The Cost of Staying
by Amy Tam https://x.com/amytam01/status/2023593365401636896
Every technical person I know is doing the same math right now. They won’t call it that. They’ll say they’re “exploring options” or “thinking about what’s next.” But underneath, it’s the same calculation: how much is it costing me to stay where I am?
Not in dollars. In time. There’s a feeling in the air that the window for making the right move is shrinking—that every quarter you spend in the wrong seat, the gap between you and the people who moved earlier gets harder to close. A year ago, career decisions in tech felt reversible. Take the wrong job, course correct in eighteen months. That assumption is breaking down. The divergence between people who repositioned early and those still weighing their options is becoming visible, and it’s accelerating.
I see this up close. I’m an investor at Bloomberg Beta, and I spend most of my time with people in transition: leaving roles, finishing programs, deciding what’s next. I’m not a career advisor, but I sit at the intersection of “what are you leaving” and “what are you chasing.”
The valuable skill in tech shifted from “can you solve this problem” to “can you tell which problems are worth solving and which solutions are actually good.” The scarce thing flipped from execution to judgment: can you orchestrate systems, run parallel bets, and have the taste to know which results matter? The people who figured this out early are on one arm of a widening K-curve. Everyone else is getting faster at things that are about to be done for them.
The shift from execution to judgment is happening everywhere, but the cost of staying and the upside of moving look completely different depending on where you’re sitting.
FAANG
Here’s the tradeoff people at big tech companies are running right now: the systems are built, the comp is great, and the work is… fine. You’re increasingly reviewing AI-generated outputs rather than building from scratch. For some people, that’s a gift—it’s leverage, it’s sustainable, it’s a good life. The tradeoff is that “fine” has a cost that doesn’t show up in your paycheck.
The people leaving aren’t unhappy. They’re restless. They describe this specific feeling: the hardest problems aren’t here anymore, and the organization hasn’t caught up to that fact. The ones staying are making a bet that stability and comp are worth more than being close to the frontier. The ones leaving are making a bet that the frontier is where the next decade of career value gets built, and every quarter they wait is a quarter of compounding they miss.
Both bets are rational. But only one of them is time-sensitive.
Quant
Quant still works. Absurd pay, hard problems, immediate feedback. If you’re good, you know you’re good, because the P&L doesn’t lie.
The tradeoff that’s emerging: the entire quant toolkit (ML infrastructure, data obsession, statistical intuition) turns out to be exactly what AI labs and research startups need—same muscle, different problem. The difference is surface area. In quant, you’re optimizing a strategy. In AI, you’re building systems that reason. Even the quant-adjacent world is feeling it: the most interesting work in prediction markets and stablecoins is increasingly an AI infrastructure problem. One has a ceiling. The other doesn’t, or at least nobody’s found it yet.
Most quant people are staying, and they’re not wrong to. But the ones leaving describe something specific: they hit a point where the intellectual challenge of finance felt bounded in a way it didn’t before. They’re not chasing money. They’re chasing the feeling of working on something where the upper bound isn’t visible.
Academia
This is where the tradeoff is most painful, because it shouldn’t be a tradeoff at all.
Publishing novel results used to be the purest form of intellectual prestige. You did the work because the work was beautiful. That hasn’t changed. What changed is that the line between what you can do at a funded startup and what you can do in a university lab is blurring, and not in academia’s favor. A 20-person research startup can now do in a weekend what takes an academic lab a semester, because compute costs money that universities don’t have.
The most ambitious PhD students I talk to aren’t choosing between academia and industry. They’re choosing between theorizing about experiments and actually running them. The pull toward funded startups and labs isn’t about selling out. It’s about wanting to do the science, and the science requires resources that academia can’t provide.
The people staying in academia for the right reasons (open science, long time horizons, genuine intellectual freedom) are admirable. But they should know that the clock is ticking differently for them too: the longer the compute gap widens, the harder it becomes to do competitive work from inside a university.
AI Startups (Application Layer)
If you’re building products on top of models, you already know the feeling: the clever feature you shipped in March gets commoditized by a model update in June. The ground moves every quarter, and your moat evaporates.
The tradeoff here is between chasing what’s exciting and building what’s durable. The founders who are thriving right now stopped caring about model capabilities and started caring about the things models can’t take away: data moats, workflow capture, integration depth. It’s less fun to talk about at a dinner party. It’s where the actual companies get built.
The people making the sharpest moves in this world are the ones who got excited about plumbing—not the demo, not the pitch, not the capability. The ugly, boring infrastructure that makes a product sticky independent of which model sits underneath it.
Research Startups: The New Center of Gravity
This is where the K-curve is most visible.
Prime Intellect, SSI, Humans&—10-30 people doing genuine frontier research that competes with organizations fifty times their size. This would have been impossible three years ago. It’s happening now because the tools got good enough that a small number of people with great judgment can outrun a bureaucracy with more resources.
