Unlike the early web or telecommunications, which required massive infrastructure scaling before delivering value, AI represents a probabilistic computer deployed on top of a mature global network, enabling instant distribution to hundreds of millions of users. (2:30)
Dismissals of AI applications as superficial GPT wrappers mirror early critiques of relational database software; top products like Cursor layer dozens of specialized models and capture proprietary domain data that compounds switching costs. (10:00)
The traditional software application may ultimately dissolve into raw centralized data layers operated on directly by autonomous AI agents, using crypto networks as native financial rails. (5:00)
Venture capital backing remains vital for early-stage startups primarily to manufacture initial momentum in path-dependent, winner-take-all markets and to bridge the executive skill deficit of first-time technical founders. (15:00)
Detailed notes
## Background
Published on October 24, 2026, as an audio podcast by Ben Horowitz and Andrew Chen on RSS and web platforms, this episode explores what it takes to build enduring companies as artificial intelligence reshapes technology. The description highlights discussions around whether tiny teams can build billion-dollar businesses, how software applications might dissolve into autonomous agents, and why the hardest parts of entrepreneurship—navigating uncertainty and overcoming self-doubt—remain fundamentally unchanged.
## TL;DR
Ben Horowitz and Andrew Chen examine the structural differences between AI as a probabilistic computer and previous network-driven tech waves, predicting that standalone software applications will increasingly give way to autonomous agents acting directly on raw data. Horowitz addresses defensibility in AI, emphasizing that competitive moats come from complex technical layering and customer inertia rather than simple baseline models. Drawing on his tenure leading Loudcloud through the dot-com crash, Horowitz argues that great venture-backed founders are defined not by an absence of weaknesses, but by possessing an exceptional, asymmetric superpower and the psychological stomach to execute least-bad decisions amid existential crises.
## Atmosphere & style
The dialogue has the candid, battle-hardened atmosphere of a seasoned wartime CEO delivering practical mentorship without corporate polish or academic abstraction. Horowitz speaks with blunt, street-smart pragmatism—freely using colorful language and self-deprecating war stories to demystify the raw emotional panic of running a company. Chen serves as an incisive, grounded interlocutor, steering the exchange from sweeping macroeconomic paradigms into an empathetic, intense examination of founder psychology, imposter syndrome, and decision-making under fire.
## Bulletin
- `02:30` Unlike the early web or telecommunications, which required massive infrastructure scaling before delivering value, AI represents a probabilistic computer deployed on top of a mature global network, enabling instant distribution to hundreds of millions of users.
- `10:00` Dismissals of AI applications as superficial GPT wrappers mirror early critiques of relational database software; top products like Cursor layer dozens of specialized models and capture proprietary domain data that compounds switching costs.
- `05:00` The traditional software application may ultimately dissolve into raw centralized data layers operated on directly by autonomous AI agents, using crypto networks as native financial rails.
- `15:00` Venture capital backing remains vital for early-stage startups primarily to manufacture initial momentum in path-dependent, winner-take-all markets and to bridge the executive skill deficit of first-time technical founders.
## Key takeaways
- `00:00` Entrepreneurs are defined not by career planning, but by an unavoidable, compulsive drive to bring an idea into the world.
- `02:30` AI functions as a probabilistic computer rather than a network, bypassing the adoption lulls and infrastructure constraints that plagued the 1990s dot-com era.
- `05:00` While AI enables tiny teams to generate enormous revenue, ever-expanding human desires historically force organizations to scale up headcount and operational complexity.
- `07:30` Software may shift away from discrete standalone applications toward direct execution by autonomous AI agents querying centralized data.
- `10:00` AI application defensibility relies on layering proprietary domain-specific models and accumulating compounding user data rather than relying purely on foundational models.
- `11:00` Displacing an established software incumbent requires an order-of-magnitude 10x improvement, as customer inertia is nine-tenths of competitive advantage.
- `11:47` Decentralized crypto networks provide the programmatic financial rails required for autonomous agents that lack legal identity to access traditional banking systems.
- `15:00` Venture firms add value primarily through path-dependent momentum, using institutional signaling to attract engineering talent, customers, and follow-on capital.
- `17:45` First-time technical founders face an acute CEO skill deficit, requiring hands-on operational coaching to evaluate executive hires and manage governance.
- `20:00` CEO failure is driven by crises of confidence and operational hesitation rather than a lack of raw intelligence.
- `22:38` Andreessen Horowitz evolved from providing functional agency services into building specialized industry verticals as software expanded from 15 to 150 winners per year.
- `26:50` Generative AI entertainment represents a novel artistic medium requiring native storytelling paradigms, distinct from translating film techniques or social media protagonist dynamics.
- `30:00` Navigating existential corporate crises requires choosing and executing the least-bad available option immediately, rather than freezing over painful fallout.
- `32:50` Effective executive mentorship relies on pragmatic, granular behavioral scripts rather than rigid generic management platitudes.
- `37:04` Breakout technologies and great founders are defined by asymmetric, world-class superpowers in a critical dimension rather than a well-rounded lack of weaknesses.
- `39:46` Impostor syndrome and intense fear are universal byproducts of taking entrepreneurial risk, requiring founders to focus strictly on execution rather than emotion.
## Detailed walkthrough
### `00:00` What Makes an Entrepreneur and the Early Days
Andrew Chen introduces the conversation by reflecting on the core struggles of company building. AI may make creating companies technically easier, but it doesn't change the emotional toll, the uncertainty, or the sheer grit required to succeed as a founder. Andrew brings up his very first meeting with Ben Horowitz years ago, visiting the offices of Loudcloud turned Opsware. Andrew vividly remembers that every single conference room was named after a hip-hop artist, signaling right away that Ben was a unique individual running an uncommon company culture. At that time, Opsware was on the verge of being sold to Hewlett-Packard (HP), and Andrew asked Ben what he planned to do next. Ben told him quite plainly: he was going to retire because he was exhausted.
```mermaid
flowchart LR
accTitle: Ben Horowitz's Career Trajectory
accDescr: Progression from Opsware to intended retirement to co-founding Andreessen Horowitz.
A["Loudcloud / Opsware<br/>(Sold to HP)"] --> B["Intended Retirement<br/>('I was tired')"]
B --> C["New Idea Strikes<br/>('Must get it out')"]
C --> D["Andreessen Horowitz<br/>(A16Z)"]
```
Of course, Ben did the complete opposite of retiring. Asked why he pivoted away from rest, Ben explains that he simply had another idea, and life fundamentally centers on getting your ideas out into the world. In Ben's view, that compulsive drive is the absolute definition of an entrepreneur: someone who believes in an idea so intensely that pursuing it feels unavoidable.
> “You're not a great entrepreneur unless you feel like you don't have a choice but to build a company because otherwise you're going to quit anyway.” — Ben Horowitz
- Opsware conference rooms were famously named after hip-hop artists.
- Opsware was sold to HP right around the time Andrew and Ben first met.
- Ben initially intended to step down and retire simply because he was completely worn down.
- That decision was reversed because an entrepreneur fundamentally cannot resist pursuing a compelling new idea.
---
`02:30` AI as the Probabilistic Computer vs. Networks
Andrew asks Ben to contrast the current AI wave with previous tech cycles he lived through, including desktop software, the web, mobile, and cloud computing. Ben notes that the current excitement mirrors the earliest days of the internet: every single morning, you check what’s happening and find some mind-blowing, awesome new development. Beyond that daily feeling of wonder, however, the underlying mechanics of AI are structurally different.
Throughout tech history, technological evolution has swung between computers and networks—microwaves, integrated circuits, mainframes, SNA, personal computers, local area networks (LANs), smartphones, and the internet. These two categories have radically different adoption curves, even as they reinforce each other. In this paradigm, AI represents the computer, not the network. Specifically, it is a probabilistic computer, in contrast to the deterministic computers of the past.
Networks require scale before they deliver value, making them vulnerable to premature hype cycles. Ben revisits the great dot-com bust that triggered the telecom collapse: in 1996, Netscape held a staggering 90% browser market share, yet that represented only 55 million total users across the entire web. Half of them were tethered to slow dial-up connections. Deploying an internet application back then was extraordinarily expensive—demanding Sun Microsystems servers, EMC storage arrays, and costly infrastructure—yet the addressable market topped out around 30 million people. Because the network had not yet matured, the financial model collapsed under its own weight before eventually rebounding. AI behaves in reverse: because the global internet infrastructure is already fully built out, a working AI product can immediately distribute to hundreds of millions of users overnight.
> “I think in this era, AI is the new computer, like very clearly.” — Ben Horowitz
> “We've seen companies grow faster than any companies we've ever seen because you don't need more people on the network.” — Ben Horowitz
- In 1996, Netscape held roughly 90% browser market share, serving approximately 55 million users worldwide.
- Roughly half of internet users during that period connected via dial-up connections.
- Early web startups spent heavily on Sun Microsystems hardware and EMC storage despite facing a tiny total addressable user base.
- Networks suffer adoption lulls until scale is reached, whereas AI builds directly on top of ubiquitous global connectivity.
---
`05:00` Tiny Teams, Expanding Human Desires, and the Disappearance of Applications
Andrew highlights how shockingly capable Large Language Models (LLMs) are at writing code, raising the possibility that software creation could become as effortless as publishing web content. This raises questions: will we see solopreneur-run, billion-dollar companies, or hyper-lean teams taking over industries? Ben agrees that a tiny team can achieve staggering reach and generate massive sums of money—pointing out that companies like Instagram might well have reached billion-dollar status with very few staff had they not been acquired by Meta.
However, Ben questions whether such lean operations can endure without expanding. Historically, consumer standards continuously ratchet up; when people receive more, they demand even more. Ben cites economist John Maynard Keynes, who famously predicted in the 1930s that once humanity's basic physical needs—food, shelter, and clothing—were met, people would only need to work 15 hours a week. Keynes proved spectacularly wrong because human appetite expanded into brand-new categories: suddenly households needed a car, then a car for every family member, then a television, then a television in every room, air conditioning, and smartphones. Satisfying ever-expanding customer demands almost always forces an organization to scale.
When Andrew asks about trends like "fast fashion" SaaS clones or bespoke code generated dynamically at runtime, Ben shifts focus to a deeper architectural shift: the fading relevance of the standalone "application." Just as human fashion trends still converge despite endless options, software might move away from discrete, specialized applications entirely. In an agent-driven world, people and enterprises will simply organize their raw data and set autonomous AI agents loose to execute tasks directly against it.
> “I think what people need does tend to expand almost infinitely. So to fulfill that need, sometimes you have to build a bigger company, but we'll see.” — Ben Horowitz
> “My question on kind of software is, like how much will be like what we consider software today and how much will just be like, okay, we'll just use the AI differently and it won't even be an application.” — Ben Horowitz
> “The GPT wrapper thing was always a bit of a dumb idea in terms of not to build one, but to call it that.” — Ben Horowitz
- Andrew and Ben debate whether AI enables enduring billion-dollar solopreneurs or if scaling headcounts remains necessary.
- John Maynard Keynes predicted in the 1930s that basic necessity fulfillment would drop workweeks to 15 hours.
- Human demand continually expands to accommodate new technologies (cars, multi-room TVs, smartphones, air conditioning).
- The classic concept of a software application may be replaced by centralized data layers paired with direct AI agent execution.
- Ben dismisses the common critique of "GPT wrappers" by comparing it to early dismissals of database applications when declarative SQL first arrived.
`10:00` AI App Defensibility, Proprietary Data, and Switching Costs
Ben Horowitz dismisses the popular critique that applications built on top of large language models are merely superficial wrappers with no defensible value. He compares this skepticism to the early days of enterprise software, when detractors claimed new software categories had nothing to them because they were just “an RDBMS wrapper.” In reality, modern AI applications are layering deep technical complexity on top of base foundation models. Horowitz points to Cursor as a prime example, noting that it runs roughly 14 specialized, sophisticated models layered above the underlying foundational model. When a product interacts continuously with high-end software engineers solving complex programming and design problems, it absorbs an immense volume of unique domain data that no competitor possesses, creating genuine architectural and data differentiation.
Beyond technical depth, Horowitz stresses that customer inertia and distribution constitute nine-tenths of the law in software competition. Comparing leading tools today, he observes that the gap between products like ChatGPT and Grok is relatively marginal for someone starting from scratch, and it does not yet exhibit an inherent network effect. However, once a user adopts an interface and incorporates it into their daily habits, they are exceptionally unlikely to switch unless presented with an order-of-magnitude leap in capability. As Horowitz puts it: “So like you don't take out a competitor with an equally good product or a slightly better; it's like a 10 times better product.” Once an AI company seizes the market, that capture compounds because the winner aggregates deeper, more specific behavioral data about its user base.
- Dismissals of AI apps mirror early dismissals of relational database software as simple RDBMS wrappers.
- Cursor utilizes approximately 14 specialized models on top of foundational AI models to power its developer workflows.
- Continuous interaction with elite programmers funnels unique technical context into Cursor's proprietary model layer.
- Displacement requires a 10x superior experience because routine usage creates high psychological and operational inertia.
---
`11:47` Technological Phases: From Web Servers to Agentic Crypto Economies
Andrew Chen frames the current AI market as a classic initial land grab, where the single most important product attribute is simply the “it works” feature, while unique user experience paradigms and collaborative network effects emerge later. Horowitz agrees, explaining that when a new computing platform solves previously unsolvable problems, it creates immense value before traditional network effects take hold. Looking forward, Horowitz suggests that the rise of crypto may become the financial rails for autonomous agents. Because artificial intelligences cannot legally apply for credit cards or open conventional bank accounts, a decentralized crypto network offers a programmatic medium for value exchange, enabling an agent-to-agent economy with its own native financial network effects.
Chen and Horowitz trace how founder skill profiles shift across technology cycles. In the earliest days of the consumer internet, building something as conceptually simple as the Yahoo directory required Stanford PhDs because there were no commercial application servers, off-the-shelf load balancers, or pre-existing infrastructure designed to scale software to 500 million people—previously, the largest systems were built merely for the size of IBM’s internal workforce. Over the course of a decade, the emergence of the LAMP stack made programming accessible to non-computer-science majors, ultimately paving the way for design- and UX-led founders to build companies like Airbnb. Horowitz points out that while pure language model breakthroughs have plateaued recently, reinforcement learning is scaling rapidly and driving massive advancements in code generation. Every subsequent foundational breakthrough will trigger an architectural realignment across both the application and infrastructure layers.
