Paul Graham On Startups, Ambition, and Great Founders

Paul Graham On Startups, Ambition, and Great Founders — Y Combinator

Y Combinator

AI-assisted summary of Paul Graham On Startups, Ambition, and Great Founders — original source. Check the original for full context.

Source published:

Key ideas

  • Paul Graham argues that startup fundamentals remain unchanged across technological eras, even as modern YC founders tackle far heavier engineering challenges such as intercontinental ballistic cargo and on-demand oncology research. (1:22)
  • Contrary to assumptions that abundant capital and AI agents have killed the lean startup methodology, compute inference costs drop approximately 30x annually at constant quality, enabling deep-tech founders to bootstrap early milestones via simulations and expert validation. (10:00)
  • Modern artificial intelligence inverted historical computer science expectations by leaping straight to broad human fluency like a bluffing undergraduate rather than building up from narrow insect-level perfection, creating a jagged frontier where models tackle pure mathematics while struggling with basic local business hours. (12:35)
  • Relentless shipping speed remains the single most reliable predictor of startup success; despite powerful AI coding tools, human conceptual clarity continues to be the primary operational bottleneck. (15:47)

Detailed notes

## Background Published by Y Combinator via RSS on September 5, 2026, this episode features YC Visiting Partner Vivian Shen sitting down with co-founder Paul Graham at the accelerator's original Mountain View office. According to the show description, the conversation explores the enduring fundamentals of startups, the evolution of artificial intelligence, founder ambition, and the essential qualities that define great entrepreneurs.

## TL;DR Reflecting on Y Combinator's 47th batch and over two decades of investing, Paul Graham argues that while modern founders tackle heavier, frontier problems—ranging from rocket cargo logistics to on-demand cancer research—the core mechanics of company building remain unchanged. He examines founder psychology, asserting that day-to-day execution is powered by an acute fear of failure and an innate formidability rather than dutifulness or credential-seeking. Graham also pushes back against claims that lean startup practices or batch models have become obsolete in the age of AI, illustrating how modern models display a 'jagged frontier' of human-like competence mixed with blustery imperfection. Ultimately, he maintains that the next trillion-dollar companies will not be defined by trending themes or expensive tooling, but by founders who possess relentless speed, determination, and conceptual clarity.

## Atmosphere & style The conversation carries an intimate, reflective, and unpretentious tone, feeling like a candid fireside chat inside the historic cradle of seed investing. Graham speaks with characteristic self-deprecating humor and sharp philosophical clarity, tossing off memorable analogies about model trains and fantasy novels while dissecting decades of founder data. Shen's conversational, measured prompting allows Graham to casually dismantle Silicon Valley hype, creating an atmosphere that feels grounded, pragmatic, and rich with hard-won operational wisdom.

## Bulletin - `01:22` Paul Graham argues that startup fundamentals remain unchanged across technological eras, even as modern YC founders tackle far heavier engineering challenges such as intercontinental ballistic cargo and on-demand oncology research. - `10:00` Contrary to assumptions that abundant capital and AI agents have killed the lean startup methodology, compute inference costs drop approximately 30x annually at constant quality, enabling deep-tech founders to bootstrap early milestones via simulations and expert validation. - `12:35` Modern artificial intelligence inverted historical computer science expectations by leaping straight to broad human fluency like a bluffing undergraduate rather than building up from narrow insect-level perfection, creating a jagged frontier where models tackle pure mathematics while struggling with basic local business hours. - `15:47` Relentless shipping speed remains the single most reliable predictor of startup success; despite powerful AI coding tools, human conceptual clarity continues to be the primary operational bottleneck.

## Key takeaways - `00:00` The core mechanics of company building remain identical regardless of underlying platform shifts, whether microprocessors, internal combustion engines, or modern AI. - `01:22` Skeptics have claimed YC has peaked since 2008, but recent cohorts build technically heavier and more ambitious ventures than early batches dominated by lightweight consumer software. - `03:52` Search incumbents like Google were not displaced through direct frontal competition, but by the paradigm shift of users querying AI directly for answers rather than sifting through web links. - `05:00` Day-to-day founder motivation is primarily driven by the raw fear of failure and public humiliation rather than calculated dreams of becoming a billionaire. - `07:14` Deep personal ambition is largely inborn; founders who appear to become more ambitious during an accelerator typically possessed the trait all along but had been conditioned to suppress it. - `07:14` Founding a startup cannot function as a resume credential because failure carries no institutional prestige, while success becomes an all-consuming life's work. - `07:14` Paul Graham defines a 'formidable' founder operationally as someone who consistently finds a way to get what they want in any situation. - `10:00` Capital-intensive hard-tech founders can still bootstrap early stages by calibrating output to their funding level—using simulations and designs to reach convincing validation before large capital raises. - `12:35` AGI is not a discrete finish line but a wide, uneven 'smear' that humanity is currently traversing, where systems surpass human capability in complex domains while failing at basic tasks. - `15:47` The primary economic divergence for modern AI startups is operational expense: budgets have shifted from purely engineering salaries toward substantial recurring hardware and inference costs. - `15:47` The YC batch experience delivers value by resolving founder isolation, crowdsourcing peer solutions to technical blockers, and providing an internal marketplace of rapid early adopters. - `17:50` The accelerator batch structure was originally an accidental experiment devised in 2005 to test angel investing on college students instead of traditional summer software internships. - `17:50` The next trillion-dollar company will not emerge from trending market categories or thematic niches, but strictly from the exceptional caliber and determination of its founders.

## Detailed walkthrough ### `00:00` The 47th Batch and the Enduring Fundamentals of Startups

Sitting down at the original Y Combinator headquarters in Mountain View, Vivian Shen introduces Paul Graham just before he delivers his traditional address to the incoming batch. Shen notes that this marks the 47th YC batch and the 21st year of the accelerator's existence. Graham admits he had lost track of the exact tally, remarking with self-deprecating humor that having given the talk so many times explains why he does not feel anxious about it. Even after two decades, Graham insists on preparing his batch remarks from scratch every single time: a couple of hours before the presentation, he catches himself realizing he ought to plan what he is going to say, opens a blank text file, and starts jotting down notes.

When asked if there are fresh revelations or new themes compared to the preceding winter batch, Graham observes that the core playbook of company building remains untouched by passing trends. Whether founders are building around microprocessors, artificial intelligence, or the internal combustion engine, the essential mechanics do not shift. As Paul Graham puts it: “Most of starting a startup is always the same, right? In, you know, microprocessors or AI or like the internal combustion engine, it's always the same stuff.”

`01:22` Beyond the "Good Old Days": Intercontinental Ballistic Cargo and Cancer Research

Responding to a recent post where Graham argued that modern YC startups tackle far heavier problems than those from the program's early history, he pushes back against nostalgic critics who claim YC has "jumped the shark." Graham recalls that critics first started claiming YC had peaked as early as 2008. Because detractors cannot credibly argue YC was never good, their only rhetorical avenue of attack is to insist that it used to be great but has now fallen off. In reality, Graham argues, the early days featured consumer web apps like Reddit—which, while undeniably valuable, pale in technical weight compared to the ambitious engineering under development today.

