参考文献

共 279 条，编号与各章正文中的 [n] 一致；每条下方列出引用它的章节，点击跳到该章第一次引用处。

[1] Kaplan J, McCandlish S, Henighan T, et al. Scaling Laws for Neural Language Models. arXiv:2001.08361, 2020.被引用于：第 1 章（2 处） · 第 2 章（6 处） · 第 6 章（10 处） · 第 9 章（2 处）

[2] Hoffmann J, Borgeaud S, Mensch A, et al. Training Compute-Optimal Large Language Models (Chinchilla). arXiv:2203.15556, 2022.被引用于：第 1 章 · 第 3 章（3 处） · 第 6 章（4 处） · 第 9 章（2 处）

[3] Touvron H, Lavril T, Izacard G, et al. LLaMA: Open and Efficient Foundation Language Models. arXiv:2302.13971, 2023.被引用于：第 3 章 · 第 4 章 · 附录 A

[4] Sardana N, Portes J, Doubov S, Frankle J. Beyond Chinchilla-Optimal: Accounting for Inference in Language Model Scaling Laws. ICML 2024 (arXiv:2401.00448).被引用于：第 3 章 · 第 6 章

[5] Achiam J, Adler S, Agarwal S, et al. GPT-4 Technical Report. arXiv:2303.08774, 2023.被引用于：第 1 章 · 第 4 章 · 第 5 章 · 第 7 章（2 处） · 第 8 章 · 附录 A

[6] Yang A, Xiao B, Wang B, et al. Baichuan 2: Open Large-Scale Language Models. arXiv:2309.10305, 2023.被引用于：第 4 章

[7] Peebles W, Xie S. Scalable Diffusion Models with Transformers (DiT). ICCV 2023.被引用于：第 4 章 · 第 6 章

[8] Esser P, Kulal S, Blattmann A, et al. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis (SD3). arXiv:2403.03206, 2024.被引用于：第 4 章

[9] Clark A, de las Casas D, Guy A, et al. Unified Scaling Laws for Routed Language Models. ICML 2022.被引用于：第 4 章 · 第 6 章（6 处）

[10] Tworkowski S, Staniszewski K, Pacek M, Wu Y, Michalewski H, Miłoś P. Focused Transformer: Contrastive Training for Context Scaling. arXiv:2307.03170, 2023.被引用于：第 6 章

[11] Chen Y, Qian S, Tang H, et al. LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models. ICLR 2024 (arXiv:2309.12307).被引用于：第 6 章

[12] Muennighoff N, Rush A, Barak B, et al. Scaling Data-Constrained Language Models. NeurIPS 2023 (arXiv:2305.16264).被引用于：第 4 章（3 处） · 第 6 章（2 处）

[13] Goyal S, Maini P, Lipton Z C, Raghunathan A, Kolter J Z. Scaling Laws for Data Filtering—Data Curation Cannot Be Compute Agnostic. CVPR 2024.被引用于：第 4 章 · 第 6 章（4 处）

[14] Gao L, Schulman J, Hilton J. Scaling Laws for Reward Model Overoptimization. ICML 2023 (arXiv:2210.10760).被引用于：第 4 章 · 第 6 章

[15] Ouyang X, Ge T, Hartvigsen T, Zhang Z, Mi H, Yu D. Low-Bit Quantization Favors Undertrained LLMs: Scaling Laws for Quantized LLMs with 100T Training Tokens. arXiv:2411.17691, 2024.被引用于：第 4 章 · 第 6 章（2 处）

[16] Zhang B, Luo L, Chen Y, et al. Wukong: Towards a Scaling Law for Large-Scale Recommendation. arXiv:2403.02545, 2024.被引用于：第 4 章 · 第 6 章（2 处）

[17] Yuan Z, Yuan H, Li C, et al. Scaling Relationship on Learning Mathematical Reasoning with Large Language Models. arXiv:2308.01825, 2023.被引用于：第 4 章

[18] Chung H W, Hou L, Longpre S, et al. Scaling Instruction-Finetuned Language Models. arXiv:2210.11416, 2022.被引用于：第 4 章

[19] Tan H, Xu X, Ying C, et al. ManiBox: Enhancing Spatial Grasping Generalization via Scalable Simulation Data Generation. arXiv:2411.01850, 2024.被引用于：第 4 章

[20] Snell C, Lee J, Xu K, Kumar A. Scaling LLM Test-Time Compute Optimally Can Be More Effective than Scaling Model Parameters. arXiv:2408.03314, 2024.被引用于：第 4 章 · 第 6 章 · 附录 A（2 处）

[21] Michaud E J, Liu Z, Girit U, Tegmark M. The Quantization Model of Neural Scaling. NeurIPS 2023.被引用于：第 5 章（5 处） · 第 6 章（2 处）

[22] Sharma U, Kaplan J. Scaling Laws from the Data Manifold Dimension. JMLR 2022 (arXiv:2004.10802, 2020; arXiv 版题为 A Neural Scaling Law from the Dimension of the Data Manifold).被引用于：第 5 章（3 处） · 第 6 章（6 处）

[23] Pope P, Zhu C, Abdelkader A, Goldblum M, Goldstein T. The Intrinsic Dimension of Images and Its Impact on Learning. ICLR 2021 (arXiv:2104.08894).被引用于：第 5 章（2 处）

[24] Levina E, Bickel P J. Maximum Likelihood Estimation of Intrinsic Dimension. NeurIPS 2004.被引用于：第 5 章

[25] Facco E, d'Errico M, Rodriguez A, Laio A. Estimating the Intrinsic Dimension of Datasets by a Minimal Neighborhood Information. Scientific Reports 7:12140, 2017 (arXiv:1803.06992).被引用于：第 5 章

[26] Liu Y, Liu Z, Gore J. Superposition Yields Robust Neural Scaling. NeurIPS 2025 (arXiv:2505.10465).被引用于：第 5 章（3 处）

[27] Wei J, Tay Y, Bommasani R, et al. Emergent Abilities of Large Language Models. TMLR 2022 (arXiv:2206.07682).被引用于：第 1 章 · 第 8 章（2 处）

[28] Schaeffer R, Miranda B, Koyejo S. Are Emergent Abilities of Large Language Models a Mirage? NeurIPS 2023 (arXiv:2304.15004).被引用于：第 1 章 · 第 8 章（2 处）

[29] Pearl J, Mackenzie D. The Book of Why. Basic Books, 2018.被引用于：第 11 章

[30] Rosenblatt F. The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain. Psychological Review, 1958.被引用于：第 9 章（2 处）

[31] Vapnik V, Chervonenkis A. On the Uniform Convergence of Relative Frequencies of Events to Their Probabilities. Theory of Probability and its Applications, 1971.被引用于：第 9 章

[32] Cortes C, Jackel L D, Solla S A, Vapnik V, Denker J S. Learning Curves: Asymptotic Values and Rate of Convergence. NeurIPS 1993.被引用于：第 9 章（2 处）

[33] Banko M, Brill E. Scaling to Very Very Large Corpora for Natural Language Disambiguation. ACL 2001.被引用于：第 9 章（2 处）

