This represents an 81% reduction in physical die area, proving that high-level front-end architectural optimizations translate directly into highly compact, routable, and performant silicon implementation.
Qwen3.8-Max 在一次连续自主运行中把芯片版图面积压缩了 81%，且验证结果落地到真实可布线的物理设计层面，不只是停留在算法层的优化。这类案例值得在"RSI 工具层证据"的清单里和 Anthropic 8× 代码产出、Astra 数学证明并列看待——中国厂商在同一条自动化研发曲线上给出了独立可验证的证据。
We think there is opportunity for AI to more fully automate what has traditionally been a very human-intensive experimental loop
Jeff Dean 亲口对 NYT 说的这句话，来自一位在 Google 工作 27 年、参与过搜索核心基础设施和 Gemini 多模态模型的资深人物——他的表态本身就是行业信号：当最了解"人类主导科研有多慢"的人开始押注全自动实验闭环，说明这不是外部炒作，而是内部人对趋势的判断。
progress has traditionally relied on slow, sequential human iterations, creating a significant bottleneck
Discovery Loop 官方新闻稿把"人类是科研进度的瓶颈"这句话说得毫不含糊。这是判断这家公司战略定位的关键句——它不是在做"AI 辅助科研工具"，而是把人类的顺序迭代本身当作需要被优化掉的系统缺陷。
which would cut human iteration out of the loop entirely.
这句话直接点名了 Discovery Loop 的终极野心——不只是加速科研，而是让 AI 参与"创造更强 AI"这个环节本身，把人类从迭代循环里彻底移除。这是 EP.97 故事线 C 论证"RSI 正在从叙事变成组织形态"最直接的证据：Jeff Dean、Sanjay Ghemawat 等人离开 Google，创办的公司名字本身就是 RSI 的定义（Discovery Loop = 发现闭环）。
once harness design becomes an executable search space, a strong coding agent can exploit the same design space human engineers use
这句话描述的 Meta-Harness（用 coding agent 自动搜索、优化 harness 代码本身）是 RSI 在"工具层"最具体的落地案例：模型不是在改自己的权重，而是在改写包裹自己的运行系统，而这恰好是人类工程师原本要做的工作。EP.97 故事线 C 把这类证据归类为"工具层/架构层 RSI"，与"规范层 RSI"（模型自主设定目标）明确区分。
speeding up one part of a process often just shifts the bottleneck elsewhere: overall pace is capped by the parts that haven’t sped up
Anthropic 自己引用 Amdahl 定律给 RSI 叙事踩了刹车：即使编码和实验环节完全自动化，组织整体速度仍然会被没有加速的环节（比如人类代码审查、方向判断）卡住。这是判断"RSI 到底能带来多大实际提速"时最重要的限定条件，也是 EP.97 反复强调"本期观察到的一切仍停留在工具层/架构层 RSI，规范层 RSI 尚无公开证据"的直接依据。
Two human researchers, over about a week, recovered roughly 23% of that gap; the agents recovered 97% over 800 cumulative hours and used roughly $18,000 in compute.
这组对比数据是本篇最关键的"能力端 RSI"证据：同一个开放式 AI 安全研究问题，人类专家一周只能填补 23% 的性能差距，Agent 集群靠 800 小时算力（约 1.8 万美元）填补了 97%——且假设/实验/迭代全部由 Agent 自主设计，人类只定义了问题和评分标准。这也印证了 EP.97 故事线 C 里 Astra 用约 2000 美元攻克数学难题的模式：用远低于人力成本的算力换取此前需要顶尖人才才能达成的结果。
today, Anthropic engineers on average ship 8x as much code per quarter as they did from 2021-2025
这是 Anthropic 首次用内部一手数据（而非公开基准）证明 RSI 已经在"工具层"生效：不是模型能力测评分数上升，而是公司自身研发速度的真实提升。EP.97 故事线 C 用这个数字论证"资本开支即 RSI 押注"——DeepMind 首席战略官的表态不是空谈，Anthropic 自己就是活案例。
And when there is a federal financial aid review, or ”audit” as it’s called, of your institution, what they’ll do is they’ll go and take a sample of classes, and that they’ll look to see what happened in that course? Were there group interactions? Were there individual interactions? That they’ll look to see what happened and then they’ll look to see, does it meet the regulations? And have you developed faculty? Have you let them know that these are the expectations of them? That they’re looking for those sorts of things, and did it actually have an effect in the courses?