Where Does the Bottleneck Move When Intelligence Becomes Cheap?

I've been thinking about a question that goes beyond scaling laws: what happens when intelligence becomes abundant, but the physical world remains expensive?

A capable model can increasingly produce code, proposals, designs, mathematical arguments, and scientific hypotheses at marginal costs that were once unimaginable. Yet a proposed solution and a working solution are two different economic objects. The second still needs evidence, execution, and often physical infrastructure.

Solving one bottleneck changes which bottleneck matters next.

This is not a claim that all intelligence will become cheap, or that every problem reduces to verification. It is a hypothesis about where constraints—and therefore research opportunities—might migrate as model capabilities improve.

1. The scaling paradox: intelligence spreads, frontier production concentrates

AI is becoming more accessible to users and developers. A small team can build products and run experiments using models it could never train itself. But training frontier foundation models can demand extraordinary capital, compute, data pipelines, and engineering coordination.

That creates a paradox: the application of intelligence may democratize while the production of frontier intelligence concentrates. Distillation and efficient architectures have historically offered ways to narrow the gap, but tighter access to high-quality model outputs could make independent training and original feedback environments more valuable.

I would not conclude that small labs are doomed. New algorithms can change the compute frontier; open models can alter bargaining power. My narrower prediction is that research freedom will increasingly depend on which experiments a lab can actually run—not merely which ideas its researchers may propose.

2. AI for Science: the bottleneck moves from hypotheses to evidence

Drug discovery is a useful stress test. Stronger models may accelerate literature review, target prioritization, molecular design, and experiment planning. But a candidate molecule is not a medicine. Wet-lab replication, toxicology, animal studies, clinical recruitment, and regulatory evidence remain costly and slow.

This matters for where economic value accumulates. Better prediction may increase the number of promising candidates, yet simultaneously raise demand for experimental capacity. If validation cannot scale, the result could be a larger queue of hypotheses rather than proportionally more approved drugs.

Materials science and chemistry face similar constraints: measurement precision, failed experiments, equipment interoperability, and the difficulty of turning simulation results into reproducible physical outcomes.

The scarce resource may shift from generating candidates to obtaining trustworthy ground truth.

This is a prediction, not a universal law. Sometimes better models reduce experimentation by filtering bad hypotheses. The empirical question is whether model progress lowers total validation cost faster than it expands the set of ideas worth testing.

3. Robotics and manufacturing: reality has an API problem

In robotics, a better foundation model may improve perception, planning, and generalization. Yet successful operation in an uncontrolled environment still depends on contact dynamics, latency, sensing, safety, recovery, and access to real-world feedback.

This makes physical interaction both a bottleneck and a potential data advantage. A lab that can repeatedly execute actions, detect failures, attribute causes, and update a policy has something more valuable than an impressive demo: an improvement loop.

Semiconductor design illustrates a related but distinct boundary. AI can help explore architectures and optimize design choices, while manufacturing advanced equipment introduces precision optics, materials, metrology, supply chains, and engineering constraints that cannot simply be wished away by more gradient descent.

My interest here is not merely building specialized robots or tools around today's models. It is asking how foundation agents can learn from consequences in environments without cheap, perfect verifiers.

4. After the smartphone: do stronger agents need brain–computer interfaces?

I also wonder whether the next interface shift will be driven by increasingly capable AI. Typing, speech, screens, and gestures were designed for systems that required frequent explicit instructions. If agents can do much more, our interaction patterns could change radically.

One possibility is a higher-bandwidth interface: electromyography, wearable sensors, and eventually brain–computer interfaces for more continuous expression of intent. Another possibility is almost the opposite: a highly autonomous agent may need only occasional, low-bandwidth human direction.

Does stronger AI increase demand for higher-bandwidth human interfaces—or decrease it by requiring fewer instructions?

I don't know the answer. And neural signals are not equivalent to readable thoughts: reliable decoding, comfort, safety, and individual calibration remain serious challenges. But this is exactly the kind of uncertain question that feels worth investigating, rather than treating BCI as an inevitable successor to the phone.

5. Quant, energy, and infrastructure: not every bottleneck is a model problem

Quantitative finance is another interesting case. More powerful models can search for signals, but any discovered edge interacts with competition, market impact, transaction costs, latency, and shifting market behavior. A predictive advantage is not automatically a durable trading strategy.