[34] Sutton R. The Bitter Lesson. 2019-03-13. http://www.incompleteideas.net/IncIdeas/BitterLesson.html被引用于：第 9 章

[35] Hestness J, Narang S, Ardalani N, et al. Deep Learning Scaling is Predictable, Empirically. arXiv:1712.00409, 2017.被引用于：第 6 章 · 第 9 章（3 处）

[36] Belkin M, Hsu D, Ma S, Mandal S. Reconciling Modern Machine Learning Practice and the Classical Bias–Variance Trade-off. PNAS 2019.被引用于：第 9 章（2 处）

[37] Nakkiran P, Kaplun G, Bansal Y, Yang T, Barak B, Sutskever I. Deep Double Descent: Where Bigger Models and More Data Hurt. ICLR 2020.被引用于：第 9 章（2 处）

[38] Xiao C, Cai J, Zhao W, et al. Densing Law of LLMs. Nature Machine Intelligence 7(11):1823–1833, 2025 (doi:10.1038/s42256-025-01137-0; arXiv:2412.04315, 2024).被引用于：第 9 章 · 第 10 章

[39] Aschenbrenner L. Situational Awareness: The Decade Ahead. situational-awareness.ai, 2024-06.被引用于：第 10 章（2 处）

[40] Villalobos P, Ho A, Sevilla J, Besiroglu T, Heim L, Hobbhahn M (Epoch AI). Position: Will We Run Out of Data? Limits of LLM Scaling Based on Human-Generated Data. ICML 2024 (arXiv:2211.04325).被引用于：第 1 章 · 第 10 章

[41] Gholami A, Yao Z, Kim S, Hooper C, Mahoney M W, Keutzer K. AI and Memory Wall. IEEE Micro, 2024 (arXiv:2403.14123).被引用于：第 10 章

[42] Tong S, Fan D, Nguyen J, et al. Beyond Language Modeling: An Exploration of Multimodal Pretraining. arXiv:2603.03276, 2026.被引用于：第 4 章（6 处） · 第 6 章（7 处）

[43] ERNIE Team, Baidu. ERNIE 5.0 Technical Report. arXiv:2602.04705, 2026.被引用于：第 4 章

[44] Hu S, Liu X, Han X, et al. Predicting Emergent Abilities with Infinite Resolution Evaluation. ICLR 2024 (arXiv:2310.03262).被引用于：第 7 章 · 第 8 章（3 处）

[45] Ruan Y, Maddison C J, Hashimoto T. Observational Scaling Laws and the Predictability of Language Model Performance. arXiv:2405.10938, 2024.被引用于：第 1 章 · 第 7 章 · 第 8 章（2 处）

[46] Ganguli D, Hernandez D, Lovitt L, et al. Predictability and Surprise in Large Generative Models. FAccT 2022 (arXiv:2202.07785).被引用于：第 8 章

[47] Shumailov I, Shumaylov Z, Zhao Y, Papernot N, Anderson R, Gal Y. AI Models Collapse When Trained on Recursively Generated Data. Nature 631, 2024.被引用于：第 6 章 · 第 8 章 · 第 11 章

[48] Hubinger E, Denison C, Mu J, et al. Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training. arXiv:2401.05566, 2024.被引用于：第 8 章（2 处）

[49] Burns C, Izmailov P, Kirchner J H, et al. Weak-to-Strong Generalization: Eliciting Strong Capabilities with Weak Supervision. ICML 2024 (arXiv:2312.09390).被引用于：第 8 章（2 处）

[50] OpenAI. Learning to Reason with LLMs. OpenAI Blog, 2024-09-12. https://openai.com/index/learning-to-reason-with-llms/被引用于：第 4 章 · 第 6 章 · 第 8 章 · 第 9 章（2 处） · 第 11 章 · 附录 A（3 处）

[51] Kwa T, West B, Becker J, et al. Measuring AI Ability to Complete Long Tasks. METR, 2025-03-19 (arXiv:2503.14499; 博客页现题为 Measuring AI Ability to Complete Long Software Tasks). https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/被引用于：第 11 章

[52] LeCun Y. Lex Fridman Podcast #416 Transcript. Lex Fridman Podcast, 2024-03-07. https://lexfridman.com/yann-lecun-3-transcript/ ; LeCun Y. LinkedIn note on text vs sensory data bandwidth, 2024-01. https://www.linkedin.com/posts/yann-lecun_ive-made-that-point-before-llm-1e13-activity-7156484065603280896-QH63被引用于：第 11 章

[53] Meinke A, Schoen B, Scheurer J, Balesni M, Shah R, Hobbhahn M (Apollo Research). Frontier Models are Capable of In-Context Scheming. arXiv:2412.04984, 2024.被引用于：第 8 章（2 处）

[54] Ouyang L, Wu J, Jiang X, et al. Training Language Models to Follow Instructions with Human Feedback. NeurIPS 2022 (arXiv:2203.02155).被引用于：第 8 章 · 附录 A（2 处）

[55] Open X-Embodiment Collaboration. Open X-Embodiment: Robotic Learning Datasets and RT-X Models. arXiv:2310.08864, 2023.被引用于：第 11 章

[56] Black K, Brown N, Driess D, et al. π0: A Vision-Language-Action Flow Model for General Robot Control. arXiv:2410.24164, 2024.被引用于：第 11 章

[57] Driess D, Xia F, Sajjadi M S M, et al. PaLM-E: An Embodied Multimodal Language Model. ICML 2023 (arXiv:2303.03378); Kim M J, Pertsch K, Karamcheti S, et al. OpenVLA: An Open-Source Vision-Language-Action Model. CoRL 2024 (arXiv:2406.09246).被引用于：第 11 章

[58] Chollet F. OpenAI o3 Breakthrough High Score on ARC-AGI-Pub. ARC Prize Blog, 2024-12-20. https://arcprize.org/blog/oai-o3-pub-breakthrough被引用于：第 4 章 · 第 8 章 · 第 9 章 · 附录 A（3 处）

[59] Huh M, Cheung B, Wang T, Isola P. The Platonic Representation Hypothesis. arXiv:2405.07987, 2024.被引用于：第 6 章

[60] Sharma M, Tong M, Korbak T, et al. Towards Understanding Sycophancy in Language Models. ICLR 2024 (arXiv:2310.13548).被引用于：第 8 章（2 处） · 第 11 章

[61] Ho A, Besiroglu T, Erdil E, et al. Algorithmic Progress in Language Models. NeurIPS 2024 (arXiv:2403.05812). https://epoch.ai/publications/algorithmic-progress-in-language-models ; Epoch AI. Trends in AI (dashboard). https://epoch.ai/trends被引用于：第 6 章（2 处） · 第 10 章（3 处）

[62] Goldman Sachs Research. GS SUSTAIN: Data Center Power Demand: The 6 Ps Driving Growth and Constraints. Goldman Sachs, 2025-11-10. https://www.goldmansachs.com/insights/goldman-sachs-research/data-center-power-demand-the-6-ps-driving-growth-and-constraints被引用于：第 10 章

