Research Freedom in the Scaling Era

I've been thinking a lot about what it means to do meaningful AI research in the era of scaling and foundation models.

We often celebrate research freedom: the ability to explore ambitious ideas, challenge existing paradigms, and pursue questions simply because they are interesting.

But there's a distinction we don't discuss enough:

The freedom to explore an idea is not the same as having the resources to test it.

Without sufficient compute, access to frontier models, high-quality data, and experimental infrastructure, many ambitious ideas cannot even be properly evaluated.

This makes me wonder: as the resources required for frontier AI research become increasingly concentrated in a handful of labs, what does research freedom actually mean?

1. Scaling is changing who can explore the frontier

For years, smaller labs have narrowed the gap with frontier models through distillation, architectural innovation, and training efficiency.

But this path may be getting harder.

As access to frontier models and their reasoning traces becomes more restricted, catching up may increasingly depend on independent pretraining, large-scale RL, and proprietary data and feedback loops.

Compute is no longer just an accelerator of research. In some areas, it determines which research questions are feasible to investigate.

This doesn't mean smaller labs will stop producing important breakthroughs. Algorithmic efficiency, evaluation, and new research paradigms can still change the frontier.

But the set of questions that researchers can realistically explore is becoming increasingly dependent on their access to infrastructure.

2. Generation is scaling. Verification may be the harder problem.

One of the most interesting questions in recursive self-improvement (RSI) is not simply whether AI can generate better ideas.

It's whether AI can reliably determine which ideas are actually better.

Coding has automated tests. Formal mathematics has proof checkers. These provide relatively scalable feedback, even though correctness alone does not guarantee usefulness or scientific significance.

Now consider scientific discovery and robotics.

A model can propose a new molecule, a physical hypothesis, or a robot control policy. But determining whether it works may require expensive experiments, noisy physical interactions, or months of observation.

Generating hypotheses may become increasingly cheap. Obtaining reliable evidence may remain expensive.

If we want AI systems to continuously improve beyond code and mathematics, we'll need to scale not only intelligence, but also the systems that produce trustworthy feedback.

Perhaps one of the next major bottlenecks in RSI is building environments where improvements can be reliably measured.

3. Foundation models may redefine what research is valuable

As foundation models become more capable, some specialized architectures and pipelines may become obsolete.

But this doesn't mean scientific research becomes less important.

It means the questions worth asking may change.

• What failures persist as models scale?
• Which capabilities remain difficult to evaluate?
• How can agents learn from real-world outcomes?
• What experimental infrastructure enables reliable self-improvement?

These questions connect foundation models, reinforcement learning, AI for Science, and embodied intelligence.

The frontier may increasingly depend not only on how powerful models become, but also on how effectively they interact with reality, receive feedback, and improve.

4. The changing economics of research also change the value of a PhD

This is where the technical discussion becomes personal.

We're asked to write five-year research plans in a field where predicting the next twelve months is already difficult.

We encourage ambitious, long-term research. Yet publication pressure, deadlines, and resource constraints often push young researchers toward safer, incremental projects.

At the same time, frontier labs increasingly offer something many academic groups cannot: direct access to the strongest models and the infrastructure needed to test ambitious ideas at scale.

This creates a difficult tension.

A PhD offers time to think deeply, develop scientific judgment, and pursue independent questions. These remain valuable.

But if the experiments needed to investigate those questions are increasingly accessible only within a few industrial labs, the traditional relationship between academic freedom and frontier research becomes more complicated.

The question is no longer simply whether academia or industry offers more freedom.

It's whether intellectual freedom without sufficient experimental capacity is enough—and whether proximity to the frontier is worth giving up some independence.

For young researchers, the opportunity cost of a PhD deserves serious consideration. Not because the degree has become meaningless, but because the research landscape may change substantially during those years.

5. A personal reflection

I still love exploring new ideas. Scientific curiosity is one of the reasons I entered research in the first place.

But I'm also increasingly excited about working close to the Pareto frontier of AI capabilities, where ambitious ideas can actually be tested at scale and meaningful improvements can push the limits of what models can do.

I don't think research freedom and proximity to the frontier are necessarily opposites.

But when they conflict, the trade-off is becoming harder to ignore.

Ultimately, I don't just want the freedom to imagine what might work.

I want the opportunity to find out.

Maybe the next frontier of AI research isn't just scaling intelligence. It's scaling our ability to discover, verify, and improve it.

