深度调研Deep Research
长篇调研与配套登记册都在这里：每项标类型与日期，结论挂公开来源；证据分级、限度明示，未决事项明确标注。Long-form research and its companion registers: each item is labelled by type and date, every claim is traceable to public sources; evidence is graded, limitations are disclosed, and unresolved questions remain explicitly marked.
这里是深研：一篇一个题目，给出完整研究、证据等级与限度。要看每周汇编，去周刊 →；要追每日动态，去新闻 →。This is Research: one topic per piece, with the full study, its evidence grade and its limitations. For the weekly digest, see the Weekly →; for the daily stream, see News →.
按发布日期，新的在前。点上面的领域只看这一类。Newest first. Pick an area above to see only that area.
模型、智能体与生产系统的结构、失效与设计取舍The structure, failure modes and design trade-offs of models, agents and production systems
规则、执法、评估与企业治理的真实状态The real state of rules, enforcement, evaluation and corporate governance
AI 给出的数学宣称与可核查性Mathematical claims made by AI, and how checkable they are
药物研发、临床转化与医疗 AI 的证据The evidence on drug discovery, clinical translation and medical AI
就业、人才与职业结构的数字与口径The numbers and calibers behind employment, talent and occupational structure
融资、估值与成本结构的数字与口径The numbers and calibers behind funding, valuations and cost structures
《队里有个内鬼，监测为什么抓不全》The Saboteur on the Team: Why the Monitors Couldn't Catch Them All
被指示暗中破坏的队友几乎从未被识破；只读公开频道的监测器抓到三到四成，置信度从未过半。A teammate instructed to sabotage was almost never identified; monitors reading only the public channel caught 30 to 40 percent, and their confidence never rose above half.
2026-11-07约 9 分钟About 10 min跨区：AI 安全与治理Also: AI Safety & Governance读全文 →Read →
《通用聊天机器人给出个体化医疗处置建议的责任边界：万博文诉 OpenAI 案追踪》The Liability Line for Individualized Medical Advice from a General-Purpose Chatbot: Tracking Wan v. OpenAI
一名长期服用抗癫痫药的用户在减药期间死亡，其父母起诉 OpenAI 的非正常死亡与产品责任案。本页分账追踪诉状指控、独立事实与责任边界三条线，不下致因与责任结论。a wrongful death and product liability suit against OpenAI, brought by the parents of a long-term user of an anti-seizure medication who died during a medication taper. The piece keeps three threads separate throughout, the complaint's allegations, the independently verifiable facts, and the liability frame, and draws no conclusion on causation or liability.
2026-10-31约 21 分钟About 31 min跨区：AI 与医疗Also: AI & Healthcare读全文 →Read →
《同一天的三本账》Three Ledgers, One Day
2026 年 10 月 26 日一天之内：Anthropic 公开 S-1、Meta 至多 950 亿美元融资计划、微软向约 6 万名合格员工发出自愿离职要约。逐页对文件：950 亿是计划上限，6 万是合格人数，S-1 的主新闻在留空处。Three filings landed on one day, October 26, 2026: Anthropic's public S-1, Meta's financing plan of up to $95 billion, and Microsoft's voluntary exit offer to some 60,000 eligible employees. Read against the documents: the $95 billion is a planned ceiling, the 60,000 an eligibility count, and the S-1's headline sits in its blanks.
2026-10-29约 15 分钟About 18 min跨区：AI 系统与架构Also: AI Systems & Architecture多源已核Multi-source verified读全文 →Read →
《Manus 过五关：40 亿美元估值，从来没成交过》The Gauntlet: Manus, the $4B Valuation That Never Closed
决定总评的两个数字都没有坐实：约 40 亿美元估值是谈判期目标价，从未被任何一方证实；约 5 亿美元年化收入只有一条报道链，没有任何申报文件支撑。Neither number the case rests on survives checking: the roughly $4 billion valuation was a pre-close target and was never confirmed by anyone, and the roughly $500 million revenue run rate rests on a single reporting chain with no filing behind it.