[63] Chen L, Wu P, Chitta K, Jaeger B, Geiger A, Li H. End-to-End Autonomous Driving: Challenges and Frontiers. IEEE TPAMI, 2024 (arXiv:2306.16927); Elluswamy A (Tesla). Building Foundation Models for Autonomy. Keynote, CVPR 2023 Workshop on Autonomous Driving (WAD), 2023. https://cvpr2023.wad.vision/被引用于：第 11 章

[64] Li H, Zheng W, Wang Q, et al. (StepFun). Predictable Scale: Part I, Step Law — Optimal Hyperparameter Scaling Law in Large Language Model Pre-training. arXiv:2503.04715, 2025. https://github.com/step-law/steplaw被引用于：第 4 章（6 处） · 第 6 章

[65] Bahri Y, Dyer E, Kaplan J, Lee J, Sharma U. Explaining Neural Scaling Laws. PNAS 121(27), 2024 (arXiv:2102.06701, 2021).被引用于：第 5 章（4 处） · 第 6 章（9 处）

[66] Tay Y, Dehghani M, Abnar S, et al. Scaling Laws vs Model Architectures: How does Inductive Bias Influence Scaling? Findings of EMNLP 2023 (arXiv:2207.10551, 2022).被引用于：第 6 章（11 处）

[67] Henighan T, Kaplan J, Katz M, et al. Scaling Laws for Autoregressive Generative Modeling. arXiv:2010.14701, 2020.被引用于：第 6 章（6 处）

[68] Hernandez D, Kaplan J, Henighan T, McCandlish S. Scaling Laws for Transfer. arXiv:2102.01293, 2021.被引用于：第 6 章（2 处）

[69] Hutter M. Learning Curve Theory. arXiv:2102.04074, 2021.被引用于：第 5 章（3 处） · 第 6 章（4 处）

[70] Maloney A, Roberts D A, Sully J. A Solvable Model of Neural Scaling Laws. arXiv:2210.16859, 2022.被引用于：第 5 章（2 处） · 第 6 章（3 处）

[71] Bansal Y, Ghorbani B, Garg A, et al. Data Scaling Laws in NMT: The Effect of Noise and Architecture. ICML 2022 (arXiv:2202.01994).被引用于：第 6 章（5 处） · 第 7 章

[72] Aghajanyan A, Yu L, Conneau A, et al. Scaling Laws for Generative Mixed-Modal Language Models. ICML 2023 (arXiv:2301.03728).被引用于：第 6 章（10 处）

[73] Amari S, Fujita N, Shinomoto S. Four Types of Learning Curves. Neural Computation 4(4):605–618, 1992.被引用于：第 9 章

[74] Rosenfeld J S, Rosenfeld A, Belinkov Y, Shavit N. A Constructive Prediction of the Generalization Error Across Scales. ICLR 2020 (arXiv:1909.12673).被引用于：第 6 章 · 第 9 章

[75] Lovelace J, Belardi C, Kundurthy S, Sudhakar S, Weinberger K Q. Prescriptive Scaling Laws for Data Constrained Training. arXiv:2605.01640, 2026.被引用于：第 4 章

[76] Besiroglu T, Erdil E, Barnett M, You J. Chinchilla Scaling: A replication attempt. arXiv:2404.10102, 2024.被引用于：第 3 章 · 第 6 章（4 处） · 第 7 章（2 处）

[77] Pearce T, Song J. Reconciling Kaplan and Chinchilla Scaling Laws. TMLR 2024 (arXiv:2406.12907).被引用于：第 2 章 · 第 3 章 · 第 7 章（2 处）

[78] Weng L. Scaling Laws, Carefully. Lil'Log, 2026-06-24. https://lilianweng.github.io/posts/2026-06-24-scaling-laws/被引用于：第 7 章 · 第 9 章

[79] Bordelon B, Canatar A, Pehlevan C. Spectrum Dependent Learning Curves in Kernel Regression and Wide Neural Networks. ICML 2020 (arXiv:2002.02561).被引用于：第 5 章（2 处）

[80] Spigler S, Geiger M, Wyart M. Asymptotic Learning Curves of Kernel Methods: Empirical Data vs. Teacher-Student Paradigm. J. Stat. Mech. 2020 (arXiv:1905.10843).被引用于：第 5 章（2 处）

[81] Cagnetta F, Petrini L, Tomasini U M, Favero A, Wyart M. How Deep Neural Networks Learn Compositional Data: The Random Hierarchy Model. Physical Review X 14, 2024 (arXiv:2307.02129).被引用于：第 5 章

[82] Elhage N, Hume T, Olsson C, et al. Toy Models of Superposition. Transformer Circuits Thread, Anthropic, 2022-09-14. https://transformer-circuits.pub/2022/toy_model/index.html被引用于：第 5 章（2 处）

[83] Johnson W B, Lindenstrauss J. Extensions of Lipschitz Mappings into a Hilbert Space. Contemporary Mathematics 26:189–206, 1984.被引用于：第 5 章

[84] Delétang G, Ruoss A, Duquenne P-A, et al. Language Modeling Is Compression. ICLR 2024 (arXiv:2309.10668).被引用于：第 6 章（3 处） · 第 8 章 · 第 11 章

[85] Silver D, Sutton R S. Welcome to the Era of Experience. 2025-04 preprint (essay, in: Designing an Intelligence, MIT Press).被引用于：第 11 章（2 处）

[86] Guo D, Yang D, Zhang H, et al. (DeepSeek-AI). DeepSeek-R1 Incentivizes Reasoning in LLMs through Reinforcement Learning. Nature 645:633–638, 2025 (arXiv:2501.12948; arXiv 版题为 DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning).被引用于：第 4 章 · 第 8 章 · 第 11 章 · 附录 A（2 处）

[87] Yue Y, Chen Z, Lu R, et al. Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? NeurIPS 2025 (Oral) (arXiv:2504.13837).被引用于：第 11 章（2 处）

[88] Khatri D, Madaan L, Tiwari R, et al. (Meta, UT Austin 等). The Art of Scaling Reinforcement Learning Compute for LLMs (ScaleRL). arXiv:2510.13786, 2025.被引用于：第 4 章 · 第 6 章（2 处） · 第 11 章

[89] Novikov A, Vũ N, Eisenberger M, et al. (Google DeepMind). AlphaEvolve: A Coding Agent for Scientific and Algorithmic Discovery. arXiv:2506.13131, 2025.被引用于：第 11 章

[90] Prime Intellect. Environments Hub: A Community Hub To Scale RL To Open AGI. Prime Intellect Blog, 2025-08-27. https://www.primeintellect.ai/blog/environments被引用于：第 11 章

[91] Lu K. The Only Important Technology Is The Internet. Essay, 2025-07. https://kevinlu.ai/the-only-important-technology-is-the-internet被引用于：第 11 章

[92] Rae J. Compression for AGI. Stanford MLSys Seminar, 2023; Sutskever I. An Observation on Generalization. Simons Institute, 2023.被引用于：第 11 章

[93] Generalist AI. GEN-0: Embodied Foundation Models That Scale with Physical Interaction. Generalist AI Blog, 2025-11-04. https://generalistai.com/blog/gen-0被引用于：第 4 章 · 第 11 章（2 处）

[94] Zheng R, Niu D, Xie Y, et al. (NVIDIA GEAR). EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data. arXiv:2602.16710, 2026.被引用于：第 4 章（2 处） · 第 11 章