As an example of this newfound gravity, Graham cites a recent office hours session with a startup developing intercontinental ballistic cargo—effectively an ICBM designed not to detonate, but to land and deliver physical freight across the planet at rocket speed. He points out that founders are also taking on medical challenges like curing cancer. Rather than relying solely on preventative vaccines, one current batch company focuses on therapeutic interventions by delivering on-demand clinical and scientific research for cancer patients. Graham views this as inflicting a "death of a thousand cuts" on a complex disease that rarely yields to a single universal cure.

```mermaid flowchart TD accTitle: Approaches to Complex Startup Problems accDescr: Contrasting traditional consumer ideas with modern high-ambition batch concepts like rocket cargo and cancer therapies.

subgraph EarlyBatch["Early Era Focus"] A["Consumer Web Software\n(e.g., Reddit)"] end

subgraph ModernBatch["Modern Hard-Tech Ambition"] B["Intercontinental Ballistic Cargo\n(ICBM that lands freight)"] subgraph CancerTherapy["Cancer Treatment Frontier"] C["Preventative Vaccines"] D["On-Demand Patient Research\n('Death of a thousand cuts')"] end end

EarlyBatch -.->|"Evolution toward higher ambition"| ModernBatch ```

- Critics have consistently claimed YC "jumped the shark" in cycles since roughly 2008 because they must concede it was once great to argue it has declined. - Early batches were heavily dominated by lighter software ventures, whereas recent founders tackle frontier aerospace and specialized oncology.

`03:52` "Frighteningly Ambitious" Ideas and the Disruption of Google

Graham reflects on how the oncology startup originated directly from within the YC alumni network. The team had originally assembled to assist Sid Sijbrandij, the co-founder of GitLab, through his own cancer diagnosis. Sijbrandij approached his treatment with the analytical intensity of an operator—an experience he later described in an essay about going "founder mode on my cancer." Graham had thought at the time that someone ought to turn that exact protocol into a venture for everyone, and he was delighted to discover that the very team who assisted Sijbrandij entered YC to build that company.

This leads Shen to revisit Graham’s 2012 essay, *Frighteningly Ambitious Startup Ideas*, many of which have materialized over the ensuing decade. Chief among them was the creation of a "new Google." Graham had predicted in the essay that a search incumbent cannot be toppled through a frontal assault; rather, an attacker must wait for an underlying paradigm shift that renders the incumbent's core model obsolete. Graham explains that OpenAI, which was incubated at YC as one of Sam Altman's side projects, achieved this breakthrough: when users need information, asking an AI directly replaces the intermediary chore of sifting through web links. As Paul Graham describes the shift: “What I really wanted from searches was not web pages. I wanted information. Why not just ask the AI for the information, right?”

`05:00` Ambition, Fear of Failure, and the Model Train Set

Turning to founder psychology, Shen asks when Graham realized that exceptional technical talent is useless without intense ambition. Graham explains that he and the founding YC team recognized this before YC even launched because of their own grueling experiences as startup founders. Dutifulness alone collapses when confronted with the terrifying obstacles of entrepreneurship. While observers imagine founders are driven by visions of becoming billionaires, Graham argues that everyday execution is driven almost entirely by the raw terror of failure and humiliation.

To illustrate, Graham uses the analogy of a child tending to an intricate model train set. When a server crashes or an operational disaster looms, the founder is not calculating equity values; they are scrambling to stop the engine from falling off the edge of the table. Founders routinely keep their heads down for a decade guarding their fragile creation, only to be caught off-guard when a financing round suddenly renders them extraordinarily wealthy. As Paul Graham puts it: “You know what actually motivates founders day to day? The fear of failure.”

- Being purely dutiful fails because startup friction is too severe for conscientiousness alone to sustain. - Day-to-day founder motivation is dominated by the terror of public catastrophe, operational breakdown, and looking like a fool. - Graham notes that he is occasionally the first person to inform founders that they are billionaires after multiplying their share count by their latest valuation.

`07:14` Innate Drive, Credentialism, and What It Means to Be "Formidable"

Shen asks if founders can acquire ambition during a YC batch. Graham notes that true transformations are exceedingly rare because drive is largely inborn. Sam Altman, for instance, was visibly formidable before he was even accepted into the very first batch. The few founders who appear to become more ambitious during YC usually possessed the trait all along; they had simply been conditioned by pushy parents or regimented schooling to be compliant and suppress their independent desires.

Graham sharply criticizes the trend of treating a startup as a resume credential or prestige badge. At an elite university like Harvard, an admitted student can slide through by selecting an easy major and emerge with an impressive stamp. Startups, by contrast, offer no easy paths; Graham likens joining a startup for prestige to getting into Harvard and being forced to study theoretical physics. A failed startup provides virtually zero reputational cachet, while a successful one becomes an all-consuming life's work rather than a transient milestone.

Closing on the quality he values most, Graham unpacks the word "formidable," a term originating from the private vocabulary he shares with Jessica Livingston. As Paul Graham explains the operational definition: “And I described it in an essay. I said, I think that it's someone who gets what they want. That's the test, right? Do you get what you want?”

`10:00` Formidability and Whether the Lean Startup Is Dead

Investors seek out formidable founders because a formidable person is someone who consistently finds a way to get what they want in any situation. When an investor backs a founder, their financial incentives are completely intertwined through equity ownership; if the founder succeeds in bending the world to their will, the investor wins alongside them. As Paul Graham puts it: “If they get what they want, you get what you want.”

Addressing Patrick Collison’s recent remark at Startup School suggesting that the lean startup methodology might be obsolete—arguing that abundant capital and AI-driven parallel agent workflows change the math—Graham remains skeptical. He openly admits that he has never read Eric Ries’s *The Lean Startup*, joking that “Tolkien didn't read all those other fantasy books people wrote.” Defining lean startups simply as beginning with very little money focused on a tight problem, Graham dismisses the idea that building cheaply is dead. While token inference is currently expensive because of severe GPU shortages, underlying compute prices continually drop by roughly 30x annually at equivalent quality, meaning technology inevitably gets cheaper over time.

Even for capital-intensive domains like aerospace, Graham maintains that founders can bootstrap early milestones. Rather than attempting to manufacture an actual rocket on day one, founders adjust their output to their current funding: they build simulations, design specifications, and present data to domain experts. Reaching convincing early proof points unlocks subsequent venture rounds. Vivian Shen cites Philip from StarCloud, who raised funding simply by drafting a white paper and booking a rocket launch. Graham notes that while booking a launch creates undeniable conviction, an established expert can leverage personal credibility that an unknown recent graduate does not possess.

- Token prices drop roughly 30x per year at constant inference quality levels. - Early milestones for deep-tech companies rely on software designs, expert validation, and booking commitments before heavy capital deployment.

```mermaid flowchart TD accTitle: Capital Calibration Process accDescr: How founders make progress on limited capital by matching milestones to their budget. subgraph BudgetPhase ["Stage 1: Budget Calibration"] A["Limited Starting Capital"] --> B["Build Design & Run Simulations"] B --> C["Review with Domain Experts"] end subgraph FundingPhase ["Stage 2: Milestone & Expansion"] C --> D["Convincing Milestone (e.g. Booking a Launch)"] D --> E["Raise Tranche from Investors"] E --> F["Advance to Physical Build"] end ```

