Researchers let Claude Code discover AI scaling algorithms that humans probably wouldn't have designed 研究人员让Claude Code发现了人类很可能不会设计出的AI扩展算法

The most significant development in AI this week isn't a larger model or a slicker demo. It’s a quiet paper showing how an algorithm designed to control AI reasoning was itself designed by an AI, cutting computational costs by a staggering 70% for no loss in accuracy. And it did it for forty dollars.

Let that sink in. A coding agent, essentially an advanced autocomplete for programmers, was set loose on the problem of making its own cousins more efficient. It independently engineered a new algorithm, one that the paper suggests human researchers "probably wouldn't have designed." This isn't just an incremental optimization. It's a proof-of-concept for a new kind of R&D, where the process of scientific discovery is automated, and it’s happening at a price point that makes a university hackathon look like a moonshot.

The research, involving the usual suspects like Google and Meta alongside the University of Maryland, uses a framework called AutoTTS. Think of it not as a single tool, but as a virtual laboratory. The coding agent (Claude, in this case) isn't just generating code; it’s iterating, testing, and refining strategies in a simulated environment to achieve a meta-goal: improving the performance of another AI task, in this case, "self-consistency" for reasoning. The result is an algorithm that decides, on a step-by-step basis, whether to keep chaining out logical thoughts or to cut its losses and commit to an answer, saving a massive amount of compute.

The first takeaway is the brutal economics. The entire experimental search cost $40 and took about 2.5 hours. Compare that to the millions spent on training runs for frontier models, or the six-figure salaries and years of grant proposals funding human-led algorithmic research. We’ve entered an era where you can, for the price of a mediocre takeout dinner, outsource a specific, high-value R&D problem to an autonomous agent that will likely deliver a novel solution humans overlooked. The ROI is obscene. Research labs with their supercomputers and PhD teams now face competition from a script running on a cloud instance. What is the strategic value of a human researcher’s intuition when a brute-force, agentic search can explore a vast design space more thoroughly and cheaply?

This leads to the deeper, more unsettling implication: the obsolescence of a certain kind of human ingenuity. For decades, the art of algorithm design has been a hallmark of human intelligence—elegant shortcuts, clever heuristics, insights that feel almost aesthetic. The "Aha!" moment of a researcher seeing a simpler path. Now, we’re witnessing the automation of that "Aha!" The agent here didn’t use elegance; it used scale and systematic trial. It found a solution that is functionally superior by metrics we care about (compute, accuracy) but may lack the intuitive explainability a human would craft. We’re moving from algorithms as human art to algorithms as machine-mined resources. The value shifts from the designer to the design environment—whoever controls the best AutoTTS-like frameworks and has the cheapest compute wins.

This isn't just about efficiency. It's about the trajectory of AI development itself. We are building the tools that will build the next generation of tools. This is the recursive loop everyone hypothesized, and now it’s demonstrably here, operating at a scale that’s both impressive and humbling. The paper found its solution in 160 minutes. How many months or years of a postdoc’s time would it have taken to stumble upon the same logic? And would they have even been incentivized to look? Academic publishing rewards novel tasks, not necessarily radical efficiency gains on existing ones. The agent doesn’t care about publishing in NeurIPS; it optimizes for the objective function you give it.

Let’s be clear-eyed about what this is not. This is not a general intelligence solving general problems. It’s a highly specialized agent solving a well-defined optimization problem within a sandboxed code-generation environment. It’s brilliant narrow AI, not HAL 9000. The "discovery" is contingent on the quality of the objective function and the simulation. If the metrics are wrong, the agent will gleefully optimize for the wrong thing. The risk isn’t Skynet; it’s deploying an algorithm that’s 70% cheaper but has some bizarre, undiscovered failure mode because no human ever really understood why it works.

So, where does this leave the field? It accelerates a bifurcation. On one side, we’ll see more "AI-for-AI" research, where agents constantly refine and compete to produce more efficient components, creating a Cambrian explosion of specialized algorithms. This will supercharge progress in narrow domains. On the other, we face a profound identity crisis for AI research as a human endeavor. Do we celebrate the efficiency, or mourn the potential loss of the human touch that often leads to more robust, generalizable understanding?