[95] Li F-F. From Words to Worlds: Spatial Intelligence is AI's Next Frontier. 2025-11-10. https://drfeifei.substack.com/p/from-words-to-worlds-spatial-intelligence ; World Labs. Marble: A Multimodal World Model. World Labs Blog, 2025-11-12. https://www.worldlabs.ai/blog/marble-world-model被引用于：第 11 章

[96] AgiBot-World-Contributors, Bu Q, Cai J, et al. AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems. IROS 2025 (arXiv:2503.06669).被引用于：第 11 章

[97] NVIDIA. Cosmos World Foundation Model Platform for Physical AI. arXiv:2501.03575, 2025.被引用于：第 11 章

[98] Deng X, Da J, Pan E, et al. (Scale AI). SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks? arXiv:2509.16941, 2025; OpenAI. Why SWE-bench Verified No Longer Measures Frontier Coding Capabilities. OpenAI, 2026-02-23. https://openai.com/index/why-we-no-longer-evaluate-swe-bench-verified/ ; OpenAI. Separating Signal from Noise in Coding Evaluations. OpenAI, 2026-07-08. https://openai.com/index/separating-signal-from-noise-coding-evaluations/被引用于：第 11 章

[99] Wang H, Mang Q, Cheung A, Sen K, Song D. How We Broke Top AI Agent Benchmarks: And What Comes Next. UC Berkeley RDI Blog, 2026-04. https://rdi.berkeley.edu/blog/trustworthy-benchmarks-cont/ ; Wang H, Li H, Mang Q, Cheung A, Sen K, Song D. Do Androids Dream of Breaking the Game? Systematically Auditing AI Agent Benchmarks with BenchJack. arXiv:2605.12673, 2026.被引用于：第 11 章

[100] Pandey A, Ontiveros J E, Patel D. How AI Labs Are Solving the Power Crisis: The Onsite Gas Deep Dive. SemiAnalysis, 2025-12-30. https://newsletter.semianalysis.com/p/how-ai-labs-are-solving-the-power被引用于：第 10 章

[101] Lee G, Greenbaum L (Goldman Sachs Global Institute). Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out. Goldman Sachs, 2026-05-01. https://www.goldmansachs.com/insights/articles/tracking-trillions-the-assumptions-shaping-scale-of-the-ai-build-out被引用于：第 10 章（3 处）

[102] Karpathy A. Post on X ("In era of pretraining, what mattered was internet text … in this era of reinforcement learning, it is now environments …"), 2025-08-27. https://x.com/karpathy/status/1960803117689397543被引用于：第 11 章（3 处）

[103] Parker-Holder J, Fruchter S. Genie 3: A New Frontier for World Models. Google DeepMind Blog, 2025-08-05. https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/被引用于：第 11 章

[104] Porian T, Wortsman M, Jitsev J, Schmidt L, Carmon Y. Resolving Discrepancies in Compute-Optimal Scaling of Language Models. NeurIPS 2024 (Spotlight) (arXiv:2406.19146).被引用于：第 2 章 · 第 3 章 · 第 6 章（3 处） · 第 7 章（2 处）

[105] Cagnetta F, Raventós A, Ganguli S, Wyart M. Deriving Neural Scaling Laws from the Statistics of Natural Language. ICML 2026 (arXiv:2602.07488).被引用于：第 5 章（4 处） · 第 6 章（2 处）

[106] Qwen Team, Alibaba. Qwen2 Technical Report. arXiv:2407.10671, 2024.被引用于：附录 A（2 处）

[107] Qwen Team, Alibaba. Qwen2.5 Technical Report. arXiv:2412.15115, 2024.被引用于：第 3 章 · 附录 A

[108] Qwen Team, Alibaba. Qwen3 Technical Report. arXiv:2505.09388, 2025.被引用于：附录 A（2 处）

[109] DeepSeek-AI. DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model. arXiv:2405.04434, 2024. (MLA + DeepSeekMoE，236B / 21B，8.1T token)被引用于：附录 A

[110] DeepSeek-AI. DeepSeek-V3 Technical Report. arXiv:2412.19437, 2024.被引用于：附录 A

[111] Kimi Team (Moonshot AI). Kimi k1.5: Scaling Reinforcement Learning with LLMs. arXiv:2501.12599, 2025.被引用于：附录 A

[112] Kimi Team (Moonshot AI). Kimi K2: Open Agentic Intelligence. arXiv:2507.20534, 2025.被引用于：附录 A

[113] Touvron H, Martin L, Stone K, et al. Llama 2: Open Foundation and Fine-Tuned Chat Models. arXiv:2307.09288, 2023.被引用于：第 3 章 · 附录 A

[114] Llama Team, AI @ Meta. The Llama 3 Herd of Models. arXiv:2407.21783, 2024.被引用于：第 10 章 · 附录 A（2 处）

[115] Meta AI. The Llama 4 Herd: The Beginning of a New Era of Natively Multimodal AI Innovation. Meta AI Blog, 2025-04-05. https://ai.meta.com/blog/llama-4-multimodal-intelligence/ ; Llama 4 Model Card. https://github.com/meta-llama/llama-models/blob/main/models/llama4/MODEL_CARD.md被引用于：附录 A（2 处）

[116] Zeng A, Liu X, Du Z, et al. GLM-130B: An Open Bilingual Pre-trained Model. ICLR 2023 (arXiv:2210.02414).被引用于：附录 A

[117] Team GLM (Zhipu AI & Tsinghua). ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools. arXiv:2406.12793, 2024.被引用于：附录 A

[118] GLM-4.5 Team (Zhipu AI). GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models. arXiv:2508.06471, 2025.被引用于：附录 A（2 处）

[119] ByteDance Seed. Doubao-1.5-pro. ByteDance Seed, 2025-01-22. https://seed.bytedance.com/en/doubao_1_5_pro被引用于：附录 A（2 处）

[120] Patel D, Wong G. GPT-4 Architecture, Infrastructure, Training Dataset, Costs, Vision, MoE. SemiAnalysis, 2023-07-10. https://newsletter.semianalysis.com/p/gpt-4-architecture-infrastructure被引用于：附录 A

[121] OpenAI. Introducing GPT-5 / GPT-5 System Card. OpenAI, 2025-08-07. https://openai.com/index/introducing-gpt-5/被引用于：附录 A（2 处）

[122] xAI. Grok 4. xAI, 2025-07-09. https://x.ai/news/grok-4被引用于：第 11 章 · 附录 A（2 处）

[123] Anthropic. Claude 3.7 Sonnet and Claude Code. Anthropic, 2025-02-24. https://www.anthropic.com/news/claude-3-7-sonnet ; Anthropic. Claude 3.7 Sonnet System Card. Anthropic, 2025-02. https://www.anthropic.com/claude-3-7-sonnet-system-card ; Anthropic. Claude's Extended Thinking. Anthropic, 2025-02-24. https://www.anthropic.com/research/visible-extended-thinking被引用于：附录 A（2 处）

