Home
›
VC
›
Venture Capital Is Good for Society and Bad for Most Founders
5 May 2026
Venture Capital Is Good for Society and Bad for Most Founders
风险投资对社会有益,对大多数创始人不利。两半都在实证文献中有据可查。大多数关于风险投资的写作只处理其中一半;这篇同时处理两半。
By Doug Scott, with Claude Opus 4.7, ChatGPT, Grok, and Gemini · 5 May 2026
Author note
Joint byline: Doug Scott (publisher, prompter, editor) with Claude Opus 4.7, ChatGPT, Grok, and Gemini contributing — all four AI tools fed into the work; Claude Opus 4.7 was the synthesiser that pulled it together. AI-generated, no human expert review. Empirical claims carry primary citations; readers making consequential decisions should follow the citations to the original sources.
风险投资对社会有益,对大多数创始人不利。 这一系统资助了其他形式的资本无法以同等规模资助的突破。它为这些突破付费的方式,是诱使许多聪明人接受一项个人期望值为负的财务赌注。两句话都已在实证文献中得到记录。多数关于风险投资的写作只讲第一句,不讲第二句。
对大多数创始人而言,创业之路是一项风险调整后 为负的财务赌注 — 对典型创始人而言财务上是不利的。整个创始人群体的平均回报可能为正,因为罕见的极端离群点支撑了整个分布。但典型创始人,在期望值意义上、在风险调整之后,过得比同期受薪工作的反事实更差。非财务方面 — 自主性、经验、可选性 — 仍可能值得。但对典型创始人来说,财务的那部分不值。
社会一面,三个数字
美国风险投资对 LP 的合并净回报:在长期视角下约为 14.3% (剑桥协会美国风险投资指数 ),领先于公开市场股票基准[强]。
经过风险投资管道的技术包括个人电脑、智能手机、商用互联网、现代生物技术、锂离子电池堆栈、大规模太阳能、mRNA 疫苗,以及当前 AI 浪潮所基于的基础模型。其中多项技术的奠基都依赖公共研发 — 支持 mRNA 平台研究的 BARDA 与 NIH 资金、互联网早期的 DARPA、电池与太阳能领域的 ARPA-E,以及基础模型继承的几十年学术研究。诚实的论点不是风险投资发明了这些技术。而是,风险投资在把底层的技术可能性转化为能够快速规模化、招聘并部署的公司方面,异常有效。这是一项不同的、也更难反驳的主张。
替代方案都试过。政府主导的研究基金擅长为早期发现提供资金;在公司组建上较弱。与国家利益绑定的战略资本能产生工业规模;在高方差押注上更慢。在社会账本上,风险投资系统的产出超过了它在资本中的份额。
个人一面,三个数字
Hall 与 Woodward(2010),《American Economic Review》 ,22,004 家风投支持的公司 :中位数创始人在公司生命周期内的总收入,低于他们在同一些年里从事受薪工作所能挣得的金额[强]。平均值为正,仅仅是因为罕见的极端离群点拉起了分布。典型创始人承受了财务上的损失。
Cooper、Woo、Dunkelberg(1988) ,2,994 名创业者:33% 把自己的成功概率评为 100% ,而五年存活率的基准更接近 50%[强]。创始人系统性地高估自身的概率。
Freeman 等(2019) :在 242 名创业者样本中,创始人报告的抑郁症患病率为对照人群的 2 倍、ADHD 为 6 倍、物质使用障碍为 3 倍、双相情感障碍为 10 倍 [中等;方向性模式在多项相邻调查中得到复现,包括英国 Sifted 与欧盟 Cerevity 的数据]。该研究的研究对象是创业者,并非专门是有风投支持的创始人;不能确立因果。这些升高的患病率究竟反映了选择(具备这些特征的人在创始人中被过度代表)、处境(创始人经历产生或加重了这些状况),还是兼而有之,仍有争议。
合成图景:大多数创始人相对受薪反事实承担财务损失,他们预期自己不会承担,而所进入的群体在精神健康方面的报告患病率显著更高;究竟是选择还是处境的份额,仍有争议。
为什么两半同时为真
这个结构性论证有四步。
基金经济学。 风险投资的回报遵循幂律。一只典型的 25 笔投资基金至少需要一个极端离群点,才能把资本翻数倍归还。
选择压力。 一只需要离群点的基金负担不起仅仅"还不错"的项目。它必须拒绝可靠盈利的公司,转而接受投机性公司。这不是错误;这就是数学。
创始人选择。 向基金路演的创始人,必须说服合伙人公司处于上端。承认中位数结果的创始人不会被投。选择机制是有意奖励过度自信的。
招募叙事。 要让正在路演的创始人群体足够大,以便寻找离群点能够奏效,更广义的文化必须说服许多人:他们可能就是那个离群点。Pitch deck、创始人播客、加速器的营销、成功故事的人物专访 — 这就是基金经济学在任何合伙人见到任何创始人之前,先在潜在创始人身上施加的作用。
系统产出的总量,取决于一个个读者被说服:基准比率不适用于他们。其中大多数读者错了;系统需要他们仍然行动起来。一个个传话者并没有撒谎。在局部理性的决策聚合起来之后,所呈现出的便是这一模式。
你正在做的决定
如果你是潜在创始人 ,你正在决定是否用五到十年的工作生命(以及一份已知的受薪替代选项),换取上文提到的基准比率。中位创始人的财务结果会更糟。典型创始人对自己成功概率的估计将远高于真实比率。你将加入的群体,在精神健康方面的报告患病率显著更高,且因果有争议。这些都没有告诉你具体的你是不是那个离群点。但当你决定时,这三件事都应当摆在台面上。
如果你是风险投资人 ,你的模型要求招募叙事持续产出尝试者群体。这一叙事产生了真实的福利成本,集中落在那些不是离群点的创始人身上。你目前的激励机制并未对这些成本定价;基金记下了回报;创始人吸收了其余部分。这种不对称是结构性的,数学并不要求它必须保持隐形。
如果你是政策制定者 ,风险投资系统所体现的取舍是真实的,且不易被现有可选的替代方案改进。把它当作单边的来行动 — 无论是采取吹捧立场,还是采取批判立场 — 都会比把两半都视为真实更糟糕地形成政策。
诚实的总结就是开头那句话。它对社会有益,对大多数创始人不利。两半并不抵消。一位读者读完后被说服:风险投资对人类整体净正面 — 在现有证据下他的推理是正确的。一位读者由此推论出"我个人应该进入风险投资体系" — 他正在做出一个不同的决定,关于他自己,而不是关于这个体系 — 而那正是招募信息所设计要去影响的决定。
下一步去哪里
带原始文献引用的实证细节:The reality of being a founder 。
"招募叙事"作为独立成篇的论证:For prospective founders 。
结构性机制的详细版本:The power law and what it forces 。
分析长篇版(约 26,000 字;PDF 与 zip):VC: most fail, most suffer, some win lots 。
How this piece was made
Architect: Doug Scott. Builders and checkers: Claude (Anthropic), ChatGPT (OpenAI), Grok (xAI), Gemini (Google). The publication was produced by Doug running parallel conversations with the four AI tools and routing work between them by hand — pasting Claude's output into ChatGPT for critique, taking ChatGPT's pushback to Grok for a different angle, feeding Gemini the result and asking what was missing. The cross-critique was a loop Doug routed manually; the AI tools did not talk to each other directly. Doug held the publication's intent, decided which output to keep and which to feed where next, and decided when the loop had converged. He did not edit the prose, check citations against primary sources, or verify the model math. No human expert with relevant domain expertise reviewed any of this work before publication. A specialist reader is invited to find errors the manual AI cross-critique did not catch — see the corrections page .
