已知限制与参考文献
本书的边界条件与全部引用来源。
已知限制
本书的账本立场要求把未闭合的部分写清楚，而不是让它们混在正文里不被察觉。以下各项是本稿已知、但尚未解决的限制，读者在引用相关结论时应当同时引用这一节。
一、引用覆盖不完整。本书的量化基准与产业层数据已逐条核实并给出来源，包括 Epoch 的趋势率与能力指数、METR 时间跨度、ARC-AGI 各代分数与评分方法学、ICLR 2025 的模仿学习数据 scaling law、Open X-Embodiment、V-JEPA 2、台积电 CoWoS 产能与英伟达占比、GE Vernova 燃机在手订单、变压器交期与并网排队、HBM 三家份额、四大云厂商 capex，以及英伟达、台积电、博通、美光、三星电子、SpaceX、OpenEvidence、Chegg 的财报口径。仍未逐条落到一手文件的，主要是 Part 2 中一级市场标的的融资轮次与投后估值（Etched、Mercor、Cohere、Harvey、Fireworks、World Labs、Generalist AI、General Intuition、Figure 等），以及 Part 4 韩国一节的市场数据——后者现用来源中有三处为维基类转述，须换交易所与公告原文。在补齐之前，本书对这两类具体数字的引用应被视为待核。
二、两个承重参数缺少自有数据点。 Wright's Law 的学习率借自锂电、光伏、汽车与飞机（本书取 15% 基准 / 20% 乐观 / 25% 极乐观三情景，全书时间窗按基准情景给出），人形机器人本身没有任何自有学习曲线数据；各层场景的装机容量（10 万 / 50 万 / 300 万 / 1 亿台）是量级猜测。§3.3.7 附的敏感性分析表明：由此得出的结构性判断（飞轮由台数驱动）在这两个参数的合理范围内稳健，但时间落点不稳健，$20,000 售价的到达年份在 2030 至 2038 之间移动——分别对应 25% 与 15% 学习率情景的两端。
三、拟合的可复现性不足。 §3.1.2 中 MMLU 的三参数 logistic 与 ARC-AGI-2 的 logit 回归，其横轴（累计 OOM）的零点未在文中定义，逐点数据亦未附表。这意味着第三方无法复现这些拟合，而「账本可每季度用公开数据更新」这一承诺正需要这种可复现性。这是本书与自身方法论标准之间最直接的一处差距。因此本书不由这些拟合导出任何具体落点判断（见判断一），只保留不依赖复现的方向性结论。
四、时间线预测不含框架的增量信息。 §3.1.3 已声明：凡按历史趋势率折算的落点，其信息量等同于日历时间上的线性外推。四把尺子的贡献在单位换算与饱和诊断，不在年份。唯一独立于趋势外推的是 §3.1.5 的物理约束论证（133 GW 方案在当前电网交付与融资结构下经济上不可行）与由此得出的减速情景。读者若只取本书的年份而不取其条件，就取走了本书最弱的部分。
五、反方文献的覆盖仍不对称。 §3.1.7 前的反方文献一节现已收录四条：度量假象（Schaeffer 等）、时间跨度的外部效度（Ord）、METR 官方的三条限制，以及 VLA 泛化与失败恢复的反向结果，另收录 AI 产业叙事的系统性怀疑论（Narayanan & Kapoor 2025）与机器人基础模型失败模式的实证研究。但语言模型一侧的反方来自同行评议文献，物理 AI 一侧的反方多来自方法改进类论文的问题陈述部分，两者的证据地位并不对等。§3.3 的乐观程度应据此打折。 §3.3 的打折已在 §3.3.3 六「证据地位分栏」中执行，其结论是：§3.2–3.3 的框架按 A 档证据成立，时点按 C 档证据不可用。仍未对等的是：语言模型侧的反方来自同行评议文献，物理 AI 侧的反方仍多来自方法改进类论文的问题陈述部分。
六、未闭合事项统一清单。以下六项本稿尚未闭合，读者引用相关结论时应一并引用本条。前五项待补充，第六项不由本书决定：① §3.1.2 logistic 拟合的横轴零点定义与逐点数据；② §3.1.2「智能 vs OOM」四曲线拟合图的复现数据；③ §1.1.6 token 单价的端点口径核实（两个端点须为同一能力档模型）；④ §1.1.5 由每员工软件支出基数推算到 2030 年总量的路径（基数本身已有来源，见 [158]）；⑤ §3.2.1 人形机器人 BOM 成本排序的 BofA 原始出处；⑥ METR 任务集在 16 小时以上的测量不可靠（228 项任务中仅 5 项超过 16 小时，METR 官方声明该阈值以上的读数不足以支持精确比较或外推）——这是 §3.1 与 §5.4 指标① 的硬边界，只能等 METR 扩充任务集与公布新方法。
另有三项范围声明。其一，Part 4 只交付 4.1–4.2，不构成对 Situational Awareness 第四章地缘论证的完整回应，其结论不进入 Part 5 的判断链；§4.2.4 记录的是 §3.4 风险线的一次真实兑现，虽不进入判断链，仍建议与 §3.4 并读。其二，Part 1 与 Part 2 的产业数字（公司财报、融资轮次、投后估值、市场份额，以及 §4.2 韩国市场数据）未逐条落到一手文件，本书对这两类数字的引用应视为待核。其三，§1.1.4.3 提出的利率竞争解释至今未被分离，凡依赖应用层死亡率的结论（判断六）均应据此打折。
参考文献
Works Cited
1. "2025 YouTube Statistics." Teleprompter, teleprompter.com/blog/2025-youtube-statistics. Accessed 16 Aug. 2026.