[124] MiniMax. MiniMax-01: Scaling Foundation Models with Lightning Attention. arXiv:2501.08313, 2025. (Lightning Attention + Softmax 混合 + MoE，456B 总 / 45.9B 激活、32 专家；训练 1M、推理外推 4M；含多模态 MiniMax-VL-01)被引用于：第 4 章 · 第 6 章 · 附录 A

[125] MiniMax. MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention. arXiv:2506.13585, 2025. (首个开源大规模混合注意力推理模型；CISPO；512×H800 约三周完成全量 RL)被引用于：附录 A（2 处）

[126] MiniMax. MiniMax-M2 (Model Card & Repository). MiniMax, 2025. https://huggingface.co/MiniMaxAI/MiniMax-M2 (MoE 230B 总 / 10B 激活；coding / agentic、交错思考；SWE-bench Verified 69.4)被引用于：附录 A

[127] Yang G, Hu E J, Babuschkin I, et al. Tensor Programs V: Tuning Large Neural Networks via Zero-Shot Hyperparameter Transfer (µP / µTransfer). NeurIPS 2021 (arXiv:2203.03466).被引用于：第 4 章 · 第 6 章 · 第 7 章（2 处）

[128] Xie S M, Pham H, Dong X, et al. DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining. NeurIPS 2023 (arXiv:2305.10429).被引用于：第 4 章

[129] Ye J, Liu P, Sun T, Zhan J, Zhou Y, Qiu X. Data Mixing Laws: Optimizing Data Mixtures by Predicting Language Modeling Performance. ICLR 2025 (arXiv:2403.16952).被引用于：第 4 章 · 第 6 章（3 处）

[130] Liu Q, Zheng X, Muennighoff N, et al. RegMix: Data Mixture as Regression for Language Model Pre-training. ICLR 2025 (arXiv:2407.01492).被引用于：第 4 章 · 第 6 章（3 处）

[131] Kumar T, Ankner Z, Spector B F, et al. Scaling Laws for Precision. arXiv:2411.04330, 2024.被引用于：第 3 章 · 第 4 章 · 第 6 章（4 处）

[132] Busbridge D, Shidani A, Weers F, Ramapuram J, Littwin E, Webb R. Distillation Scaling Laws. ICML 2025 (arXiv:2502.08606).被引用于：第 6 章（3 处）

[133] Yin Y, Zhao Y, Zheng M, et al. Towards Precise Scaling Laws for Video Diffusion Transformers. CVPR 2025 (arXiv:2411.17470).被引用于：第 4 章（2 处） · 第 7 章

[134] Feinberg V. Gemini Pretraining (guest lecture, Princeton COS 568), 2025-04. https://vladfeinberg.com/assets/2025-04-24-princeton-talk.pdf被引用于：第 6 章（2 处）

[135] Caballero E, Gupta K, Rish I, Krueger D. Broken Neural Scaling Laws. ICLR 2023 (arXiv:2210.14891).被引用于：第 5 章

[136] Sorscher B, Geirhos R, Shekhar S, Ganguli S, Morcos A S. Beyond Neural Scaling Laws: Beating Power Law Scaling via Data Pruning. NeurIPS 2022 (arXiv:2206.14486).被引用于：第 6 章（5 处）

[137] Ludziejewski J, Krajewski J, Adamczewski K, et al. Scaling Laws for Fine-Grained Mixture of Experts. ICML 2024 (arXiv:2402.07871; arXiv 版署名顺序为 Krajewski J, Ludziejewski J，二人为共同一作).被引用于：第 6 章（5 处）

[138] Tao C, Liu Q, Dou L, et al. Scaling Laws with Vocabulary: Larger Models Deserve Larger Vocabularies. NeurIPS 2024 (arXiv:2407.13623).被引用于：第 6 章（3 处）

[139] Frantar E, Riquelme C, Houlsby N, Alistarh D, Evci U. Scaling Laws for Sparsely-Connected Foundation Models. ICLR 2024 (Spotlight) (arXiv:2309.08520).被引用于：第 6 章（5 处）

[140] Isik B, Ponomareva N, Hazimeh H, Paparas D, Vassilvitskii S, Koyejo S. Scaling Laws for Downstream Task Performance in Machine Translation. ICLR 2025 (arXiv:2402.04177; arXiv 版题为 Scaling Laws for Downstream Task Performance of Large Language Models).被引用于：第 6 章（4 处）

[141] Mayilvahanan P, Wiedemer T, Mallick S, Bethge M, Brendel W. LLMs on the Line: Data Determines Loss-to-Loss Scaling Laws. ICML 2025 (arXiv:2502.12120).被引用于：第 6 章（2 处）

[142] Wen K, Hall D, Ma T, Liang P. Fantastic Pretraining Optimizers and Where to Find Them. ICLR 2026 (arXiv:2509.02046).被引用于：第 4 章 · 第 6 章（3 处）

[143] Bordelon B, Atanasov A, Pehlevan C. How Feature Learning Can Improve Neural Scaling Laws. ICLR 2025 (arXiv:2409.17858).被引用于：第 6 章（2 处）

[144] Dohmatob E, Feng Y, Yang P, Charton F, Kempe J. A Tale of Tails: Model Collapse as a Change of Scaling Laws. ICML 2024 (arXiv:2402.07043).被引用于：第 6 章

[145] Paquette E, Paquette C, Xiao L, Pennington J. 4+3 Phases of Compute-Optimal Neural Scaling Laws. NeurIPS 2024 (arXiv:2405.15074).被引用于：第 6 章

[146] DeepSeek-AI. DeepSeek LLM: Scaling Open-Source Language Models with Longtermism. arXiv:2401.02954, 2024.被引用于：第 4 章 · 第 6 章（3 处）

[147] Cagnetta F, Kang H, Wyart M. Learning Curves Theory for Hierarchically Compositional Data with Power-Law Distributed Features. ICML 2025 (arXiv:2505.07067).被引用于：第 5 章（2 处）

[148] Cagnetta F, Wyart M. Towards a Theory of How the Structure of Language Is Acquired by Deep Neural Networks. NeurIPS 2024 (arXiv:2406.00048).被引用于：第 5 章（2 处）

[149] Silver D, Hubert T, Schrittwieser J, et al. A General Reinforcement Learning Algorithm That Masters Chess, Shogi, and Go through Self-Play. Science 362(6419):1140–1144, 2018 (预印本 arXiv:1712.01815，题为 Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm).被引用于：第 11 章

[150] Wei J. Asymmetry of verification and verifier's rule. Blog, 2025-07-15. https://www.jasonwei.net/blog/asymmetry-of-verification-and-verifiers-law被引用于：第 11 章

[151] DeepSeek-AI. DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models. arXiv:2512.02556, 2025.被引用于：第 4 章 · 第 11 章（2 处）

[152] Patwardhan T, Dias R, Proehl E, et al. GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks. arXiv:2510.04374, 2025; OpenAI. Introducing GPT-5.2. OpenAI, 2025-12-11. https://openai.com/index/introducing-gpt-5-2/被引用于：第 11 章

[153] METR. Recent Frontier Models Are Reward Hacking. METR Blog, 2025-06-05. https://metr.org/blog/2025-06-05-recent-reward-hacking/被引用于：第 4 章 · 第 11 章