`12:35` The Jagged Frontier: Bullshitting Undergrads vs. Perfect Flies

When examining how modern artificial intelligence has surprised researchers, Graham contrasts today’s large language models with the 1980s symbolic AI paradigm he studied as an undergraduate, which he dismisses as an unworkable dead end. Decades ago, computer scientists assumed artificial intelligence would evolve hierarchically, starting with narrow, flawless biological models: first building a "perfect fly" that executed simple insect instincts flawlessly, then leveling up to mice, cats, monkeys, and eventually human-level reason.

```mermaid flowchart LR accTitle: Historical AI Expectation vs Modern Reality accDescr: Comparing the 1980s bottom-up perfection model with modern top-down human approximation. subgraph Historical ["1980s Expectation"] F1["Fly Brain<br/>(Narrow & Perfect)"] --> F2["Mammalian Brain<br/>(Mice & Cats)"] F2 --> F3["Human Intelligence"] end subgraph Reality ["Modern LLM Reality"] R1["Full Human Scope<br/>(Flawed & Hallucinating)"] --> R2["'Bullshitting Undergrad'"] R2 --> R3["Iterative Refinement<br/>toward Perfection"] end ```

Instead of ascending from narrow perfection to broad intelligence, modern generative systems did the exact reverse: they leaped immediately to the breadth of full human language, but behaved with the unreliable bluster of a student bluffing through an essay. Paul Graham highlights this inversion: “So instead of starting with perfect and then working your way up to human, you start with human and then work your way towards perfect.”

This uneven distribution of competence is what researchers term the "jagged frontier." Graham observes that while models can make inroads on unsolved frontiers in pure mathematics, he still struggles to get an assistant to reliably report local restaurant operating hours. When asked for his operating definition of Artificial General Intelligence (AGI), Graham sticks with the Turing test, noting that having to revisit the formal definition was proof that the milestone was drawing near. He rejects the notion of AGI as an abrupt finish line, explaining: “The finish line is actually this smear and we're on it.” Because the threshold has breadth and dimension, some technical capabilities are already far past human equivalence, while rudimentary day-to-day retrieval remains stranded in the weaker parts of the smear.

`15:47` Pace of Shipping, AI Costs, and the YC Batch Advantage

Despite the widespread adoption of AI tools, Graham insists that the primary benchmark for predicting startup success remains unchanged: relentless shipping speed. AI tooling accelerates routine coding tasks, but human founders must still formulate the underlying product insights. Graham points out: “Everyone always asks me, like, with all this AI, is anything still the same? And the answer is so far, almost everything is exactly the same.” Many founders in the active YC batch continue to ship too slowly because the constraint is conceptual clarity rather than mechanical keystrokes.

The single major departure from historical software economics is the composition of operational expenses. Software startups traditionally spent their budgets almost entirely on engineering salaries, laptops, and basic infrastructure. Today, hardware compute and model inference bills have introduced intense balance-sheet demands, with early-stage companies frequently burning tens of thousands of dollars per day on tokens and GPU rentals.

Evaluating the structural value of the Y Combinator batch experience, Graham identifies three decisive advantages: - Combating the inherent isolation of founding a company by giving founders peer colleagues without surrendering corporate autonomy. - Accelerating technical troubleshooting by allowing founders to crowdsource immediate solutions to engineering blockers from batchmates facing identical hurdles. - Access to the internal marketplace known as the "YCGDP," where batch peers act as immediate early adopters who test products, decide rapidly, and provide candid, high-velocity feedback.

`17:50` The Accidental Origins of the Batch and the Next Trillion-Dollar Company

The batch model itself was an accidental discovery born out of Graham and his co-founders’ desire to set up a dedicated seed-stage angel firm. In 2005, the venture market was bifurcated between individual angels writing personal checks and venture capital firms conducting late-stage rounds with onerous legal terms. To learn how to invest on standardized paperwork, YC ran an experiment to fund a group of college students over the summer as an alternative to corporate software internships at companies like Microsoft.

Graham recalls assuming that undergrads would tolerate amateur investors because they were amateur founders: “We thought, all right, we're going to always do investing that way.” The experimental cohort quickly transformed the students into real founders and the organizers into institutional investors, solidifying the synchronized summer batch format. Addressing persistent criticisms that YC has grown too big, Graham emphasizes that skeptics made the exact same claims when classes reached 40 startups. By dividing large batches into localized groups of 70 companies, YC reproduces the same intimacy and cohort dynamics that existed in 2012.

When asked where the next trillion-dollar company will originate, Graham insists that future market caps will not be dictated by trendy sector ideas like dog-walking apps or specific AI niches, but by the caliber of the people leading them. Paul Graham concludes: “The answer to your question, where are the next giant company comes from is it comes from the right founders, right?” Armed with twenty years of longitudinal data from Y Combinator, he asserts that top-performing founders look and behave exactly as they did two decades ago, driven by the same enduring traits of determination, speed, and formidability.

## Notable quotes > `05:00` You know what actually motivates founders day to day? The fear of failure. — Paul Graham

> `00:00` Most of starting a startup is always the same, right? In, you know, microprocessors or AI or like the internal combustion engine, it's always the same stuff. — Paul Graham

> `03:52` What I really wanted from searches was not web pages. I wanted information. Why not just ask the AI for the information, right? — Paul Graham

## Mentioned Paul Graham, Vivian Shen, Y Combinator, Reddit, Sid Sijbrandij, GitLab, Frighteningly Ambitious Startup Ideas, Google, OpenAI, Sam Altman, Jessica Livingston, Patrick Collison, Startup School, Eric Ries, The Lean Startup, Philip, StarCloud, Alan Turing, Microsoft, YCGDP

## Video description <p>YC Visiting Partner Vivian Shen sits down with Paul Graham at the original YC office in Mountain View to talk about startups, AI, ambition, and what makes great founders.</p>

[Source](https://anchor.fm/s/8c1524bc/podcast/play/125222193/https%3A%2F%2Fd3ctxlq1ktw2nl.cloudfront.net%2Fstaging%2F2026-8-4%2F431209191-44100-2-a7e65b9e8e208.mp3)

中文简报

保罗·格雷厄姆谈创业、雄心与优秀创始人 回顾Y Combinator第47期项目以及二十多年的投资历程，保罗·格雷厄姆（Paul Graham）指出，尽管当下的创始人正在应对更硬核、更前沿的问题——从火箭货运物流到按需癌症研究——但创办公司的核心机制依然未变。他剖析了创始人的心理，坚称日常执行是由对失败的强烈恐惧和与生俱来的强大气场驱动的，而非循规蹈矩或对资历的追求。格雷厄姆还驳斥了在人工智能时代精益创业实践或批次模式已经过时的说法，并阐释了现代模型如何表现出一种“参差前沿”——既具备类人的胜任力，又夹杂着浮夸的缺陷。他最终坚持认为，下一批万亿美元级的公司不会由热门趋势或昂贵工具来决定，而是属于那些具备极快迭代速度、坚定决心和清晰概念构想的创始人。 保罗·格雷厄姆指出，尽管当代YC创始人正在应对洲际弹道货运和按需肿瘤学研究等远为硬核的工程挑战，但贯穿各个技术时代的创业基本面依然没有改变。 (1:22) 与充裕资本和AI智能体已导致精益创业方法论过时的假想相反，在保持质量不变的前提下，计算推理成本每年约下降30倍，这使深科技创始人能够通过模拟和专家验证以自筹资金的方式达成早期里程碑。 (10:00) 现代人工智能打破了计算机科学的历史预期，它并未从昆虫级别的狭隘完美逐步向上构建，而是像个虚张声势的本科生一样直接跃升至广泛的人类语言流畅度，从而形成了一个参差的前沿——模型能攻克纯数学难题，却在处理基本的本地营业时间时遇到困难。 (12:35) 持续不懈的产品交付速度依然是创业成功最可靠的单一预测指标；尽管拥有强大的AI代码编写工具，人类的概念清晰度仍是主要的操作瓶颈。 (15:47)