The real takeaway from this $40 experiment is a shift in the locus of innovation. The next breakthrough might not come from a genius in a lab, but from a well-prompted agent in a bargain-basement cloud GPU cluster. The competitive edge is no longer just about having the biggest model, but about having the most efficient, autonomous R&D pipeline. This paper is a shot across the bow: the future of AI improvement is being automated, and the first movers are already getting results for pocket change. The rest of the industry should be checking its pockets.

四十美元，十六分钟——这个数字组合在一起，本身就构成了一则极具冲击力的科技新闻。当马里兰大学、谷歌、Meta等机构的研究人员让Claude Code这样的编程代理，通过AutoTTS框架“独立”发现了一种新算法，而该算法竟能将大语言模型推理时的计算成本削减70%，同时准确率不降时，我们看到的不仅是一次技术优化。我们瞥见的，是“工具”正在快速进化为“合作者”，甚至“发现者”的模糊身影。

这则新闻里，最耐人寻味的不是那70%的节省，尽管这在商业部署层面是惊人的性价比。真正值得玩味的是“独立发现”和“人类可能设计不出来”这两个表述。这意味着，AI代理在庞大的搜索空间里，找到了一条优化路径，而这条路径或许超出了人类研究者基于现有范式和直觉所划定的常规思路。我们习惯于将AI视为强大的计算器和模仿者，但这一次，它更像是一个在实验室角落里，通过无数次自我博弈和调整，偶然（但又必然地）“悟”出新解法的初级研究员。那40美元和160分钟，与其说是计算成本，不如说是雇佣了一位不眠不休、绝对专注、且毫无思维定式的“实习科学家”的试用期费用。

这立刻引出一个尖锐的问题：未来，我们研究前沿的“发现权”将如何归属？当算法优化的核心环节，其设计思路可以由AI自动搜索生成时，人类研究者的角色是什么？是提出宏观问题和定义搜索边界的“导师”，还是最终为结果背书和定性的“审核员”？这次发现的是一种“控制算法”，用于管理推理时的计算分配，属于“元算法”层面的创新。如果AI能持续在“元”层面进行探索和优化，那等于它在学习如何更好地学习和思考。这已不是简单的工具使用，而是工具开始介入对工具本身的改进循环。Scaling Law（规模法则）的优化空间，可能正从人类工程师的白板，转移到这些自动搜索框架的沙盒之中。

AutoTTS（或类似的代码搜索框架）的价值，在此刻显得尤为清晰。它不再是一个简单的代码补全或Bug修复工具，而是变成了一个“算法进化的军备竞赛”中的加速器。它证明了，给定一个清晰的目标（如减少计算量）和评价标准（准确率），AI可以通过系统的、大规模的探索，开辟出新的可能性。这可能会加剧AI领域的马太效应：拥有强大AutoTTS类工具和算力的团队，其模型迭代和优化速度将呈指数级领先。竞争的前沿，将从“谁的人才更聪明”，部分转向“谁的自动化研究工具更高效”。

当然，我们必须保持一份冷静的怀疑。所谓的“独立发现”，其边界依然是由人类通过AutoTTS的框架、奖励函数（准确率）和搜索空间预先设定的。它更像是在一个精心布置的、规则明确的沙盘里寻找最优解，而非在无垠的未知领域进行真正的原创性探索。那“人类可能设计不出来”的论断，也可能源于人类思维的固有惰性或计算力的局限，而非绝对的智能壁垒。AI在这里展现的是强大的优化与搜索智能，但离提出全新概念、构建全新理论体系的“科学革命”式创新，还有本质距离。

然而，即便承认这些限制，这则新闻的警示与启示依然强烈。它标志着AI从“执行层”向“设计层”的渗透又深了一步。对于产业界，这意味着模型压缩和推理优化的成本将被极大压低，高性能AI的普惠化可能加速。对于学术界，这或许会催生一门新的交叉学科：如何设计更智能的、用于科学发现的“AI研究者”。我们可能正站在一个拐点上，未来的重大突破，将越来越多地源于人与AI智能体的深度协作，甚至是一些意想不到的、由AI主导的“偶然发现”。

那40美元，买到的不是一行优化的代码，而是一个未来的预演：在AI的辅助下，解决AI自身带来的问题（如高计算成本），其效率和创造性可能远超预期。这场关于智能的竞赛，规则正在被悄然改写。