Energy and compute infrastructure operate on different timelines. Data centers, chips, power delivery, cooling, and fabrication capacity are physical systems. Even if model algorithms improve quickly, scaling their deployment can remain constrained by capital expenditure and engineering lead times.

That is why I am cautious about turning a ranking of AI-adjacent industries—chips, drug discovery, materials, brain interfaces, fusion, or quantum—into a single forecast. They have very different verification loops, regulatory burdens, and routes to commercialization. The relevant question is not only how much AI can help, but which bottleneck survives after it helps.

6. Capital, inequality, and who gets to explore

There is a social consequence to all of this. If advanced intelligence is cheap to consume but expensive to produce, access to the frontier may remain uneven. A handful of organizations might control critical training infrastructure while millions of people build on their outputs.

That could generate enormous productivity gains and new opportunities. It could also concentrate research agenda-setting and the returns to infrastructure. Both outcomes can coexist.

For a young researcher, this is not abstract political economy. It changes the feasible research agenda, collaboration opportunities, and even the relative value of an academic position versus an industrial lab. Intellectual independence matters; so does being able to test an idea.

7. What would change my mind?

I would revise this view if smaller groups repeatedly pushed the foundation-model frontier with modest resources; if AI-driven simulation displaced large portions of physical validation without sacrificing reliability; if agents proved highly effective with almost no new interface hardware; or if verification became as cheap and scalable as generation across many scientific domains.

Those are not inconveniences to explain away. They are observations that should update the hypothesis.

8. A personal conclusion

I still care about open-ended research and the pleasure of finding a question no one has framed quite right. I also find it exciting to work near the Pareto frontier of model capabilities, where even small improvements can matter at extraordinary scale.

These ambitions are not necessarily incompatible. But they explain why I increasingly think about research environments in terms of both intellectual freedom and experimental capacity.

Perhaps the most useful question for the coming years is not simply, “What can AI automate?” It is: once AI makes something cheap, what becomes scarce next—and who will be able to study it?

中文版｜当智能变得廉价，瓶颈将转移到哪里？

最近，我一直在思考一个超越 Scaling Laws 本身的问题：当智能变得越来越充裕，而物理世界依然昂贵时，会发生什么？

模型正在以过去难以想象的边际成本生成代码、方案、设计、数学论证和科学假设。但一个看起来有希望的方案，与一个真正有效的方案，是两种截然不同的经济对象。后者仍然需要证据、执行，以及许多情况下无法绕开的物理基础设施。

解决一个瓶颈，会改变下一个真正重要的瓶颈是什么。

这不意味着所有智能都会廉价，也不意味着一切问题最终都归结于验证。我更愿意把它视为一个假设：随着模型能力提高，限制条件与研究机会将向哪里迁移？

1. Scaling 的悖论：智能使用日益普及，前沿生产却可能更加集中

AI 对开发者和普通用户越来越容易获得。小团队可以利用自己根本无力训练的模型构建产品、开展研究；但训练最前沿基础模型，又需要巨额资本、算力、数据流水线和复杂的工程组织。

于是出现一种悖论：智能的应用可能越来越民主化，而前沿智能的生产可能越来越集中。 蒸馏和高效架构过去帮助追赶者缩小差距，但如果高质量模型输出的访问不断受限，独立训练能力和原创反馈环境就可能变得更有价值。

这不是说小实验室注定出局。算法创新可以移动算力前沿，开源模型也可以改变竞争结构。我更具体的预测是：科研自由将越来越取决于你实际能开展哪些实验，而不仅是你可以自由提出哪些想法。

2. AI for Science：瓶颈从提出假设转向获得证据

药物研发是一个很好的试金石。更强的模型可能加速文献检索、靶点筛选、分子设计和实验规划。但候选分子还不是药物。湿实验重复验证、毒理、动物实验、临床招募和监管证据依旧昂贵而缓慢。

这决定了产业价值可能落在哪里：更好的预测模型增加了有希望的候选方案，也可能增加实验设施的需求。如果验证能力无法同步扩展，结果可能是等待检验的假设排得更长，而不是新药数量同比例增长。