```mermaid
flowchart TD
accTitle: Evolution of Tech Shifts and Founder Archetypes
accDescr: Diagram showing the progression from the early internet infrastructure phase to modern AI agent workflows.
subgraph EarlyWeb["Early Web Era"]
A["Stanford PhDs & Systems Engineers"] -->|"Custom servers & load balancers"| B["Core Web Infrastructure (e.g., Yahoo)"]
end
subgraph MatureWeb["Mature Web & Cloud"]
C["LAMP Stack & Cloud (AWS)"] -->|"Lowers barrier to entry"| D["UX & Design-Led Founders (e.g., Airbnb)"]
end
- Early internet scale was unprecedented; enterprise software had previously only been engineered to support internal corporate headcounts like IBM's.
- The progression from low-level systems engineering to the LAMP stack allowed product- and UX-focused founders to dominate consumer web waves.
- Traditional LLM capabilities are experiencing a plateau, whereas reinforcement learning techniques are rapidly advancing coding capabilities.
- AI agents lack legal identity to access traditional banking rails, making crypto a viable mechanism for autonomous agent-to-agent transactions.
---
`15:00` Venture Capital Evolution and the Dynamics of Path-Dependent Momentum
Horowitz reflects on how venture capital operated virtually unchanged from 1970 until Andreessen Horowitz was founded in 2009. Historically, building a technology company required massive capital outlays to assemble direct enterprise sales forces and internal professional services organizations for companies like Sun Microsystems or Tandem Computers. Investments were expensive, failure rates were high, and venture firms were structurally designed not to evolve. The advent of modern internet infrastructure, cloud platforms like AWS, and developer tools completely transformed corporate formation, culminating in companies like Twitter launching with just five people on laptops. As AI accelerates this trajectory, it will force yet another evolution in venture investing.
Despite teams requiring less capital to launch, Horowitz argues that the core function of venture capital remains rooted in navigating complex adaptive, path-dependent systems. When two rival companies start with the identical idea and equally talented teams, the winner is determined by who aggregates momentum first. Securing backing from a top-tier venture firm acts as an institutional signal that attracts premier engineering talent, which in turn leads to a superior product, more customer acquisitions, increased revenue, and faster follow-on financing. As Horowitz explains: “And so a big role of a VC is to kind of help you get that momentum.” The underlying operational features of companies will shift, but the imperative for venture firms to manufacture momentum and assist in institutional company-building remains constant.
```mermaid
flowchart LR
accTitle: Path-Dependent Startup Momentum Flywheel
accDescr: Flowchart illustrating how top-tier venture signaling catalyzes talent, product quality, customers, and follow-on capital.
A["Top-Tier VC Backing"] -->|"Market signaling"| B["Elite Engineering Talent"]
B -->|"Velocity & polish"| C["Superior Product"]
C -->|"Adoption & retention"| D["Customer Growth & Revenue"]
D -->|"Accelerated rounds"| E["Follow-on Capital"]
E -->|"Reinvest in lead"| A
```
- The institutional VC operational model remained essentially static from 1970 through the late 2000s.
- Early enterprise giants (Sun Microsystems, Tandem) required heavy overhead, direct sales organizations, and professional services arms.
- The shift to AWS enabled five-person teams on laptops to build massive scale consumer platforms like Twitter.
- Early startup competition is path-dependent; initial signaling and momentum dictate which team captures market share.
---
`17:45` Startup Stakes and the CEO Skill Deficit at Seed Stage
Addressing why wildly profitable, bootstrapped, or high-performing startups still actively pursue venture capital, Horowitz points out the stark asymmetry of entrepreneurial outcomes. Building a meaningful company that improves the world alongside a chosen team is an unmatched professional high, but the downside is brutal. When a venture fails, founders face profound personal misery after convincing their smartest peers to leave stable jobs and losing money invested by close family. Speaking directly to the emotional toll of failure, Horowitz remarks: “You hired all the smartest people you know. You went out and raised money. Your mom invested in the company. She's pissed. It sucks. It really does.” To mitigate that risk, founders seek every available advantage to secure market leadership.
Looking ahead to the upcoming Speedrun Demo Day in October—where founders pitch to roughly 1,000 investors in San Francisco—Chen asks how early-stage teams should evaluate prospective seed investors. Horowitz emphasizes that first-time founders almost never understand how to execute the CEO role. While technical founders excel at product, they routinely lack the specialized operational frameworks required to manage an executive team or establish corporate governance. As a founder, you are forced to evaluate and hire leaders for roles you have never performed yourself—such as hiring a Chief Financial Officer when you have no prior experience with internal control structures or interacting with Wall Street. A high-value seed investor must serve as an operational guide who can help a first-time founder bridge this acute executive skill deficit.
- Speedrun Demo Day features companies presenting to approximately 1,000 angel investors and seed funds in San Francisco.
- The emotional misery of startup failure is magnified by recruiting close friends and taking capital from personal connections.
- First-time technical founders face an immediate handicap because they have never operated as enterprise CEOs.
- Critical CEO responsibilities include evaluating and managing executive hires in foreign domains, such as CFOs handling corporate control structures and public market communication.
`20:00` The Founder’s Crisis of Confidence and Useless VC Advice
Founders face an unrelenting barrage of management challenges where they lack the expertise to evaluate what they are seeing, much like trying to judge a Japanese interpreter when you do not speak a word of Japanese yourself—everyone sounds convincing on the surface. Adding to that, employees have personal career ambitions, and deep-pocketed competitors like Mark Zuckerberg might dangle billion-dollar packages in front of key talent. Ben Horowitz notes that when founders fail at the CEO job, it almost never stems from a lack of raw intelligence; it comes down to a crisis of confidence. A founder knows they do not know what they are doing, knows their lack of knowledge is hurting the company, yet must make critical calls anyway. Hesitation is fatal: it makes the entire organization anxious, grinds execution to a halt, and breeds toxic internal politics as others step into the power vacuum.
Horowitz emphasizes that an investor’s greatest value is helping a founder navigate that uncertainty to reach high-quality, timely decisions, while providing tangible momentum in hiring and fundraising. Unfortunately, the market is saturated with poor advice from people who have never run a company. Horowitz laments non-operator investors and angel investors who speak with massive unwarranted certainty, such as ordering an early-stage founder to instantly recruit an entire executive team—a move that will blow up an early-stage startup. He mocks the superficial platitudes frequently handed down to founders, such as being told the big secret is to hire high performers. As Horowitz puts it:
> “My plan was to hire a bunch of fucking morons, but now I'm going to hire A players.”
Andrew Chen points out that this dynamic highlights one of the core historical strengths of the Silicon Valley ecosystem over the past two decades. While venture capital in Europe and elsewhere historically drew from private equity and finance professionals who dabbled in venture, the Bay Area pioneered a model dominated by former operators who have actually built and managed companies.
- **The trap of executive hesitation**: When a founder vacillates because they lack certainty, the company interprets hesitation as weakness, slowing down velocity and opening the door for political infighting.
- **Operator-led investing**: Silicon Valley fundamentally differentiated itself from finance-heavy European VC ecosystems by staffing venture firms with former tech founders and executives who understand the operational reality of company-building.
---
`22:38` The Evolution of Andreessen Horowitz: From Network-as-a-Service to Verticalization
Reflecting on how Andreessen Horowitz (a16z) has transformed over time, Horowitz recalls that the traditional venture capital model was designed to serve the venture capitalists rather than the entrepreneurs. The old standard was assigning a single clever investor to a startup's board. While that investor had their own personal rolodex and could offer helpful product advice for the first three months, that advice had a short shelf life. Beyond the initial product definition, the knowledge gap between a founder working on the problem day and night and a board member visiting once a month or quarter becomes an unbridgeable gulf. Looking back at his own experience as a founder, Horowitz realized that what founders truly lacked was operational confidence and institutional support to feel like legitimate CEOs.
To solve this, a16z built dedicated functional agency teams inside the firm. If a founder needed enterprise customers, the firm built a briefing network connecting them directly to senior purchasing executives at corporations like FedEx or Citibank. If a founder needed to hire a Chief Financial Officer but had never interviewed an executive, a16z arranged sit-downs with five top-tier CFOs so the founder could learn what world-class execution looked like, what competencies to test for, and how to structure the interview process before making a hiring decision.
```mermaid
flowchart TD
accTitle: Evolution of the a16z Venture Firm Model
accDescr: Chart showing the three eras of a16z from the agency model to market verticalization and late-stage multinational scaling.
subgraph Era1["Phase 1: Operational Firm Model"]
A["Traditional VC: Lone Board Member"] --> B["a16z Agency Model"]
B --> B1["Executive Briefings: FedEx, Citibank"]
B --> B2["Talent Coaching: Learn how to hire CFOs"]
end
subgraph Era2["Phase 2: Vertical Expansion"]
C["Marc Andreessen (2011):<br/>'Software is Eating the World'"] --> D["Market Expands: 15 to 150 winners/year"]
D --> E["Dedicated Vertical Practices"]
E --> E1["Infrastructure & Applications"]
E --> E2["Bio, Crypto, Games, Defense"]
end
subgraph Era3["Phase 3: Scaling Mega-Private Companies"]
F["Startups stay private longer"] --> G["International Expansion: e.g. Japan"]
F --> H["Multi-Product & Multi-Channel Scale"]
end
Era1 --> Era2
Era2 --> Era3
```
The firm’s second major structural evolution followed Marc Andreessen’s 2011 thesis that "software is eating the world." Historical venture wisdom dictated that only about 15 venture-backed companies a year would ever reach $100 million in revenue, meaning a firm's sole strategy was squeezing into as many of those 15 as possible. But if software was revolutionizing every sector of the global economy, that pool was expanding from 15 to 150 companies annually. To capture this without diluting partner focus, a16z verticalized, creating specialized investment practices in infrastructure, enterprise applications, bio, crypto, gaming, and defense and public safety. Now, Horowitz explains, the firm is adapting to an era where startups remain private much longer and scale to massive valuations, requiring firm infrastructure that can orchestrate international expansion—such as landing a company’s first three commercial contracts in Japan—and assist with complex multi-product and multi-channel distribution.
---
`26:50` AI as a New Medium and the Psychology of Founding
Transitioning to audience questions, Horowitz outlines why generative AI video and entertainment should not be evaluated as mere software automation for existing film and television pipelines, but as an entirely novel artistic medium. Drawing a historical parallel to the birth of cinema, he observes that early filmmakers simply planted a stationary camera in front of a stage play and recorded the performance. It took decades of experimentation for cinema to discover its own distinct visual language, techniques, and narrative structures. Horowitz predicts AI entertainment will experience the same evolutionary trajectory: while initial experiments focus on rendering individual VFX shots or augmenting podcasts, a new cohort of creators will eventually invent native storytelling conventions built specifically for AI.
Horowitz also contrasts AI storytelling with modern social media. On platforms like Instagram or TikTok, the storytelling format forces the creator to be the protagonist, which automatically repels anyone who is not naturally exhibitionist or narcissistic. AI eliminates that constraint by giving storytellers the power to construct autonomous characters, entirely separate from their own identity, to tell novel stories. As Horowitz observes:
> “So there was kind of like plays and then there was film and then there's going to be AI film.”
When asked about managing the crushing psychological pressure of being a founder, Horowitz emphasizes that an entrepreneur's sanity depends on aggressively narrowing focus down to what is directly within their control. Founders routinely become paralyzed by the infinite list of things they cannot fix. Looking back at his own tenure running Loudcloud and Opsware, Horowitz reflects that his most pivotal decisions were not intellectually complicated, but psychologically agonizing to execute. As a prime example, during the brutal dot-com collapse, Horowitz took Loudcloud public just 18 months after founding it, carrying only $2 million in trailing twelve-month revenue while projecting an audacious leap to $75 million. Conventional market wisdom dictated that an IPO under those metrics was suicidal, yet facing financial extinction, surviving required executing a plan that looked irrational on paper.
- **The social media bottleneck**: By requiring the creator to be the centerpiece of the narrative, contemporary social media filters out creative individuals who prefer not to make themselves the main character.
- **Psychological courage over intellectual complexity**: The hardest choices in company building rarely involve intricate analysis; they demand the stomach to execute counterintuitive, terrifying decisions in the face of public skepticism and extreme downside risk.
`30:00` Making the Least-Bad Decision and Running at the Problem
Reflecting on the near-collapse of Loudcloud, Ben Horowitz explains that taking the company public was an agonizing choice, but the only alternative after speaking to virtually every private market investor was bankruptcy. He emphasizes that the magic of navigating a company through an existential crisis lies in not hesitating when faced with terrible options. Many peers folded simply because they froze when confronted with choices that looked unappealing. Founders frequently stumble into psychological traps by agonizing over consequences: they know they must re-architect the product, but dread missing their quarterly targets or struggling to fundraise; or they know an executive must be let go, yet avoid the confrontation because it will disrupt operations in the short term.
```mermaid
flowchart TD
accTitle: Tough Decision Framework under Crisis
accDescr: A flowchart illustrating the founder choice between facing painful immediate action versus hesitating and risking collapse.
A["Crisis / Difficult Reality"] --> B{"Founder Psychology"}
B -->|"Hesitates out of fear"| C["Delays re-architecture or executive firing"]
B -->|"Trusts eyes & executes"| D["Takes least-bad option immediately"]
C --> E["Catastrophic Failure / Bankruptcy"]
D --> F["Endures painful fallout, but company survives"]
```
Horowitz insists that avoidance is fatal and that founders must run directly at the problem, no matter how miserable the fallout promises to be. As Ben put it, “All that is true, but it's not as bad as leaving that motherfucker in place.” When making these calls, dwelling on the painful aftermath is unproductive; during Loudcloud’s IPO, had he allowed himself to obsess over every punishing consequence that would follow—all of which did happen—he joked he would have shot himself in the head. The discipline of a chief executive comes down to assessing reality clearly, identifying the single best action available right now, and executing it without hesitation.
- Loudcloud’s choice came down strictly to an IPO or bankruptcy after exhausting private financing options.
- Hesitation on re-architecting software or firing heads of sales stems from fearing immediate quarterly hits, which often guarantees company failure.
- The standard for action is not whether a move is inherently good, but whether it is the best available move on the board.
`32:50` The Mechanics of Executive Coaching and Mentorship
Horowitz stresses that nobody can be formally trained to be a CEO, which is why *The Hard Thing About Hard Things* was written around specific historical war stories rather than generalized step-by-step formulas. “There are no five steps,” he states; leadership is profoundly situational. Mentorship should not consist of rigid instruction, but rather an ongoing dialogue with someone who has already made every mistake and can think through operational friction in granular detail.