2. "AgiBot World." arXiv, arxiv.org/html/2503.06669v2. Accessed 16 Aug. 2026.
3. "AI Competitiveness: How the EU Compares to the US and China." Institut der deutschen Wirtschaft, Fact Sheet 13, 2026, www.iwkoeln.de.
4. "AI Spending Tracker 2026." Value Add VC, valueaddvc.com/ai-spending. Accessed 16 Aug. 2026.
5. Alibaba. Qwen3.8-Max. 3 Aug. 2026.
6. Andreessen Horowitz (a16z). "AI Market Commentary Series." a16z, a16z.com. Accessed 16 Aug. 2026.
7. "Anthropic Accuses Alibaba of Illicitly Accessing Its AI Models." Bloomberg, 24 June 2026, www.bloomberg.com/news/articles/2026-06-24/anthropic-accuses-alibaba-of-illicitly-accessing-its-ai-models. Accessed 16 Aug. 2026.
8. "Anthropic Says Trump Admin Has Lifted Export Controls on Claude Fable 5 and Mythos 5." CNBC, 30 June 2026, www.cnbc.com/2026/06/30/anthropic-says-trump-admin-has-lifted-export-controls-on-claude-fable-5-and-mythos-5.html. Accessed 16 Aug. 2026.
9. "ARC-AGI-3 Scoring Methodology." ARC-AGI-3 Docs, ARC Prize Foundation, docs.arcprize.org/methodology. Accessed 16 Aug. 2026.注：2026 年 4 月 14 日起人类基线由「十人中次优」改为「每关中位人类」，RHAE 按 (human actions / AI actions)² 计算。本书 η 数据横跨该变更日，引用时应注意口径差异。
10. "ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence." ARC Prize Foundation, Mar. 2026, arcprize.org/media/ARC_AGI_3_Technical_Report.pdf. Accessed 16 Aug. 2026.
11. ARK Invest. "Big Ideas 2025." ARK Investment Management, 2025, www.ark-invest.com/big-ideas-2025.
12. Armstrong, J. Scott, editor. Principles of Forecasting: A Handbook for Researchers and Practitioners. Springer, 2001.
13. Aschenbrenner, Leopold. "Situational Awareness: The Decade Ahead." Self-published, June 2024, situational-awareness.ai/wp-content/uploads/2024/06/situationalawareness.pdf.
14. "ASML Raises 2026 Revenue Forecast to €43–45 Billion on Surging AI Chip Demand." MLQ News, 15 July 2026, mlq.ai/news/asml-raises-2026-revenue-forecast-to-43-45-billion-on-surging-ai-chip-demand/.
15. "ASML Reports €8.8 Billion Total Net Sales and €2.8 Billion Net Income in Q1 2026." ASML, 2026, www.asml.com/en/news/press-releases/2026/q1-2026-financial-results.
16. Assran, Mahmoud, et al. "V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning." arXiv, 11 June 2025, arxiv.org/abs/2506.09985.
17. Bass, Frank M. "A New Product Growth for Model Consumer Durables." Management Science, vol. 15, no. 5, 1969, pp. 215–227.
18. Bhatia, Ashish. "Sam Altman Says We're in the Singularity—What Does He Actually Mean?" Forbes, 28 July 2026, www.forbes.com/sites/ashishbhatia/2026/07/28/sam-altman-says-were-in-the-singularity-what-does-he-actually-mean/. Accessed 16 Aug. 2026.
19. "Big Tech's AI Spending Plans Reach $725 Billion." Tom's Hardware, www.tomshardware.com/tech-industry/big-tech/big-techs-ai-spending-plans-reach-725-billion. Accessed 16 Aug. 2026.
20. "BotQ." Figure, figure.ai/news/botq. Accessed 16 Aug. 2026.
21. "Broadcom Inc. Announces Second Quarter Fiscal Year 2026 Financial Results." Broadcom Investor Relations, 3 June 2026.
22. "Building a Synthetic Motion Generation Pipeline for Humanoid Robot Learning." NVIDIA Developer Blog, developer.nvidia.com/blog/building-a-synthetic-motion-generation-pipeline-for-humanoid-robot-learning/. Accessed 16 Aug. 2026.
23. "Business Electricity Price in China." Statista, Apr. 2026.
24. "Chegg, Inc. Quarterly Report (10-Q), Q1 2026." U.S. Securities and Exchange Commission, 2026. Q1 2026 收入 6,330 万美元、同比 −48%.
25. "China AI Model Export Controls." TechRepublic, www.techrepublic.com/article/news-apac-china-ai-model-export-controls/. Accessed 16 Aug. 2026.
26. "China Is Considering Export Controls on AI Technologies, Including Banning Local Companies from Using TSMC, Report Claims." Tom's Hardware, www.tomshardware.com/tech-industry/artificial-intelligence/china-is-considering-export-controls-on-ai-technologies-including-banning-local-companies-from-using-tsmc-report-claims-restrictions-would-also-advanced-ai-models-training-data-and-overseas-acquisitions. Accessed 16 Aug. 2026.
27. Chollet, François. "On the Measure of Intelligence." arXiv, 2019, arxiv.org/abs/1911.01547.
28. Christensen, Clayton M. The Innovator's Dilemma: When New Technologies Cause Great Firms to Fail. Harvard Business School Press, 1997.
29. "The Circular Economy of AI." The Register, 4 Nov. 2025, www.theregister.com/2025/11/04/the_circular_economy_of_ai/. Accessed 16 Aug. 2026.