[154] Liu M, Diao S, Lu X, et al. ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models. arXiv:2505.24864, 2025.被引用于：第 4 章 · 第 11 章（2 处）

[155] Hu J, Liu M, Lu X, et al. BroRL: Scaling Reinforcement Learning via Broadened Exploration. arXiv:2510.01180, 2025.被引用于：第 4 章 · 第 6 章 · 第 11 章（2 处）

[156] Zeng Z, Ivison H, Wang Y, et al. RLVE: Scaling Up Reinforcement Learning for Language Models with Adaptive Verifiable Environments. ICML 2026 (arXiv:2511.07317).被引用于：第 4 章（2 处） · 第 11 章（2 处）

[157] Gerstgrasser M, Schaeffer R, Dey A, et al. Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data. arXiv:2404.01413, 2024.被引用于：第 8 章（2 处） · 第 11 章

[158] Shafayat S, Tajwar F, Salakhutdinov R, Schneider J, Zanette A. Can Large Reasoning Models Self-Train? arXiv:2505.21444, 2025.被引用于：第 11 章

[159] Song Y, Zhang H, Eisenach C, Kakade S, Foster D, Ghai U. Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models. ICLR 2025 (arXiv:2412.02674).被引用于：第 11 章

[160] Zhao A, Wu Y, Yue Y, et al. Absolute Zero: Reinforced Self-play Reasoning with Zero Data. arXiv:2505.03335, 2025.被引用于：第 11 章

[161] Lin F, Hu Y, Sheng P, Wen C, You J, Gao Y. Data Scaling Laws in Imitation Learning for Robotic Manipulation. ICLR 2025 (arXiv:2410.18647).被引用于：第 4 章 · 第 11 章

[162] Brohan A, Brown N, Carbajal J, et al. RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. arXiv:2307.15818, 2023.被引用于：第 11 章

[163] Wei A, Agarwal A, Chen B, Bosworth R, Pfaff N, Tedrake R. Empirical Analysis of Sim-and-Real Cotraining of Diffusion Policies for Planar Pushing from Pixels. IROS 2025 (arXiv:2503.22634).被引用于：第 4 章 · 第 11 章

[164] Assran M, Bardes A, Fan D, et al. V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning. arXiv:2506.09985, 2025.被引用于：第 11 章

[165] Baniodeh M, Goel K, Ettinger S, et al. (Waymo). Scaling Laws of Motion Forecasting and Planning — A Technical Report. arXiv:2506.08228, 2025.被引用于：第 4 章（2 处） · 第 11 章（2 处）

[166] METR. Time Horizon 1.1. METR Blog, 2026-01-29. https://metr.org/blog/2026-1-29-time-horizon-1-1/被引用于：第 4 章 · 第 11 章

[167] Patel D. Ilya Sutskever — We're moving from the age of scaling to the age of research. Dwarkesh Podcast, 2025-11-25. https://www.dwarkesh.com/p/ilya-sutskever-2被引用于：第 11 章

[168] Roose K, Newton C. Google's Gemini 3 Is Here: A Special Early Look (interview with Demis Hassabis). Hard Fork, The New York Times, 2025-11-18.被引用于：第 11 章（2 处）

[169] Amodei D. Machines of Loving Grace: How AI Could Transform the World for the Better. Essay, 2024-10. https://www.darioamodei.com/essay/machines-of-loving-grace被引用于：第 11 章

[170] Patel D. Andrej Karpathy — AGI is still a decade away. Dwarkesh Podcast, 2025-10-17. https://www.dwarkesh.com/p/andrej-karpathy被引用于：第 11 章

[171] Sutskever I. Sequence to Sequence Learning with Neural Networks: What a Decade. NeurIPS 2024 Test of Time Award talk, Vancouver, 2024-12-13. https://neurips.cc/virtual/2024/test-of-time/105032 ; YouTube 转载：https://www.youtube.com/watch?v=1yvBqasHLZs被引用于：第 1 章 · 第 9 章

[172] He K, Zhang X, Ren S, Sun J. Deep Residual Learning for Image Recognition. CVPR 2016 (arXiv:1512.03385).被引用于：第 2 章 · 第 9 章

[173] Almeida D. Scaling Laws, Honestly. Complete Skeptic (Substack), 2026-07-04. https://www.completeskeptic.com/p/scaling-laws-honestly (博客仅署名 "Diogo"，作者在文中自述曾在 OpenAI 从事 LLM 优化；姓氏 Almeida 见二手报道：https://interestingengineering.substack.com/p/23)被引用于：第 7 章

[174] Hägele A, Bakouch E, Kosson A, Ben Allal L, Von Werra L, Jaggi M. Scaling Laws and Compute-Optimal Training Beyond Fixed Training Durations. NeurIPS 2024 (Spotlight) (arXiv:2405.18392).被引用于：第 7 章

[175] Dey N, Zhang B C, Noci L, et al. Don't be lazy: CompleteP enables compute-efficient deep transformers. NeurIPS 2025 (arXiv:2505.01618).被引用于：第 4 章 · 第 7 章

[176] Patel D, Nishball D, Ontiveros J E. Multi-Datacenter Training: OpenAI's Ambitious Plan To Beat Google's Infrastructure. SemiAnalysis, 2024-09-04. https://semianalysis.com/2024/09/04/multi-datacenter-training-openais/被引用于：第 10 章

[177] Patel N. Why IBM's CEO doesn't think current AI tech can get to AGI (Decoder 播客访谈 IBM CEO Arvind Krishna). The Verge, 2025-12-01. https://www.theverge.com/podcast/829868/ibm-arvind-krishna-watson-llms-ai-bubble-quantum-computing被引用于：第 10 章

[178] Epoch AI. Data on AI Companies (revenue reports 数据表). Epoch AI, 2026 (2026-09-24 版). https://epoch.ai/data/ai-companies被引用于：第 10 章

[179] International Energy Agency. Energy and AI. IEA, 2025-04-10. https://www.iea.org/reports/energy-and-ai被引用于：第 10 章（4 处）

[180] Patel D. Dylan Patel — Deep dive on the 3 big bottlenecks to scaling AI compute. Dwarkesh Podcast, 2026-03-13. https://www.dwarkesh.com/p/dylan-patel被引用于：第 10 章（4 处）

[181] Brittain B. US judge approves Anthropic's \(\$\)1.5 billion settlement of copyright lawsuit. Reuters, 2026-07-20. https://www.reuters.com/world/us-judge-approves-anthropics-15-billion-settlement-copyright-lawsuit-2026-07-20/被引用于：第 10 章

[182] California State Legislature. SB 53: Transparency in Frontier Artificial Intelligence Act (Chapter 138, Statutes of 2025). 2025-09-29. https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202520260SB53被引用于：第 10 章

[183] Das K. GPU Cost Monitoring: Track Utilization and Attribute AI Spend (引 Cast AI《2026 State of Kubernetes Optimization Report》). Cast AI Blog, 2026-07-23. https://cast.ai/blog/gpu-cost-monitoring-kubernetes/被引用于：第 10 章