To illustrate, Horowitz recounts advising a founder whose primary dilemma was: *"My CTO is an asshole."* While conventional advisors might react impulsively, Horowitz broke the dynamic down functionally:
1. **Define the objective:** The founder did not want to fire the CTO, because the CTO was an exceptional engineer who executed and shipped products far better than anyone before him.
2. **Clarify the boundary:** The founder could not fire him, could not lose him by picking a counterproductive fight, but also could not allow the behavior to continue.
3. **Examine the behavior:** The founder explained that the CTO had just made a young woman in finance cry during an internal interaction.
4. **Deliver actionable feedback:** Horowitz scripted the exact conversation the founder needed to have, framing the issue not around morality or generic rules, but around the CTO's long-term career effectiveness:
- Acknowledge that he is a brilliant Director of Engineering who delivers software on time.
- Explain that a true Chief Technology Officer must orchestrate the entire company across cross-functional departments to secure the resources tech needs.
- Clarify that walking into a room and making someone cry means he will never get what he needs from her, her manager, or her department head.
- Give him the choice: develop organizational effectiveness, or accept that the company will eventually hire someone above him who can.
Horowitz contrasts this targeted, pragmatic approach with abstract management clichés, arguing: “You kind of want somebody who can have a conversation like that where you can go, okay, that's actually some functional advice that makes sense that I can use as opposed to just 'no asshole rule,' you know, or like some dumb ass shit that they fucking heard at Harvard Business School.”
`35:22` Adapting to Platform Shifts and Managing Firm Obsolescence
Transitioning a firm through era-defining technology shifts often requires making painful personnel decisions. Horowitz reflects that while starting companies is mechanically easier now than during the dot-com era, pushing an incumbent organization forward is uniquely grueling because it requires moving past loyal people whose capabilities belonged to a prior era. Venture capital undergoes the same brutal cycles; Horowitz notes that Andreessen Horowitz no longer invests in consumer internet mobile apps or traditional SaaS businesses, leaving categories of internal domain experts with obsolescent skillsets unless they transform into world-class specialists in emerging areas.
The firm's operational machinery had to reinvent itself in the same manner. The conventional tech marketing models that worked a decade ago stopped delivering results, especially at a16z's current scale. Surviving meant rebuilding marketing into a direct media operation and building out a massive podcast network. Horowitz admits that telling long-time performers who did fantastic work in the past that the firm must pivot away from their skillset is the part of his job he enjoys the least, but executing it is non-negotiable for institutional survival.
`37:04` Technological Inflection Points and Asymmetric Strengths
When evaluating emerging technologies, Horowitz cautions against linear extrapolations that ignore underlying mechanics. The industry learned this lesson with 3D printers, which generated massive hype around the promise of printing arbitrary physical objects at home. However, 3D printing did not follow Moore’s Law—adding raw compute power did not accelerate printing speeds or physical material science—capping its growth rate and making it difficult to build generational breakout companies in the space. Conversely, critics initially dismissed generative AI because models hallucinate, a flaw Horowitz argues is solvable and often advantageous in creative applications.
| Technology Wave | Initial Market Criticism / Assumption | Underpinning Reality | Breakout Outcome |
|:--- |:--- |:--- |:--- |
| **3D Printing** | Assumed to follow exponential software curves ("print anything"). | Bound to physical material properties, not Moore's Law. | Failed to produce massive standalone venture returns. |
| **Generative AI** | Dismissed for output hallucinations and factual inaccuracies. | Compute scaling curves hold; hallucinations are solvable or beneficial. | Unlocks natural language computing interfaces. |
| **Original iPhone (2007)** | Worse phone calls, inferior physical keyboard compared to BlackBerry. | Packed a GPS and camera into a pocket computer. | Enabled novel platforms (e.g., Uber) impossible on a PC. |
Horowitz draws a parallel between technology bets and founding talent: breakthroughs never require perfection across every dimension; they require an overwhelming, asymmetric spike in one critical area. The original 2007 iPhone was roundly criticized for poor cellular reception, an awkward touch keyboard compared to BlackBerry, and computing power far behind desktop PCs—yet its combination of a GPS and pocket camera enabled entirely new markets like Uber that could never exist on a PC.
Investors must evaluate founders through the exact same lens. “When we think about entrepreneurs, what I always say is like, we're not looking for lack of weakness,” Horowitz explains. “There's something wrong with every entrepreneur, including all of you, including me, like there's something like horribly wrong, but that's not the question—the question is, are they world-class at a thing, are they the best in the world at something, do they have a secret that nobody else has?” Assessing an AI interface that allows people to talk to a computer in plain English requires focusing on the magnitude of that single superpower, not its marginal deficiencies.
`39:46` Mastering Self-Doubt: Advice to First-Year Founders
Offering final guidance to founders embarking on year-one projects, Horowitz stresses the discipline of honoring the original conviction that drove them into entrepreneurship. Founders inevitably experience moments of profound uncertainty, but spending mental energy agonizing over whether a founding bet was correct yields zero return. The primary danger an entrepreneur faces is not external skepticism, but internal paralysis.
Horowitz notes that even the most celebrated founders suffered from crippling impostor syndrome early in their journeys. As Horowitz observed, “I spent a lot of time with Mark Zuckerberg when he was 22 years old, and he had a tremendous doubt about whether he was going to make it.” Founders must realize that internal fear is an inherent byproduct of taking massive risks rather than evidence of impending failure. Success comes down to separating emotional turmoil from operational execution: “It's not what you feel, it's what you do. Just focus on what you do. You're going to feel scared. If your guts aren't boiling, you're not even trying at this thing.”
## Notable quotes
> `00:00` You're not a great entrepreneur unless you feel like you don't have a choice but to build a company because otherwise you're going to quit anyway. — Ben Horowitz
> `39:46` It's not what you feel, it's what you do. Just focus on what you do. — Ben Horowitz
## Mentioned
Ben Horowitz, Andrew Chen, Marc Andreessen, Mark Zuckerberg, John Maynard Keynes, Andreessen Horowitz, Loudcloud, Opsware, Hewlett-Packard, Netscape, Sun Microsystems, EMC, Tandem Computers, FedEx, Citibank, IBM, Meta, Instagram, Twitter, TikTok, Airbnb, Uber, BlackBerry, Cursor, ChatGPT, Grok, AWS, Speedrun, The Hard Thing About Hard Things, iPhone
## Video description
What does it take to build a company when AI is changing the rules this quickly? Ben Horowitz argues that technology may make companies easier to start, but the hardest parts of entrepreneurship, making decisions under uncertainty, building momentum, and managing your own doubts, aren't going away.
Speedrun's Andrew Chen sits down with Ben to discuss why AI feels different from previous technology shifts, whether billion-dollar companies can be built with tiny teams, and why Ben thinks the application itself may eventually become less important as agents work directly against data.
They also get into defensibility in AI, what founders should look for in an investor, how venture capital is changing, and why the best entrepreneurs aren't defined by a lack of weaknesses, but by having a strength or insight that's genuinely exceptional.
Ben Horowitz argues that technology cycles reward asymmetric structural advantages—such as AI scaling instantly over established global networks—while long-term founder success depends on decisive execution, managing executive self-doubt, and securing early path-dependent momentum.
AI Mechanics, Defensibility, and Architectural Shifts (2:30)
Horowitz distinguishes AI from prior computing shifts by characterizing it as a probabilistic computer running on top of already mature global networks. While early internet platforms struggled with high hardware costs and a tiny addressable user base—Netscape held a 90% browser share in 1996 with only 55 million global users, half on dial-up—AI products can distribute instantly to hundreds of millions without waiting for network infrastructure to catch up. Furthermore, Horowitz dismisses claims that application-layer AI companies are superficial wrappers, comparing them to early dismissals of relational databases. Products like Cursor, which uses approximately 14 specialized models above base foundational layers, aggregate proprietary behavioral data and build high switching costs from routine user inertia.
Looking forward, Horowitz anticipates that standalone software applications may give way to centralized enterprise data operated directly by autonomous agents. Because artificial agents lack legal standing to hold traditional bank accounts or credit cards, he suggests decentralized crypto networks could serve as the native financial rails for an emerging agent-to-agent economy.
Path-Dependent Startup Momentum and Venture Evolution (15:00)
Because cloud infrastructure and developer tools lowered startup launch costs from the multi-million-dollar outlays required by Sun Microsystems in the 1980s to five-person teams on laptops, Horowitz argues the modern venture capital function centers on path-dependent systems. When competing startups launch identical concepts with equivalent talent, initial signaling from top-tier investors attracts premier engineers, producing superior products, faster customer acquisition, and accelerated follow-on financing. Venture firms essentially act as catalysts for this momentum flywheel.
To institutionalize that support, Andreessen Horowitz evolved from the traditional model of a lone board member into a specialized agency structure. After Marc Andreessen's 2011 thesis predicted that annual software winners generating $100 million in revenue would expand from 15 to 150, the firm split into vertical practices—including crypto, bio, gaming, and defense. The firm established dedicated operational teams to orchestrate corporate briefings with buyers like FedEx and Citibank, train first-time founders on executive hiring, and navigate late-stage international expansion.
Founder Psychology and Crisis Execution (30:00)
Horowitz contends that CEO failure rarely stems from intellectual deficiencies, but rather from a crisis of confidence when navigating unfamiliar terrain. First-time technical founders face an acute skill deficit when evaluating executive functions they have never performed, such as corporate financial controls. When founders hesitate because they lack certainty, organizations interpret delay as weakness, which halts execution velocity and invites internal politics. According to Horowitz, an effective board or seed advisor must help the founder bridge this executive gap and make difficult calls rapidly.
Under existential pressure, founders frequently freeze because every accessible option carries painful fallout. Recalling Loudcloud’s decision to pursue an IPO with only $2 million in trailing revenue during the dot-com crash, Horowitz notes that going public was widely perceived as irrational, yet it represented the only alternative to bankruptcy. He maintains that survival depends on running directly at problems—such as dismissing poor executive fits or executing painful code re-architectures—and selecting the best available option on the board rather than dwelling on unavoidable negative consequences.
Pragmatic Management and Valuing Asymmetric Strengths (37:04)
Leadership coaching, Horowitz asserts, requires granular, situational problem-solving rather than abstract management formulas. Advising a founder dealing with a high-performing but abrasive technical executive, Horowitz avoided moralizing rules and instead scripted feedback tied directly to career effectiveness: demonstrating that alienating colleagues prevents an executive from securing cross-functional resources, leaving the company no choice but to hire a manager above them if behavior remains unchanged.
Similarly, Horowitz cautions against judging breakthrough technologies or founders by their deficiencies. Just as 3D printing failed to deliver venture-scale returns because physical materials did not follow exponential Moore's Law scaling, transformative breakthroughs succeed because of an overwhelming, asymmetric spike in capability. The original iPhone featured poor call quality and an awkward keyboard relative to BlackBerry, but its combination of GPS and camera unlocked unprecedented platforms like Uber. Investors and founders must evaluate whether an asset possesses a genuine superpower, such as natural language computing interfaces, rather than demanding an absence of flaws.
While modern developer tools and AI models dramatically reduce the friction of company formation, Horowitz emphasizes that building enduring enterprises still requires founders to navigate severe psychological self-doubt, personnel obsolescence, and brutal trade-offs. The broader question remains how rapidly autonomous agents and reinforcement learning will displace traditional software applications, reshaping startup defensibility and enterprise organizational design.
