冲向痛处：Ben Horowitz 谈犹豫、信心，以及把公司带出深渊

Asked about a lesson from a pilot that his friend Shaka Senghor suggested, Horowitz recounts asking a pilot what really happened in the JFK Jr. crash. The pilot told him that all plane crashes are a series of bad decisions, none terrible alone, that add up to one terrible outcome. In this case the crash involved something like 17 bad decisions in a row, starting with treating arrival as the priority.

Horowitz draws the mirror lesson: one decision leads to another, so if you can let go of sunk cost you escape many bad paths. Success, he says, is a small hard thing that doesn't look high-impact but leads to the next small hard thing. The press prefers a neat narrative about a smart move and a happy ending, but it never actually works that way.

He describes how Senghor spent seven years in solitary confinement yet changed deep beliefs about himself there. Senghor's conclusion, which Horowitz found striking, was that the loss of rights after 19 years in prison was nothing compared with what he did to himself. Horowitz applies it to CEOs: systems, bias or people undercutting you are small compared with what you do if you believe their story about you.

Horowitz says the destructive move for a CEO is hesitation, and it happens because both choices are horrible. Unlike a business school case study, real decisions look like rearchitecting a product and missing a quarter versus not rearchitecting and knowing the architecture is wrong. Avoiding the subject, he says, makes the whole company nervous because staff already know it needs fixing.

His own example was taking Loudcloud public in March 2001 with $2 million in trailing twelve-month revenue at 18 months old. It was obviously a bad idea, he says, but the alternative given the private markets was bankruptcy, and many CEOs who hesitated went bankrupt. He knew the press would pillory him, and it did, including a Businessweek story called "The IPO From Hell."

The psychological muscle of a great leader, he says, is to look into the abyss and say "that way is slightly better." The same applies to firing the head of sales: you can compute every bad consequence of acting, but not acting is usually worse.

He can't coach founders into being decisive, he says, but he can confirm that what they see is real and help with the conversations they fear. Using a football analogy, an athlete who doesn't trust their eyes is a step slower. In one example, a CEO wanted to address a CTO who had made a junior finance colleague cry; Horowitz suggested framing it as specific effectiveness across the organization, not "you're a jerk."

Horowitz says the goal is to be liked and respected in the long run, not the short run. He pushes his own firm to be respectful rather than merely "entrepreneur friendly," because a board member must say what founders don't want to hear. If everyone agrees with a decision, he says, you added no value; leadership is making the decision most people don't like.

On whether to found a company, he recalls Citigroup CEO John Reed telling him the only reason is an irrational desire, because it isn't worth the money. When Loudcloud sold for $1.6 billion, Horowitz says it still didn't feel worth it. Doing it for money is a bad reason; you need a drive to build something larger than yourself or you won't survive the bad things.

He notes that a product that forces you to build a company is a good start, and that Hewlett-Packard started more abstractly, asking what the next engineer on the bench needed. Whatever the path, there must be some purpose beyond making money.

On Databricks, he says six people including professor Ion Stoica (a name he gives phonetically) asked for $200,000. He knew Spark competed with well-funded Hadoop companies and worried academics think too small, so he offered $10 million and told them to build a real company or stay in school. He later made Ali Ghodsi CEO, which he calls good luck.

Horowitz recalls telling Ali Ghodsi, newly CEO, that you don't make people great; you find people who make you great. A VP of engineering can develop engineers, but a CEO knows little about being CFO or head of HR. The company also can't afford a CEO who spends time on that instead of fast, high-quality decisions and direction.

Managerial leverage, he says, is simple: if you are the one with ideas for what a department should do and you push the leader forward, there is no leverage. Leverage is when the executive tells you what to do and how to move the company. When you find yourself asking why aren't we doing this, it is time to make a change, and he says Ghodsi does this without hesitating.

On when founders should be replaced, he says it starts with confidence. Expensive early mistakes erode confidence, which leads to hesitation, which makes senior people jump into the void and turns the company political. Most of a16z's work, he says, is helping founders climb the confidence and competency curves together, and sometimes it helps to be used to failing, since the median CEO test score is more like 18 than 90.

Horowitz says organizational design, goals and OKRs need only an eighth-grade education. The hard part is emotional: a reorg redistributes power, so someone good loses it and is angry. He says this is why his first book normalized how painful the job is, and that the usual narrative of a genius idea and smart hires hides constant, costly mistakes at every famous company.

He notes Airbnb's Brian Chesky learned this the hard way, including deferring to senior hires, a discussion that fed into the "founder mode" talk. Horowitz says you can't defer as CEO because you know what the company should do even if the finance head knows finance better. He adds he too puts off decisions when things are going well.

The design of a16z, he says, is partly about confidence: a 600-person firm builds networks so a founder can call any CEO or official on day one. The "CEO barbecue" at his house brought portfolio CEOs together with figures like Larry Page and Mark Zuckerberg so that they would go home feeling like they belong. Founders can build the next product, he says, which almost no professional CEO manages.

Asked for counterintuitive lessons, Horowitz says common wisdom keeps flipping. The old advice to hire a team of senior executives quickly after product-market fit is wrong, he says, because you start deferring and executives build empires and politics. You must hire at a pace you can integrate and manage.

The reaction, he says, has swung to never hiring experienced people, which is also bad. He recalls telling a Databricks cofounder with a computer science PhD that the company needed someone who knows sales, because territories, rep profiles and organization are learned by trial and error. He is annoyed by podcast snippets of advice and says you have to understand the next several layers of each question.