Reading summary

Y Combinator co-founder Paul Graham argues that despite technological shifts from microprocessors to artificial intelligence, the fundamental mechanics of building enduring startups remain anchored in founder psychology, relentless execution speed, and innate formidability. Technological Shifts and Expanding Startup Ambition (1:22) Graham counters persistent claims since 2008 that YC has passed its prime by pointing out that modern cohorts tackle far more complex engineering problems than early consumer software ventures like Reddit. Current batch companies pursue deep-tech frontiers, including intercontinental ballistic cargo systems designed to deliver physical freight across the globe via landing rockets, as well as on-demand clinical and scientific research platforms that offer therapeutic interventions for cancer patients. Psychology, Fear of Failure, and Formidability (5:00) Technical expertise collapses without intense ambition, yet Graham stresses that daily founder performance is not powered by abstract dreams of wealth, but by a visceral fear of failure and public humiliation. Comparing founders to children protecting fragile model train sets from tipping off a table, he observes that many remain heads-down for years before realizing financing rounds have made them wealthy. Because genuine drive is largely inborn rather than taught, Graham cautions against treating startups as prestige badges or resume credentials, defining the most critical founder trait as being 'formidable'—an operational ability to consistently bend circumstances to get what one wants. Capital Calibration and the Inverted AI Frontier (10:00) Skeptical of assertions that the lean startup model is dead, Graham contends that early-stage teams can still validate ambitious concepts on limited budgets by producing simulations, technical white papers, and expert reviews before deploying large capital pools. While early-stage burn rates now frequently include tens of thousands of dollars per day for GPU rentals and inference tokens, Graham notes that compute prices historically decline by roughly 30x annually at equivalent output quality. Meanwhile, AI capabilities have evolved in reverse compared to 1980s symbolic AI expectations: modern language models leap immediately to broad, human-like linguistic competence with uneven factual reliability—a 'jagged frontier' where systems can contribute to advanced mathematics while stumbling on local business hours. Enduring Execution and the Batch Engine (15:47) Even with widespread adoption of automated coding tools, shipping velocity remains the central predictor of startup traction because conceptual clarity, rather than raw coding throughput, remains the binding bottleneck. Graham identifies the core institutional value of the YC batch model as relieving founder isolation, accelerating peer-to-peer technical troubleshooting, and providing an immediate testing market via the internal network. By clustering large cohorts into smaller units of 70 companies, the accelerator preserves early cohort dynamics while reinforcing that trillion-dollar outcomes depend on founder caliber rather than trending sector themes. Graham concludes that the origins of the next trillion-dollar enterprise will not hinge on topical software categories, but on the timeless attributes of top-performing founders. However, reconciling rapid experimental iteration with the mounting capital demands of frontier AI infrastructure remains an evolving operational tension for early-stage ventures.

精炼阅读摘要

Paul Graham 认为，从微处理器、内燃机到人工智能，创业的核心机制与对伟大创始人的评判标准在过去二十年间始终未变；驱动日常创业执行的关键动力是对失败的恐惧与“达成目标的能力”（formidability），而未来的万亿美元级企业仍将取决于具有极快交付速度的创始人，而非特定热门赛道。 底层法则的延续与深科技雄心 (0:00) 在 Y Combinator 成立第 21 年及第 47 期开营之际，Paul Graham 指出，无论是基于微处理器还是人工智能，创立公司的底层逻辑完全一致。针对外界自 2008 年以来不断出现的 YC“已走下坡路”的论调，Graham 反驳称，批评者往往陷入怀旧，但事实上早期批次主要集中于 Reddit 等轻量级消费软件，而当前批次正在攻关技术体量极重的深科技难题。例如，有团队正在研发不携带弹头、以极速在全球降落送货的洲际弹道货运技术，亦有团队针对癌症展开按需临床研究，通过个性化治疗手段对复杂疾病实施“千刀万剐”式的打击。这一癌症项目源于协助 GitLab 联合创始人 Sid Sijbrandij 抗癌的团队实践。此外，Graham 回顾其 2012 年提出的“极其雄心勃勃的创业想法”，强调颠覆搜索巨头 Google 无法靠正面对抗，而是依赖底层范式转移——如同 OpenAI 让用户直接向 AI 获取确切信息，从而取代翻阅网页链接的中间步骤。 失败恐惧与“令人敬畏”的特质 (5:00) 在创始人心理层面，Graham 认为仅凭尽职负责（dutifulness）无法承受创业的巨大阻力，日常推动创始人前进的核心动力并非成为亿万富翁的幻想，而是对公开失败与陷入难堪的极度恐惧。他将创业比作照料一套精密的模型火车：当服务器崩溃或运营遇险时，创始人不是在计算股权价值，而是在竭力防止列车从桌角跌落；许多创始人埋头苦干十年，往往直到融资后由他人计算估值，才意识到自己已身价过亿。与此同时，Graham 强调真正的雄心很大程度上源于天生，极少能在孵化器中后天习得。他严厉批评将创业视为简历镀金的精英主义风气，指出在名校选修水课可轻松混得文凭，但在创业中失败毫无声誉溢价，成功则要求全盘投入。在 Graham 与 Jessica Livingston 的语境中，投资人最看重的品质是“令人敬畏”（formidable），其操作定义为“无论身处何种境地，总能想方设法得到自己想要结果的人”；由于股权利益高度绑定，创始人的成功即意味着投资人的成功。 精益创业存废之争与 AI 参差前沿 (10:00) 针对 Patrick Collison 提出充裕资本与 AI 智能体并行工作流可能使精益创业模式过时的观点，Graham 表示怀疑。他将精益创业界定为以极少资金专注解决紧凑问题，并指出尽管当前 GPU 短缺导致模型推理 Token 成本高昂，但在同等质量下算力成本正以每年约 30 倍的速度下降，低成本构建依然有效。即使在航天等高资本领域，早期创始人也能通过软件模拟、专家论证或预订火箭发射等方式，在极低资本消耗下确立信任里程碑以撬动后续融资。谈及现代人工智能，Graham 将其与 1980 年代无功而返的符号 AI 进行对比：过去学界预期计算机会从无瑕疵的“完美苍蝇”昆虫本能逐级演进至人类智力，现代大语言模型却完全相反，直接跃迁至涵盖全部人类语言但充满幻觉的“胡说八道的本科生”状态，再向精确性逐步修正。这一能力分布构成了所谓的“参差前沿”（jagged frontier）——AI 可以在前沿纯数学领域取得突破，却难以稳定回答本地餐馆营业时间。Graham 认为通用人工智能（AGI）不是一条清晰的终点线，而是一片人类正在经历的模糊过渡带。 交付速度瓶颈与社群规模网络 (15:47) 尽管生成式 AI 大幅缩短了日常编码时间，但 Graham 强调决定创业成败的首要指标依然是持续的交付速度（shipping speed）。当前许多创始人交付依然缓慢，症结不在于敲击键盘的速度，而在于缺乏对产品深层洞见的认知清晰度。在财务结构上，传统软件企业主要支出为人力薪酬与基础设备，而当前初创企业由于高昂的模型推理与硬件开销，每天往往需要烧掉数万美元用于购买 Token 和租用 GPU。在孵化模式方面，Graham 阐述了 YC 批次机制不可替代的协同效应：它打破了创始人单打独斗的孤独感，让同侪能即时众包解决工程阻碍，并依托内部互助市场“YCGDP”形成高频反馈与首批付费种子用户。YC 最初在 2005 年通过资助本科生暑期项目试水标准化种子投资，如今通过将数百家公司的大班拆解为每组 70 家的本地小组，维持了早期的亲密社群黏性。Graham 总结道，未来的万亿美元级企业绝不会由特定概念决定，而是完全取决于创始人的执行力和坚韧品质。 Graham 展现了对创业基础动力学极强的连续性信念，坚信无论外部工具和热门叙事如何更迭，认知清晰度与创始人的原始驱动力依然是决定商业生死的唯一核心。然而，随着 AI 早期推理成本和深科技硬件开支显著抬升，算力成本年降 30 倍的长期趋势能否在短期资本压力下支撑新一代创始团队跑通商业闭环，仍是深科技创业面临的现实考验。

Detailed notes — zh

## Background 本期播客由 Y Combinator 于 2026 年 9 月 5 日在 RSS / 网页平台发布。节目中，YC 客座合伙人 Vivian Shen 在加州山景城的 YC 原始办公室与保罗·格雷厄姆（Paul Graham）展开对谈，围绕初创企业、AI、远大雄心以及顶尖创始人的核心特质进行了深入交流。 ## TL;DR 保罗·格雷厄姆反驳了外界对 YC 走下坡路的批评，指出当前批次涵盖洲际弹道货运与癌症治疗等项目，雄心远超以往。他剖析了顶尖创始人内心真正的驱动力是对失败和出丑的恐惧，而非物质财富，并强调“拿到想要的结果（formidable）”是评估创业者的核心标准。此外，格雷厄姆探讨了 AI 从人类高级功能向底层演进的戏剧性现实与 AGI 的模糊边界，重申初创公司制胜的关键仍是极致的发版速度，并坚信未来的万亿美元公司必将来自具备顶尖精神特质的创始人本身。 ## Atmosphere & style 对谈在 YC 最早的旧办公室进行，氛围随性松弛却充满思想密度与反思色彩。格雷厄姆的叙事风格极具个人特色，善用“模型火车快要掉下桌子”或“托尔金不看奇幻小说”等生动隐喻，幽默中带着直切本质的犀利。两人在怀旧回顾与对当下 AI 浪潮的客观审视之间自如切换，整体节奏舒缓自信，传递出一种看透商业周期与技术迭代后的从容与笃定。 ## Bulletin - `00:00` 保罗·格雷厄姆指出本期 YC 项目的雄心远超以往，硬科技比例大幅提升，涵盖洲际弹道货运及基于个性化科研的癌症疗法等重工业与前沿医疗领域。 - `05:00` 驱动优秀创始人日常拼搏的核心推手并非亿万富翁的物质野心，而是对失败、系统宕机及成为笑柄的极致恐惧。 - `13:00` 现代 AI 的演进路径彻底颠覆了早期设想，呈现出从人类水准的模糊粗糙走向完美，而非从完美模拟低等生物逐步进化的特征；AGI 终点线更像是一片充满维度的模糊涂痕。 - `16:00` 尽管 AI 显著改变了算力和 Token 等成本结构，但衡量创业公司成败的核心预测指标依然是团队发布新东西的速度。 ## Key takeaways - `00:00` 无论技术浪潮如何从微处理器、内燃机演进至 AI，创办初创企业的底层核心逻辑始终保持不变。 - `00:00` 外界针对 YC‘大不如前’的批评早在 2008 年就已出现，但如今入选项目的技术雄心和严肃程度远胜早期的互联网项目。 - `00:00` 以个性化按需研究攻克癌症的团队入选 YC，其模式源于此前协助 GitLab 联合创始人 Sid Sijbrandij 以‘创始人模式’成功抗癌的实践。 - `03:45` 颠覆 Google 这类巨头无法正面强攻，必须依靠外部技术底层发生根本性剧变，OpenAI 满足了用户直接获取信息而非检索网页的本质需求。 - `05:00` 创始人通常专注于护住眼前的突发危机，往往是在经历长达十年的埋头执行并完成高估值融资后，才偶然意识到自己已积累巨额财富。 - `07:22` 创业绝非可以用来美化简历的短期跳板，其残酷程度堪比在名校被迫攻读理论物理，没有任何轻松划水的途径。 - `07:22` 顶尖创始人最关键的特质是‘Formidable’，其本质衡量标准在于无论面临多大阻碍，都能始终拿到自己想要的结果。 - `10:00` 反驳‘精益创业已死’的论调，尽管短期 Token 成本高昂，但推理成本每年呈数十倍暴跌，科技永远越变越便宜，创始人依然能以低成本分阶段推进里程碑。 - `13:00` AI 的现实发展呈现‘参差不齐的前沿’，它既能解出复杂的数学未解难题，又会在餐馆营业时间等常识检索上出现低级错误。 - `16:00` AI 时代创业公司唯一的根本性变化在于成本结构，从过去的员工薪资为主转变为每日消耗数千美元的 GPU 算力与 Token 账单。 - `16:00` YC 训练营模式的核心优势在于驱散创业孤独、提供即时的邻座技术互助，以及形成天然的‘YC 内循环生产总值’作为早期用户基本盘。 - `18:30` YC 创立之初选定三个月夏季批次纯属偶然，原本只是为了给大学生提供暑期实习的替代方案并借此积累早期天使投资经验。 - `18:30` 下一家万亿美元级别的公司并非由特定点子决定，而是必然诞生于具备强悍精神风貌的下一批顶尖创始人手中，且这类人才绝不会被 AI 替代。 ## Detailed walkthrough ### `00:00` YC 第 47 期与当下更严肃的创业项目 YC 客座合伙人 Vivian Shen 与保罗·格雷厄姆（Paul Graham，简称 PG）坐在位于加州山景城（Mountain View）最初的 YC 办公室中展开对谈。这是 Y Combinator 成立的第 21 个年头，也是 YC 迎来的第 47 期训练营。尽管 PG 已经为批次学员做过数十次开场演讲，但他坦言自己每次依然是从空白文档开始重新梳理思路，通常在演讲前两小时猛然想起“糟糕，我马上要给 YC 讲话了”，随后打开文本文件开始打字。不过，演讲的核心逻辑其实未曾大变，PG 强调：“Most of starting a startup is always the same, right? In, you know, microprocessors or AI or like the internal combustion engine, it's always the same stuff.”（创办一家初创企业的大部分事情总是相同的，无论面对的是微处理器、人工智能还是内燃机，底层逻辑从来都是那一套。） 针对近期 PG 发文提到“现在的 YC 资助的项目比所谓的‘过去的黄金岁月’要严肃得多”，他解释了自己对“过去的美好时光”这种说法的反感。自 2008 年左右开始，外界就频繁批评 YC 已经“走下坡路（jumped the shark）”，因为批评者无法全盘否定 YC 曾经的辉煌，只能退而求其次声称它现在大不如前。PG 指出，早期虽然诞生了像 Reddit 这样极其有价值的公司，但它们终究算不上跨大洲弹道运输这类重工业硬科技。 在当前批次中，创业项目的雄心有了质的飞跃。例如有团队正在做“洲际弹道货运（intercontinental ballistic cargo）”——运作机制类似洲际弹道导弹（ICBM），区别仅在于它最终是安全着陆卸货而非爆炸。在医疗领域，则有直面癌症治疗的团队。针对攻克癌症，行业既有研发疫苗的一端，也有聚焦个性化疗法的一端。当前批次的一家初创公司选择为癌症患者提供定制化的按需研究（on-demand research），对癌症施加“千刀万剐之刑（death of a thousand cuts）”，因为癌症并非那种靠单一特效药就能彻底解决的简单疾病。这家团队此前曾协助 GitLab 创始人 Sid Sijbrandij 成功抗癌，当时 Sid 凭借“创始人模式（founder mode）”对待自己的癌症，PG 当时就设想应该有人成立一家公司把 Sid 的抗癌经验产品化普惠大众，而如今正是当年协助 Sid 的团队站出来创办了这家企业。 - 当前为 YC 举办的第 47 期，也是机构成立以来的第 21 年。 - 外界关于“YC 不如从前”的论调早在 2008 年就已出现。 - 本期代表性硬核项目：洲际弹道货运（ICBM 式降落送货）以及基于定制科研的癌症治疗服务。 - GitLab 联合创始人 Sid Sijbrandij 曾以“创始人模式”与癌症抗争并取得积极成果。 --- ### `03:45` “可怕的雄心”与 OpenAI 的颠覆方式 谈及雄心壮志，Vivian 提到 PG 在 2012 年发表的经典文章《可怕的雄心》（Frighteningly Ambitious Ideas）。文章中列出的许多在当时看似不可思议的想法，如今已逐一实现，其中最为典型的就是“造一个新的 Google”。PG 在当年就预言过，要击败 Google 绝不能正面强攻，必须等待外部技术环境发生根本性剧变，导致现有的商业和搜索模型彻底过时。 OpenAI 的崛起正是这种颠覆的最佳印证。PG 发现自己在深度体验过 OpenAI 的产品后，彻底放弃了传统搜索：“What I really wanted from searches was not web pages. I wanted information. Why not just ask the AI for the information, right?”（我从搜索中真正想要的并不是网页，而是信息。既然如此，为什么不直接问 AI 获取信息呢？）这一曾经作为 Sam Altman 众多副业之一、在 YC 孵化出的技术平台，如今已经成为重塑整个互联网入口的核心力量。 --- ### `05:00` 驱动创始人的核心动力：并非百亿身家，而是对失败的恐惧 当被问及何时意识到优秀的创始人不仅需要出众的技术天分，还必须具备极大的雄心时，PG 表示这在创办 YC 之前他就深有体会。创立企业的过程极度痛苦煎熬，仅凭“尽忠职守（dutifulness）”这类责任心根本无法抵御扑面而来的巨大阻碍，必须有更深层的力量作为引擎。对于常人以为的“想成为亿万富翁”的物质野心，PG 认为那绝非创始人日复一日坚持的真正推手。 在日常运作中，真正驱使创业者拼尽全力的其实是“对失败的极致恐惧”。创业者每天满脑子想的都是害怕陷入灾难、害怕变成外人眼里的笑柄，或是服务器宕机导致系统崩溃。PG 用模型火车作比喻：当服务器宕机时，创始人绝对不会盘算“只要修好我就能成为亿万富翁”，他们唯一的念头是“天哪，我的模型火车头快要从桌子边缘掉下去了，我必须冲过去接住它”。正是这种不想让自己心爱的模型散架的本能，让创始人们埋头苦干整整十年。往往到了最后，当算清最后一轮融资估值下的持股价值时，他们才会猛然惊醒自己已经身家数十亿。PG 甚至多次充当了第一个为创始人做算术并告知对方“按照这轮估值，你已经是个十亿级富豪了”的角色，而创始人们的第一反应往往只是惊讶地附和“啊，还真是”。 ```mermaid flowchart TD accTitle: 创始人内在驱动力与财富结果演进路径 accDescr: 展示创始人从日常恐惧失败到十年专注投入，最终意外获得巨大财富的心智模型。 subgraph Daily_Drive["日常驱动力 (Day-to-Day)"] F1["对失败与出丑的恐惧"] F2["突发危机（如服务器宕机）"] F1 --> F2 F2 --> Focus["护住'模型火车'不摔碎<br/>（专注化解眼前灾难）"] end subgraph Long_Term["长期投入 (10 Years)"] Focus --> Work["长达十年的埋头苦干与持续执行"] end subgraph Outcome["最终结果 (Surprise Outcome)"] Work --> Calc["新一轮高估值融资交割"] Calc --> Wealth["突然意识到已成为亿万富翁"] end ``` --- ### `07:22` 创业不是镀金文凭，以及何为“不好对付的人（Formidable）” 关于创始人能否在 YC 期间从胸无大志蜕变为雄心勃勃，PG 认为这极其罕见，因为这种特质很大程度上是与生俱来的。例如 Sam Altman 在初次申请、尚未被录取时就展现出了令人侧目的强大气场。现实中极少数看似“后天觉醒”的案例，往往只是因为当事人从小被规训得过分顺从，听从父母或权威的要求而压抑了天性，但这并不代表他们缺少野心，只是被训练得不敢表露而已。 Vivian 谈到如今许多人甚至把“创过业”当作简历上的一枚镀金徽章。PG 驳斥了这种荒谬逻辑：如果创业成功，那将直接变成你毕生的事业，你根本不需要也不可能再去收集其他徽章；而如果创业失败，一枚“失败创业者”的标签根本谈不上什么含金量。早在 YC 发展初期，就有人像申请哈佛大学一样为了名气而申请 YC，但两者有着本质差异：“With startups there are no easy majors so right it's like applying to Harvard and being forced to study theoretical physics, it's really rough if you're not very very hard working and clever and determined.”（创业可没有什么容易划水的专业；这就像考进哈佛后被迫去学理论物理，如果你不是极其勤奋、聪明且坚毅，过程会非常痛苦。）如果仅仅是为了让自己显得很酷，去创业可以说是效率最低、最残酷的方式，你必须在数年漫长且残酷的煎熬中熬过去，才可能换来一丝外界的认可。 最后，PG 解析了他经常用来形容顶尖创始人的词汇——“Formidable（令人敬畏的/不好对付的）”。这个词源于他与妻子 Jessica Livingston 在创立 YC 前的私下习惯用语，用来定义某一类极其罕见的人格特质。PG 在文章中曾给出最直接的评判标准：“I think that it's someone who gets what they want. That's the test, right? Do you get what you want?”（我认为它指代的是那种总能拿到自己想要的结果的人。这就是检验的标准：你能否如愿以偿？） - 创业无法成为简历上的普通跳板：成功了便成为终身事业，失败了也毫无光环可言。 - 相比哈佛等名校可以通过选修简单专业混过关，创业没有任何简易路径，难度堪比主修理论物理。 - 评估顶尖创始人的核心标尺“Formidable”的终极测试：无论面对何种阻碍，始终都能拿到最终结果（Gets what they want）。 ### `10:00` 投资人为何偏爱强悍的创始人，以及“精益创业”真的已死吗？ 保罗·格雷厄姆（Paul Graham）深入剖析了投资人为什么极度偏爱具备“强悍（formidable）”特质的创始人。他指出，如果你无法得到自己想要的东西，那你根本算不上强悍；一个真正强悍的人，能够在任何复杂局势下达成自己的目标。投资人选择这类人，本质上是因为双方的利益高度捆绑：投资人持有公司股份，创始人同样持有股份，“如果他们得到了自己想要的，你也就能得到你想要的；这就是为什么投资人渴望那些强悍的人，因为投资人能分到创始人所获得成果的一杯羹。” 随后，主持人 Vivian Shen 转换话题，提到 Stripe 联合创始人帕特里克·科利森（Patrick Collison）最近在 Startup School 上抛出的观点——由于如今融资极其容易，加上 AI 让团队可以并行指派大量智能体（Agents）同步推进多项任务，“精益创业（Lean Startup）”模式可能已经走向消亡。对此，格雷厄姆直言不讳地表示自己其实并不确切了解何为“精益创业”。他风趣地调侃道，虽然 YC 资助的项目形态可能与《精益创业》那本书所描述的内容有重合，但他从未读过那本书，并打比方说：“托尔金可不会去读别人写的那些奇幻小说。” 针对“用极少的资金在极聚焦的领域起步”这一模式是否过时的质疑，格雷厄姆明确予以反驳。他承认眼下由于 GPU 短缺，Token 成本确实昂贵，但这属于供需带来的短期现象。在给定的推理水平下，模型推理成本正在以惊人的速度暴跌（比如每年降价约 30 倍），而且 Token 的质量也在不断提升。科技总是越来越便宜的，因此创始人依然能靠极少资金启动创业。即使是高昂的火箭创业项目（例如 StarCloud 仅凭白皮书和预订发射就完成了融资），创始人也完全可以根据预算调整推进手段——先画出设计图、做计算机模拟或找权威专家验证，拿到了里程碑成果再去说服投资人。“只要你能达到某种里程碑，你就能说服投资人给你更多资金。” ### `13:00` AI 发展的意外转向与“模糊的 AGI 终点线” 回顾自己在 20 世纪 80 年代学习人工智能的经历，格雷厄姆感叹如今 AI 的发展路径完全出乎所有人的预料。当年学者们所研究的 AI 根本行不通，简直像个笑话。当时大家设想的技术演进路线是从“底层完美”逐步往上推：先造出一只完美模拟苍蝇大脑的机器——它能做的事虽少，但每个动作都精准完美，之后再循序渐进地升级到老鼠、猫、猴子，最终迈向人类。 然而现实却戏剧性地颠倒了过来。现在的 AI 诞生时就直奔人类的高阶功能，却表现得像一个在期末论文里胡编乱造的本科生（“like an undergrad trying to bullshit his way through a paper”）。格雷厄姆犀利地概括了这种预期与现实的落差：“你没有从完美出发一步步走向人类水准，而是从人类水准出发，一步步走向完美。”如今的 AI 虽然能解决著名的数学未解难题，却连附近餐馆几点开门这种基础事实都会答错，生动地展现了所谓“参差不齐的前沿（jagged frontier）”。 对于 AGI（通用人工智能）的定义，格雷厄姆认为阿兰·图灵提出的图灵测试至今依然非常精妙。过去人们总以为 AGI 是一道清晰的终点线，跨过去就一目了然；但真正站在跟前时才发现，这道线其实是有厚度和维度的。 ```mermaid flowchart LR accTitle: AGI 终点线的认知转变 accDescr: 对比 1980 年代将 AGI 视作单根细线与如今将其视作具有宽度的模糊地带 subgraph Historical["20 世纪 80 年代设想"] A["明确而单一的终点线"] --> B["跨过即实现 AGI"] end subgraph Modern["现实发展状态"] C["具备厚度的宽幅涂层 / 抹痕"] C --> D["黎曼猜想等高阶推理 (已跨越)"] C --> E["餐馆营业时间等常识检索 (滞后中)"] end ``` - 格雷厄姆关于 AGI 的核心评述：“终点线实际上是一片涂痕，而我们正身处其中（The finish line is actually this smear and we're on it）。” ### `16:00` AI 时代创业的不变法则：发版速度与 YC 生态红利 尽管大模型带来了前所未有的工具红利，但在格雷厄姆看来，创业的核心法则并未动摇。他坚持认为，预测一家初创公司能否成功的最佳指标依然是“他们发布新东西的速度（the pace that they ship new stuff）”。即便工具再强大，本期 YC 训练营中依然有许多团队动作不够快。因为发版速度不仅取决于代码生产效率，更取决于创始人构思好点子的速度。 格雷厄姆强调，在 AI 浪潮下，商业的底层逻辑几乎全无变化。唯一发生的实质性剧变在于公司的成本结构：过去初创公司最大的开销就是员工薪资，笔记本电脑和日常开销相比之下极其便宜；而如今，创始人每天都要面临成千上万美金的 Token 消耗和巨额的 GPU 账单。 针对 YC 训练营（Batch）能为创始人提供何种竞争优势，格雷厄姆总结了三点核心价值： 1. **驱散创业孤独**：创业虽然拥有无限的上升空间与决策自主权，但本质上是极度孤独的。训练营创造了一种大家同处一室、拥有大量“同事”却各自为自身事业奋斗的独特氛围。 2. **高效的技术互助**：同批次创始人面临的技术难题往往高度重合，遇到瓶颈时只需跨过桌子询问邻座，就能迅速获得现成的解决方案。 3. **“YC 内循环生产总值（YC GDP）”**：无论团队做出什么产品，都能在训练营内部找到第一批乐意迅速做决策的早期种子用户，其他创始人至少会认真听取你的商业推介。 ### `18:30` YC 诞生的偶然机缘与万亿美元创始人的特质 谈及创立 Y Combinator 的渊源，格雷厄姆透露这纯粹是一系列偶然的产物。最初他们只是想建立一家针对早期天使投资的投资机构（Angel Firm），采用标准化的投资文件，填补当时大型 VC 与个人天使之间的空白。因为没有投资经验，他们决定一次性资助一批初创公司来“练手”。他们观察到大学生暑期通常去微软等大公司实习，于是萌生了提供暑期实习替代方案的想法——这也是 YC 训练营选在夏季举办、周期为期三个月的根本原因。他们当时的心态十分放松，觉得面对一群在校大学生，“就算我们不是真正的投资人，他们也不会介意，因为他们也不是真正的创始人”。然而在实际运转中，投资人和创始人都迅速完成了蜕变，这一偶然摸索出来的批次模式（Batch）便被永久沿用了下来。 面对外界多年来指责“YC 规模太大、已经走下坡路（jumped the shark）”的论调，格雷厄姆淡然回应：早在每期只有 40 家公司时就有人这么说了。如今 YC 将大型批次切分成大约 70 家公司一组的独立小组，在体验上与 2012 年每期 70 家公司时并无二致。 在访谈的最后，当被问及“下一家万亿美元级别的公司会诞生在哪里”时，格雷厄姆给出了极具个人风格的斩钉截铁的结论： > “下一家万亿美元公司来自哪里？答案是它来自正确的创始人。它不是来自某个特定的点子……它来自下一批万亿美元级别的创始人。” 他坚信，未来的优秀创始人绝不会被 AI 机器人所取代。基于 YC 过去 20 年积累的数据样本，顶级创始人的核心精神风貌与 20 年前完全一致，再过 20 年也同样不会改变。 ## Notable quotes > `18:30` 下一家万亿美元公司来自哪里？答案是它来自正确的创始人。它不是来自某个特定的点子……它来自下一批万亿美元级别的创始人。 — Paul Graham > `10:00` 只要你能达到某种里程碑，你就能说服投资人给你更多资金。 — Paul Graham ## Mentioned Paul Graham, Vivian Shen, Y Combinator, Reddit, Sid Sijbrandij, GitLab, Sam Altman, Google, OpenAI, Jessica Livingston, Patrick Collison, Stripe, StarCloud, J. R. R. Tolkien, Alan Turing, Microsoft ## Video description <p>YC Visiting Partner Vivian Shen sits down with Paul Graham at the original YC office in Mountain View to talk about startups, AI, ambition, and what makes great founders.</p> [Source](https://anchor.fm/s/8c1524bc/podcast/play/125222193/https%3A%2F%2Fd3ctxlq1ktw2nl.cloudfront.net%2Fstaging%2F2026-8-4%2F431209191-44100-2-a7e65b9e8e208.mp3)

Source quotes