30. "Cohere: $7B Valuation, $240M ARR, Enterprise AI." Value Add VC, 2026, valueaddvc.com/blog/cohere-valuation-revenue-2026-enterprise-ai-240m-arr.
31. "Companies." Y Combinator, www.ycombinator.com/companies. Accessed 16 Aug. 2026.
32. "Cosmos World Foundation Models." NVIDIA Blog, blogs.nvidia.com/blog/cosmos-world-foundation-models/. Accessed 16 Aug. 2026.
33. Cotra, Ajeya. "Draft Report on Biological Anchors for AI Timelines in 2020." Open Philanthropy, 2020, www.openphilanthropy.org/research/draft-report-on-biological-anchors-for-ai-timelines-in-2020/.
34. "Custom Silicon Inflection 2026: Hyperscaler ASICs vs. Nvidia GPU." Introl, introl.com/blog/custom-silicon-inflection-2026-hyperscaler-asics-nvidia-gpu. Accessed 16 Aug. 2026.
35. "Data Scaling Laws." Data Scaling Laws project site, data-scaling-laws.github.io. Accessed 16 Aug. 2026.
36. "DeepSeek V4-Flash 正式版上线公告" ["DeepSeek V4-Flash Official Release Announcement"]. DeepSeek, 31 July 2026.
37. "DeepSeek-V4." DeepSeek, 24 Apr. 2026, deepseek.ai/deepseek-v4. Accessed 16 Aug. 2026.
38. "DexMimicGen." arXiv, arxiv.org/html/2410.24185v1. Accessed 16 Aug. 2026.
39. DROID: A Large-Scale In-the-Wild Robot Manipulation Dataset. droid-dataset.github.io. Accessed 16 Aug. 2026. 〔待补全〕
40. "ECI Documentation — FAQ." Epoch AI, epoch.ai/data/eci-documentation/faq. Accessed 16 Aug. 2026.
41. "EDGAR Full-Text Search: Form D Filings Referencing Figure AI, Etched, Mercor and Anysphere." U.S. Securities and Exchange Commission, efts.sec.gov/LATEST/search-index. Accessed 16 Aug. 2026.
42. "Ego4D." Robotics Center, roboticscenter.ai/datasets/ego4d/. Accessed 16 Aug. 2026.
43. "Egocentric Data Collection for Robot Training." Unidata, unidata.pro/blog/egocentric-data-collection-for-robot-training/. Accessed 16 Aug. 2026.
44. "Egocentric Datasets for Robotics." Labellerr, labellerr.com/blog/egocentric-datasets-robotics/. Accessed 16 Aug. 2026.
45. "EgoScale Robot Training Framework." Labellerr, labellerr.com/blog/egoscale-robot-training-framework/. Accessed 16 Aug. 2026.
46. "EgoScale." NVIDIA Research, GEAR Lab, research.nvidia.com/labs/gear/egoscale/. Accessed 16 Aug. 2026.
47. "Electricity Monthly Update: End Use." U.S. Energy Information Administration, May 2026, www.eia.gov/electricity/monthly/update/end-use.php.
48. "Elon Musk, SpaceX IPO, AI Coding Startup Cursor Acquisition." Fortune, 16 June 2026, fortune.com/2026/06/16/elon-musk-spacex-ipo-ai-coding-startup-cursor-acquisition/. Accessed 16 Aug. 2026.
49. "Epoch Capabilities Index." Epoch AI, epoch.ai/benchmarks/eci. Accessed 16 Aug. 2026.
50. "Etched Raises $300M at a $10.3B Valuation to Scale Production of Frontier Scale Inference Hardware." GlobeNewswire, 23 July 2026, www.globenewswire.com/news-release/2026/07/23/3332366/0/en/etched-raises-300m-at-a-10-3b-valuation-to-scale-production-of-frontier-scale-inference-hardware.html.
51. Etched. "Etched Emerges from Stealth with Working Chip, $800M Raised, and Over $1B in Customer Contracts." GlobeNewswire, 30 June 2026, www.globenewswire.com/news-release/2026/06/30/3319922/0/en/Etched-Emerges-From-Stealth-With-Working-Chip-800M-Raised-and-Over-1B-in-Customer-Contracts.html. Accessed 16 Aug. 2026.
52. "Europe Needs a Strategy to Close the Artificial Intelligence Compute Gap." Bruegel, 2026, www.bruegel.org/analysis/europe-needs-strategy-close-artificial-intelligence-compute-gap.
53. "Failing Forward." arXiv, 2026, arxiv.org/abs/2605.08434.
54. "Figure AI Valuation (2026): $39B — The Biggest Pre-Revenue Bet in Robotics." Teahose, 2026, www.teahose.com/guides/figure-ai-valuation.
55. "Fireworks AI: Revenue, Valuation & Funding." Sacra, sacra.com/c/fireworks-ai/. Accessed 16 Aug. 2026.
56. "Five Big Tech Giants Have $1.65 Trillion in AI Debt Hidden from Their Balance Sheets, Nikkei Analysis Finds." Tech Startups, 23 July 2026, techstartups.com/2026/07/23/five-big-tech-giants-have-1-65-trillion-in-ai-debt-hidden-from-their-balance-sheets-nikkei-analysis-finds/. Accessed 16 Aug. 2026.
57. "Form D: Etched SPV, LLC." U.S. Securities and Exchange Commission, EDGAR, 27 Jan. 2026, www.sec.gov/Archives/edgar/data/2107488/000210748826000001/primary_doc.xml.
58. "Form D: Figure AI Inc." U.S. Securities and Exchange Commission, EDGAR, 13 Mar. 2024, CIK 0002014185.