精炼阅读摘要
Ben Horowitz 与 Andrew Chen 在对话中指出，人工智能是一场概率型计算机革命，但企业能否胜出仍取决于创始人在面对不确定性与心理危机时的决断力，以及由顶级风险投资所催化的路径依赖型势头。
AI作为概率型计算机与即时分发 (2:30)
Horowitz 将技术演进划分为计算机与网络两个交替周期，并将当前的人工智能定义为一种有别于以往确定性计算的“概率型计算机”。网络需要达到一定规模后才能产生实际价值，例如在1996年，Netscape 虽占据约90%的浏览器市场份额，但全球互联网用户仅有约5500万，其中约半数依赖拨号连接，高昂的硬件投入与狭小的市场导致了后来的网络泡沫破裂。相比之下，AI 建立在现已成熟的全球互联网络之上，其产品能够在一夜之间向数亿用户分发，从而创下了前所未有的企业增长速度。
防御壁垒、专有数据与软件形态重构 (5:00)
针对外界关于 AI 应用只是基础大模型简单套壳（GPT wrapper）的质疑，Horowitz 驳斥称这种论调就像早期将企业软件贬低为关系型数据库管理系统（RDBMS）套壳一样片面。以编程工具 Cursor 为例，它在基础大模型之上构建了约14个精细的专业模型，通过与顶尖工程师的持续互动积累了独有的领域设计数据与深厚架构壁垒。同时，软件竞争的关键在于用户惯性与分发能力，用户一旦在日常工作流中建立习惯，除非竞争对手能提供10倍以上的性能跃升，否则不会轻易切换。
在组织规模与软件架构层面，虽然大语言模型大幅提升了代码编写效率，但 Horowitz 援引凯恩斯在1930年代对每周15小时工时预测的落空指出，人类对汽车、电视等新产品的欲望会不断膨胀，服务于扩张需求的企业往往仍需扩大规模。未来的软件架构可能逐渐淡化独立的“应用程序”概念，转变为以集中化的底层数据直接驱动自主 AI Agent 执行任务。由于 AI 缺乏法定身份开设传统银行账户，去中心化加密网络可能成为 Agent 经济的原生金融轨道。
风投本质：路径依赖与高管能力赤字 (15:00)
Horowitz 认为，从1970年到2009年 a16z 成立前，传统风投模式基本停滞，主要针对 Sun Microsystems 等需要庞大直销与专业服务团队的沉重架构。随着云计算与开发工具的成熟，极少数人组成的轻量团队也能打造大规模平台。然而，早期创业竞争是一个复杂的自适应系统，在点子和技术团队同质化的情况下，市场由谁能率先建立“势头”决定。顶级风投的背书能够作为关键信号，吸引顶尖工程师并带来产品优势、客户增长和后续融资的飞轮效应。
首次创业的技术型创始人通常具备出色的产品研发能力，但普遍缺乏担任 CEO 所需的管理经验和公司治理框架，尤其在招聘和评估从未接触过的高管职位（如需要对接华尔街与设计内控机制的 CFO）时面临明显短板。Horowitz 强调，高价值的种子期投资人核心作用在于充当运营向导，帮助创始人弥补这一关键的管理能力赤字，而非给出泛泛而谈的抽象建议。
创始人的心理危机与“最不坏”决断 (20:00)
技术创始人转型为合格 CEO 的最大障碍往往不是智商不足，而是内心的信心危机。当创始人因不了解某些领域而陷入犹豫摇摆时，组织内部会感知到软弱，不仅拖慢执行速度，还会滋生政治内斗。Horowitz 指出，创始人在绝境中不能因恐惧痛苦的后果而回避冲突。例如当年面对互联网泡沫破灭时，Loudcloud 在过去12个月收入仅200万美元的情况下顶着市场质疑强行推进 IPO，成为避免破产的唯一生路。
优秀的管理辅导应当是针对具体痛点的务实博弈，而非商学院式教条。以技术过硬却性格暴躁的 CTO 为例，成熟的沟通方案是将其行为与个人职业发展的有效性挂钩，指出其人际冲突会切断部门间的资源获取，促使其在提升组织协同能力与接受外部管理层介入之间做出选择。作为商业机构，企业在应对技术周期演进时必须果断淘汰旧范式团队，正如 a16z 停止投资传统 SaaS 和移动应用，将市场营销重组为直接媒体与播客网络一样，战略迭代必须承担人员换轨的阵痛。
非对称优势与首年创始人的心理防线 (37:04)
在评估新技术与团队时，Horowitz 警惕忽视底层规律的盲目外推。例如3D打印技术因受限于物理材料学而无法遵循摩尔定律，最终未能在风投市场形成巨大的商业回报；而生成式 AI 具备算力扩展曲线支撑，早期被诟病的“幻觉”在创意领域反而具有独特价值。正如初代 iPhone 虽在通话和键盘输入上逊于黑莓，却凭借 GPS 和摄像头的非对称优势催生了 Uber 等全新商业模式一样，技术突破和创始人选拔的核心均在于单一维度的压倒性特长，而非无缺点的平衡。
对于首年起步的创始人，Horowitz 指出甚至连22岁时的 Mark Zuckerberg 也曾深陷严重的自我怀疑。承受巨大风险所必然伴随的恐惧与翻江倒海的压力是创业的常态，创始人必须将情绪体验与实际执行彻底剥离，停止无谓的内耗，全力聚焦于眼下唯一可行且最关键的行动。
本场对话展现了从技术范式到企业生存法则的务实分析，指出 AI 正重构软件形态，但初创企业的根本壁垒依然依赖于极致的专有数据、创始人在极端压力下的决断力以及关键特长的非对称爆发；留给行业的未解难题则是 Agent 经济与加密货币轨道的实际融合节奏，以及极简团队能否在不断膨胀的人类欲望面前长期维持竞争力。
Detailed notes — zh
## Background
该内容由 Ben Horowitz 与 Andrew Chen 于 2026 年 10 月 24 日在播客及音频平台联合发布。节目围绕 AI 快速重构商业规则的背景展开，探讨了在技术降低建司门槛的当下，创始人如何应对不确定性决策、团队规模极限、AI 应用护城河以及风险投资演进等深层挑战。
## TL;DR
Ben Horowitz 深入剖析了当今 AI 浪潮与早期互联网周期的本质差异，指出 AI 作为概率型计算机直接架构在成熟网络基础设施之上，使得产品能够瞬间实现海量分发。他驳斥了将 AI 应用视为脆弱“包装器”的偏见，强调围绕数据飞轮与工作流构建专有模型的重要性，并预测未来软件可能会从传统应用形态转向由 Agent 与数据层直接交互的新范式。此外，Ben 结合创办 Opsware、LoudCloud 逆势上市以及打造 a16z 的亲身经历，指明顶级创始人从来不是完美无瑕的全才，而是具备极致长板、能在绝境与剧烈自我怀疑中保持果断决策力的行动者。
## Atmosphere & style
对话洋溢着硅谷老将特有的沉稳、硬核与快人快语的实战气息，完全剔除了浮于表面的鸡汤与理论说教。Andrew Chen 提问精准且具时代前瞻性，Ben Horowitz 的回答则极富穿透力，既有对技术史演进逻辑的冷静拆解，又有回忆当年绝境求生、直面公司濒临破产与艰难决策时的坦率与血性。整场交谈在幽默讽刺与严肃深邃之间张弛有度，听众能深刻体会到创业维艰背后那种胃部翻腾的心理重压，以及唯有靠坚定行动才能破局的宿命感与决绝。
## Bulletin
- `03:00` AI 与早期互联网的根本差异在于其作为“概率型计算机”直接构建于成熟的全球宽带基础设施之上，使新产品无需等待网络渗透即可瞬间触达数十亿用户，实现前所未有的扩张速度。
- `09:00` 所谓的“GPT 包装器”偏见忽视了软件历史逻辑，类似于早期嘲讽企业软件只是关系型数据库的外壳；真正的防御壁垒建立在专有工作流数据回流、专有模型群构建以及用户习惯的深度锁定之上。
- `15:00` 针对团队规模微型化与资金门槛降低的趋势，风险投资的核心价值已转向在复杂自适应系统中为初创企业注入打破僵局的“势能”，帮助创始人跨越认知盲区与 CEO 信心危机。
- `37:05` 评估技术突破与创始人时不应追求全能或挑剔缺陷，而应聚焦于某一维度的“非对称长板”；真正的突破只需在一个关键特性上做到颠覆性极致即可重塑未来商业版图。
## Key takeaways
- `00:00` 真正的创业者被不可抗拒的创造冲动所驱动，只有当创业成为一种别无选择的必然本能时，创始人才能抵御中途退出的巨大阻力。
- `03:00` 计算史是“计算机”与“网络”交替共生的演进史；AI 是一种全新的概率型计算机，与 1990 年代受制于网络规模不足的 Dot-com 时代相比，AI 拥有零延迟的全球分发通道。
- `05:00` 人类需求的无限膨胀意味着极小团队维持长期竞争力的难度极高，未来软件很可能不再以传统固化应用程序形态存在，而是演化为自主 Agent 直连核心数据的新范式。
- `10:00` 打破“GPT 包装器”偏见的核心在于数据复利与工作流嵌入，率先占领用户心智和习惯能够形成排他性壁垒，击败对手必须拿出 10 倍好于现有方案的产品体验。
- `12:35` AI 发展正处于模型分化期：纯通用 LLM 的文本生成已出现平台期迹象，而基于强化学习的代码生成与逻辑推理仍在经历指数级演进，引发底层架构与应用层的全面重构。
- `20:00` 初创企业面临的最大致命伤往往不是智商不足，而是创始人在面临不确定性时的决策信心崩溃；推迟决断会迅速引发组织恐慌并滋生内部政治斗争。
- `22:30` 泛滥的业余 VC 建议常给早期公司带来灾难，a16z 围绕帮助创始人建立 CEO 掌控感打造了系统性组织，并先后经历了垂直化与协助后期企业全球化的战略跃迁。
- `26:00` AI 视频与娱乐绝非传统影视流程的降本增效工具，而是一种全新的独立叙事媒介，不仅摆脱了传统视听语法，更打破了社交媒体以创作者自身为主角的展示逻辑。
- `30:00` 在遭遇极端危机时，过度推演未来灾难只会让人陷入绝望瘫痪，创始人必须看清客观现实，毫不迟疑地咬牙执行当下唯一的破局选项。
- `32:00` CEO 的管理工作高度情境化且不可被流程化，优秀的导师不是发布教条化指令的教官，而是能结合具体人际与业务场景引导实战对话的同行者。
- `35:46` 创投机构与成熟企业必须具备自我革命的决心，a16z 停止投资传统 SaaS 和消费级 App 并转向媒体化直达模式，证明组织演进需要跨越对过往功勋模式的路径依赖。
- `37:05` 优秀的技术与创始人不需要面面俱到，初代 iPhone 的突破性源于口袋化的相机与 GPS，投资者寻找的是拥有独家秘密与绝对极致长板的领袖。
- `39:35` 最致命的怀疑永远来自创业者内心的自我动摇，顶尖创始人同样会经历持续的恐慌，战胜心魔的唯一方式是将注意力从内在情绪彻底转向手头的具体行动。
## Detailed walkthrough
### `00:00` 创业者的宿命感与创立 a16z 的契机
对真正的创业者而言，创办公司绝非一项经过轻重权衡的普通职业选择，而是一种近乎不可抗拒的本能冲动。面对创业过程中无法回避的内心疑虑与外界压力，创始人必须依靠压倒一切的信念去行动。
回顾早年经历，Andrew Chen 提到了多年前初次拜访 Ben Horowitz 的场景。当时 Ben Horowitz 的公司刚从 LoudCloud 更名为 Opsware，整个办公室的会议室全部以嘻哈歌手（hip-hop artists）命名，展现出极其独特的极客与街头融合文化。彼时 Opsware 即将以巨资出售给惠普（HP），当 Andrew 问及功成名就后有何打算时，身心俱疲的 Ben 坦言自己准备直接退休。然而事实完全相反，Ben 不仅没有退隐，反而转身与 Marc Andreessen 共同创立了风投机构 Andreessen Horowitz（a16z）。
解释这种转变时，Ben Horowitz 谈到创业者骨子里的驱动力：生活很大程度上就是把脑海中的想法变成现实。在 Opsware 出售之后，他又萌生了全新的构想，而作为天生的创业者，他根本无法抑制去实现它的渴望。正如 Ben Horowitz 直言：
> “You're not a great entrepreneur unless you feel like you don't have a choice but to build a company because otherwise you're going to quit anyway.”
---
### `03:00` 计算与网络的交替：为什么 AI 浪潮的爆发速度远超互联网早期
回顾计算行业的发展脉络，技术演进始终围绕着“网络”与“计算机”两条主线的交替共生展开。从微波通信与集成电路，到大型主机（Mainframes）与 SNA 架构，再到个人电脑（PC）与局域网（LAN），乃至智能手机与移动互联网，两者具有截然不同的技术特征和采纳曲线（adoption curves），却又彼此剧烈赋能。
```mermaid
flowchart TD
accTitle: 计算与网络的技术协同演进
accDescr: 展示历史上计算机与网络两条主线的交替协同，以及当今AI作为概率型计算机直接架构在现有互联网之上的逻辑。
subgraph HistoricalPairs["历史周期：计算机与网络的协同"]
direction TB
C1["计算机主线：集成电路 / 主机 / PC"] <--> N1["网络主线：微波 / SNA / 局域网"]
C2["移动计算：智能手机"] <--> N2["互联网络：消费级互联网"]
end
subgraph ModernEra["当今时代：AI 浪潮"]
direction TB
NetExists["全球互联网基础设施（数十亿高带宽终端用户）"]
AICompute["AI：全新的『概率型计算机』"]
NetExists -->|"零等待分发通道"| AICompute
AICompute -->|"产品上线即达海量用户"| HyperGrowth["前所未见的用户扩张与变现速度"]
end
HistoricalPairs -.->|"范式迭代"| ModernEra
```
AI 与 1990 年代互联网初期的共通之处在于：技术日新月异，从业者每天醒来都能在网络上看到令人惊艳的全新突破。但在底层逻辑上，两者有着本质差异：互联网是网络基础设施，而 AI 则是全新的“计算机”——一种有别于以往确定性计算机（deterministic computers）的“概率型计算机（probabilistic computer）”。
互联网在 2000 年前后经历毁灭性的 Dot-com 泡沫破裂与电信崩盘，根源在于网络效应的采纳曲线早期被过度炒作，而实际入网人数严重不足。在 1996 年的网景（Netscape）时期，即便坐拥 90% 的浏览器市场份额，网景全球也仅有 5,500 万用户，且其中半数仍在使用极慢的拨号上网（dial-up）。当时开发者耗费巨资采购 Sun Microsystems 服务器和 EMC 存储设备，最终却只能触达两三千万用户，根本无法支撑起商业闭环。
反观 AI，它诞生在互联网基础设施早已极其成熟且普及的时代。开发者无需等待底层网络搭建或用户入网，产品只要上线即可瞬间触达全球数十亿用户。这使得 AI 公司的增长斜率超越了以往任何科技浪潮。
- 1996 年 Netscape 占有约 90% 的浏览器市场，但其绝对用户体量仅约 5500 万人。
- 互联网早期成本高昂，开发者必须重度依赖 Sun Microsystems 服务器与 EMC 存储硬件。
- 互联网早期面临“网络规模不足”的硬约束，而当下的 AI 产品无需等待网络渗透，上线即面向全网。
---
### `05:00` 极小团队的超级公司、凯恩斯预言与软件终局
大语言模型（LLM）在代码编写上的惊人表现引发了全行业的思考：写软件是否会变得像在社交网络上制作内容一样门槛全无？这是否意味着未来将诞生只由极少数人甚至单打独斗的超级个体（solopreneurs）运营的十亿美元级独角兽？
Ben Horowitz 认为，凭借极少数人员触达海量用户并赚取巨额利润在今天是完全可行的，类似昔日 Instagram 在被 Meta 收购前用极小团队撬动巨大规模的案例完全可以重现。但企业能否凭借微型团队维持长期竞争力则要打上问号。从商业历史来看，人类对产品的标准与需求永远在水涨船高。经济学家约翰·梅纳德·凯恩斯（John Maynard Keynes）曾在 1930 年代做过著名预言，认为随着衣食住行等基础需求得到满足，人们每周的工作时间将缩减至 15 小时。然而现实却截然相反——人类的需求几乎呈无限膨胀态势，从必须有一辆车，到家庭成员人手一辆车；从拥有一台电视机，到每个房间都配备电视机，再到后来的智能手机与空调。正如 Ben Horowitz 所指出的：
> “I think what people need does tend to expand almost infinitely. So to fulfill that need, sometimes you have to build a bigger company, but we'll see.”