[18:30] Paul Graham: 下一家万亿美元公司来自哪里？答案是它来自正确的创始人。它不是来自某个特定的点子……它来自下一批万亿美元级别的创始人。 [10:00] Paul Graham: 只要你能达到某种里程碑，你就能说服投资人给你更多资金。

Source quotes

[5:00] Paul Graham: You know what actually motivates founders day to day? The fear of failure. [0:00] Paul Graham: Most of starting a startup is always the same, right? In, you know, microprocessors or AI or like the internal combustion engine, it's always the same stuff. [3:52] Paul Graham: What I really wanted from searches was not web pages. I wanted information. Why not just ask the AI for the information, right?

Original transcript

[0:00] So we're here in Mountain View today with PG himself at the original offices doing your talk that you do for the YC batch. I looked it up. This is actually the 47th YC batch. You know, I've been trying to calculate that number and I wasn't sure exactly what it was. No wonder I'm not too worried about this talk. You know, I've done it that many times. It changes a little bit. Yes, it's the 21st year also of YC. You know, the weird thing is I make the talk from scratch every time. I probably say the same things over and over. But I always feel like I should – I think, oh, my God. You know, think a couple hours beforehand. Oops, I better – I'm just talking to YC. I better think in advance a little bit about what to say. And so I open up a text file and start typing things, and that's how it always is. Any tidbits from today that are new from winter? No, most of the things, most of starting a startup here, this is an actually interesting point. Most of starting a startup is the same. Most of starting a startup is always the same, right? In, you know, microprocessors or AI or like the internal combustion engine, it's always the same stuff. Good. Well, in that case, let's run through some questions. Okay. So you recently posted that the startup's YC funds now are much more serious than they were in the supposedly good old days. Yeah. What did you mean by that? Well, what I meant by supposedly good old days is people are always saying that YC has jumped the shark. You know, and like I remember when people started saying this, which is in about like 2008. Right. YC is always this thing that was great, but now it's not what it was. Right. Because they can't say it was always terrible. They have to admit it was once great. So if they want to attack it, they have to say it's fallen. Right. It was good, but now it's not. It's always been. It's always not good anymore. old days because the old days are what people claim used to be good. But like in the old days, there was a lot of things like Reddit. which, valuable as it is, is not exactly... intercontinental ballistic cargo, right? Which I did office hours with recently. How about that for an idea? Like it's like an ICBM except instead of exploding, it lands. It's ambitious. And drops stuff off. Yeah. That's serious, right? You don't get much more serious than an intercontinental ballistic cargo. Yeah. What were some of the other serious ones? Yeah. Well, curing cancer. There's a startup in this batch. I mean, there's all sorts of different approaches to curing cancer. And I say, fund many of them because like, anyone working would be good. So like one way to approach curing cancer is vaccines. That's one end, right? The opposite end is therapies, right? And so there's a startup in this batch that is... basically doing on-demand research for people with cancer. And what I like about it is they're sort of inflicting on cancer the death of a thousand cuts, which may be the way to defeat it because it's notoriously... It's notoriously not the sort of thing where you just like come up with a cure for cancer. I mean, maybe you do if these vaccines work. Maybe you do, right? Or maybe the way you fight it is the death of a thousand cuts. And they're actually really interesting because they helped another YC. Yeah, right. It's part of the whole YC community. They helped Sid from GitLab deal with his cancer, which he has done very successfully. I mean, he approached it like a startup, right? And it's funny because I remember thinking after that, you know, I always have this process going of what would be a good startup idea, right? And I remember after that saying, you know, somebody should start a startup to do for everyone what Sid did for himself. And now, not only is this startup happening, it's the people who actually did it for Sid. So it couldn't be better. I was so delighted when I realized that one was happening. Yeah, and I think his essay that I went founder mode on my cancer. Yeah, that's a very good way of describing what he did. is cancer. So along that vein, I was rereading the essay you wrote in 2012, where you coined the term frighteningly ambitious. Frighteningly ambitious ideas. A lot of those ideas have actually gotten done since then. Like which ones? Well, a new Google, that was one of them, right? And it turns out, look, And I sort of predicted this, right? That the way you make a new Google is not by attacking them head on. You have to wait for things to change so much that their model is obsolete. And that's when you can do it. And that's what OpenAI is, right? Like I realized very early on, like after using OpenAI, like, whoa, I don't use Google anymore. I don't use searches. What I really wanted from searches was not web pages. I wanted information. Why not just ask the AI for the information, right? Everyone's now had this experience. And it was incubated at YC. Yeah, yeah, yeah, yeah. one of Sam's many side projects. Yes. [5:00] And so along the same veins of around ambitious as a term, I am curious when you figured out why a founder needs to be not just unusually talented and technical ability or something else, but they actually have to be really ambitious as well. When did I figure out that founders have to be really ambitious? I think I already knew that before we started YC, because when we started YC, we already had all this experience of doing a startup ourselves. Right. And so we knew that it was grueling. and you need something to drive you through all those obstacles. And honestly, being dutiful is not enough. The obstacles are too fearsome for mere dutifulness to do it. So there's got to be something driving you and you call it ambition. And part of ambition, I assume, is wanting to be a billionaire someday. You'd be surprised, actually. Tell me more. You know what actually motivates founders day to day? The fear of failure. even though they're doing this thing that if they succeed they'll be super rich that does motivate them but the thing that motivates them at that at any given moment is the fear of like disaster looking like a fool the server crashing right