slop, he argues, taste wins: how much your perspective differs from what an LLM would write for the same prompt, while still resonating with others.

Lenny notes that most workplace tools help teams collaborate, whereas AI mostly improves individuals. Komoroske, citing his friend Ethan Mollick, says people may not tell their managers they can do their jobs twice as fast. Duct tape is easy to rig into anything but hard to scale, so usage sits in the long tail below the organization's awareness.

That, he says, explains the apparent paradox that few breakout successes exist beyond the labs while AI is used everywhere. He talks to Claude about 20 times a day, with projects loaded with context including his Bits and Bobs, and says he could not do his current job without such a conversation partner. He asks it for ten examples, critiques, or help layering in a concept such as contextual integrity.

He describes it as an extremely well-read but slightly naive friend who never makes you feel dumb, letting you explore questions you would not spend an expert's expensive time on. He stresses this is for exploring a domain and then checking with experts, not for legal advice. He calls it an electric bike for idea spaces, though he thinks the phrase probably originated with someone else.

Komoroske says he is hyper-optimistic, not cynical: describing how organizations really work lets you change them. Kayfabe, a wrestling term, means something everyone knows is fake but acts as real. In small doses it is useful optimism, the lubricant that lets teams attempt hard things.

The trouble comes as organizations grow. A project marked yellow gets reported up as green because the owner is confident of fixing it first, a reasonable self-defensive white lie, and the same happens at each layer until leadership is orders of magnitude from ground truth. Anyone who wants to press the "ground truth button" is told it would shatter everything, so they are pulled into complicity.

At the extreme, he says, people simply turn off the part of their brain that disagrees and earnestly believe the story, and the organization becomes a zombie. His remedy is to let disconfirming evidence arrive in small, survivable doses instead of one big moment. Lenny notes some leaders tell their reports to bring bad news immediately, and Komoroske adds that powerful people are easily surrounded by confirming evidence and that the root asymmetry is that you cannot make your boss look dumb.

Fighting entropy or organizational dynamics head-on is a losing game, Komoroske says, so he favors tactics that look playful: plant lots of small seeds that cost little, and water the ones that start growing. Don't try to predict beforehand which seed becomes the oak. Admitting you don't know the answer is fine if the experiment is cheap and the planter enjoys it.

He contrasts the builder mindset, where you make a plan and manipulate things to fit it, with gardening. A builder's result cannot exceed the effort put in, while a garden can compound, and done properly it looks like luck because you are farming for miracles. He looks for things with compounding loops, like ecosystems and network effects, where a tiny open-source tool can attract developers and users on its own.

The catch is that emergence makes you look unserious, and when something works people call it luck. Promotion to director at Google freed him, he says, because he no longer had to make his impact legible. His advice to PMs is to put about 70% of effort on work everyone agrees is valuable so nobody asks what the team does, and use the other 30% to let people, such as a junior PM with a quirky idea, plant seeds.

Lenny describes this as cover fire for a team, and Komoroske adds that the squirrel digging up the acorn is the main threat, so leaders need credibility to protect space. He tries to see seeds of greatness in everyone, naming a person's superpower early in mentoring so nudges feel like "be more" rather than "be different."

Komoroske's slime-mold deck, which Perplexity's co-founder cited to Lenny, argues that even with good, hardworking people, coordination cost rises with the square of headcount. Founders steer a small company like a sports car, but a big company is a big rig and driving it like a sports car damages the engine. In a big rig you should pivot less, invest in program management and leave slack in plans.

The alternative is a swarm of sports cars, which gives autonomy but visible incoherence. He says Apple chose the first approach and prizes coherence, while AWS has the second with many different ways of doing things yet a powerful, anti-fragile suite. Organizations that value employee autonomy are more like slime molds than they realize, and that can be a strength.

On how he finds metaphors, he says he is an external processor who tries frames in conversation and watches for an "aha." If people from very different backgrounds, say sales and engineering, both react, the idea likely has broad reach. A metaphor connects nine of ten dots and invites the listener to connect the last, which lets you say controversial things, and the slime-mold deck worked partly because it makes slime sound good.

