AI Interview Wiki
主题综述

数字台账 · Numbers That Matter

教 AI 经济学与规模感的硬数字。同一主题族内按时间排序以显趋势线；相互冲突或彼此更新的数字相邻放置并标注对照；最亮眼的 19 条附逐字引用。

更新日志

推理成本与算力经济

成本曲线：同等能力的推理价格以每年 10x—100x 的速度崩塌，而总消耗按 Jevons 悖论指数上升；算力与收入的换算关系首次有了公开口径。

"We exited at 23 at 2 billion in ARR, so 200 megawatts, 2 billion. We exited at 24 at 6 billion, so 6 billion, 600 megawatts. And we exited last year a little over 20 billion, 20 billion, 2 gigawatts. Actually, it's been accelerating."
「我们 23 年收官时 ARR 是 20 亿美元——200 兆瓦对 20 亿；24 年收官 60 亿——600 兆瓦对 60 亿；去年收官略超 200 亿——2 吉瓦对 200 亿。而且这个比例还在加速。」
Sarah Friar (OpenAI CFO) · Episode 12 - State of the AI Industry
"The cost of a million tokens on an O1 equivalent model in December of 24 was about 26 bucks. And then in November of 25, it was 30 cents. So we saw another 88X drop, not 88%, you know, 88 times cheaper in 11 months for that next generation of models."
「o1 等级模型一百万 token 的价格，24 年 12 月约 26 美元，25 年 11 月只要 30 美分——又一次 88 倍的下降。不是 88%，是便宜 88 倍，只用了 11 个月。」
Elad Gil · From SaaS to AI-First
"And the entire weights of the model would be about 100 gigabytes. And, you know, with some distortion, they remember the entire Internet. ... Whereas this KVCache thing, you take a few tens of kilobytes of article and it becomes those 80 gigabytes of brain state."
「整个模型的权重大约是 100GB。而且，他们在一定程度的失真下，记住了整个互联网。……而这个 KVCache 的东西，你用几十 KB 的文章，它就变成那 80GB 的脑状态。」
Dan Biderman (Engram) · Memory and Continual Learning: Engram's Dan Biderman and Jessy Lin
"And you're like, wait a minute, this is the same open source model with the same NVIDIA hardware. And there's a 5x performance difference and a multiple x throughput difference. And the way to just simply understand that is these companies are paying the margins of the cloud providers and running on top and making money."
「您会想，等一下，这是相同的开源模型，使用相同的 NVIDIA 硬件。而且还有 5 倍的性能差异和多个 x 的吞吐量差异。简单理解这一点的方法是，这些公司正在支付云供应商的利润并在其上层运行，获得利润。」
Eric Vishria (Benchmark) · Eric Vishria - A Decade of Lessons Investing in Software & Hardware

训练与 CapEx

两条主线：数据质量 > 数据数量（专家数据溢价成为行业定价）；capex 从美元计价换成吉瓦计价，回本算术仍没有答案。

"10 gigawatts is like $400 billion, something like that. And that $400 billion will have to be largely funded by their offtake, right, their revenues, which is growing exponentially. It has to be funded by their capital, the money they've raised through equity, and whatever debt they can raise."
「10 吉瓦大约就是 4000 亿美元。这 4000 亿主要要靠他们的承购——也就是指数增长中的收入——再加上股权融资和能募到的债务来买单。」
Jensen Huang · NVIDIA: OpenAI, Future of Compute, and the American Dream
"If you invest $150 billion in NVIDIA chips, that's about $300 billion of data center investments. And to pay that back, the person using the compute needs to earn a 50% gross margin. So there's about $600 billion of revenue that needs to get generated. ... The question behind the question was, is the customer's customer healthy?"
「买 1500 亿美元的英伟达芯片，约等于 3000 亿美元的数据中心投资；要回本，用这些算力的人得按 50% 毛利挣出约 6000 亿美元收入。……问题背后的问题是：客户的客户健康吗？」
David Cahn (Sequoia) · 20VC: Sequoia's David Cahn on The Winners and Losers in AI
"You could have gotten something that was ChatGPT 3.5 level maybe back in 2018 or 2019 with a couple people."
「你本可以在 2018 或 2019 年用几个人得到一些达到 ChatGPT 3.5 水平的东西。」
John Schulman (Thinking Machines) · John Schulman on dead ends, scaling RL, and building research institutions
"Training neural nets and LLMs specifically is a huge amount of code. But all of that code is actually complexity from efficiency. It's just because you need it to go fast. If you don't need it to go fast and you just care about the algorithm, then that algorithm actually is 200 lines of Python."
「训练神经网络、尤其是 LLM，代码量巨大。但那些代码其实都是效率带来的复杂度——只是因为你需要它跑得快。如果不需要快、只关心算法本身，那这个算法其实就是 200 行 Python。」
Andrej Karpathy · Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI
"We built in the last 15 months more Azure capacity than we built in the first 15 years."
「过去 15 个月我们建成的 Azure 容量，超过了最初 15 年建成的总和。」
Satya Nadella (Microsoft CEO) · The Rise of the Full-Stack Builder and Hyper-Leveraged Generalist
"…you can see the jumps just on 1800 tasks with about 500K in compute. These are pretty dramatic. Corporate log going from 4.7% to 26.6%. But notice that this is just the Apex Agents dataset we gave it, and it actually generalized incredibly well to GDPVal and Apex V1, which doesn't have these data rooms…"
「……你能看到在大约 500K 计算下的 1800 个任务的跳跃。这些变化相当显著。企业日志从 4.7% 上升到 26.6%。但注意，这仅仅是我们提供的 Apex Agents 数据集，实际上它在 GDPVal 和 Apex V1 上泛化得非常好，即使在没有这些数据房间的情况下……」
Brendan Foody (Mercor)（转录将 corporate law 误作 Corporate log） · RL Environments Explained: How AI Agents Learn Real-World Work