like If you running some online service and your server crashing you not thinking I got to deal with this because if I deal with it I become a billionaire You're thinking, oh no, the server's crashing. It will be a disaster. The engine of my model train set is falling off the edge of the table. I need to go and save it. That's what you're thinking at any given time. You just don't want your model train set to break. You put your head down and work on the model train set for 10 years, and then you lift your head up and like, holy shit. Like if I add up the value of all my shares at the last round valuation, I'm a billionaire. But it takes people by surprise when it happens. Yeah, that's great. Sometimes I'm the first to tell them. Like I do the math. I'm like, wait, according to your last round, you're a billionaire. And like, oh, yeah, I am. That's the way we want it. And I'm curious, can you tell me a story about a founder where when they started YC, they were maybe not ambitious enough, and then they became very ambitious during the batch. Those stories are rare because most of this quality is inborn. And so, like, you know, Sam Altman famously was extremely formidable the first time I met him before we even accepted him And that more the rule than the exception So the one case where you see founders it not that they lack ambition It's more like they've been trained not to show it. Because when you're young, you're often in situations where you're supposed to be obedient. You're not supposed to go off and do what you want. You're supposed to do what people tell you. Maybe you have pushy parents who tell you to do what they want, right? And so you've never let yourself go. That's the one case. But even then, I would argue that the person was ambitious. They'd just been trained not to show it. Yeah. Actually, that's a good point because I think nowadays a lot of people think about startups like another badge that they should collect along their resume. Only if it fails. If it works, it's like your life's work. You don't have any more badges, right? And so what kind of badge is a failed startup? Not much of a badge. I think if people think that, they're not thinking it through, right? Like honestly, we would get, it was quite early on in YC's history when people started applying to YC for the prestige, right? Like they would apply to Harvard hoping they would get in. But the thing is if you apply to Harvard and get in you should do it because no matter how stupid you are there easy majors you can choose and sort of slide through you get a harvard degree and everybody will be impressed with you with startups there no easy major so right it like applying to harvard and being forced to study theoretical physics it's really rough if you're not very very hard working and clever and determined and so these people trying to get why trying to do yc as a credential they had no idea what they were asking for. They don't realize like this is not, this is a really, really unsuitable credential. There are many easier ways to get credentials. Yeah. Yeah. If you want to seem cool, like starting a startup is just about the least efficient way to do it. It's brutally hard. Yep. And you won't seem cool for years of years of brutal hardness before anyone will think you're cool. When you were describing some of the founders before you used the word formidable, can you tell us a little bit more about how that word came to be, what it means to you as well? Well, a lot of the language of Y Combinator is Maya and Jessica's private language, right? And so we already had this word before we started YC, like for a certain kind of person. And I described it in an essay. I said, I think that it's someone who gets what they want. That's the test, right? Do you get what you want? [10:00] Because how formidable are you if you don't get what you want? So it's someone who gets what they want in any situation. If you invest in them and you have stock in their company and they have stock in this company, your interests are aligned. If they get what they want, you get what you want. That's why investors want people who are formidable. Pretty good. Yeah, because they get a part of what the founder gets. Changing tracks a little bit, at Startup School, Patrick Collison recently said that lean startups may be dead. Yeah. Did he? And part of that is because he says, you know, it's really easy to raise money these days. AI lets you work on many things in parallel. You could, you know, spin out a bunch of agents and work on a bunch of things. Do you agree with him or would you have a rebuttal? I've never really been clear what a lean startup is. I mean, we already had a kind of startup that we funded in YC, and maybe it has a lot of overlap with what was described in this book, The Lean Startup, but I haven't actually read it. And so, I mean, why would I? Right? Tolkien didn't read all those other fantasy books people wrote. And so I don't exactly know what a lean startup is, but if you tell me what qualities you might think might be obsolete, maybe I could answer that. Let's say starting with... Just a little bit of money on a very focused area. you mean it might be dead starting with a little bit of money. Mm-hmm. I don't think so. I think there's a lot of things you can do for cheap. I mean, tokens are very expensive at the moment. But that's just because there's a shortage of GPUs. Token prices will... Certainly inference prices at any given level of inference go down dramatically, like 30x a year or something like that. So even token prices are somewhat misleading because you're getting higher quality tokens over time. So I think, no, everything, technology always gets cheaper. I think you can still start a startup on not much money. Even if it's a rocket startup or any of these? It's harder to start a rocket startup on no money. But even a rocket startup, you can start on not much money because what you do is you adjust for how much money you've got. You can't build an actual rocket. What you can build is you can build a design for a rocket and you can simulate it or show it to experts. If it's convincing enough, you can go to investors and get a bit of money to do the next stage. So you just do what you can on the money you got. As long as you can get to some kind of milestone, then you can convince investors to give you more. Yeah, I think Philip from StarCloud was saying that they basically just wrote a white paper and booked a launch and they raised the money for StarCloud. Nothing as convincing as booking a launch, right? It's got to happen if you booked a launch. But he's a huge famous expert. in the stuff he works on, right? So he comes with automatic cred. It's easier for him to do that than some kid who just graduated from college. So let's transition to talk a little bit about AI more broadly. I'm curious what surprised you about AI. Well... I was studying AI back in the 1980s, a very different kind of AI that would never have worked. I mean, it was a joke. So I've been thinking about this a lot, and I spent a lot of time thinking about what AI would eventually be like. And the way we thought it was all going to play out was you start out with – A perfect fly. Right. You know, it doesn't do much. It's like a fly, like a fly's brain. Right. It can only do what a fly does, but it does it as well as a fly. Right. And then you work your way up from flies to mice, you know, to cats. And