59. "G42 Receives U.S. Approval for Advanced AI Chip Exports." G42, 20 Nov. 2025, www.g42.ai/resources/news/g42-receives-us-approval-advanced-ai-chip-exports-enabling-full-scale-deployment-trusted-ai-infrastructure.
60. "GEN-0." Generalist AI, generalistai.com/blog/gen-0. Accessed 16 Aug. 2026.
61. "General Intuition Bags $320M Series A at $2.3B to Build the AI That Learns to Act from Gamers." Tech Funding News, 2026, techfundingnews.com/khosla-wrote-two-cheques-in-three-months-now-general-intuition-wants-a-third/.
62. "Global DRAM and HBM Market Share: Quarterly." Counterpoint Research, counterpointresearch.com/en/insights/global-dram-and-hbm-market-share. Accessed 16 Aug. 2026.
63. "Graduation Rate from Seed to Series A." Carta, 2026, carta.com/data/newsletter-graduation-rate-from-seed-to-series-a/. Accessed 16 Aug. 2026.
64. "Grid Interconnection Queue Statistics 2026." Axis Intelligence, axis-intelligence.com/grid-interconnection-queue-statistics/. Accessed 16 Aug. 2026.
65. Griliches, Zvi. "Hybrid Corn: An Exploration in the Economics of Technological Change." Econometrica, vol. 25, no. 4, 1957, pp. 501–522.
66. "Groq and Nvidia Enter Non-Exclusive Inference Technology Licensing Agreement to Accelerate AI Inference at Global Scale." Groq Newsroom, groq.com/newsroom/groq-and-nvidia-enter-non-exclusive-inference-technology-licensing-agreement-to-accelerate-ai-inference-at-global-scale. Accessed 16 Aug. 2026.
67. Gu, Qiao, et al. "SAFE: Multitask Failure Detection for Vision-Language-Action Models." arXiv, 11 June 2025, arxiv.org/abs/2506.09937.
68. "Harvey AI Valuation 2026: $11B and $190M ARR." Value Add VC, 2026, valueaddvc.com/blog/harvey-ai-valuation-revenue-2026-legal-ai-11b.
69. "How Can We Get Enough Data to Train…" It Can Think, Substack, itcanthink.substack.com/p/how-can-we-get-enough-data-to-train. Accessed 16 Aug. 2026.
70. "How Many Hours of Video Are Uploaded to YouTube Every Minute?" SOAX, soax.com/research/how-many-hours-of-video-are-uploaded-to-youtube-every-minute. Accessed 16 Aug. 2026.
71. "Huawei Ascend 950 vs NVIDIA B300 and B200 for LLM Inference: TPS, Memory Bandwidth, and Cost Comparison." Spheron, 2026, www.spheron.network/blog/huawei-ascend-950-vs-nvidia-b300-b200-llm-inference-2026/.
72. "Huawei Ascend AI Chip Roadmap and System-Level Performance Data." Convequity, convequity.substack.com/p/huawei-ascend-ai-chip-roadmap-and. Accessed 16 Aug. 2026.
73. "Huawei Debuts Atlas 350 on Ascend 950PR with In-House HBM, Touting 2.8X H20 Performance." TrendForce, 23 Mar. 2026, www.trendforce.com/news/2026/03/23/news-huawei-debuts-atlas-350-on-ascend-950pr-with-in-house-hbm-touting-2-8x-h20-performance/.
74. "Humanoid Robot Data Collection Cost." DataXPower, dataxpower.com/blog/humanoid-robot-data-collection-cost. Accessed 16 Aug. 2026.
75. ICONIQ Capital. "ICONIQ Growth: State of AI Report Series." ICONIQ Capital, www.iconiqcapital.com/growth. Accessed 16 Aug. 2026.
76. IDC. 2025 年度中国云端 AI 加速器市场报告 [China Cloud AI Accelerator Market Report, 2025]. IDC, Apr. 2026. Qtd. in "China's Domestic AI Chip Push, 2026." OpenAxo, openaxo.com/innovation/china-domestic-ai-chip-2026. Accessed 16 Aug. 2026.
77. "Isaac GR00T." Pebblous Blog, blog.pebblous.ai/project/AgenticAI/isaac-groot/en/. Accessed 16 Aug. 2026.
78. Jacobides, Michael G., Thorbjørn Knudsen, and Mie Augier. "Benefiting from Innovation: Value Creation, Value Appropriation and the Role of Industry Architectures." Research Policy, vol. 35, no. 8, 2006, pp. 1200–1221.
79. Kokotajlo, Daniel, et al. AI 2027. AI Futures Project, 2025, ai-2027.com.
80. "Leopold Aschenbrenner's Situational Awareness Fund Fire Sale." CNBC, 31 July 2026, www.cnbc.com/2026/07/31/leopold-aschenbrenner-situational-awareness-fund-fire-sale.html. Accessed 16 Aug. 2026.
81. Lin, Fanqi, et al. "Data Scaling Laws in Imitation Learning for Robotic Manipulation." Proceedings of the International Conference on Learning Representations, 2025, arxiv.org/abs/2410.18647. Oral presentation.
82. Marcus, Gary. "Marcus on AI." Substack, garymarcus.substack.com. Accessed 16 Aug. 2026.
83. "Measuring AI Ability to Complete Long Software Tasks." METR, 19 Mar. 2025, metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/.
84. "Meta Q2 2026 Capex AI Buildout." TNW, thenextweb.com/news/meta-q2-2026-capex-ai-buildout. Accessed 16 Aug. 2026.