材料科学和化学也有类似限制：测量精度、失败实验、仪器协同，以及仿真结果能否转化为可重复的物理结果。

稀缺资源可能从生成候选方案，转向获取可信的真实反馈。

但这并非普遍规律。更好的模型也可能通过排除无效假设减少实验需求。值得验证的是：模型进步降低总体验证成本的速度，是否快过了它增加值得检验的想法的速度？

3. Robotics 与制造：真实世界的 API 远不够可靠

在机器人领域，基础模型的进步可能改善感知、规划和泛化。但在不受控环境中完成任务，仍然涉及接触动力学、时延、传感、安全、失败恢复，以及真实反馈的获取。

这让物理交互既是瓶颈，也是潜在的数据优势。能够反复执行动作、检测失败、分析原因、更新策略的实验室，比只做出精彩 Demo 的团队多了一项重要资产：可运行的改进闭环。

半导体设计说明了另一个不同的边界。AI 可以帮助探索架构、优化设计，但先进制造设备还依赖精密光学、材料、计量、供应链和工程系统，不能简单归结为再多做一点梯度下降。

我真正感兴趣的，并不只是围绕当前模型限制构建专用机器人或工具，而是：Foundation Agent 如何在缺乏廉价、完美验证器的环境里，从行动后果中学习？

4. 手机之后：更强的 Agent 真的需要脑机接口吗？

我也在想，下一次交互范式变化是否会由越来越强的 AI 推动。打字、语音、屏幕和手势，是为了需要频繁明确指令的系统设计的。当 Agent 能承担更多任务，交互方式可能发生根本变化。

一种未来是更高带宽的接口：肌电、可穿戴传感器，甚至脑机接口，让人更连续地表达意图。另一种未来几乎相反：高度自主的 Agent 只需要人类偶尔给出低带宽的高层目标。

更强的 AI 会提高高带宽人机接口的需求，还是因为需要更少指令，反而降低它？

我不知道答案。而且神经信号不等于可以直接读取的思想：可靠解码、舒适度、安全性和个体校准都仍是严肃难题。正因为如此，我觉得这比直接断言「脑机必然取代手机」更值得研究。

5. 量化、能源与基础设施：不是每个瓶颈都属于模型

量化金融也很有意思。更强的模型可以寻找信号，但任何 Alpha 都会与竞争、市场冲击、交易成本、时延以及不断变化的市场行为相互作用。预测优势并不自动等于可持续收益。

能源和算力基础设施则遵循不同时间尺度。数据中心、芯片、供电、冷却和晶圆制造都是物理系统。即使模型算法飞速改善，部署能力仍可能受到资本支出和工程建设周期限制。

因此，我对把 AI 芯片、药物、材料、脑机、核聚变和量子等产业简单排出一个统一前景榜单保持谨慎。这些领域的验证闭环、监管难度和商业化路径非常不同。关键不仅是 AI 能提供多大帮助，还在于帮助之后，哪个瓶颈仍然存在。

6. 资本、不平等与谁有机会探索

这一切还涉及社会层面。如果先进智能使用起来越来越便宜，但生产前沿智能依然昂贵，那么接近技术前沿的机会可能仍然分布不均。少数组织可能控制关键训练设施，而无数人只能基于它们的输出做应用。

这可能创造巨大的生产率提升和创业机会，也可能集中科研议程的制定权与基础设施收益。两件事并不矛盾。

对年轻研究者而言，这不是抽象的政治经济学。它直接改变可行的研究问题、合作机会，甚至大学科研岗位与工业实验室的相对价值。思想独立很重要，能够真正检验自己的想法同样重要。

7. 什么证据会让我改变看法？

如果越来越多的小团队以有限资源反复推动基础模型能力前沿；如果 AI 仿真能够在不牺牲可靠性的前提下大量替代真实实验；如果高自主 Agent 几乎不需要新型交互硬件就能取得很好效果；或者多个科学领域的验证成本下降到接近生成成本——这些都应该让我修正目前的判断。

这些不是需要强行解释的反例，而是应该真正更新假设的证据。

8. 一点个人思考

我仍然珍视开放式探索，以及发现一个别人尚未准确提出的问题所带来的乐趣。但我同样会为接近模型能力的帕累托前沿而兴奋：在那里，微小的改进也可能产生巨大的影响。

两种追求未必矛盾。但这解释了为什么我越来越习惯同时用两个尺度衡量科研环境：思想自由与实验能力。

也许未来几年最值得追问的，不只是「AI 能自动化什么」，而是：当 AI 让某件事变得廉价，接下来什么会变得稀缺？又是谁有机会研究它？

Sources & framing / 参考与写作说明