eventually you get monkeys and then humans. Right. But at every point you're you're perfect. You're doing what they do well. like an undergrad trying to bullshit his way through a paper. Something plausible sounding. Full human stuff, like not even this fly stuff. Full human, but like terrible, right? And so you start, and so instead of starting with perfect and then working your way up to human, you start with human and then work your way towards perfect. So it's sort of the opposite, the opposite quality ended up, ended up being the one that got optimized. That, nobody expected that. it would be like a bullshitting undergraduate, right? Just a slot machine. Yeah, and it still has a lot of that flavor, doesn't it? Yes. I hear all these stories about how AI is solving famous open problems in math, and I can't even get it to answer questions about when restaurants are open. That's what's called the jagged frontier. Yeah. They must know things I don't know, or they have research versions. Maybe they can find out when restaurants are open. making them sound very simple. So what is the official PG definition [15:00] of AGI. Oh, well, honestly, I mean, could you do better than the Turing test? The Turing, I think Turing was pretty good, right? One of the reasons, one of the ways I knew we were getting close to AGI was I actually had to go and look up the definition of the Turing test. That in itself is a piece of evidence, like, what was the Turing test again? Are we there already? I think what's happened now is we thought back in the day, that AI would be sort of like this finish line, and AGI would be this finish line, we'd cross it, it would be very clear. But when you're standing on the finish line, you realize it has width. From back in the 1980s, it looks like it's just a line going across the horizon. But here you realize, okay, some bits of AI are way across the finish line, some bits are like maybe in the middle of it, who knew the line had dimension? The finish line is actually this It turns out to be a smear and we're on it. And that's why you can't get your restaurant times. Yeah, because they're in a bad part of the smear. Whereas the Ryman hypothesis, maybe we're already there. Okay, changing back to talk about startups. So the other thing you've said is that the best predictor of success for a startup is the pace that they ship new stuff. Yeah. With AI, does that change? Do you think there's a new metric? No, no. I mean, we have all these powerful AI tools and there's still a lot of startups in this batch that are not shipping fast enough. And so obviously there's variation in shipping speed. It's not just the rate at which you can produce things. You have to think of these ideas first, right? Everyone always asks me, like, with all this AI, is anything still the same? And the answer is so far, almost everything is exactly the same. The only weird new thing is that companies have these giant AI bills. It used to be startups' big cost was salaries. That was all that mattered, salaries, right? Everything else was cheap by comparison because it was just laptops and people. And now suddenly you've got to pay for GPUs. You know, giant, giant. I mean tens of thousands of dollars a day in tokens. I'm also curious, how does being in a batch like YC give founders an advantage? There's a bunch of reasons. One is starting a startup is normally a very lonely business. You have all this upside, you have complete control over what you're doing. The downside is you don't have any colleagues. This is like working in an office with all these colleagues, except you're all working on your own startup. So there's that. You have colleagues. You can also learn from what the other startups are doing at the same time you are. So you have some technical problem. Probably somebody in the batch had that problem. You can just ask them how they solved it, right? And another advantage is what we call the YCGDP, which is you have some sort of product. You can sell it to the startups in your batch. Almost no matter what it is, you can find some. kind of users you want. Early adopters who decide quickly, right? And they sort of have to at least listen to your pitch. How did you come up with the idea for YC? Uh, gradually. Um, initially the idea was just going to be, we were, we, we didn't have the batch idea at first. I, the part we had at first was to do, was to be an angel firm. Because at that point there were VC firms who would do these giant late stage round. And there were angel investors who were individual people investing their own money, but there was no, there were no angel firms. and use standardized paperwork. And then having decided that, we thought, all right, we need to learn how to be investors since we don't know how to be investors. We'll just fund a whole bunch of startups all at once. And so... we thought to ourselves, well, all these college students get summer jobs working at Microsoft. Wouldn't they like to start startups instead? We'll create this alternative to getting a summer job. That's why the batch takes place during the summer. It turns out to be the perfect length, but that was just an accident. And so we thought, we'll fund a whole bunch of startups. We'll replace summer jobs. It'll be for current undergrads. They won't mind if we're not real investors because they're not real founders, right? Just a bunch of college students. And it turned out in the course of that, we became real investors and they became real founders very quickly. And this batch thing was so great, this thing that we just discovered by accident. We thought, all right, we're going to always do investing that way. were not changed since it's grown. It's changed very little. It's basically... like the same thing but more of it. You know, like the problems the startups have are pretty much the same. There's a lot more of them. But they, you know, they live in small groups that are just like individual bits of YC. So one of the things people saying, when people say YC is jump the shark, one of the things they always say is, oh, it's so big now. Well, people used to say that back when we had 40 startups in the batch, right? They would say, oh, come [20:00] The old days are gone, there's 40 startups in the batch, right? People are always saying that, and if you divide YC up into little bits where each startup is part of a pot of 70 startups, then it's just like they were back in 2012 when there were 70 startups. Where do you think the next trillion dollar company comes from? What are the The founders of it are formidable. In fact, that is the answer to your question. The answer to your question, where are the next questions? 20 giant company comes from is it comes from the right founders right it's not some particular idea I mean it probably is some idea rather than others it's probably not dog walking right maybe who knows maybe I mean startups ideas are very mutable but it comes from the next Trillion dollar founders is where it comes from. And they probably have good ideas. Right? Whatever they're working on is probably promising. Do you think the founders in the future are going to look very different from the ones in the past? They're going to be robots. I hope not. No, no, no. Founders, they still, I mean, we now have 20 years of data. They still look exactly like they used to 20 years ago. Why should they change 20 years from now? Great. Thank you so much. Thank you.