[184] Jiang Z, Lin H, Zhong Y, et al. MegaScale: Scaling Large Language Model Training to More Than 10,000 GPUs. NSDI 2024 (arXiv:2402.15627).被引用于：第 10 章

[185] Springer J M, Goyal S, Wen K, et al. Overtrained Language Models Are Harder to Fine-Tune. arXiv:2503.19206, 2025.被引用于：第 3 章

[186] Lourie N, Hu M Y, Cho K. Scaling Laws Are Unreliable for Downstream Tasks: A Reality Check. Findings of EMNLP 2025 (arXiv:2507.00885).被引用于：第 1 章 · 第 4 章 · 第 8 章

[187] Lin S, Hilton J, Evans O. TruthfulQA: Measuring How Models Mimic Human Falsehoods. ACL 2022 (arXiv:2109.07958).被引用于：第 8 章

[188] Ren R, Basart S, Khoja A, et al. Safetywashing: Do AI Safety Benchmarks Actually Measure Safety Progress? NeurIPS 2024 (arXiv:2407.21792).被引用于：第 8 章（2 处）

[189] Anthropic. Introducing 100K Context Windows. Anthropic, 2023-05-11. https://www.anthropic.com/news/100k-context-windows被引用于：附录 A（2 处）

[190] xAI. Open Release of Grok-1. xAI, 2024-03-17. https://x.ai/news/grok-os (314B MoE 基座模型的权重与架构，Apache 2.0)被引用于：附录 A（2 处）

[191] Z.ai. GLM-4.6: Advanced Agentic, Reasoning and Coding Capabilities. Z.ai Blog, 2025-09-30. https://z.ai/blog/glm-4.6被引用于：附录 A（2 处）

[192] Gemini Team, Google. Gemini: A Family of Highly Capable Multimodal Models. arXiv:2312.11805, 2023.被引用于：附录 A（3 处）

[193] Gemini Team, Google. Gemini 1.5: Unlocking Multimodal Understanding across Millions of Tokens of Context. arXiv:2403.05530, 2024.被引用于：附录 A（3 处）

[194] Comanici G, Bieber E, Schaekermann M, et al. (Gemini Team, Google). Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities. arXiv:2507.06261, 2025.被引用于：附录 A（2 处）

[195] Qwen Team. Qwen3.5: Towards Native Multimodal Agents. Alibaba Cloud Blog, 2026-02-17. https://www.alibabacloud.com/blog/602894 ; Qwen Team. Qwen3.8-Max: A New Bar for Coding and Cowork. Qwen Blog, 2026-08-02. https://qwen.ai/blog?id=qwen3.8 ; 开放权重模型卡：https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B被引用于：附录 A（3 处）

[196] DeepSeek-AI. DeepSeek-V4 Preview: Entering the Era of Affordable Million-Token Context. DeepSeek News, 2026-04-24. https://www.deepseek.com/en/news/v4-preview/ ; DeepSeek-AI. Introducing DeepSeek-V4.1-Flash. DeepSeek News, 2026-09-10. https://www.deepseek.com/en/news/deepseek-v4-1-flash/被引用于：第 11 章 · 附录 A（3 处）

[197] Kimi Team (Moonshot AI). Kimi K2.5: Visual Agentic Intelligence. arXiv:2602.02276, 2026.被引用于：附录 A

[198] MiniMax. MiniMax M3: Frontier Coding, 1M Context, Native Multimodality — All in One Model. MiniMax, 2026-06-01. https://www.minimax.io/blog/minimax-m3 ; 模型卡：https://huggingface.co/MiniMaxAI/MiniMax-M3 ; Lai X, Xu W, Yang Y, et al. MiniMax Sparse Attention. arXiv:2606.13392, 2026.被引用于：附录 A（3 处）

[199] GLM-5 Team (Z.ai). GLM-5: from Vibe Coding to Agentic Engineering. arXiv:2602.15763, 2026. 模型卡：https://huggingface.co/zai-org/GLM-5被引用于：第 11 章 · 附录 A（2 处）

[200] ByteDance Seed. Seed2.0 正式发布. 字节跳动 Seed 博客, 2026-02-14. https://seed.bytedance.com/zh/blog/seed2-0-%E6%AD%A3%E5%BC%8F%E5%8F%91%E5%B8%83被引用于：附录 A

[201] Meta Superintelligence Labs. Introducing Muse Spark: Scaling Towards Personal Superintelligence. Meta AI Blog, 2026-04-08. https://ai.meta.com/blog/introducing-muse-spark-msl/ ; Meta Superintelligence Labs. Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device. Meta AI Research, 2026-08-10. https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model被引用于：附录 A（2 处）

[202] OpenAI. Introducing GPT-5.5. OpenAI, 2026-04-23. https://openai.com/index/introducing-gpt-5-5/ ; OpenAI. GPT-6 Astra: A New Generation of Intelligence. OpenAI, 2026-09-03. https://openai.com/index/gpt-6-astra/被引用于：第 11 章（2 处） · 附录 A

[203] Anthropic. Introducing Claude Opus 4.6. Anthropic, 2026-02-05. https://www.anthropic.com/news/claude-opus-4-6 ; Anthropic. Introducing Claude Fable 5.1 and Claude Mythos 5.1. Anthropic, 2026-09-01. https://www.anthropic.com/claude-fable-and-mythos-5-1 ; Anthropic. Claude Opus 5.5. Anthropic, 2026-09-22. https://www.anthropic.com/claude-opus-5-5被引用于：附录 A

[204] Google DeepMind. Gemini 3.1 Pro — Model Card. Google DeepMind, 2026-02-19. https://deepmind.google/models/model-cards/gemini-3-1-pro/被引用于：附录 A

[205] SpaceXAI (formerly xAI). Introducing Grok 4.7. SpaceXAI, 2026-09-21. https://x.ai/news/grok-4-7被引用于：第 11 章 · 附录 A

[206] Zhang C, Bengio S, Hardt M, Recht B, Vinyals O. Understanding Deep Learning Requires Rethinking Generalization. ICLR 2017 (arXiv:1611.03530).被引用于：第 9 章

[207] Loog M, Viering T, Mey A, Krijthe J H, Tax D M J. A Brief Prehistory of Double Descent. PNAS 117(20):10625–10626, 2020 (arXiv:2004.04328).被引用于：第 9 章

[208] Coates A, Lee H, Ng A Y. An Analysis of Single-Layer Networks in Unsupervised Feature Learning. AISTATS 2011, PMLR 15:215–223 (先见于 NIPS 2010 Deep Learning Workshop).被引用于：第 9 章（2 处）

[209] Ng A. Focus on the Future, Learn From the Past. The Batch (DeepLearning.AI), 2024-12-18. https://www.deeplearning.ai/the-batch/focus-on-the-future-learn-from-the-past被引用于：第 9 章

[210] Li H, Zheng W, Wang Q, et al. Predictable Scale: Part II, Farseer: A Refined Scaling Law in Large Language Models. arXiv:2506.10972, 2025.被引用于：第 4 章（2 处）

[211] Besnier V, Cao A-Q, Ramzi E, et al. (valeo.ai). How Far Can 5,500 Hours of Driving Take You? A Scaling Law Analysis of Video Diffusion Models. ECCV 2026 DriveX Workshop (arXiv:2608.28404).被引用于：第 4 章