面对快速克隆的“快时尚 SaaS（Fast Fashion SaaS）”或运行时根据用户需求个性化动态生成的定制软件（bespoke software），Ben 提出了更为根本的范式转移：未来的软件甚至可能不再以传统“应用程序（Application）”的形态存在。传统应用程序的核心逻辑很可能会逐步消亡，取而代之的是由底层数据组织与面向数据直接工作的 AI Agent 所构成的新架构。
```mermaid
flowchart LR
accTitle: 软件架构范式的转变
accDescr: 对比传统SaaS应用形态与未来以数据和智能体为核心的新型架构演变。
subgraph LegacyModel["传统软件范式 (SaaS)"]
direction TB
App["固化应用程序 UI / 逻辑"]
User["用户手动操作应用"]
Data1["数据库"]
User --> App --> Data1
end
subgraph AgenticModel["未来软件范式 (Agentic Data)"]
direction TB
Data2[("核心数据层 / 知识组织")]
Agents["自主 AI 智能体 (Agents)"]
Intent["用户意图 / 自然语言任务"]
Intent --> Agents
Agents <-->|"动态执行 / 查询 / 交互"| Data2
end
LegacyModel -.->|"范式更替"| AgenticModel
```
---
### `09:00` AI 应用的防御壁垒与所谓的“GPT 包装器”偏见
针对行业内关于 AI 应用缺乏网络效应、沦为脆弱的“GPT 包装器（GPT Wrappers）”的普遍质疑，Ben Horowitz 认为这种批评极其短视且缺乏历史纵深感。
把基于底层模型构建的应用贬低为“Wrapper”，在逻辑上完全站不住脚。Ben 回溯了关系型数据库（Relational Database）与 SQL（声明式语言）诞生之初的商业历史：当时 SQL 允许用户直接用声明式语言告诉数据库该做什么，也有大量声音质疑“既然能直接向数据库下达指令，为何还需要复杂的企业软件？”但事实证明，正是在关系型数据库的底层能力之上，才诞生了 SAP、Siebel 和 Salesforce 等数十亿美元规模的企业级巨头。底层技术的标准化从来不会扼杀上层应用，关键在于创业者如何围绕数据、工作流与实际业务场景构建起深度的防御壁垒。围绕所谓 GPT 包装器的争论，Ben 总结道：
> “The GPT wrapper thing was always a bit of a dumb idea in terms of not to build one, but to call it that.”
### `10:00` AI 应用的护城河、数据复利与“圈地阶段”
针对外界认为“AI 应用不过是基础大模型外的一层薄包装（Wrapper）”的贬低论调，Ben Horowitz 进行了直接驳斥。他回顾了关系型数据库（RDBMS）早期的历史，当时很多人也曾不屑地嘲弄某些产品“不过是个关系型数据库的外壳”，但事实证明这种看法完全错误。在当下的 AI 生态中，以代码编辑器 Cursor 为例，它已经在基础模型之上构建了多达 14 个高度精密的专有模型。这些模型深度融入了最顶尖程序员的日常工作流，每天都在持续吸收关于软件架构与程序设计的海量独家高价值数据，而这正是任何竞争对手都不具备的核心壁垒。
谈及竞争格局与产品壁垒时，Ben 强调软件竞争往往遵循“占有即拥有九成主权（possession is nine-tenths of the law）”的铁律。只要率先拿下客户，即使产品差异并不悬殊也能立于不败之地。例如 ChatGPT 与 Grok 相比，对于从零开始的用户而言两者的差距并不大，界面也极其简单直观，但由于用户已经形成了使用习惯，便极难发生迁移。Ben Horowitz 明确指出：“你不可能靠一个同样优秀或者稍微好一点的产品来干掉竞争对手，你必须拿出好上十倍的产品（It's like a 10 times better product）。”拿下市场份额不仅能直接锁定用户，还能在 AI 时代获得更多特定用户的数据反馈循环。
Andrew Chen 总结称，当前的市场正处于典型的“圈地阶段（land grab phase）”，在这个时期，最关键的产品特性仅仅是“它能用（it works）”，随后才会逐渐过渡到独特的交互设计（UX）创新和协作网络效应。Ben 进一步补充指出，当一台具备全新能力的新计算平台能够解决以往无人能解的难题时，哪怕暂时没有网络效应也具备巨大价值。此外，由于 AI Agent 无法合法申请信用卡，未来 Crypto 很可能会扮演 AI 之间的原生货币网络角色，从而在价值交换过程中催生出全新的网络效应。
```mermaid
flowchart LR
accTitle: AI 应用的数据复利与防御壁垒演进
accDescr: 展示 AI 应用如何从早期功能可用到积累数据、构建专有模型并形成防守壁垒。
subgraph S1["第一阶段：圈地与留存"]
A["可用性突破<br/>('It Works')"] --> B["先发占领用户<br/>习惯锁定"]
end
subgraph S2["第二阶段：模型与数据专有化"]
B --> C["高价值工作流交互"]
C --> D["专有数据回流<br/>构建多层模型群"]
end
subgraph S3["第三阶段：长期护城河"]
D --> E["10倍体验壁垒<br/>难以被替代"]
E --> F["潜在网络效应<br/>(Crypto原生价值交换)"]
end
```
### `12:35` 创始人背景的演变轨迹与技术周期的平台期
回顾互联网的发展历程，Andrew 和 Ben 探讨了创始人背景与技能重心的时代迁移。在万维网（Web）诞生之初，由于缺乏应用服务器和负载均衡器等基础构件，软件扩展到数亿人规模是前所未有的工程难题，甚至连写一个像 Yahoo 这样的链接目录，都不得不由斯坦福大学的计算机博士亲自上阵手写 Web 服务器。随后行业历经十余年演进，LAMP 架构的普及使得非计算机专业的开发者也能自如编写代码，并最终催生了像 Airbnb 这样由设计和用户体验（UX）主导的标志性应用。
这种创始人角色的迁移规律正在当前的 AI 浪潮中重演。虽然早期充斥着顶尖的计算机博士，但主流开发者构建 AI 应用的门槛已经较五年前大幅降低。不过，行业眼下面临的一个核心变量是技术突破的“平台期（Plateau）”何时到来。Ben 指出，目前纯语言模型（LLM）的发展已经出现了明显的平台期放缓迹象，而强化学习（RL）分支则正在飞速扩展，驱动编程等能力大幅跃升，但常规聊天机器人的体验变化并不剧烈。
Ben Horowitz 强调了技术范式转移对创业周期的深远影响：“每一次技术突破都是一次底层架构的彻底重构，它会重塑整个应用层以及整个基础设施层（every time there's a breakthrough, that's an architectural change. It'll change the entire application layer, the whole infrastructure layer）。”创业公司的机遇窗口正是精准排列在这些“技术突破”与“平台期收敛”的交替节点之上。
- **早年工程壁垒**：早期软件的最大服务对象仅为 IBM 等大型企业的内部员工规模，向 5 亿大众提供高并发服务是当时前所未见的系统挑战。
- **现阶段模型分化**：纯文本生成的通用 LLM 迎来边际效应递减，而基于强化学习的代码生成与推演能力则在呈指数级扩张。
### `15:00` 风险投资的演变逻辑：在复杂自适应系统中为创业者注入“势能”
Andrew Chen 提出，随着社交媒体、播客的普及以及 AI 赋能小团队打造庞大企业，风险投资（VC）这一资产类别本身的运作方式是否正在面临重构。Ben 回忆称，从 1970 年代到 2008 年 a16z 创立之前，风投行业的组织模式近乎保持了四十年未变。早先支持像 Sun Microsystems 或 Tandem Computers 这样的公司极其昂贵且试错成本极高，企业必须搭建庞大的专业服务网络与直销团队；随后互联网降低了建司门槛，到了 Twitter 时代，只需五个人带着笔记本电脑和 AWS 即可起步。
针对“初创公司团队更小、资金需求更低”的未来趋势，Ben 分析认为 VC 的核心底层功能并未改变——即在充满路径依赖的复杂自适应系统中，为企业赋予至关重要的“势能（Momentum）”。当市场上两家创业公司拥有完全相同的点子和同样出色的团队时，胜负的关键取决于谁先建立起势能滚雪球效应。顶级 VC 的背书能让公司迅速招揽最聪明的工程师、赢得第一批种子客户、拿到更多资金，从而以更快的节奏完成下一轮融资并终结比赛。
即使是已经实现高度盈利、看似不再需要外部资金的企业，依然极度渴望顶尖风投的助力。Ben 极其直白地道出了创业失败的真实痛苦：“你招揽了你认识的所有最聪明的人，你四处奔波募集了资金，甚至你妈妈也投资了公司——结果她现在怒火中烧，这种感觉糟透了，真的太痛苦了（You hired all the smartest people you know. You went out and raised money. Your mom invested in the company. She's pissed. It sucks. It really does）。”创业者承担着背水一战的心理与声誉重压，任何能切实提升成功概率的外部力量都具有无法抗拒的吸引力。
对于即将在 Demo Day 面对上千名投资者的早期创始人，Ben 针对“如何挑选种子轮合作伙伴”给出了核心建议：缺乏大公司治理经验的创始人最欠缺的正是“如何胜任 CEO 职责”。创始人往往从未担任过 CFO，既不懂财务控制架构，也不知道如何同华尔街对话，却不得不去考核与招募这类高端复合型高管。因此，优秀的早期投资伙伴必须能够提供跨越认知盲区的实战赋能，帮助他们完成从技术专家到企业领袖的蜕变。
### `20:00` 创业者的信心危机与不靠谱的投资人建议
Ben Horowitz 深入剖析了当一名创业 CEO 的残酷现实：这不仅是一份难度极高的工作，而且通常你根本不知道该怎么做。他打了个生动的比方，就像听日语同声传译一样，“如果你不懂日语，听上去所有人都说得挺好”，然而现实中你得面对各种员工的职业野心，甚至还要应付“Mark Zuckerberg 正在给你的核心骨干开出数亿美元挖角”这类极端压力。Ben 认为，当他看到创始人败下阵来时，核心症结往往并非智商，而是信心彻底崩溃——你清楚地知道自己不会做、知道决策可能会伤害公司，但你别无选择，必须做出决定。一旦创始人陷入犹豫不决，整座组织就会陷入恐慌，业务节奏被拖垮，政治斗争也会随之滋生，因为各路人马都会试图去填补那个权力空白。
正因如此，创始人最需要的是能帮你建立决断信心的投资人，以及能在招募和下一轮融资中提供真正势能的合伙伙伴。紧接着，Ben 毫不留情地讽刺了风投圈泛滥的业余指导。他直言不讳地嘲讽道：“那些连一天 CEO 都没当过的混蛋，居然指导别人怎么做 CEO，给出的建议简直荒唐至极。”比如劝一个早期初创公司立刻招满整套高管班子，这种瞎指挥只会把公司直接整垮。对于坊间最常见的鸡汤建议，Ben 戏谑地分享了他最“喜欢”的一条高论：“我得到的建议里最经典的是：‘关键是要招 A 级人才。’我当时就心想：‘好吧，我原本的计划是招一帮彻头彻尾的白痴，但既然你这么说了，我现在改招 A 级人才。’”
```mermaid
flowchart TD
accTitle: 创始人信心动摇引发的连锁反应
accDescr: 展示创始人缺乏决策信心如何导致组织迟缓与内部政治滋生
A["面临未知问题与高管挖角等外部压力"] --> B["创始人产生信心危机"]
B --> C["犹豫不决、推迟决策"]
C --> D["团队陷入焦虑与不安"]
C --> E["业务执行全面减速"]
C --> F["产生权力空白，滋生内部政治"]
D & E & F --> G["公司陷入恶性循环"]
```
Andrew Chen 顺势指出，湾区创投生态与欧洲等地区截然不同：欧洲的许多 VC 往往脱胎于私募基金或纯金融背景，而湾区的强项在于拥有大量真正操盘过业务的前实业家（former operators）。Ben 赞同这一趋势，不过也补充道，虽然行业整体在向实操型投资者靠拢，但这种能力“绝非人人具备，段位层次也大有不同”。
> “Having the confidence to make that decision when you don't know what you're doing and you don't know if that decision is right, but not like hesitating. And then, you know, when you hesitate, it makes the company nervous. It slows everything down. It creates politics because people try and fill that void.” —— Ben Horowitz
---
### `22:30` a16z 的三次演进：从构建 CEO 信心到垂直化与全球化
回顾 Andreessen Horowitz（a16z）的发展历程，Ben 解释了创立这家机构的核心初衷：传统风险投资对 LP 来说是不错的产品，但对创业者却并不够友好。过去典型的模式是给公司董事会安排一个聪明人，但除了个人的有限人脉外，几乎提供不了系统性支持；而且在早期产品探索期过后（约三个月后），天天泡在业务里的创始人与一个月甚至一季度才来开一次董事会的投资人之间，认知鸿沟已经变得巨大。Ben 从自己当年创办 Opsware 等公司的痛苦经历出发，思考创始人真正需要什么——归根结底还是“如何让你真正具备掌控感，像一个真正的 CEO”。
为了解决这个问题，a16z 围绕企业客户拓展与高级人才招揽建立了体系化的支持网络。例如，如果创始人需要接触 FedEx、Citibank 的高管，机构有专门的团队负责对接；如果要招聘 CFO，机构会直接为你引荐五位顶尖 CFO，手把手教你如何设计面试考题与考察维度，让创始人迅速实现技能跃升。随后，随着 Mark Andreessen 在 2011 年发表著名的《软件正在吞噬世界》，a16z 迎来了第二次重大战略跃进：以往创投行业认为每年只有 15 家公司能跑到 1 亿美元年营收，但如果软件重塑各行各业，这个数字将扩大到 150 家。为了在扩张的同时不失传统精品 VC 的敏锐与深度，a16z 率先推行了“垂直化战略”，分别在基础设施（infra）、应用（apps）、加密（crypto）、生物科技（bio）、游戏（games）以及国防安全等领域建立了具备深度专业能力的垂直投资部门。
到了今天，随着私有公司估值体量越来越庞大，a16z 正迈向第三阶段演化：重点帮助后期非上市公司建立跨国扩张与多元销售能力。Ben 提到，机构如今正在协助企业高效打入国际市场，例如利用 a16z 的品牌背书直接帮助企业落地日本市场并敲定前三笔标杆大单，并帮助企业应对多产品线、多渠道拓展的复杂运营挑战。
> “Our takeaway from that was, okay, if that's going to happen, then kind of VC up until that point was this idea that there were 15 companies in a given year that would ever get to 100 million in revenue. And the whole idea of the firm would be to kind of get into as many of those 15 as you could. But we're like, well, if software is going to eat the world, that's not going to be 15, that's going to be 150. And so how do you actually scale a VC firm you know kind of without losing all the things that make it a VC firm?” —— Ben Horowitz
---
### `26:00` AI 影视是全新媒介；18 个月 200 万营收逆势 IPO 的心理试炼
在进入问答环节后，话题转到了 AI 与内容创作的前景。Ben 斩钉截铁地下了断言：AI 视频与娱乐绝不仅仅是传统影视制作流程的降本增效，而是一种“全新的叙事媒介”。他援引了电影早期的演进历史——在电影诞生的最初期，创作者们只是把摄像机架在固定位置拍摄舞台剧，但电影最终发展出了完全独立于话剧的蒙太奇与叙事语法。Ben 预测 AI 影视也会经历完全相同的蜕变，绝不会止步于用 AI 生成几段传统电影场景。
不仅如此，Ben 还对比了社交媒体与 AI 叙事的本质差异。社交媒体的逻辑是“叙事中的主角永远是创作者自己”，这把那些低调内敛、不自恋或不愿意出镜展示自己的人天然排斥在外；而 AI 则赋予了每个人创造独立角色并讲述虚构故事的自由。Ben 告诫创业者，不要一味去照搬电影、电视的固定模式，而应探索只有 AI 能带来的全新叙事表达形式。
谈及 CEO 在极端困境中如何应对心理重压时，Ben 强调，关键在于把全部精力收束回“你当下唯一能掌控的事情上”，因为最击垮人的永远是无力感。他指出，自己当年做 CEO 时最自豪、也是代价最惨痛的决策，往往并不是逻辑上最复杂的抉择，而是“在心理上极难咬牙做出的决定”。例如，当年 Loudcloud 逆势 IPO 时，公司成立仅仅 18 个月，过去 12 个月的过去滚动营收（trailing 12 months revenue）只有区区 200 万美元，却顶着外部看来极度荒唐的预期（预测营收将暴涨到 7500 万美元）强行挂牌上市。在常人看来这显然是“糟糕透顶的主意”，但面对泡沫破裂的资本寒冬，硬着头皮走下去却是公司活命的唯一路径。
- **AI 媒介的历史参照**：如同早期电影不是“架台机器录话剧”，AI 视频绝非单纯的特效生产工具，而是会孕育出一批全新的独立创作者。
- **叙事模式的解放**：突破社交媒体“创作者即唯一主角”的展示自恋约束，AI 允许普通用户操纵非自身角色开展多元叙事。
- **Loudcloud 极端 IPO 数据**：
- 成立时间：仅 18 个月
- 过去 12 个月（TTM）实际营收：200 万美元
- 当时预期的远期营收目标：7500 万美元