2026-10-31约 14 分钟About 19 min跨区：AI 系统与架构Also: AI Systems & Architecture读全文 →Read →
《企业 AI 智能体系统架构评估报告 · 示范案例》Enterprise AI Agent System Architecture Assessment Report · Demonstration Case
总体不合格。七层全不合格；十关 0 过、3 部分、7 不过；红线三条触发，另有一条临界；整改清单 18 项。an overall Fail. All seven layers Fail; the ten gates come out 0 Pass, 3 Partial, 7 Fail; three red lines are triggered and one more is borderline; the remediation list runs to 18 items.
2026-10-30约 30 分钟About 39 min跨区：AI 安全与治理Also: AI Safety & Governance读全文 →Read →
《AI 给出的证明由谁来证明》Who Checks the Proofs AI Writes?
两份 ETH 稿都卡在第四关人审：机器验证关一家只证较宽界（部分）、一家主定理可重跑，第五关判定还没有人签字。Both ETH papers are stuck at the fourth gate, human review: at the machine gate one proved only the wider bound (partial) and the other's main theorem is rebuildable; at the fifth gate, no one has signed a verdict yet.
2026-10-28约 13 分钟About 18 min跨区：AI 安全与治理Also: AI Safety & Governance读全文 →Read →
《AI 监管合规管理现状深研》The State of AI Regulation and Compliance: A Deep Dive
规则总量达到历史最高，依 AI 专门法作出的罚款仍为 0，真实执法几乎全走一般法轨道。Rulebooks are at a record high, yet fines under AI specific law still stand at zero; nearly all real enforcement runs on general law.
2026-10-26约 33 分钟About 45 min跨区：AI 基础设施与经济Also: AI Infrastructure & Economics读全文 →Read →
《AI 公司为什么进军生物科技》Why AI Companies Are Moving Into Biotech
判读标尺只有一条：想法问题正被 AI 变便宜，证明问题不会；四家动作硬进展为 0 产出披露、约 27 亿美元融资对 0 个临床候选药、唯一 I 期只有安全性、工作台与实名客户已落地。One yardstick: idea problems are getting cheaper fast, proof problems are not; the four moves show zero outputs disclosed, about $2.7B raised against zero clinical candidates, one Phase I with safety data only, and a live workbench with named customers.
2026-10-26约 30 分钟About 39 min跨区：AI 基础设施与经济Also: AI Infrastructure & Economics读全文 →Read →
《三个数字，三回事》Three Numbers, Three Different Things
三个数是三种量：就业相对降幅、求职成功率、单公司招聘占比，不能并置成一个结论。The three numbers measure three different things, a relative employment decline, a job finding rate, and one company's hiring share, and cannot be stacked into one conclusion.
2026-10-24约 10 分钟About 14 min读全文 →Read →
《黄仁勋说对了一半》Jensen Huang Got It Half Right
同一句「50%」三口径三答案：本科来源 57%、工作地 40.6%（卡内基 2025 年样本）、最顶尖层 12%（MacroPolo 2.0 版，样本与年份不同）。One "50 percent", three answers by caliber: 57 percent by undergraduate origin and 40.6 percent by workplace (Carnegie 2025 sample), 12 percent at the very top layer (MacroPolo 2.0, a different sample and year).
2026-10-22约 13 分钟About 18 min跨区：AI 基础设施与经济Also: AI Infrastructure & Economics读全文 →Read →
《工业 AI 七层单点失效》Industrial AI, Seven Layers: Which Single Layer Fails Catastrophically?
门禁层居首（27/30，与工具层同分）：它不制造错误，但让所有错误合法通行。Access Control ranks first (27/30, tied with the Tool layer): it creates no errors, but lets all of them pass as authorized.
2026-10-19约 22 分钟About 27 min跨区：AI 安全与治理Also: AI Safety & Governance读全文 →Read →
《研究实验室登记册》Research Labs Register
企业 26 家、高校与研究机构 20 家、独立机构 7 家；美国 22 家、中国 21 家。26 corporate labs, 20 universities and institutes, 7 independent bodies; 22 in the US and 21 in China.
约 11 分钟About 11 min跨区：AI 基础设施与经济Also: AI Infrastructure & Economics读全文 →Read →这一类暂时没有。Nothing here yet.