26% 的研发由 AI 主导：Anthropic 用三项指标打开实验室的门

2026 年 9 月 17 日，Anthropic 发布了一份不太像公司公告的公告：《理解前沿实验室内部 AI 研发节奏的测量方法》。它没有发布新模型，而是提出三项可被外部跟踪的指标——AI 有多少研发是 AI 自己做的、智能体的行为被监督到什么程度、算力如何分配——并附上了 2026 年 8 月自家的一份快照。这件事的价值不在于数字有多大，而在于它第一次把「实验室内部到底发生了什么」拆成了可以被第三方核查的问题清单。

把实验室内部拆成可核查的问题清单

一、一个能读懂百分比的前置刻度：Epoch AI 的自动化等级

Anthropic 用了 Epoch AI 提出的「自动化等级」（Automation Level，AL）刻度，从 AL0（无 AI 参与）到 AL5（AI 完全自主运行，无人在环）。中间两级是关键：AL3 是「AI 协作」——在人类紧密指导下完成大量工作；AL4 是「AI 主导」——给定高层提示即可端到端完成大部分任务，人类只做监督。据 2026 年 8 月的快照：Claude 在任何被测的 AI 研发子集上都还没有达到完全自主；它「主导」了 Anthropic 全部 AI 研发工作的 26%；达到或超过「AI 协作」这一级的比例高于 90%。把这三个数字放在一起，才能真正理解「26%」意味着什么：不是 AI 在做四分之一的活，而是四分之一的活已经不需要人类给出每一步的方向。

# 1. The scale that makes the number readable: Epoch AI's Automation Level
AUTOMATION_LEVELS = {
    "AL0": "no AI involvement",
    "AL3": "AI collaborates: large chunks of work under close human direction",
    "AL4": "AI leads: most of the task end-to-end from a high-level prompt, human supervises",
    "AL5": "AI operates fully autonomously, no human in the loop",
}

# Anthropic R&D Automation Index methodology, in one sentence:
# catalogue every kind of AI R&D work, rate how automated each task is, aggregate.
SNAPSHOT_AUG_2026 = {
    "fully_autonomous_subsets": 0,      # Claude is at AL5 for no measured subset
    "claude_leads": 0.26,               # AL4
    "at_or_above_collaborates": ">0.90",  # AL3+
}

def progress_is_news(prev, cur):
    return "automation moved a level, not a percentage point" if prev != cur else "flat"

二、为什么单一百分比不能当指标：Anthropic 承认的两个障碍

报告罕见地写清了自己的局限。第一，缺少通用方法论，因此目前各实验室的数字无法直接横向比较。第二，也是最微妙的一点：他们用自己的模型去评估自己的系统，这意味着「裁判」模型可能与被评估的模型共享同样的失败模式。Anthropic 给出的缓解路径有三条：由独立第三方核查；由其他开发者的模型核查，并设置护栏以限制竞争敏感数据的交换；以及把这些测量值当作触发条件，例如在模型被用于进一步 AI 研发之前，强制设置一段固定的测试窗口。对任何要做内部 AI 治理的团队来说，这里可复制的原则很朴素：如果数字无法被验证，那就把方法公开。

// 2. Why a single percentage is not a metric - the two stated obstacles
const reportingGaps = {
  methodology: "no common methodology across labs, so numbers are not comparable yet",
  judgeBias: "labs evaluate their own systems with their own models, so the judge "
             + "can share the same failure modes as the model under review",
};

const mitigationsAnthropicProposes = [
  "verification by an independent third party",
  "verification by other developers' models, with guardrails limiting the exchange of competitively sensitive data",
  "using the measures as a trigger for stronger requirements, such as a fixed testing window before a new model is used for further AI R&D",
];

// Design takeaway: if you cannot verify the number, publish the method.