> “The thing that overwhelms you is what you can't do and like all the problems and everything. When I think about like the decisions I made as CEO, the ones that I'm most proud of and would have cost us most weren't actually difficult decisions. They were just difficult to make.” —— Ben Horowitz
### `30:00` 绝境中的果断决断与直面问题的心态
在回顾创立 LoudCloud 的艰难时期时，Ben Horowitz 谈到了公司上市的背景。当时他们在私人融资市场上几乎接触了每一位投资人，最终摆在面前的只有两个选择：上市或者破产。破产显然是一个糟糕得多的结局，但即便是选择上市，在当时也是极其艰难且看似很差的选项。这种决策的精髓在于即便它看起来非常糟糕，但毫不迟疑地执行：“To be able to do such a bad choice like going public and not hesitate, just go, okay, I'm going to do it. That was the magic of that decision.” 那段时期有无数公司正因为犹豫不决而最终走向倒闭。创业者必须看清现实，做当下唯一能做的事，绝不能被恐惧瘫痪。
很多创始人在心理防线崩溃往往源于逃避痛苦的抉择。例如明知产品架构必须彻底重构，但因为重构会导致融资困难、业绩未达标、无法按时推出功能，就陷入犹豫，结果导致整家公司陷入绝境；或者明知必须开除销售负责人，却害怕面对激烈的谈话而选择视而不见。Ben 直言，开除高管确实会带来可怕的短期阵痛——没有销售主管、季度业绩可能崩盘、责任全落到自己头上，但放任问题继续发酵的代价远比这些更加惨重：“All that is true. But it's not as bad as leaving that motherfucker in place.” 在极其糟糕的逆境中，去过度推演所有负面后果毫无意义，关键在于明确眼下最好的出路，然后立刻去执行。
```mermaid
flowchart TD
accTitle: 创始人应对危机的决策路径
accDescr: 展示创始人面对危机时的错误犹豫与正确果断执行逻辑。
subgraph Hesitation["逃避与犹豫（致命陷阱）"]
A["面临严峻抉择\n（重构架构 / 开除高管 / 极端筹资）"] --> B["畏惧短期阵痛\n（无法融资 / 谈话尴尬 / 丢掉业绩）"]
B --> C["犹豫不决、转移视线"]
C --> D["彻底崩盘与破产"]
end
subgraph Execution["直面现实（破局之道）"]
E["看清客观现实"] --> F["找到当下唯一的最佳选项"]
F --> G["毫不迟疑地执行\n（承担短期代价）"]
G --> H["保留生机，绝处逢生"]
end
```
- **破产与上市的抉择**：LoudCloud 当时已见遍了所有私募市场投资者，唯有上市才能避免破产。
- **心理内耗的危害**：过度计算未来所有可能发生的坏事只会让人绝望，创始人必须聚焦于“我当下能做什么”。
---
### `32:00` 无法被流程化的 CEO 角色与导师的真正价值
成为 CEO 没有现成的培训体系，这也是 Ben 撰写《创业维艰》（*The Hard Thing About Hard Things*）的初衷。管理工作是高度情境化的（Situation-specific），根本不存在照本宣科的“五步法”。优秀的商业导师不是发布指令的教官，而是能够凭借深厚的实战经历、与创始人探讨棘手难题的对话者。
为了阐明情境化指导的意义，Ben 分享了一位创始人向他求助的真实案例。该创始人抱怨公司的 CTO 是个“混蛋”（asshole）。Ben 没有给出泛泛的建议，而是层层剖析痛点：
1. 创始人并不想解雇 CTO，因为该 CTO 的执行力与产出远超以往任何人。
2. 创始人真正苦恼的是如何与对方沟通，既不想激怒他导致其离职，又不能任由恶劣行径继续。
3. 冲突的具体诱因是：该 CTO 在与财务部一位年轻女性沟通时，言语过激将对方骂哭了。
Ben 给出的实操建议是，创始人应直接坐下来与 CTO 摊牌：“Look, you're a great director of engineering. Amazing, ship products on time and so forth. But you're not like really effective as a CTO and like, at some point I'm going to have to hire somebody who is an effective CTO until you can do that.” 真正的 CTO 不仅要管好研发团队，更必须跨部门协同并调动全公司的资源支持技术发展。如果 CTO 在会议中把初级员工骂哭，他就永远无法从对方及其上级那里获得所需的支持。Ben 尖锐地批评了哈佛商学院推崇的所谓“混蛋法则”（no asshole rule）等教条理论，强调只有紧密结合公司、产品与团队实际情况的指导才具备可操作性。导师也不一定必须是资深元老，哪怕只是走在前面一步的同行也能提供极大帮助。
---
### `35:46` 时代变迁与 Andreessen Horowitz 的自我迭代
推动一家成型公司的演进往往极为艰难。公司里常有功勋卓著的老员工，但随着时代转轨，他们的技能组合可能已无法适应新的产业浪潮。这种残酷性在风险投资行业体现得尤为明显：Andreessen Horowitz（a16z）如今已经不再投资消费级互联网产品、移动 App 或传统的 SaaS 企业。这意味着过去在这些领域积累深厚经验的专家如果不能在全新方向上建立世界级的专业度，就会面临被时代淘汰的危机。
此外，早期的品牌营销策略对现在的 a16z 来说也已经失效。随着机构体量的急剧扩大，他们必须建立全新的核心能力，转而采用直达受众（Go-direct）的战略，打造庞大的播客网络与媒体影响力。Ben 坦言，面对老员工不得不说明“虽然你过去立下汗马功劳，但现在我们必须转向新道路”，是他最不喜欢做的事情，但作为掌舵人必须硬着头皮推进。
---
### `37:05` 评估技术的突破点与创始人的“非对称长板”
在评估新兴技术时，很多人容易误判技术的发展轨迹。其中一个典型的误区是盲目外推（Extrapolate）。例如当年 3D 打印机风靡一时，许多人认为它能打印万物，但事实证明 3D 打印并不遵循摩尔定律，算力的增长并不能直接使其性能获得指数级飞跃，因此难以催生出万亿级的巨头。反观 AI，许多人抓住“幻觉”（Hallucination）等缺陷不放，但实际上这些缺陷不仅相对容易解决，在很多生成场景中甚至是不可或缺的优势。
突破性创新往往不需要在所有维度全面胜出，只要在一个关键维度上做到极致即可。如同 2007 年初代 iPhone 面世时，按当时的测评标准，它作为手机的通话质量糟糕，作为电脑远不如传统 PC，虚拟键盘体验也被黑莓（BlackBerry）碾压。但它把相机和 GPS 装进了口袋，而这种能力是 PC 至今都无法孕育出 Uber 的原因：“You only need one feature, right? Or one thing that kind of makes a breakthrough and creates a new future. Like it just has to be unbelievable at something. It doesn't have to be better than the last thing at everything.”
在识别人才时也是同样的逻辑。投资机构寻找的从来不是毫无瑕疵的人，而是拥有极高长板的创业者：“We're not looking for lack of weakness. Like every, there's something wrong with every entrepreneur.” 关键不在于他们有什么毛病，而在于他们是否在某一领域是世界顶尖的、是否掌握了独一无二的秘密，以及核心优势的量级到底有多大。就像今天能直接用自然语言（英语）与计算机交互本身，就是一项足以重构未来的巨大突破。
| 观察维度 | 常见认知误区 | 核心评估标准 | 典型代表案例 |
|:--- |:--- |:--- |:--- |
| **新技术爆发力** | 纠结于技术初期的缺陷，或盲目向所有行业外推 | 是否具备某个不可替代的突破性优势（突破点） | 3D 打印（未遵循摩尔定律） vs. iPhone（GPS与相机构建 Uber 基础） |
| **创始人素质** | 追求全能、挑剔创始人的性格或能力短板 | 是否在特定领域拥有世界级能力与绝对长板 | 是否拥有非对称认知与他人无法复制的“秘密” |
---
### `39:35` 克服创始人心魔：关注行动而非内心的恐惧
对于处于创业第一年的创始人们，Ben 给出的最终建议是：全力让自己最初的豪赌获得回报。既然最初相信自己并迈出了这一步，就要坚持下去。在创业旅程中，创始人会无数次怀疑当初的决定、质疑自己的信念，但沉溺于怀疑毫无用处。
最沉重、最危险的怀疑从来不是外界施加的冷嘲热讽，而是创始人施加给自己的内省怀疑。Ben 强调，即使是顶尖人物也无法免俗：“I spent a lot of time with Mark Zuckerberg when he was 22 years old, and he had a tremendous doubt about whether he was going to make it. Everybody has that. It's not what you feel, it's what you do. Just focus on what you do.” 恐惧和内心的翻江倒海是不可避免的生理反应——如果创业时感受不到那种肠胃翻腾的紧迫感，说明根本没有全力以赴。创始人唯一要做的，就是无视内心的动摇，把所有专注力投入到眼前的行动中去。
## Notable quotes
> `00:00` You're not a great entrepreneur unless you feel like you don't have a choice but to build a company because otherwise you're going to quit anyway. — Ben Horowitz
> `05:00` I think what people need does tend to expand almost infinitely. So to fulfill that need, sometimes you have to build a bigger company, but we'll see. — Ben Horowitz
> `09:00` The GPT wrapper thing was always a bit of a dumb idea in terms of not to build one, but to call it that. — Ben Horowitz
> `10:00` It's like a 10 times better product — Ben Horowitz
> `12:35` every time there's a breakthrough, that's an architectural change. It'll change the entire application layer, the whole infrastructure layer — Ben Horowitz
> `30:00` To be able to do such a bad choice like going public and not hesitate, just go, okay, I'm going to do it. That was the magic of that decision. — Ben Horowitz
> `30:00` All that is true. But it's not as bad as leaving that motherfucker in place. — Ben Horowitz
> `37:05` You only need one feature, right? Or one thing that kind of makes a breakthrough and creates a new future. Like it just has to be unbelievable at something. It doesn't have to be better than the last thing at everything. — Ben Horowitz
## Mentioned
Ben Horowitz, Andrew Chen, Marc Andreessen, Andreessen Horowitz, Opsware, Loudcloud, HP, Netscape, Sun Microsystems, EMC, Instagram, Meta, John Maynard Keynes, SAP, Siebel, Salesforce, Cursor, ChatGPT, Grok, IBM, Yahoo, Airbnb, Tandem Computers, Twitter, AWS, Mark Zuckerberg, FedEx, Citibank, Uber, BlackBerry
## Video description
What does it take to build a company when AI is changing the rules this quickly? Ben Horowitz argues that technology may make companies easier to start, but the hardest parts of entrepreneurship, making decisions under uncertainty, building momentum, and managing your own doubts, aren't going away.
Speedrun's Andrew Chen sits down with Ben to discuss why AI feels different from previous technology shifts, whether billion-dollar companies can be built with tiny teams, and why Ben thinks the application itself may eventually become less important as agents work directly against data.
They also get into defensibility in AI, what founders should look for in an investor, how venture capital is changing, and why the best entrepreneurs aren't defined by a lack of weaknesses, but by having a strength or insight that's genuinely exceptional.
[Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/d9bff268-e096-49a4-a2a0-3ad436b025be/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=d9bff268-e096-49a4-a2a0-3ad436b025be&feed=JGE3yC0V)
Source quotes
[0:00] Ben Horowitz: You're not a great entrepreneur unless you feel like you don't have a choice but to build a company because otherwise you're going to quit anyway.
[5:00] Ben Horowitz: I think what people need does tend to expand almost infinitely. So to fulfill that need, sometimes you have to build a bigger company, but we'll see.
[9:00] Ben Horowitz: The GPT wrapper thing was always a bit of a dumb idea in terms of not to build one, but to call it that.
[10:00] Ben Horowitz: It's like a 10 times better product
[12:35] Ben Horowitz: every time there's a breakthrough, that's an architectural change. It'll change the entire application layer, the whole infrastructure layer
[30:00] Ben Horowitz: To be able to do such a bad choice like going public and not hesitate, just go, okay, I'm going to do it. That was the magic of that decision.
[30:00] Ben Horowitz: All that is true. But it's not as bad as leaving that motherfucker in place.
[37:05] Ben Horowitz: You only need one feature, right? Or one thing that kind of makes a breakthrough and creates a new future. Like it just has to be unbelievable at something. It doesn't have to be better than the last thing at everything.