Komoroske calls them nerd clubs and his secret weapon. He started the first at Google, facing a dozen groups working on one problem domain whose strategies undermined each other, and which he thought would only produce expensive pageantry in executive reviews. He created a secret club, deliberately named so only people who wanted to be there joined, and set clear norms: it is optional, collaborative, and "yes, and."

If someone says something you think is dumb you can ignore it, or reply that you wonder whether a different lens applies, which makes it about you. He says people build on the most interesting ideas, so convergence signals value. He adds one to three new perspectives per week while avoiding adding a jerk, in what he compares to novelty search, and says an idea lab like this cannot be steered towards an outcome.

To start one, he says begin with the smallest seed of people already engaged, keep adding likely fans, and make sure there is always conversation even when you are not there. He nudges people to share their ideas in the group, makes sure live sessions always reach quorum, and sends notes afterwards so absentees feel FOMO. His summary is that a community with nobody speaking is dead, and one with one person speaking doesn't yet realize it's dead.

Komoroske's productivity advice is to "play yourself like a fiddle." Every task has an activation energy and a burst when you finish, so he chains tiny tasks and gives a hard one a short window, since too much space makes it harder. He also prefers always rules to sometimes rules for self-control, and has done a Peloton workout every day since the pandemic began, checking the box with a meditation on commute days.

He types very fast, a skill he credits to transcribing 150 hours of interviews for his undergraduate thesis on Wikipedia's editor community. That lets him capture ideas as they arrive, like a butterfly collector, and he writes down a rough 30-minute draft whenever inspiration hits. His life philosophy is to do things that give you energy and that you are proud of, judged from ten years on in front of people you respect.

He also describes the Hallmark card fallacy: when you finally absorb a cliché such as "the friends we made along the way," it feels profound, but others dismiss it as trite. You cannot force people to be ready, so he writes seeds that may open later. He cites mentees who return years later to thank him for advice they resented at the time.

Lenny and Komoroske link this to awe and transcendence, which he says keep you open to disconfirming evidence. Ego wants to avoid it, but he stresses that disconfirming evidence makes both systems and people stronger. He accepts that some people see him as kooky, as long as the indirect impact of his work shows it functions.

The adjacent possible, Komoroske says, is the set of actions within arm's reach that will almost certainly work. Tech tends to assume a flying leap to the end state, but because the adjacent possible is small you have full agency inside it, and each action reconfigures the world and opens the next set. Slicing decisions into small, safe steps removes risk but keeps exposure to upside, and you stop stepping down a path that does not pay off.

You need both, he says, because incremental-only follows the steepest gradient in front of you and ends in a random walk. So set a low-resolution North Star three to five years out that everyone with relevant knowledge agrees is plausible but would still be exciting. Then pick the step in your adjacent possible with the steepest gradient toward it, and repeat.

Lenny compares this to a meditation teacher's advice to point your cart in a direction and start walking. Komoroske agrees the path will be slightly inefficient, and warns against false precision, such as arguing whether a number will be 93 or 95 in five years. For compounding ideas the outcome is more like zero or a thousand, so the difference between 1,000 and 1,001 does not matter.

He closes with Tim Urban's framing that happiness is reality minus expectations, so hold expectations lightly because they are easier to change than reality. A friend's trick is to feel now the emotion a story will give you in ten years. He stresses this is not settling for mediocrity but seeking something you can be proud of while staying open to being convinced otherwise.

In the lightning round Komoroske recommends Origin of Wealth by Eric Beinhocker, on complexity economics, and a short book on thinking in systems with the line "dance with the system." His favorite recent movie is The Green Knight, which he disliked at first but could not stop thinking about, and he likes parables for the same reason. His favorite product is WebSim, and he uses Claude about 20 times a day.

He describes LLMs as little holograms of humanity, a mirror that reflects different things back at you, and has experimented with generating fake data that matches a fictional person's backstory. Lenny calls it a zip file of all human knowledge. His two life mottos are doing things that give you energy and that you are proud of, and happiness as reality minus expectations.

Picking a random note from his Compendium, he reads a reflection that aggregators make sense in the late stage of an era but curtail exploration early on. He objects that the canonical piece of software for most people is an app: monolithic, not composable and allowed to exist only if large companies permit it. He rejects a future where everyone is locked in a box with a super-God AI Clippy and hopes LLMs, as magical duct tape, let people escape the box and use software to exercise their own agency.