85. "Micron Technology, Inc. Reports Record Results for the Third Quarter of Fiscal 2026." Micron Investor Relations, 2026.
86. Moonshot AI. Kimi K3. 16 July 2026.
87. "Morgan Stanley: AI Spending Could…" Yahoo Finance, finance.yahoo.com/technology/ai/articles/morgan-stanley-ai-spending-could-180205911.html. Accessed 16 Aug. 2026.
88. Musk, Elon [@elonmusk]. "We Have Entered the Singularity." X, 4 Jan. 2026, x.com/elonmusk/status/2007831396333850868. Accessed 16 Aug. 2026.
89. Narayanan, Arvind, and Sayash Kapoor. "AI as Normal Technology: An Alternative to the Vision of AI as a Potential Superintelligence." Knight First Amendment Institute, 2025, normaltech.ai.
90. NFX. "AI Startup and Market Analysis Series." NFX, www.nfx.com. Accessed 16 Aug. 2026.
91. "Nvidia and OpenAI in Talks for up to $250 Billion AI Backstop." CNBC, 27 July 2026, www.cnbc.com/2026/07/27/nvidia-and-openai-in-talks-for-up-to-250-billion-dollar-ai-backstop.html. Accessed 16 Aug. 2026.
92. "NVIDIA Announces Blackwell Ultra B300 — 1.5X Faster than B200 with 288GB HBM3e and 15 PFLOPS Dense FP4." Tom's Hardware, www.tomshardware.com/pc-components/gpus/nvidia-announces-blackwell-ultra-b300-1-5x-faster-than-b200-with-288gb-hbm3e-and-15-pflops-dense-fp4. Accessed 16 Aug. 2026.
93. "NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2026." NVIDIA Newsroom, 2026.
94. "NVIDIA Announces Isaac GR00T N1 Open Humanoid Robot Model." Engineering.com, www.engineering.com/nvidia-announces-isaac-gr00t-n1-open-humanoid-robot-model/. Accessed 16 Aug. 2026.
95. "NVIDIA Blackwell Ultra and B300 Infrastructure Requirements." Introl, introl.com/blog/nvidia-blackwell-ultra-b300-infrastructure-requirements-2025. Accessed 16 Aug. 2026.
96. "Nvidia, OpenAI Financing: Jensen Huang, SSI." Axios, 27 July 2026, www.axios.com/2026/07/27/nvidia-openai-financing-ai-jensen-huang-ssi. Accessed 16 Aug. 2026.
97. O'Neill, Abby, et al. "Open X-Embodiment: Robotic Learning Datasets and RT-X Models." arXiv, 13 Oct. 2023, arxiv.org/abs/2310.08864.
98. "OpenAI Says Its Own AI Models Escaped…" The Hacker News, July 2026, thehackernews.com/2026/07/openai-says-its-own-ai-models-escaped.html. Accessed 16 Aug. 2026.
99. "OpenAI–Hugging Face Breach." InfoQ, Aug. 2026, www.infoq.com/news/2026/08/openai-huggingface-breach/. Accessed 16 Aug. 2026.
100. "OpenEvidence: Revenue, Valuation & Funding." Sacra, sacra.com/c/openevidence/. Accessed 16 Aug. 2026.
101. Ord, Toby. "Is There a Half-Life for the Success Rates of AI Agents?" arXiv, 8 May 2025, arxiv.org/abs/2505.05115.
102. "Physical Intelligence Valuation (2026): $5.6B Confirmed, $11B in Talks." Teahose, 2026, www.teahose.com/guides/physical-intelligence-valuation.
103. Porter, Michael E. Competitive Strategy: Techniques for Analyzing Industries and Competitors. Free Press, 1980.
104. Rogers, Everett M. Diffusion of Innovations. 5th ed., Free Press, 2003.
105. "Sam Altman, Singularity, OpenAI, Hugging Face Hack." Quartz, 27 July 2026, qz.com/sam-altman-singularity-openai-hugging-face-hack-072726. Accessed 16 Aug. 2026.
106. "The Sandbox That Didn't Hold: How OpenAI's Models Escaped through Artifactory Zero-Days." Thrive, thrivenextgen.com/the-sandbox-that-didnt-hold-how-openais-models-escaped-through-artifactory-zero-days/. Accessed 16 Aug. 2026.
107. Schaeffer, Rylan, Brando Miranda, and Sanmi Koyejo. "Are Emergent Abilities of Large Language Models a Mirage?" Advances in Neural Information Processing Systems 36, 2023, proceedings.neurips.cc/paper_files/paper/2023/hash/adc98a266f45005c403b8311ca7e8bd7-Abstract-Conference.html. Outstanding Paper, NeurIPS 2023.
108. SemiAnalysis. "AI Supply Chain and Capacity Analysis Series." SemiAnalysis, semianalysis.com. Accessed 16 Aug. 2026.
109. Song, Wenxuan, et al. "ReconVLA: Reconstructive Vision-Language-Action Model as Effective Robot Perceiver." arXiv, 2025, arxiv.org/abs/2508.10333.
110. "SOXX vs. SMH: Total Returns." TotalRealReturns, totalrealreturns.com/n/SOXX,SMH. Accessed 16 Aug. 2026.
111. "SpaceX (SPCX) Cursor Acquisition, IPO." CNBC, 16 June 2026, www.cnbc.com/2026/06/16/spacex-spcx-cursor-acquisition-ipo.html. Accessed 16 Aug. 2026.
112. "SpaceX IPO Takeaways: SPCX Closes at $161, Jumping 19% after Record Debut." CNBC, 12 June 2026, www.cnbc.com/2026/06/12/spacex-ipo-spcx-live-updates.html.