Source quotes
[0:00] Ben Horowitz: You're not a great entrepreneur unless you feel like you don't have a choice but to build a company because otherwise you're going to quit anyway.
[39:46] Ben Horowitz: It's not what you feel, it's what you do. Just focus on what you do.
Original transcript
[0:00] You're not a great entrepreneur unless you feel like you don't have a choice but to build a company. You don't take out a competitor with an equally good product. You need something dramatically better. We went public on $2 million in trailing revenue. $2 million going to $75 million. That was the forecast. That's clearly a bad idea. The doubts are going to be there. Ignore the doubts. The worst doubts aren't the ones people put on you, it's the ones you should put on yourself. If your guts aren't boiling, you're not even trying it. a company, but it doesn't make being a founder any easier. In this episode, Andrew Chen sits down with A16Z co-founder Ben Horowitz to talk about what AI changes and what it doesn't about entrepreneurship. They discuss why AI companies can reach massive audiences faster than previous technology waves, whether billion-dollar businesses can be built with tiny teams, and what actually creates defensibility when everyone has access to similar models. But they a mate, making hard decisions when you don't know if you're right, maintaining confidence through uncertainty, and continuing to bet on yourself when the doubts inevitably show up. I don't know if you remember this, but many, many years ago when I first met you, so I visited Ben at the LoudCloud turned Ben Opsware offices. Every... conference room was named after a hip-hop artist. I remember that. And I was like, all right, this is going to be a unique guy, unique culture. And I remember asking you, since you were about to sell the company to HP, what you were going to do next after. And I remember you told me that you were going to retire next. That's what you said. I was tired. And so the funny thing is, of course, Ben has done exactly the opposite of that, like, yeah, what happened? You decided not to retire. You decided to do the opposite. How did that come about? I had another idea. That's a lot of what life is about is getting your ideas out. I had an idea. And I think that's what makes an entrepreneur, right? Like you have an idea and you believe in your idea and you want to get it out. As an entrepreneur, that's who I am. That's what I got to do. I don't have a choice at that point. So you're not a great entrepreneur unless you're a entrepreneur. You feel like you don't have a choice but to build a company because otherwise you're going to quit anyway. So forget about it. You've seen so many stages of the Internet. We've had desktop software, then the web and then mobile and cloud. And then now this AI thing. What is unique and special about AI that makes it maybe similar or also just different to AI? to some of the other stuff. So the thing that's very reminiscent of the early days of the internet is when the internet first happened, like every day you would look out on the internet and there would be some very new awesome thing that was going on. And that's definitely the case with AI. So that's what's really similar, I think. Everything else is kind of different though, in that AI is, so if you think about the history of computing, of the history of the industry, always been networks and computers. And so, you know, we had microwaves and integrated circuits and, you know, mainframes and SNA and PCs and LANs and, you know, smartphones and the internet. And they're very different technologically and very different adoption curves, but then they really enhance each other. And I think in this era, AI is the new computer, like very clearly. You know, it's a probabilistic computer where we, you know, previously had deterministic computers And so, you know, in that way, it's very different than the internet, which is definitely a network and networks. Networks have these adoption curves that can start, like can get very high on the hype cycle too early because there's not enough people on the network. You can see where it's going, but until it gets there, it's not there. And so particularly with the internet, if you go back to kind of the great, dot-com crash that led to the great telecom crash. What happened was, There were everybody, every investor was like, oh, this is the biggest thing ever. I'm going to put all my money into it. But the truth was, there was only like at Netscape in 1996, we had 90% browser share and we had 55 million users. And so the whole internet wasn't that big. And then half of those people were on dial-up. So you could build, you could go build an internet app and it would cost you a lot of money because you had to use some microsystem servers. MC storage and all this. But you could get to like 30 million people or something. So good luck with that. That's not really a business. And so the whole thing kind of collapsed before it rebuilt. But eventually it was going to get there. Whereas with AI, when it works... instantly, you know, you know,
[5:00] Almost instantly, you can get to a giant number of users. We've seen companies grow faster than any companies we've ever seen because you don't need more people on the network. That's not a thing. The Internet's already here. Everybody's on it. And all you have to do is put your thing up and they can get to it. So it's very different in that way. One of the amazing things. Things that I think everyone in the industry is so shocked by is how good LLMs are at writing code. And who would have thought? And so there's this idea that maybe building software, writing software, becomes as easy as creating content on the internet. There's the idea that maybe you end up with a bunch of billion-dollar companies being built with much smaller teams. Maybe it becomes even further geographically democratized, or maybe there's going to be solopreneurs with billion-dollar companies. Do you agree with that? What risks do you think come with that? Do you think that's a lasting trend or do you think you need to still scale businesses ultimately? if they hadn't got bought by Meta, probably. And so I think you can get to... a giant audience and a ton of money with a very few number of people. I think that's for sure true. And then the question is, can you sustain it or will you get taken out by somebody who's doing more things? I would tend to think just historically that people always want more. The standard just goes up in terms of what people want. John Maynard Keynes had this very famous, like 1950. 30... 4 and 35, which basically said, as soon as people's needs are met, So once they had like shelter, food, and close, people would start to work maybe no more than 15 hours a week. And that turned out to be spectacularly wrong because, of course, then – All of a sudden, everybody needed a car, and then everybody needed a car for everybody in the family, and everybody needed a television, and everybody needed television in every room in the house. And now everybody needs a smartphone, and everybody needs air conditioning. And none of that stuff even existed when he wrote that. And so now people are working like – I'm sure nobody here is working 15 hours a week. I think what people need does tend to expand almost infinitely. So to fulfill that need, sometimes you have to build a bigger company, but we'll see. Software gets so easy to build that you kind of have fast fashion companies. SaaS, you know, like you kind of get copycats all the time that kind of get get cloned the other version that has kind of been predicted is sort of you know, sort of bespoke software. Maybe everybody uses a different, slightly different, you know, version of, you know, software that gets generated, that's instantiated at basically at runtime, that's personalized to you. I mean, this is all very sci-fi, but like, you know, maybe we'll get there. It's funny how humans are though, because like even in fashion, we do have fast fashion. I'm looking at a lot of people are wearing kind of similar stuff. My question on kind of software is, like how much will be like what we consider software today and how much will just be like, okay, we'll just use the AI differently and it won't even be an application. Right. Just ask the AI to do it. And so I think that's a little bit TBD, but I think the paradigm of an application process. maybe the thing that fades the most. You organize your data, and then you have agents that you ask to do stuff against the data. Right. And that's how business applications work, and personal applications work, too. How do you think about defensibility in kind of, especially for AI apps? You know, there's sort of an argument that, hey, a lot of the products right now, they don't really have... very much in the way of, you know, network effects, maybe they're GPT wrappers, you know, there's all these kind of critiques. And so what do you think of as kind of, you know, the durable advantages that, you know, founders should be? GPT wrapper thing was always a bit of a dumb idea in terms of not to build one, but to call it that. And what I mean, so if you go back, if you're as old as me, you remember when relational databases came out and relational declarative language, and then you just tell the database what to do, and then it does it. And so why the heck would you need SAP or in those days, Siebel or Salesforce or whatever
[10:00] Like, that's ridiculous. There's nothing there. It's just a RDBMS wrapper. But of course, that turned out to be wrong. And when I look at the applications that are built on LOMs today, they themselves have many models. I think Cursor's This is a 14 models now, like like fairly sophisticated models on top of, you know, the kind of foundation model. And so I don't think they're right. Like, I think they're pretty sophisticated kind of models and applications and so forth. And those models have a lot of data and understand, you know, like if you're talking to like super high end programmers all day about how to design their program. that's a lot of information into your system and into your model that nobody else has. And so that becomes, I think, Thank you. part of the differentiation. And then possession is nine-tenths of the law and like competing on software. So if you get the customer Like how much better... is ChatGPT than Grok right now. Like if you were starting from scratch, it's like kind of marginal. It's not like a massive differentiator and it's not a network effect yet. Like just the fact that you're using that one, I'm used to it. It's not even a complicated interface. You're just talking to it, like to switch it. But it doesn't matter if that's what you're using. That's very likely what you're going to continue to use. And to get you off that, you need something dramatically better. So like you don't, take out a competitor with an equally good product or a slightly better. It's like a 10 times better product. You know, if you take the market, I think that's probably the most obvious, most significant thing. And then, of course, with AI, if you take the market, you do get more data about your particular kind of user. Yeah, I mean, it seems like maybe the market's just going to, naturally evolve in phases. And, you know, we're kind of in this like land grab phase where just the most important feature is the it works feature, right? And then we'll get to the unique UX innovations and to all the, you know, collaboration network effects features over time. You know, when you have a new computer and you can solve problems that nobody could solve before, then that's like pretty interesting, regardless if there's a network effect. I think in the the internet era, the The network effect thing became super powerful for a lot of things. And it probably will again. But yeah, I think for right now, and it'll be interesting to see with like the rise of like crypto is probably a money network for AI. So how does an agent pay another agent or get money? And, you know, because it turns out AIs aren't allowed to apply for credit cards. So how does that work? And so, you know, does that become... a crypto network and then you have network effects from value exchange and that kind of thing. I think that'll start to emerge and other ideas like that. Yeah, I think on the phases too, it's so interesting to see how the backgrounds of founders changes so much because for all the folks that were around then, it was like, When the web first came out, you literally had Stanford PhDs building Yahoo, like a directory of links, right? Because you had to write your own web server, you had to write your own infrastructure, the whole thing. There was no such thing as, yeah. I mean, in those days, there was no application servers, no load balancers. No, it was very hard to scale software to that size. And nobody ever had built, like the biggest piece of software, like the audience was however many employees they had at IBM. 500 million people are going to use. That's right. And then it took like 10 years to get to the LAMP stack and non-CS major coders, and then eventually like Airbnb, the kind of design-dominant, UX-dominant people building the most interesting applications. So you could maybe make the argument that all these AI apps will follow some kind of a similar trajectory. We have tons of PhD CS talent right now. And then, you know, it may be maybe years before. I mean, it's very interesting, obviously, you know, these days, like I think the mainstream developers actually able to build AI apps. in a way that was not true five years ago. Yeah, and when do things plateau is a big question. So right now, the language models are... plateaued quite a bit in terms of, um, just the pure LLM and then the reinforcement learning side is scaling. So code is getting way better, but the regular kind of you know, chatbot is Not that different. How many more AI breakthroughs will we get? And that'll, you know, every time there's a breakthrough, that's an architectural change. It'll change. the entire application layer, the whole infrastructure layer. You know, when will those plateaus happen? When will those breakthroughs occur? And that's kind of how the opportunities line up. You know, you could kind of make the argument that,
[15:00] as the tech landscape has changed, like having social media, having podcasts and the whole thing, that's really changed the VC industry. And maybe the landscape and the way that you reach founders is kind of very intertwined. And so in this next phase of everything that will happen with AI and maybe technology, people being able to build big companies with small teams and that kind of thing. How do you think this might affect venture as an asset class or as a sort of just the VC industry as a whole. Yeah, so it's funny because the VC industry kind of was – The same from like... From 1970 to 2008, 2009, we started in 2009. Kind of a little bit the way VCs were set up, they were kind of built not to change. And that's just kind of how it rolled. And so companies got easier to build with the Internet. Like that was the first big change is that like if you're talking about building some microsystems or tandem computers and so forth, to build professional services. You needed a kind of direct sales force and so forth. That was a pretty complicated just corporation to build. Now, and then if you fast forward to Twitter, five people with laptops and AWS and Thank you. Very different. And then, okay, with AI, there's kind of the next... step of that. So I think that it's almost certainly going to change the VC landscape again. You know, in the old days, like that, you know, you could only do so many. Every bet was very expensive. It's hard to do. They were very likely to fail and so forth. And then kind of The hit rate, at least for the promising companies, went up. So that's a big change. And then kind of how, you know, like a big value in a VC is – okay, you have company A and company B, And they have both the exact same idea and both really good teams, which one wins. It turns out that you're kind of traveling as an entrepreneur through this complex adaptive system, you know, path dependent system. And so whoever gets more momentum ends up winning. And so a big role of a VC is to kind of help you get that momentum. So, oh, I raised money from the best VC. Therefore, like smart engineers will join my company and not their company. product, then I pick up more customers and then I have more money and I get raised the next round faster and away I go and that's how you win. If there are like... just dramatically more companies or they don't need as much money or that kind of thing, then that'll be an interesting change. But I think like a lot of it is still like a lot of the function of VC is, going to be? How do you give momentum. And that's like momentum and, you know, how do you help them build a higher quality company? How do you make that company more attractive to customers and employees and follow on investors and so forth? And that that part will be the same. But I think that everything underneath the feature set is going to be different. Well, and we often encounter companies that are. wildly profitable or doing very well, but they still want to work with VCs for all the reasons that you're Yeah, because you have to win. I mean, like the – Building, you know, the greatest thing is kind of having a great idea that makes the world a better place and doing something larger than yourself and building a team, building a company. Like if that works, like there's something better than that. Like it's it's so awesome. But if it doesn't work, that's a very high degree of misery. It's just tough. You hired all the smartest people you know. You went out and raised money. Your mom invested in the company. She's pissed. It sucks. It really does. So anything to kind of help you succeed at that, I think, is always going to be attractive. So most of the speed rent companies here in October, they're going to be doing their demo day in front of 1,000 investors in San Francisco. kind of speed dating a ton of angel investors and seed funds and all that? Do you have any tips and thoughts on how to evaluate, you know, who the best partner might be at that seed stage? The thing that you always don't have is you don't know how to do that CEO job. No founders we work with who haven't been CEOs of giant companies like know how to do that. And it turns out to be like a really complicated skill set because you're, you know, you have to select the right talent. like I had never been a CFO. I didn't know what the hell that was. Like what's a control structure? What's a, like how do you talk to Wall Street? Like how would I know any of that? Right, like so, but you still have to hire that person, which I always liken to, it's like hiring
[20:00] Japanese interpreter, if you don't know Japanese, it's like, they all sound good. And so like you have that problem and then you've got, you know, like and everybody's got their own career ambitions and, you know, and then, you know, Mark Zuckerberg's offering your people a billion dollars. Like, so it's a really hard job and you don't know how to do it. A lot of the times. like, what, When I watch founders who kind of fail at kind of doing that job, It always comes down to a crisis in confidence. So... You know you don't know what you're doing. You know it's hurting the company. But you still have to make decisions. Having the confidence to make that decision when you don't know what you're doing and you don't know if that decision is right, but not like hesitating. And then, you know, when you hesitate, It makes the company nervous. It slows everything down. It creates politics because people try and fill that void. So how do you do that? How do you kind of get to the right confidence level where you can make high quality decisions on time? If there's an investor that can help you do that, like that's invaluable. So, you know, those are probably the two things I'd look for. Somebody who can really give you momentum in hiring and getting to the next round and all that. And then somebody who you can work with who goes, wow, like this person, not only knows what they're doing, but like it's helpful because like every, by the way, every angel investor, every VC will give you advice, even if they like motherfuckers who had never been CEO. I see telling people how to be CEO and they give them these crazy ideas, you know, like you have to build a whole executive team right now. And it's like, what are you talking about? That's going to wreck the company, you know, but they have like, they have great confidence with their advice. Well, my favorite piece of advice I got is a, Here's the key. Hire A players. I was like, "Okay." My plan was to hire a bunch of fucking morons, but now I'm going to hire eight players. It's just like the level of... Like not good insight that's out there is high. And I really think it's part of a strength of the Bay Area tech ecosystem that over the last 20 years, it kind of went from... A lot of VCs, especially out in Europe and other places, are often really former finance folks. And that's really their background. And they're often coming in from maybe private equity. And so then they decide, oh, I'm going to also do some VC. And one of the big strengths, I think, of the Bay Area is that you end up working with a lot of former operators, folks. And the whole industry is kind of, move towards that model of, you know, trying to mostly get investors that have done something. Mostly. Mostly. Not universal. And there are different levels. There are levels to the game. For you, how has A16Z changed and evolved, you know, kind of the biggest beats of kind of evolution? And what are you most excited to build at the firm over the next decade? I know it's a long time, but... Yeah, I mean, so when we started, like the idea, the idea is very simple, which was, you know, the VC product we thought was very good for investors, but not nearly as good for entrepreneurs in that the kind of model was, OK, you get like a very smart person on your board. Not much more than that. Like they did. They had a network, but it was that person's network. A lot of them were very smart at product, but product advice runs out really in the first three months because when you're doing the initial kind of product. And then once you get past that, the knowledge gap between you who's been working on it all day, every day, and a person showing up at a board meeting, be it once a month or once a quarter or whatever, is just so vast that it's hard to get good advice in that way. I had thought of it as like, okay, as a founder who had tried to build a company and build a company, like, what did I actually need? And I was like, well, I could use a whole lot more than that. And the biggest thing, you know, and I go back to it, is just like confidence. And so, like, could I, like, what would make you feel like a CEO? And so, like, we did a lot of things. So, like, how do you make somebody feel like it's, well, what if you could call, like, any senior executive who might buy your product, like at FedEx or Citibank or wherever? could you do that? So we built like a whole team to make it easy to do that. And then, you know, how could you like know how to hire an executive and like what was the process for doing that? And so like, and we built a team to kind of help with that. Like, how do we introduce you to like five great CFOs and you can just sit down with them and say, hey, what would you hire in this position? And like, why would you do that? And how would you test for it? Yeah.