The daily workflow here is the clearest picture of what the upper arm looks like in practice. You’re kicking off training runs, spinning up experiments, letting things cook overnight. You come back in the morning, and your job isn’t to write code. It’s to know what to do with what came back—to have the taste to distinguish signal from noise when the system hands you a wall of results. It’s passive leverage. You set the experiments in motion, and the compounding happens whether or not you’re at your desk.
The tradeoff people are weighing: these companies are small, unproven, and many will fail. The bet is that being at the center of the frontier, with your judgment directly touching the work, compounds faster than the safety of a bigger organization, even if the specific company doesn’t make it. The skills transfer. The network transfers. The three years you spend reviewing someone else’s outputs at a big company don’t transfer the same way.
Big Model Labs: The Narrowing Frontier
The pitch “we’re building AGI” still works. It might always work on a certain type of person.
But the experience inside has shifted. The most interesting research is concentrated among a small number of senior people. Everyone else is doing important supporting work (evals, infra, product) that doesn’t feel like the frontier they signed up for. You joined to touch the thing, and you’re three layers removed from it.
The tradeoff is prestige versus proximity. A big lab on your resume still opens every door. But the people leaving are making a specific calculation: the resume value of “I was at [top lab]” is depreciating as the labs get bigger and more corporate, while the value of “I did frontier research at a place where my judgment shaped the direction” is appreciating. The window where big-lab pedigree is the best credential is closing, and the people who see it are moving.
The Clock
Every one of these tradeoffs has the same variable hiding inside it: time.
A year ago, you could sit in a comfortable seat and deliberate. The cost of waiting was low because the divergence was slow. That’s no longer true. The tools are compounding. The people who moved early are building on top of what they learned last quarter. The difference between someone who moved six months ago and someone still weighing their options is already compounding.
The upper arm isn’t closed. People are making the jump every week, and the people who are hiring them don’t care where you’ve been. They care whether you can do the work. But the math is directional: the longer you optimize for comfort, the more expensive the switch becomes—not because the opportunities disappear, but because the people who are already there are compounding, and you’re not.
The companies winning the talent war right now aren’t the ones with the best brand or the highest comp. They’re the ones where your judgment has the most surface area, where the distance between your taste and what actually gets built is zero, and where you’re surrounded by people who know things you don’t yet. The best people want to be close to others who have tricks they haven’t learned yet, at places with enough compute to actually run the experiments.
The question isn’t whether you’re smart enough. It’s that you’ve already done the math. You just haven’t acted on it.
The following content is generated by LLMs and may contain inaccuracies.
Context
This piece captures a structural shift in tech labor markets circa 2024–2025, where career optionality is compressing amid accelerating AI capabilities. It sits at the intersection of career dynamics, talent allocation theory, and the sociology of “frontier work.” The tension: traditional signals of career safety (FAANG comp, academic tenure, big lab prestige) are decoupling from proximity to where judgment-building happens. This matters because the shift from execution to orchestration—documented by economist David Autor as “task complementarity”—is happening faster than institutions can adapt, creating winner-take-most dynamics in skill accumulation.
Key Insights
The K-curve is a compounding divergence problem. Unlike previous tech cycles where skills depreciated gradually, generative AI tools create exponential productivity gaps between early adopters and laggards. Research from MIT and Stanford shows consultants using GPT-4 completed tasks 25% faster with 40% higher quality—but the variance between users widened over time. Those developing “judgment about AI outputs” compound that advantage quarterly; those executing manually fall behind non-linearly. The piece’s insight about research startups outrunning labs 50× their size reflects Coase’s theory of firm boundaries inverting: coordination costs collapsed faster than resource advantages matter.
Academia’s compute gap is a resource curse in reverse. The observation about weekend experiments versus semester timelines maps onto Brown et al.’s analysis of compute inequality in AI research. Universities can’t compete on infrastructure, but the piece misses that top labs are increasingly restricting publication to protect competitive moats—academic freedom still trades at a premium for reproducible, open work. The real cost: PhD students now optimize for “access to compute” over “intellectual community,” potentially sacrificing the collaborative serendipity that historically generated breakthrough ideas.
Open Questions
Could the K-curve collapse if AI tool improvements plateau, returning advantage to institutional stability? Or are we seeing a permanent regime change where “taste for orchestrating AI systems” becomes the dominant filter for knowledge work?
If judgment compounds faster than execution devalues, what happens to the bottom 50% of current tech workers—and does this finally force a reckoning with tech’s meritocracy mythology?