中文版｜Scaling 时代的科研自由

最近，我一直在思考：在 Scaling（规模扩展）和 Foundation Models（基础模型）的时代，究竟什么才算得上有意义的 AI 研究？

我们经常赞美科研自由：能够自由探索宏大的想法，挑战现有范式，或者仅仅因为一个问题足够有趣，就愿意投入时间去研究。

但有一个区别，我们讨论得还不够充分：

拥有探索一个想法的自由，并不等于拥有验证它所需的资源。

如果没有足够的算力、前沿模型的访问权限、高质量数据以及实验基础设施，许多宏大的研究想法甚至无法得到充分评估。

这让我不禁思考：当开展前沿 AI 研究所需的资源越来越集中在少数实验室手中时，科研自由究竟意味着什么？

1. Scaling 正在改变谁有能力探索技术前沿

过去几年，规模较小的实验室可以通过知识蒸馏、模型架构创新和训练效率优化，缩小与前沿模型之间的差距。但这条道路可能正在变得越来越困难。

随着获取前沿模型及其推理轨迹受到越来越多的限制，追赶前沿模型可能越来越依赖独立预训练、大规模强化学习，以及专有数据和反馈闭环。

算力已经不再仅仅是加速科研的工具。在某些领域，它甚至决定了哪些研究问题具有实际开展的可能性。

这并不意味着规模较小的实验室将不再产生重要突破。算法效率、评估方法和新的研究范式，仍然可能推动能力前沿的发展。

但研究者实际上能够探索的问题范围，正变得越来越依赖他们能够获得怎样的基础设施。

2. 生成能力正在 Scaling，验证可能才是更困难的问题

递归自我改进（Recursive Self-Improvement，RSI）中，一个最有意思的问题，并不只是 AI 能否产生更好的想法，而是 AI 能否可靠地判断哪些想法实际上更好。

代码有自动化测试，形式化数学有证明检查器。这些工具能够提供相对可扩展的反馈，尽管正确性本身并不意味着一个结果一定有用，或者具有重要的科学价值。

现在，把目光转向科学发现和机器人领域。

一个模型可以提出新的分子结构、物理假设，或者机器人控制策略。但要判断这些方案是否真正有效，可能需要昂贵的实验、充满噪声的物理交互，或者持续数月的观察。

生成科学假设可能会变得越来越廉价，而获得可靠证据的成本可能依然高昂。

如果我们希望 AI 系统能够在代码和数学之外持续实现自我改进，就不仅需要扩展模型智能，还需要同步扩展那些能够产生可信反馈的系统。

也许，RSI 下一阶段的重要瓶颈之一，就是如何构建能够可靠衡量改进效果的实验环境。

3. 基础模型可能重新定义什么样的研究具有价值

随着 Foundation Models 的能力不断提升，一些专门设计的模型架构和复杂 Pipeline 可能逐渐失去存在的必要。但这并不意味着科学研究本身会变得不重要。

它可能意味着，真正值得提出的研究问题正在发生变化。

• 随着模型规模和能力不断提升，哪些失败仍然持续存在？
• 哪些能力仍然难以被可靠评估？
• Agent 如何从真实世界的结果中学习？
• 什么样的实验基础设施能够支持可靠的自我改进？

这些问题将基础模型、强化学习、AI for Science 和具身智能联系在了一起。

未来 AI 能力前沿的发展，可能不仅取决于模型本身变得多么强大，还取决于它们能否有效地与现实世界交互、获得反馈，并在此基础上不断改进。

4. 科研经济结构的变化，也正在改变 PhD 的价值

到了这里，技术层面的讨论开始与个人职业选择产生联系。

我们被要求制定五年的研究计划，但在 AI 领域，即使预测未来十二个月将发生什么，也已经非常困难。

我们鼓励宏大、长期的科研探索。然而，论文发表压力、投稿截止日期以及资源限制，却常常推动年轻研究者选择那些相对安全、增量式的研究项目。

与此同时，Frontier Labs 正在越来越多地提供许多学术研究团队无法提供的东西：直接接触最强模型的机会，以及在大规模实验中验证宏大想法所需的基础设施。

这产生了一种难以回避的矛盾。

博士阶段为研究者提供深入思考的时间，帮助他们培养科学判断力，并独立探索研究问题。这些价值依然重要。

但如果研究这些问题所需的实验资源，越来越只能在少数工业实验室中获得，那么传统意义上的学术自由与前沿研究之间的关系，就变得更加复杂了。

问题不再只是学术界和工业界哪一方能够提供更多自由。

真正的问题是：如果缺少足够的实验能力，仅仅拥有思想上的自由是否足够？而为了接近技术前沿，放弃一部分研究自主权又是否值得？

对年轻研究者而言，攻读 PhD 的机会成本值得被更加认真地考虑。并不是因为博士学位已经失去了意义，而是因为在攻读博士的这几年里，整个科研格局可能发生深刻变化。

5. 一些个人思考

我依然热爱探索新的研究想法。科学好奇心，本来就是我走上科研道路的重要原因之一。

但与此同时，我也越来越向往能够接近 AI 能力帕累托前沿（Pareto Frontier）的研究环境。在那里，宏大的研究想法能够真正得到大规模实验验证，而有意义的技术改进也有机会进一步推动模型能力的极限。

我并不认为科研自由与接近技术前沿必然相互对立。但当两者发生冲突时，这种取舍正变得越来越难以忽视。

归根结底，我想要的不只是自由想象什么方法可能有效。

我更想真正拥有验证它的机会。

也许，下一阶段 AI 研究的前沿，不仅仅是继续 Scaling Intelligence（扩展智能能力），更是 Scaling 我们发现、验证和持续改进智能的能力。

Sources & context / 讨论背景