[25:00] in an interview and that kind of thing and kind of like skill you up so you knew how to do the job. So that was kind of like the first idea. And then, you know, as we got into it, one of the things, you know, and Mark wrote about it in, I think, 2011, he said, like, software is going to eat the world. Our takeaway from that was, okay, if that's going to happen, then kind of VC up until that point was this idea that there were 15 companies in a given year that would ever get to 100 million in revenue. And the whole idea of the firm would be to kind of get into as many of those 15 as you could. But we're like, well, if software is going to eat the world, that's not going to be 15, that's going to be 150. And so how do you actually scale a VC firm You know, kind of without losing all the things that make it a VC firm, like very, very personal, but can scale up. And so that was kind of the verticalization idea. So we're like, well, can we build like a firm that's better than Andreessen Horowitz original in infra, in apps, in crypto and bio and, you know, in games and, you know, and kind of defense and public safety and that kind of thing. I would say quite the major change in the firm. And then, you know, the kind of thing that we've run into now is like, Companies are getting very big as private companies, and there's a whole other layer of stuff that they need how to do. How do you go... international, like efficiently. You know, can we, walk you into Japan and get you your first three deals. You use our brand and enable you to lever off that so that you can win. How do you add channels? Because you have to be kind of multi-country, multi-channel, multi-product as you go. How do you do all that? And what can we as a firm do to enable that kind of thing? So we're playing with a lot of those ideas and other things that will hopefully make us better. Well, I think we have maybe 25 minutes to do audience questions? Yeah, so I think the biggest thing that I have concluded that I'm feeling pretty sure about is AI kind of video and entertainment and so forth is a new medium. So there was kind of like plays and then there was film and then there's going to be AI film. And I think it'll be as different from film as film most of them plays. So in the beginning of the movie industry, they would just film plays. And, you know, they like fix camera, do the play and we'll record it. And that was the idea. And then it obviously changed dramatically over the intervening decades. And I think that's going to happen. I think that's like in the beginning, it'll just be using AI to kind of generate certain scenes and movies and this kind of thing and so forth. But I think what's going to happen is whole new set of creators using the new medium to do things. You know, like it's interesting, like if you think about in the context of social media, social media is like storytelling where the main character is always you, right? So that kind of, carves out a lot of people who are either less comfortable or less narcissistic, however you want to say, who doesn't want to be the main character, you know, all over social media like that. kind of as a constraint. With AI, you can have characters that aren't you and be telling stories. So it's a new storytelling medium is the way I would think about it. And so I think it's a mistake to try and just completely replicate the movie experience or the TV experience or that kind of thing and kind of move it. And like it may start off being more interesting and like podcasts and like how do you jazz up podcasts or whatever. But I don't know exactly. But the one thing I'm pretty sure about that, I don't think we really covered in that. podcast was This is a new thing. This isn't the same. This isn't just generating the old thing automatically. This is a new thing. I think a lot of it, you know, for me, it was always like, can you bring the focus back to everything? you know, what you can do. Because like the thing that overwhelms you is what you can't do and like all the problems and everything. When I think about like the decisions I made as CEO, the ones that I'm most proud of and would have cost us most weren't actually difficult decisions. They were just difficult to make. And that was like a psychological thing. So for example, when we went public, like we went public on $2 million in trailing revenue, trailing 12 months revenue, like IPO, 2 million going to 75. That was the forecast. That's clearly a bad idea. And we were only 18 months old, 18 months from founding, 2 million trailing revenue, like,
[30:00] We went public. The other choice, and we had kind of gone through, we talked to like literally every investor in the private markets. It was bankruptcy. So that was obviously a worse choice than going public. To be able to do such a bad choice like going public and not hesitate, just go, okay, I'm going to do it. That was the magic of that decision. By the way, I can't tell you how many companies went bankrupt at that point because they didn't go do that. A lot of it is like, okay... Just do the thing you can do. And it may look horrible. But you've got to trust your eyes. You have to go do it. and not hesitate. And I see like this is where, you know, kind of people really get into trouble with their psychology. So they'll be like, oh, my God. you know, like I kind of know that we need to re-architect the product. But like that's going to make it very hard to raise money. That's going to, you know, we're going to miss our numbers. We're going to like not have any features and this and that. But I know we're. wrecked if we don't. So then they hesitate. And then they're going to be wrecked. And so it's like all these really or like something simple, like I really need to fire the head of sales. But like I don't want to have that conversation. So I'm going to look this way, you know, and that's running at the problem. Like, yeah, it's going to be bad. You're not going to have a head of sales like you may whip the quarter. There's going to be nobody to blame but you blah, blah, blah, blah. All that is true. But it's not as bad as leaving that motherfucker in place. Right. Like you got to go do it. I always kind of came back to that, that what can I do? And I better do that right now. And that's a lot how you work your way through those. Very bad. And the worst thing is like you think about like, what are all the implications of it? It doesn't matter. It's like, is that the best choice I have? If that's it, I'm doing it. You know, like if I think about all the things that are going to happen to me when I went public, and they all did happen to me, by the way, like I would have just shot myself in the head. I mean, that was... That was a bad time. When you're a CEO, when you're a founder, you're the boss. So you do have to keep that in mind. So I do think that nobody's going to train you to be a CEO. There is no training for that. And actually, the reason I wrote the book the way I did was... it does tend to be very situation specific. I spent a lot of time on like, this was my situation. And the reason for that was, So that people could go, okay, if you're there, then you do this, as opposed to here's the five steps for being a CEO. There are no five steps. I think that what you look for in a mentor is not, it's not really instruction. It's like somebody to have a conversation with who knows what they're talking about. You know, when I talk to CEOs, I'm just like, okay, like what's on your mind? What do you work on? That kind of thing. You know, we get to... like something that they're stuck on. And then I like the good thing is I've made all those mistakes before. So like I can go like this is how it went. As an example, like I had CEO and he said, well, you know, like, I need your help on this. How would you think about this? My CTO is an asshole." And I said... OK, you know, like and there's a lot of ways you could react to that. Right. You could imagine how like many advisors react to that. And I said, well, you're not asking me if you should fire him because he's a good CTO. Right. Like I mean, like everything I've heard from you is he's like executing way better than everybody you've had before. He's like, yeah, yeah, I'm not going to fire him. So I said, you're you're really asking me how do you talk to him? And he goes, yep. And I said, well, because you don't want to lose him. So you don't want to say something and piss him off. You're not going to fire him. But you don't want him to be an asshole. So how do you deal with that? And I said, well, tell me why he's an asshole. And he says, well, you know, the other day he was talking to a young woman in finance and he made her cry. And I was like, okay, that's kind of asshole-ish. And I said, look, here's the way I would have the conversation if I was here. I would just sit him down. I would say, look, you're a great director of engineering. amazing ship products on time and so forth. But you're not like... really effective as a CTO and Like, at some point I'm going to have to hire somebody who is an effective CTO until you can do that. And I, and by that, I mean, like, You can't just work with your own people effectively. You have to work with other functions. You've got to orchestrate the company to get the technology part of the company what they need. And if you go into a meeting and make like a little kid cry, like you're never getting what you want out of her or her boss or her boss's boss. Like, like that's just totally ineffective. So, like, if you want, like we can work on helping you get more effective or like like that's going to happen one day. And so, like, that's the kind of conversation I know it's a lot of detail.
[35:00] But you kind of want somebody who can have a conversation like that. where you can go, okay, that's actually some functional... advice that makes sense that I can use as opposed to just no asshole rule, you know, or like some dumb ass shit that they fucking heard at Harvard Business School. It's very situational. Like everything is situational. Everything is your company, your product, your people. How do they work? Can you get to that? And so I would like to just, people who you know, who you can talk to about that. They could be a peer. It doesn't have to be like... An old person like me, it could be like somebody who's just one step ahead of you. Yeah, so everything's so easy now compared to when I was building LoudCloud. Pushing the company forward gets harder in some ways because you have a lot of people who've been there a long time who did a good job. you know, a great job at one time, but maybe like we're in a new era that's past their skillset. Venture capital is very like that. You know, we don't invest in any consumer internet companies anymore, you know, like mobile apps or, you know, SaaS companies. Like there's a whole category of experts that we've had in the firm that, That expertise is no longer relevant. So if they don't become world-class experts in something else, that's a problem. The whole way we did marketing originally, that doesn't really work well for most companies and definitely can't work for us because we're too big. whole new skill set and going direct and like, you know, building, you know, a massive podcast network and presence and so forth. And so making those changes and just going, look, you know, You did a great job, but now we got to go this way is probably the thing that I enjoy the least. But, you know, you got to do it. Too many times I think people look at what the technology is as opposed to what it's going to be. That's tricky because sometimes you can extrapolate things that shouldn't be extrapolated. So a big example of that is when we looked at 3D printers, everybody was like, oh my God, this is the future. You're going to be able to print anything. But it turned out 3D printers don't follow Moore's law. So more compute power didn't make them better. which doesn't kind of follow that trajectory. So actually, 3D printers weren't going to change that fast or that much. So it was a little harder to kind of build a new company that was going to be giant in them. People looked at AI and they're like, oh, well, like, fine, but it hallucinates on this and that. And we're like, yeah, like, that's not that tough a problem to deal with. And sometimes the hallucinations are good. If you go, OK, these things that seem like a problem aren't really a problem. This thing, great. It's a bad phone. Like if you look at the 2007 reviews of the iPhone, it's a bad phone. It's a way worse computer than normal computer. The keyboard is much worse than a BlackBerry. Like it's like and you can go through all the things it was bad at. There was a camera and a GPS in your pocket, right? Like so. You can't build Uber on a PC even now, right? So you only need one feature, right? or one thing that kind of makes a breakthrough and creates a new future. Like it just has to be unbelievable at something. It doesn't have to be better than the last thing at everything. And so a lot of times when we think about it and we think about entrepreneurs, what I always say is like, we're not looking for lack of weakness. Like every, there's something wrong with every entrepreneur. And like, you know, including all of you, including me, like there's something like horribly wrong. But that's not the question. The question is, are they. world-class at a thing are they the best in the world at something do they have a secret that nobody else has do they have something really special so you're always looking for like what's really special what's the magnitude of the strength i'm not like Is there something wrong with them? And I think the future is a lot like that. Like, how strong is the strength? How different is this technology? Oh, we now have a computer that you can talk to in English? Like, that's kind of important. That's a hell of a thing. Any parting words of wisdom in many of the kind of year one projects here? is to like make your bet pay off. So like you believed in yourself enough to do it. You got to, there are going to be many times when you question that bet and you question that belief, but like that gets you nothing. Don't worry about it. Like it's,
[40:00] It's a nothing. You just got to do your thing and keep going and focus on what you can do and not worry about The doubts are going to be there. ignore the doubts. And the worst doubts aren't the ones people put on you. It's the ones that you put on yourself. Every entrepreneur has them. Like there's nobody... I spent a lot of time with Mark Zuckerberg when he was 22 years old, and he had a tremendous doubt about whether he was going to make it. Everybody has that. It's not what you feel, it's what you do. Just focus on what you do. you're going to feel scared. If your guts aren't boiling, you're not even trying at this thing. Don't worry about it. All right. What a way it ends. Thanks for listening to this episode of the A16Z podcast. If you liked this episode, be sure to like, comment, subscribe, leave us a rating or review and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts and Spotify. Follow us on X at A16Z A16Z.substack.com. Thanks again for listening, and I'll see you in the next episode. com forward slash disclosures. Thank you. you