留任的代价
作者：Amy Tam https://x.com/amytam01/status/2023593365401636896
我认识的每一位技术人士现在都在做同样的数学计算。他们不会这样说。他们会说自己在"探索选择"或"思考下一步"。但本质上，这是同一个计算：留在原地要花费我多少？
不是金钱。而是时间。有一种感觉在空中弥漫：做出正确选择的窗口在缩小——你在错误岗位上待的每个季度，你和那些早期转身的人之间的差距就变得更难以弥补。一年前，科技行业的职业决策似乎是可逆的。接了个错误的工作，十八个月内调整方向就行。这个假设正在瓦解。早期重新定位的人和仍在权衡选择的人之间的分化变得可见，而且在加速。
我近距离看到这一点。我是Bloomberg Beta的投资者，大部分时间都与处于过渡期的人接触：离职、完成计划、决定下一步。我不是职业顾问，但我坐在"你要离开什么"和"你在追逐什么"的交叉口。
科技行业的宝贵技能从"你能解决这个问题吗"转变为"你能判断哪些问题值得解决，哪些解决方案真正有效吗"。稀缺的东西从执行力翻转到判断力：你能编排系统、并行下注，并具有品味来判断哪些结果重要吗？那些早期弄清楚这一点的人站在不断扩大的K曲线的一臂上。其他所有人都在快速提升那些即将被自动完成的东西的能力。
从执行到判断的转变无处不在，但留任的代价和转身的上升空间看起来完全取决于你所处的位置。
FAANG
这是大科技公司人员现在的权衡：系统已构建，薪酬很好，工作是……还可以。你越来越多地审查AI生成的输出，而不是从零开始构建。对某些人来说，这是礼物——这是杠杆、可持续性、美好生活。权衡是"还可以"有一个不会出现在你薪资单上的代价。
离职的人并不是不开心。他们坐立不安。他们描述这种特定的感觉：最难的问题已经不在这里了，而组织还没有认识到这一点。留下来的人是在打赌稳定性和薪酬比接近前沿更有价值。离开的人是在打赌前沿是下一个十年职业价值的构建之地，他们等待的每个季度都是他们错失的复合增长季度。
两个赌注都是理性的。但只有其中一个具有时间敏感性。
量化投资
量化投资仍然有效。荒谬的薪酬、困难的问题、即时反馈。如果你很优秀，你就知道自己很优秀，因为损益表不会说谎。
正在出现的权衡：整个量化工具包（ML基础设施、数据迷恋、统计直觉）正好是AI实验室和研究初创公司所需的——相同的肌肉、不同的问题。区别在于表面积。在量化投资中，你优化一个策略。在AI中，你构建能够推理的系统。即使是与量化相关的世界也在感受这一点：预测市场和稳定币中最有趣的工作越来越多地是AI基础设施问题。一个有上限。另一个没有，或者至少还没有人找到。
大多数量化人才留了下来，他们没有错。但离开的人描述了一些具体的东西：他们到达了一个点，金融的智力挑战感觉到了界限，这在以前没有。他们不是在追逐金钱。他们在追逐在做某件事的感觉，其中上界是不可见的。
学术界
这是权衡最痛苦的地方，因为根本不应该有权衡。
发表新颖结果曾经是最纯粹的智力声望形式。你做工作是因为工作很美妙。这没有改变。改变的是，你在资金充足的初创公司和大学实验室中能做什么之间的界线变得模糊，而且对学术界不利。一个20人的研究初创公司现在可以在一个周末做的工作，需要一个学术实验室一个学期，因为计算成本高昂，而大学没有这样的资金。
我交谈过的最雄心勃勃的博士生不是在学术界和产业之间选择。他们在理论化实验和实际运行实验之间选择。对资金充足的初创公司和实验室的吸引力不是关于妥协。这是关于想做科学，而科学需要学术界无法提供的资源。
因为正确的原因留在学术界的人（开放科学、长期视野、真正的学术自由）是令人敬佩的。但他们应该知道，时钟对他们的嘀嗒也不同：计算差距越长，从大学内部做有竞争力的工作就越难。
AI初创公司（应用层）
如果你在模型之上构建产品，你已经知道那种感觉：你在三月份推出的聪明功能在六月份被模型更新商品化了。地形每个季度都在移动，你的护城河蒸发了。
这里的权衡是追逐令人兴奋的东西和构建持久的东西之间的权衡。现在蓬勃发展的创始人停止关心模型能力，开始关心模型无法夺走的东西：数据护城河、工作流捕获、集成深度。在宴会上谈论这些就没那么有趣了。这是真正的公司被构建的地方。
在这个世界里做出最尖锐举动的人是那些对管道感到兴奋的人——不是演示、不是宣传、不是能力。丑陋、无聊的基础设施使产品粘性独立于坐在下面的模型。
研究初创公司：重力的新中心