无法验证的数字，就要连同方法一起公开

三、智能体监督是三个数字，而不是一个

第二项测量针对的是智能体行为的监督，Anthropic 把它拆成三个不同含义的指标：覆盖率（coverage），即一个智能体的动作在执行前后经过监控的比例；复核延迟（review latency），即动作发生到被检查之间的时间，先由自动监控、再由人复核；以及升级率（escalation rate），即被升级处理的智能体活动占比。这个拆法值得直接搬进工程实践：覆盖率回答「有没有被看见」，延迟回答「来得不来得及」，升级率回答「看见了之后有没有人管」。Anthropic 自己给出的动机也很直接——就目前的监控数据看，单个智能体很少越界；但当经济体中运行着数以百万、十亿计的智能体时，罕见事件就会规律地发生。

# 3. Oversight of agents is three different numbers, not one
def oversight_metrics(agent_actions):
    return {
        # share of an agent's actions that pass through a monitor, before or after execution
        "coverage": len([a for a in agent_actions if a.monitored]) / max(len(agent_actions), 1),

        # time between an action and its review: first automated, then human
        "review_latency": {
            "automated_monitor_seconds": percentile(agent_actions, "auto_review_latency", 50),
            "human_triage_seconds":     percentile(agent_actions, "human_review_latency", 50),
        },

        # share of agent activity that escalates
        "escalation_rate": len([a for a in agent_actions if a.escalated]) / max(len(agent_actions), 1),
    }

# Anthropic's framing is the part worth stealing: individual agents rarely
# misbehave, but at millions or billions of agents, rare events happen regularly.

四、把外部的承诺变成内部今天就能跑的测试

同一份公告里，Anthropic 重申了会「嵌入来自多个组织的独立第三方评估者」，并给他们提供与内部风险评估团队相当的流程、系统与数据访问权限；这些第三方将核查安全实践、报告事件、跟踪公告中列出的关键指标。对企业团队来说，这段话的实操价值不在 Anthropic，而在你自己：把它当作一份能力清单，逐条对比你现在能给出的审计接口。流程文档有没有？事件报告有没有？监控指标能不能导出？多数团队是在审计当天才发现缺哪一项，而不是提前。

# 4. Turn an external promise into an internal test you can run today
EMBEDDED_EVALUATOR_ACCESS = {
    "processes": ["safety practice documentation", "incident reports"],
    "systems":   "comparable to what internal risk assessment teams already have",
    "data":      "the metrics described in the publication",
    "deliverables": ["verify safety practices", "report incidents",
                     "monitor key metrics"],
    "organisations": "multiple, independent of each other",
}

def gap_analysis(current_access, promised_access):
    # Whatever the difference is, it is a to-do list you can close before an
    # auditor asks for it. Most teams discover the gap during a review, not before.
    return {k: promised_access[k] for k in promised_access
            if promised_access.get(k) != current_access.get(k)}

覆盖率、延迟、升级率：看见、来得及、有人管

五、一份可以在内部直接复用的报告格式

如果你想把这件事落地，最省力的方式是照抄它的结构：报告周期、指标族、所用刻度、实测值、方法说明、已知局限，以及「一旦越过阈值就触发什么动作」。例如 AL4 占比超过 50% 就冻结模型辅助的研发流程，直到外部评估者复核完毕；覆盖率低于 100% 就先列出所有未被监控的动作类别，再谈发版。最后别忘了复核周期——Anthropic 把节奏交给季度复核，你也应该给这份报告一个到期日，否则它会退化成一张漂亮的幻灯片。

{
  "internal_ai_rd_report": {
    "period": "2026-08",
    "metric_family": "ai_led_rd",
    "scale": "AL0-AL5",
    "reported": {
      "claude_leads_share": 0.26,
      "at_or_above_collaborates_share": 0.92,
      "fully_autonomous_share": 0.0
    },
    "method": "task catalogue + automation rating + aggregation, published",
    "known_limitations": ["self-evaluation bias", "no cross-lab standard yet"],
    "threshold_actions": {
      "if_al4_share_gt_0_5": "freeze model-assisted R&D until an external evaluator reviews the pipeline",
      "if_coverage_lt_1_0": "list every unmonitored action class before the next release"
    },
    "reviewed_by": "independent evaluator + platform engineering",
    "review_due": "quarterly"
  }
}