113. "State of AI: Q2'26 Report." CB Insights, 2026, www.cbinsights.com/research/report/ai-trends-q2-2026/. Accessed 16 Aug. 2026.
114. "The State of Startup Shutdowns 2025." SimpleClosure, Dec. 2025, simpleclosure.com/blog/posts/state-of-startup-shutdowns-2025/index.html. Accessed 16 Aug. 2026.
115. "State of Venture: Q2'26 Report." CB Insights, 2026, www.cbinsights.com/research/report/venture-trends-q2-2026/. Accessed 16 Aug. 2026.
116. Teece, David J. "Profiting from Technological Innovation: Implications for Integration, Collaboration, Licensing and Public Policy." Research Policy, vol. 15, no. 6, 1986, pp. 285–305.
117. "Tesla Optimus Factory Deployment." Optimus K, optimusk.blog/blog/tesla-optimus-factory-deployment/. Accessed 16 Aug. 2026.
118. Tetlock, Philip E., and Dan Gardner. Superforecasting: The Art and Science of Prediction. Crown, 2015.
119. "Time Horizon 1.1." METR, 29 Jan. 2026, metr.org/blog/2026-1-29-time-horizon-1-1/.
120. "Transformers in 2026: Shortage, Scramble, or Self-Inflicted Crisis?" POWER Magazine, www.powermag.com/transformers-in-2026-shortage-scramble-or-self-inflicted-crisis/. Accessed 16 Aug. 2026.
121. "Trends in Artificial Intelligence." Epoch AI, epoch.ai/trends. Accessed 16 Aug. 2026.
122. "TSMC 2025 Annual Report." Taiwan Semiconductor Manufacturing Company, 2026, investor.tsmc.com/static/annualReports/2025/english/index.html.
123. "TSMC Expands CoWoS Capacity with Nvidia Booking over Half for 2026-27." DigiTimes, 10 Dec. 2025, www.digitimes.com/news/a20251210PD218/tsmc-cowos-capacity-nvidia-equipment.html.
124. "US Approves up to 70,000 Advanced AI Chips for G42, HUMAIN." Middle East AI News, www.middleeastainews.com/p/us-approves-up-to-70000-advanced. Accessed 16 Aug. 2026.
125. "Venture Capital Hits Record Highs in Q1 2026 as AI Dominance Continues: Analysis." Crowdfund Insider, Apr. 2026, www.crowdfundinsider.com/2026/04/273054-venture-capital-hits-record-highs-in-q1-2026-as-ai-dominance-continues-analysis/. Accessed 16 Aug. 2026.
126. "Vistra Corp. (VST)." Trefis, www.trefis.com/data/companies/VST. Accessed 16 Aug. 2026.
127. Willison, Simon. "OpenAI Cyberattack." Simon Willison's Weblog, 22 July 2026, simonwillison.net/2026/Jul/22/openai-cyberattack/. Accessed 16 Aug. 2026.
128. Wright, Theodore P. "Factors Affecting the Cost of Airplanes." Journal of the Aeronautical Sciences, vol. 3, no. 4, 1936, pp. 122–128.
129. Xu, et al. "APT: Action Expert Pre-Training for Vision-Language-Action Models." arXiv, 2026, arxiv.org/abs/2606.12366.
130. "YC W26 Batch Breakdown: Deep Dive on 199 Companies with Founder Data." Extruct AI, 2026, www.extruct.ai/research/ycw26/.
131. Z.ai. GLM-5.2. 13 June 2026.
132. "'7월에만 30% 넘게 폭락' 왜 하필 코스피 투자를…폭락엔 다 이유가 있다 [투자360]" ["Down More than 30% in July Alone: There Are Reasons for the Crash"]. 헤럴드경제 [Herald Business], 2026, biz.heraldcorp.com/article/10824319.
133. "[지금이뉴스] '레버리지 ETF로 개미 손실 56조원'…암울한 전망한 글로벌 투자은행" ["Retail Losses of 56 Trillion Won on Leveraged ETFs"]. YTN, 30 July 2026, www.ytn.co.kr/_ln/0545_202607301015021122.
134. "삼성전자, 2026년 2분기 잠정실적 발표" ["Samsung Electronics Announces Preliminary Q2 2026 Results"]. Samsung Newsroom Korea, 7 July 2026.
135. "씨티 '韓 개인, 레버리지 투자로 56조 손실 추정'" ["Citi: Korean Retail Investors Estimated to Have Lost 56 Trillion Won on Leveraged Investments"]. 파이낸셜뉴스 [Financial News], 29 July 2026, www.fnnews.com/news/202607291357089358.
136. "씨티 韓개미 레버리지 ETF 손실 56조 추정…하닉만 8조 물려" ["Citi Estimates Korean Retail Investors Lost 56 Trillion Won on Leveraged ETFs"]. 한국경제 [The Korea Economic Daily], 29 July 2026, www.hankyung.com/article/2026072931236.
137. "앞으로 '레버리지 몰빵' 못한다…개미들 이미 56조 손실 추산" ["No More All-In on Leverage: Retail Losses Already Estimated at 56 Trillion Won"]. 더스쿠프 [The Scoop], 2026, www.thescoop.co.kr/news/articleView.html?idxno=311075.
138. "초유의 연이틀 서킷브레이커…IMF도 넘은 7월 폭락 충격" ["Unprecedented Consecutive Circuit Breakers: July Crash Exceeds the IMF Crisis"]. Daum News, 30 July 2026, v.daum.net/v/Udy4fxPKU6.