这是K曲线最可见的地方。
Prime Intellect、SSI、Humans&——10-30人进行真正的前沿研究，与规模大五十倍的组织竞争。三年前这是不可能的。现在发生是因为工具足够好，少数具有高明判断力的人可以跑赢拥有更多资源的官僚机构。
这里的日常工作流程是上臂在实践中看起来最清晰的画面。你在启动训练运行、旋转实验、让事情一夜间进行。你早上回来，你的工作不是编写代码。这是知道如何处理返回的东西——当系统给你一堵结果时，具有品味来区分信号和噪音。这是被动杠杆。你设置实验运行，复合增长是否发生，不管你是否在办公桌前。
人们在权衡：这些公司很小、未经证实，许多会失败。打赌是在前沿中心，你的判断直接接触工作，复合速度比大型组织的安全更快，即使特定公司没有成功。技能转移。网络转移。你在大公司审查他人输出花费的三年不会以相同的方式转移。
大模型实验室：前沿变窄
“我们在构建AGI"的宣传仍然有效。它可能对某种类型的人总是有效。
但内部的体验已经转变。最有趣的研究集中在少数高级人员中。其他人都在做重要的支持工作（评估、基础设施、产品），感觉不像他们注册的前沿。你加入是为了接触这件事，你距离它有三层。
权衡是声望对邻近。大实验室在你的简历上仍然可以打开所有大门。但离开的人在做一个具体的计算：“我在[顶级实验室]“的简历价值随着实验室变得更大和更公司化而贬值，而"我在一个我的判断塑造方向的地方进行前沿研究"的价值在升值。大实验室血统是最佳证书的窗口正在关闭，看到它的人在转身。
时钟
这些权衡中的每一个都在其中隐藏着相同的变量：时间。
一年前，你可以坐在舒适的座位上深思熟虑。等待的代价很低，因为分化很慢。那不再是真的了。工具在复合。早期转身的人正在建立他们上个季度学到的东西。有人六个月前转身和有人仍在权衡选择之间的差异已经在复合。
上臂没有关闭。人们每周都在跳跃，雇用他们的人不关心你去过哪里。他们关心你是否能完成工作。但数学是方向性的：你优化舒适的时间越长，转换变得越昂贵——不是因为机会消失，而是因为已经到达那里的人在复合，而你没有。
现在赢得人才战争的公司不是那些品牌最好或薪酬最高的公司。他们是那些你的判断有最大表面积的地方，你的品味和实际构建的距离为零，你被你还没学过技巧的人包围的地方。最优秀的人想靠近其他拥有他们还没学过技巧的人，在有足够计算实际运行实验的地方。
问题不是你是否足够聪明。这是你已经做了数学。你只是还没有采取行动。
以下内容由 LLM 生成，可能包含不准确之处。
背景
这篇文章捕捉了科技劳动力市场在2024-2025年左右的结构性转变，在加速的人工智能能力中职业选择空间在压缩。它位于职业动态、人才配置理论和"前沿工作"社会学的交汇点。核心矛盾在于：传统的职业安全信号（FAANG薪酬、学术终身教职、大型实验室声誉）正在与判断力养成发生的地方脱钩。这很重要，因为从执行到协调的转变——由经济学家大卫·奥特记录为"任务互补性"——正在以制度适应的速度更快地发生，在技能积累中创造赢家通吃的动态。
关键洞见
K形曲线是一个复合性分化问题。 与以往科技周期中技能逐步贬值不同，生成式人工智能工具在早期采用者和落后者之间创造了指数级的生产力差距。麻省理工学院和斯坦福大学的研究表明，使用GPT-4的顾问完成任务的速度快25%，质量高40%——但用户之间的差异随时间扩大。那些开发出"关于人工智能输出判断力"的人每季度都在复合优势；那些手动执行的人落后的速度是非线性的。这篇文章关于研究初创企业超越其规模50倍实验室的观点反映了科斯的企业边界理论的反转：协调成本的下降速度比资源优势重要得多。
学术界的计算能力差距是反向的资源诅咒。 关于周末实验对比学期时间表的观察映射到Brown等人对人工智能研究中计算不平等的分析。大学无法在基础设施上竞争，但这篇文章没有注意到顶级实验室越来越限制出版以保护竞争优势——学术自由仍然对可复现的开放工作享有溢价。真正的代价：博士生现在为"获取计算能力"而不是"知识社群"进行优化，可能牺牲了历史上产生突破性想法的协作意外收获。
开放问题
如果人工智能工具改进进入平台期，K形曲线会崩溃吗，让优势回到机构稳定性？还是我们正在看到一个永久的政权转变，其中"协调人工智能系统的品味"成为知识工作的主导过滤器？
如果判断力的复合速度比执行贬值更快，当前科技工作者中的底部50%会发生什么——这最终是否会迫使对科技的精英统治神话进行清算？