定价·毛利·商业模型

定价锚正从软件预算（美国 ~$1T）切换到人力预算（$20T—$80T）；客服是第一个 ROI 完全可量化的用例。

"We have simple like assistant type queries where you say, you know, draft me a document that a single query can cost $20. We have like a review product where you can upload 100,000 contracts and ask the models to review them. And some of those can cost $20,000."
「我们有一些简单的助手类型查询，比如说，给我起草一份文件，而单个查询可以花费 20 美元。我们还有一个审查产品，你可以上传 100,000 份合同并要求模型进行审查。其中一些可能会花费 20,000 美元。」
Gabe Pereyra (Harvey) · Harvey Co-Founder Gabe Pereyra on the Token Pricing Reckoning Coming for AI
"We were talking with developers and companies in our portfolio and they're like, yeah, we have Developers that are spending, you know, $3,000 per month themselves, like each on Cloud Code. So it's like, wow, okay, so that's $36,000 per developer."
「我们与投资组合中的开发者和公司交流时，他们说，确实有开发者每个月自己花费 3,000 美元，像是每个人在 Cloud Code 上。所以就像，哇，好吧，那每个开发者是 36,000 美元。」
Ev Randle (Benchmark) · Benchmark's AI Bets

增长与规模

收入爬坡的新基准线：$1B→$10B 从二十年压缩到一年；agent 采用曲线（任务时长每 4-7 个月翻倍）是能力侧最重要的单一指标。

人效与组织

AI 写代码的占比、单人杠杆与组织形态：人均产出取代人头数成为核心指标，token 支出开始超过工资单。

"These numbers are just totally crazy, right? Like 4% of all commits in the world is just way more than I imagined. And like you said, it still feels like the starting point. These are also just public commits. So we actually think if you look at private repositories, it's quite a bit higher than that."
「这些数字简直疯狂——全世界 4% 的提交，远超我的想象。而且这仍然只像是起点。这还只是公开提交；看私有仓库的话，比例还要高不少。」
Boris Cherny (Anthropic) · Head of Claude Code: What happens after coding is solved
"We used to be 6,000 or over 7,000 people and we're now less than 3,000. And I didn't ask for a single dime to do all this."
「我们过去有 6000 或超过 7000 人，现在不到 3000 人。我做这一切没要一分钱。」
Sebastian Siemiatkowski (Klarna CEO) · 20VC: SaaS is Dead: Why Systems of Record Will Die in an Agentic World
"Like right now, we're spending more on tokens for our internal agents than we are on employee headcount. And I think most businesses are going to look like that."
「就在现在，我们花在内部 agent token 上的钱，已经超过了员工人头开支。我认为大多数企业都会变成这样。」
Brendan Foody (Mercor CEO) · 20VC: Mercor CEO on Why Application Layer Companies Have No Defensibility
"Anthropic engineers on average ship eight times as much code per quarter as they did compared to 2025. Coding is no longer the bottleneck."
「Anthropic 的工程师平均每季度提交的代码量是 2025 年的八倍。编码不再是瓶颈。」
Lenny Rachitsky（主持人，引 Anthropic 官方推文；Fiona 确认） · Building the most AI-pilled engineering team in the world
"And so we blogged about minions, I think in January, February, and they were doing 1,200, you know, PRs per week. And last week, I think 7,000 PRs came from minions. … And about 30% of our PRs in that week came from minions."
「所以我们在一月和二月左右写了关于 Minions 的博客，他们每周做 1200 个 PR。而上周，我想说有 7000 个 PR 来自 Minions。……在那一周，大约 30% 的 PR 来自小黄人。」
Will Gaybrick (Stripe) · Stripe's AI Strategy: Build More, Not Less

估值与融资

AI 溢价（A 轮 +30% 持续一年未消）、幂律集中与泡沫算术。

其他关键数字

评估工程、世界模型、能源与其他不肯归类但值得记住的数字。

"AI alone, 88%. ... But then they gave the AI to the doctors. The doctors improved from 73% to 76%. The AI got degraded from 88 to 76%."
「仅靠 AI，准确率 88%。……但后来他们让医生使用 AI。医生从 73% 提高到 76%。AI 从 88% 降到 76%。」
Vinod Khosla · Vinod Khosla: Predicting the Future
"All of intelligence is not fluid, capable use of language. There's so much more. This is an important part. It's like 20% or a quarter of intelligence. There's more. We're not done."
「所有智能并不是流畅、有效使用语言的能力。还有更多内容。这是一个重要的部分。它大约占智能的 20% 或四分之一。还有更多。我们还没有完成。」
Rich Sutton · Rich Sutton and Khurram Javed: Why AI Models Stop Learning, and How to Start It Again
继续读
线索
阅读线索：agent 不到 9 nines 可靠，那应该卖给企业还是消费者
线索
阅读线索：如果 coding 是 commodity，AI startup 的护城河长在哪儿
线索
阅读线索：推理成本一年跌 10×，AI app 该按什么定价
线索
摘报
线索
摘报
线索
摘报