139. "코스피, 사상 첫 이틀 연속 서킷 브레이커…이틀 만에 865조 원 사라져 [스타트경제]" ["KOSPI Hits Circuit Breakers on Consecutive Days for the First Time; 865 Trillion Won Erased in Two Days"]. YTN, 30 July 2026, www.ytn.co.kr/_ln/0102_202607300738105634.
140. "코스피/역사/2026년" ["KOSPI / History / 2026"]. 나무위키 [Namuwiki], namu.wiki/w/코스피/역사/2026년. Accessed 16 Aug. 2026.
Works Cited — 补充条目（141–161）
说明：条目 141–162 按 MLA 第 9 版格式接续主表编号排列，下一版重排时并入主表按字母序统一编号。
141. "ElevenLabs Raises $500M from Sequoia at an $11 Billion Valuation." TechCrunch, 4 Feb. 2026, techcrunch.com/2026/02/04/elevenlabs-raises-500m-from-sequioia-at-a-11-billion-valuation/. 〔原尾注 [33]，仅覆盖 ElevenLabs 部分〕
142. "Funding for AI Dominated VC in 2025." Venture Capital Journal, 2026, www.venturecapitaljournal.com/funding-for-ai-dominated-in-vc-in-2025-crunchbase/. 数据来源 Crunchbase。〔原尾注 [15]，仅覆盖总量与占比〕
143. "Google Releases Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber: A Cheaper, More Token-Efficient Flash Tier Built for Agentic Workloads." MarkTechPost, 21 July 2026, www.marktechpost.com/2026/07/21/google-releases-gemini-3-6-flash-3-5-flash-lite-and-3-5-flash-cyber-a-cheaper-more-token-efficient-flash-tier-built-for-agentic-workloads/. 〔原尾注 [9]〕
144. "How Does Replit Make Money: $525M ARR, $9B Valuation, and Effort-Based Agent Pricing." Value Add VC, 2026, valueaddvc.com/blog/how-does-replit-make-money-525m-arr-9b-valuation-and-the-ai-agent-business-model-explained. 〔原尾注 [32]〕
145. "Inference Unit Economics: The True Cost per Million Tokens." Introl, introl.com/blog/inference-unit-economics-true-cost-per-million-tokens-guide. Accessed 18 Aug. 2026. [C 档] 〔原尾注 [39]；端点口径问题仍见「已知限制」〕
146. Prabhakar, Raghu, et al. "SambaNova SN40L: Scaling the AI Memory Wall with Dataflow and Composition of Experts." arXiv, 2024, arxiv.org/abs/2405.07518. [A 档] 〔原尾注 [43]〕
147. "Startup Shutdowns Are Accelerating." Carta, carta.com/data/startup-shutdowns-q1-2024/. Accessed 18 Aug. 2026；并见 "2025 Will Likely Be Another Brutal Year of Failed Startups, Data Suggests." TechCrunch, 26 Jan. 2025, techcrunch.com/2025/01/26/2025-will-likely-be-another-brutal-year-of-failed-startups-data-suggests. 〔原尾注 [24]〕
148. "Taalas Etches AI Models onto Transistors to Rocket Boost Inference." The Next Platform, 19 Feb. 2026, www.nextplatform.com/compute/2026/02/19/taalas-etches-ai-models-onto-transistors-to-rocket-boost-inference/. 〔原尾注 [44]〕
149. "有序扩大绿电低碳算力中心规模"（新疆经济看点·年中经济观察）。石榴云／新疆日报，2026 年 8 月 10 日，转载于新疆维吾尔自治区数字化发展局官网，sfj.xinjiang.gov.cn。〔原尾注 [36]；确认自治区发改委 2026 年上半年批复 3 个绿电直连算力项目：昌吉 36%、伊吾 60%、克拉玛依〕
150. "Clarifying Limitations of Time Horizon." METR, 22 Jan. 2026, metr.org/notes/2026-01-22-time-horizon-limitations/；并见 "This Is the Most Misunderstood Graph in AI." MIT Technology Review, 5 Feb. 2026, www.technologyreview.com/2026/02/05/1132254/this-is-the-most-misunderstood-graph-in-ai/. 〔§3.1.2 METR 表注与 §3.1.7 反方文献其五〕
151. "Frontier Models with Our Harness Achieve ~99% on ARC-AGI-3 Public." Schema, 15 July 2026, schema-harness.github.io. 自报结果，未经 ARC Prize 验证。[C 档] 〔§3.1.2 ③ 与 §3.1.7 反方文献其六〕
152. "Top AI Startups That Shut Down in 2025: What Founders Can Learn." Tech Startups, 9 Dec. 2025, techstartups.com/2025/12/09/top-ai-startups-that-shut-down-in-2025-what-founders-can-learn/. 〔§1.1.4.8 失败样本对照〕
153. METR [@METR_Evals]. "We Estimate That Claude Opus 4.6 Has a 50%-Time-Horizon of around 14.5 Hours (95% CI of 6 Hrs to 98 Hrs) on Software Tasks…" X, 2026, x.com/METR_Evals/status/2024923422867030027；并见 "Exponential Progress: Claude Opus 4.6 Has 50% Time Horizon of 14.5 Hours on METR Time Horizons Benchmark." OfficeChai, 21 Feb. 2026. 〔§3.1.2 ②〕