The Longer Look is the side door. The project is the room. This publication is one of eight sites by Doug Scott published April–May 2026, written with all four AI tools — Claude Opus 4.7, ChatGPT, Grok, and Gemini — contributing substantively; Claude Opus 4.7 was the synthesiser that pulled the threads together. The analysis on this site is the door for a specific cohort: founders, government policy advisers, senior people at AI labs. The other seven sites are the rooms that door is meant to lead to. None of the eight is intended to be read alone. No human expert reviewed any of the work before publication.
All articles on this site → · Selected pieces in other languages →
Where to go next
For the full map of pieces by reader type, see the reading guide . For routing by government team, see the for-government page .
Editorial history
A dated record of substantive edits to this article. Trivial corrections (typos, link fixes) are not logged; substantive edits to argument, structure, or evidence are. The publication-level record is on the corrections page .
5 May 2026, 19:15
Third-round editorial fixes. (1) The phrase “negative expected value decision financially” was replaced with the technically correct “negative risk-adjusted financial bet — financially bad for the typical founder.” A reviewer flagged that “expected value” means the mean, and the piece's own argument is that the mean is positive (carried by outliers); the right framing is risk-adjusted, not expected-value. (2) The claim “none of these came through government R&D or philanthropic capital at comparable scale” was softened. Public R&D was foundational in several of the named technologies (notably mRNA, where US government funding pre-pandemic and during exceeded $30bn). The honest argument is that VC was unusually effective at turning underlying technical possibility into companies that scaled, not that VC invented the technologies. (3) Inline source links were added to Cambridge Associates, Hall and Woodward (2010), Cooper, Woo and Dunkelberg (1988), and Freeman et al. (2019). (4) The mental-health language was softened from “pay a measurable mental-health cost” to “enter a population with materially elevated reported mental-health conditions; whether this reflects selection, treatment, or both is contested.” Causation cannot be inferred from prevalence.
5 May 2026, 18:15
Cut from ~1,330 words to ~930 in response to a second editorial review. Opening rewritten to lead with the thesis stated directly twice in the first six lines, with the explicit risk-adjusted framing as the second beat (the wording was tightened further in the third-round fixes above). The apologetic "What this piece is not saying" section was deleted entirely. The fourth reader-typed ending (a citizen-level close) was cut as redundant; three reader-types are enough. The four mechanism steps were each compressed to a single sentence. No new material added; everything that survived was already in the previous version.
5 May 2026, 17:55
Rewritten from ~26,000 words to ~1,500 in response to a substantive editorial review. The reviewer's named thesis — "VC is probably net-positive for society but it achieves that by inducing a large number of smart people to take personally bad bets" — was promoted to the first sentence of the piece. The four-step mechanism (fund economics → selection pressure → founder selection → recruitment narrative) was made explicit. Hard reader-typed endings (founder, VC, policymaker, citizen) replaced the previous drift into general framing. The 26,000-word version was retained as the deep tier.
5 May 2026, 16:53
Initial publication as a new shorter piece extracted from the longer 26,000-word document. Written in response to reviewer feedback that the strongest argument was landing too late in the longer document and that the venture-capital body of work needed an entry-point piece in the same shape as the IHT category's lead pieces.
Share:
Copy link
Post on X
LinkedIn
Email
Share…
No tracking on these buttons. The X and LinkedIn links open in a new tab; the email link opens your mail client. The publication doesn't see the click and sets no cookies for it. What the site does and doesn't track →
← Older
The Wrong Winners Write the Books
Newer →
Both Halves of the Headline Are True