He tells a story: after he told a CEO "don't be CEO at home," meaning don't boss your family around, another VC at dinner interpreted it as "unplug sometimes." For him it shows how advice gets distorted when passed on by people who haven't done the job, and how dangerous that is.

Horowitz explains why he wrote "Good Product Manager, Bad Product Manager": PM roles differ everywhere and have no training, so people get fixated on tasks like writing specs or handling press. The job, he says, is fundamentally leadership with nobody reporting to you, which is the essence of real leadership because you can't hire, fire or promote. The goal is to get a product to market that customers love and that beats everything else.

He says the document began as an internal memo written out of frustration at Netscape, before blogging, and that many task details are dated while the mindset isn't. A big takeaway was that if you find yourself yelling at people, you probably haven't explained what you want. He adds that what he writes in five minutes often ends up better than what he writes in five weeks.

Responding to the idea that a PM isn't a mini-CEO, he says what matters is that the product works, and you work backward from there. A PM or CEO isn't the person with every idea; they collect and prioritize good ideas and get everyone aligned with high fidelity, and that isn't about being better than anyone.

On Adam Neumann, Horowitz calls him probably the most controversial investment a16z ever made, which drew accusations from stupid to sexist to racist, yet he expects it to be among the best. The principle, which he learned partly from Senghor, is not to judge a person by the worst thing that happened to them. He argues the failures at WeWork came from inexperience and a lack of anyone willing to tell Neumann the truth.

Quoting a line attributed to Al Davis, he says to coach players on what they can do. Venture capital, he says, is investing in people because the initial idea usually changes, so a16z looks for world-class strengths and then surrounds founders with people who can handle their flaws. He says Marc Andreessen is on the board and Horowitz attends board meetings because he's good at challenging a very confident founder.

He extends this logic to Cluely, saying the founders showed marketing genius and the product is good.

On Sam Altman calling AI a bubble, Horowitz notes CEOs talk with purpose while investors just need to be entertained, and that the comment is clever: it tells investors to back him alone and gives the press red meat. His own view is that when everyone thinks it's a bubble, it isn't, because a bubble needs capitulation. He recalls debating this in The Economist around 2011 or 2012.

He contrasts AI with the dot-com era, when Netscape had 90% browser share but only about 55 million people were online and building a product like Evite took 300 engineers, so unit economics didn't work. Today, he says, revenue growth is unlike anything seen, with companies going from zero to $800 million in a year. He accepts the field is early and positions may not be sustainable, but if forced to bet he says not a bubble.

Looking ahead, he is bullish on infrastructure, such as running a given open-source model cheapest with lowest latency. For foundation models, a16z's rule of thumb is that founders must be able to raise at least $2 billion without much product progress, and he says fewer than ten people in the world qualify. He thinks the application layer will be both bigger and stickier than expected.

Horowitz says the "thin wrapper" complaint echoes the 1980s, when companies like Salesforce were dismissed as thin wrappers around a relational database. He says Cursor has built 14 different models to understand how a real developer works, using many interactions. He expects a very small number of foundation-model companies and an almost unlimited number of applications.

On moats, he says the universe is fat-tailed and human behavior especially so, so getting to the essence of any domain is more complex than people expected. He sees LLMs as having asymptoted as data ran out, noting GPT-5 versus GPT-4.1 isn't scaling linearly, while reinforcement learning scales linearly but doesn't generalize, so a great programming model may be bad at math. That opens room for domain-specific products.

In the enterprise, he says, it is company data rather than internet data, so access control and semantics matter. Ten enterprises will define "customer" ten ways, which affects things like churn analysis. He adds that Waymo's hardest problems weren't sleet or rain but rare, important human behavior, such as a driver doing 75 in a 25 zone.

Asked why US leadership in AI matters, Horowitz says it starts with his view of the US and its role. He argues for a decentralized balance of power, and says the lesson of communism and fascism is that concentrating power in the state gives people no incentive and only the stick, so the system itself is the problem rather than individual dictators.

In his reading, the US system places rules above any president through the Declaration of Independence's "self-evident" truths, and he criticizes European speech restrictions justified as protecting children. He points to company creation and new ideas as evidence the US system works despite its flaws.

He notes that the countries that industrialized first and best held the power in the last century, and that AI will play that role next. The US need not be the sole winner, he says, but must be in that tier, and people abroad tell him not to over-regulate too early. He adds that crypto is an important networking technology that complements AI and says a16z's policy work is more important than its financial interests.

Horowitz describes the Paid in Full Foundation, named after the Eric B. & Rakim song, which gives pensions to early hip-hop pioneers and holds an awards event. He says the firm's ethos is "something from nothing," like both entrepreneurship and hip-hop, and that the inventors of an art form rarely get proportional benefit. Awardees include Rakim, Scarface and Roxanne Shante, and he says the impact has been large, with Rakim releasing his first album in 15 years.

His closing message to CEOs is that how you feel about yourself matters as much as anything, and that no credit is given for predicting rain, only for building an ark. Predicting you'll fail gets you nothing, so figure your way out and spend all your time on that.

In the lightning round, he recommends *The WEIRDest People in the World* and Shaka Senghor's books *Writing My Wrongs* and the forthcoming *How to Be Free*. He shares how Senghor built trust in a prison gang, such as making everyone eat lunch together and requiring honesty both inside and outside the group. His motto, from his father, is "Life isn't fair," and the useful question is what to do now.