154. METR [@METR_Evals]. "We Estimate That, on Our Tasks, Claude Opus 4.5 Has a 50%-Time Horizon of around 4 Hrs 49 Mins (95% Confidence Interval of 1 Hr 49 Mins to 20 Hrs 25 Mins)…" X, 2026, x.com/METR_Evals/status/2002203627377574113. 〔§3.1.2 ②〕
155. "Claude Mythos Shows 50% Time Horizon of 16+ Hours on METR Benchmark." OfficeChai, 2026；评测于 2026 年 3 月进行。METR 声明该阈值以上的测量「不足以支持精确比较或外推」，228 项任务中仅 5 项超过 16 小时。并见 "Task-Completion Time Horizons of Frontier AI Models." METR, metr.org/time-horizons/（「以当前任务集，16 小时以上的测量不可靠」）。〔§3.1.2 ②、§5.4 指标①〕
156. "Claude Opus 5 — ARC-AGI Results." ARC Prize, 24 July 2026, arcprize.org/results/anthropic-claude-opus-5. 经 ARC Prize 验证：ARC-AGI-3（High 档，Public Demo environments）30.16%；ARC-AGI-1 97.5%；ARC-AGI-2 90.4%（Max 档）。〔§3.1.2 ③〕
157. Berman, Jeremy [@jeremyberman]. "I Got 96.2% on ARC-AGI-3 with Opus 5, and 99.3% pass@2. The Program Is Basically Claude Code + Opus 5 (High), One Action Command, and Filesystem Logs. Almost Nothing ARC Specific." X, 2026, x.com/jeremyberman/status/2087633198822117446；代码见 github.com/jerber/arc-code. 自报结果，未经 ARC Prize 验证。[C 档] 〔§3.1.2 ③〕
158. "Average SaaS Spend per Employee in 2026: The Definitive Benchmark." Cledara, 2026, www.cledara.com/blog/average-saas-spend-per-employee-2026. 基于 87 国、180 万笔 SaaS 采购、6,800 款工具的分析。[B 档] 〔§1.1.5 每员工软件支出锚点〕
159. "OpenAI vs Anthropic Revenue Accounting Dispute: $74B Gross vs $41B Net ARR Explained." Value Add VC, 2026, valueaddvc.com/blog/openai-vs-anthropic-revenue-accounting-dispute-74b-gross-vs-41b-net-arr-explained. 含 2026-04-13 OpenAI 首席营收官内部备忘录所主张的总额→净额调整。[C 档] 〔§Part 2 第三层口径战争〕
160. Ha, David, and Jürgen Schmidhuber. "World Models." arXiv, 2018, arxiv.org/abs/1803.10122. 〔§3.2.1 广义世界模型谱系〕
161. Sutton, Richard S. "Dyna, an Integrated Architecture for Learning, Planning, and Reacting." ACM SIGART Bulletin, vol. 2, no. 4, 1991, pp. 160–163；并见 LeCun, Yann. "A Path Towards Autonomous Machine Intelligence." 2022, openreview.net/forum?id=BZ5a1r-kVsf. 〔§3.2.1 广义世界模型谱系〕
162. "ChatGPT Reaches 900M Weekly Active Users, but Gemini Adoption Continues to Heat Up: Report." Seeking Alpha, 2026；并见 "Google Closes In on Another Billion User Product with Gemini." TechCrunch, 23 July 2026, techcrunch.com/2026/07/23/google-closes-in-on-another-billion-user-product-with-gemini/. 〔§1.1.6 五 需求端 C 端规模〕
附：正文 [n] 尾注 ↔ 本表条目编号对照表
已确认对照：[1] → 88（Musk X 帖）；[6] → 110（SOXX/SMH 回报对比）；[10] → 37（DeepSeek-V4）；[11] → 36（DeepSeek V4-Flash 公告）；[12] → 131（Z.ai GLM-5.2）；[13] → 86（Moonshot Kimi K3）；[14] → 5（Alibaba Qwen3.8-Max）；[22] → 63（Carta 毕业率数据）；[25] → 114（SimpleClosure 关停报告）；[47] → 76（IDC 中国云端 AI 加速器报告）；[58] → 108（SemiAnalysis 系列）；[60] → 8（CNBC（出口管制解除报道））；[62] → 25（TechRepublic 出口管制）；[63] → 7（Bloomberg（Anthropic 指控报道））；[64] → 98（The Hacker News（模型逃逸报道））；[65] → 106（OpenAI 沙箱逃逸分析）；[66] → 99（InfoQ（OpenAI–Hugging Face 入侵））。
正文尾注与本表条目的对照（承接上段）：[9] → 143（Gemini Flash 系列发布）；[15] → 142（2025 年 AI 融资总量与占比）；[24] → 147（Carta 清盘数据）；[27] → 主表 Fortune／CB Insights 口径；[32] → 144（Replit ARR）；[33] → 141（ElevenLabs D 轮）；[36] → 149（新疆绿电直连）；[39] → 145（推理单位成本，端点口径见已知限制）；[43] → 146、[44] → 148、[45] → 主表 d-Matrix 等公司口径（C 档）；[50] → 162（ChatGPT 周活与 Gemini 月活）。至此正文全部尾注均有对应条目，不存在正文主张依赖无出处尾注的情形。
按 MLA 规范不宜入表（口头发言 / 券商转载纪要），已在正文相应位置直接交代来源性质：[26]、[48]。
〔待核〕以下尾注的条目对照尚待作者核实：[2]、[3]、[4]、[5]、[7]、[8]、[16]、[17]、[20]、[21]、[29]、[30]、[31]、[34]、[35]、[37]、[38]、[40]、[41]、[42]、[46]、[49]、[51]、[52]、[53]、[54]、[55]、[56]、[57]、[59]、[61]、[67]、[68]。