[212] Roberts N, Cho S, Gao Z, et al. Test-Time Scaling Makes Overtraining Compute-Optimal. arXiv:2604.01411, 2026.被引用于：第 3 章 · 第 4 章

[213] Sadhukhan R, Chen Z, Zheng H, Zhou Y, Strubell E, Chen B. Kinetics: Rethinking Test-Time Scaling Laws. arXiv:2506.05333, 2025.被引用于：第 4 章（2 处）

[214] Pearce T, Rashid T, Bignell D, Georgescu R, Devlin S, Hofmann K. Scaling Laws for Pre-training Agents and World Models. arXiv:2411.04434, 2024.被引用于：第 4 章（2 处）

[215] Yu Y, Huang L, Calotoiu A, Hoefler T. Scaling Laws of Global Weather Models. ICML 2026 (arXiv:2602.22962).被引用于：第 4 章

[216] Tian C, Chen K, Liu J, Liu Z, Zhang Z, Zhou J. Towards Greater Leverage: Scaling Laws for Efficient Mixture-of-Experts Language Models. arXiv:2507.17702, 2025.被引用于：第 4 章 · 第 6 章（6 处）

[217] Ludziejewski J, Pióro M, Krajewski J, et al. Joint MoE Scaling Laws: Mixture of Experts Can Be Memory Efficient. arXiv:2502.05172, 2025.被引用于：第 4 章 · 第 6 章（3 处）

[218] Cheng X, Tian R, Zeng W, et al. (DeepSeek-AI). Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models. arXiv:2601.07372, 2026.被引用于：第 4 章 · 第 6 章（2 处）

[219] Kimi Team (Moonshot AI). Kimi Linear: An Expressive, Efficient Attention Architecture. arXiv:2510.26692, 2025.被引用于：第 4 章 · 第 6 章

[220] Schwethelm K, Rueckert D, Kaissis G. How Much Is One Recurrence Worth? Iso-Depth Scaling Laws for Looped Language Models. arXiv:2604.21106, 2026.被引用于：第 4 章 · 第 6 章

[221] Prairie H, Novack Z, Berg-Kirkpatrick T, Fu D Y. Parcae: Scaling Laws For Stable Looped Language Models. arXiv:2604.12946, 2026.被引用于：第 4 章 · 第 6 章

[222] Chen M, Hui B, Cui Z, et al. Parallel Scaling Law for Language Models. arXiv:2505.10475, 2025.被引用于：第 4 章 · 第 6 章

[223] Xiong W, Liu J, Molybog I, et al. Effective Long-Context Scaling of Foundation Models. arXiv:2309.16039, 2023.被引用于：第 4 章 · 第 6 章（2 处）

[224] Shi J, Ma Q, Liu H, Zhao H, Hwang J-N, Li L. Intrinsic Entropy of Context Length Scaling in LLMs. arXiv:2502.01481, 2025.被引用于：第 4 章 · 第 6 章（2 处）

[225] Sun X, Li S, Xie R, et al. Scaling Laws for Floating Point Quantization Training. arXiv:2501.02423, 2025.被引用于：第 4 章

[226] Castro R L, Panferov A, Tabesh S, et al. Quartet: Native FP4 Training Can Be Optimal for Large Language Models. arXiv:2505.14669, 2025.被引用于：第 4 章

[227] Longpre S, Kudugunta S, Muennighoff N, et al. ATLAS: Adaptive Transfer Scaling Laws for Multilingual Pretraining, Finetuning, and Decoding the Curse of Multilinguality. ICLR 2026 (arXiv:2510.22037).被引用于：第 4 章（2 处） · 第 6 章（2 处）

[228] Prabhudesai M, Wu M, Zadeh A, Fragkiadaki K, Pathak D. Diffusion Beats Autoregressive in Data-Constrained Settings. arXiv:2507.15857, 2025.被引用于：第 4 章 · 第 6 章（3 处）

[229] Shukor M, Bethune L, Busbridge D, et al. Scaling Laws for Optimal Data Mixtures. arXiv:2507.09404, 2025.被引用于：第 4 章 · 第 6 章

[230] Kang F, Ardalani N, Kuchnik M, et al. Demystifying Synthetic Data in LLM Pre-training: A Systematic Study of Scaling Laws, Benefits, and Pitfalls. EMNLP 2025 (arXiv:2510.01631).被引用于：第 4 章 · 第 6 章 · 第 11 章

[231] Tan Z, Geng H, Yu X, et al. Scaling Behaviors of LLM Reinforcement Learning Post-Training: An Empirical Study in Mathematical Reasoning. ACL 2026 (arXiv:2509.25300).被引用于：第 4 章

[232] Cheng Z, Xie Y, Qu Y, et al. IsoCompute Playbook: Optimally Scaling Sampling Compute for LLM RL. arXiv:2603.12151, 2026.被引用于：第 4 章

[233] Brown B, Juravsky J, Ehrlich R, et al. Large Language Monkeys: Scaling Inference Compute with Repeated Sampling. arXiv:2407.21787, 2024.被引用于：第 4 章

[234] OpenAI. gpt-oss-120b & gpt-oss-20b Model Card. arXiv:2508.10925, 2025.被引用于：第 4 章

[235] Shen J, Bai H, Zhang L, et al. Thinking vs. Doing: Agents that Reason by Scaling Test-Time Interaction. arXiv:2506.07976, 2025.被引用于：第 4 章

[236] Singhi N, Bansal H, Hosseini A, et al. When To Solve, When To Verify: Compute-Optimal Problem Solving and Generative Verification for LLM Reasoning. COLM 2025 (arXiv:2504.01005).被引用于：第 4 章

[237] Liu Z, Wang P, Xu R, et al. (DeepSeek-AI). Inference-Time Scaling for Generalist Reward Modeling. arXiv:2504.02495, 2025.被引用于：第 4 章

[238] Ye Y, Huang Z, Xiao Y, Chern E, Xia S, Liu P. LIMO: Less is More for Reasoning. COLM 2025 (arXiv:2502.03387).被引用于：第 4 章

[239] Guha E, Marten R, Keh S, et al. OpenThoughts: Data Recipes for Reasoning Models. arXiv:2506.04178, 2025.被引用于：第 4 章

[240] Zhai J, Liao L, Liu X, et al. Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (HSTU). ICML 2024 (arXiv:2402.17152).被引用于：第 4 章

[241] Yan B, Liu S, Zeng Z, et al. Unlocking Scaling Law in Industrial Recommendation Systems with a Three-step Paradigm based Large User Model. WSDM 2026 (arXiv:2502.08309).被引用于：第 4 章

[242] Chai Z, Ren Q, Xiao X, et al. LONGER: Scaling Up Long Sequence Modeling in Industrial Recommenders. RecSys 2025 (arXiv:2505.04421).被引用于：第 4 章

[243] Naumann A, Gu X, Dimlioglu T, et al. Data Scaling Laws for End-to-End Autonomous Driving. CVPR 2025 Workshop on Autonomous Driving (arXiv:2504.04338).被引用于：第 4 章（2 处）

[244] Bhatnagar A, Jain S, Beazer J, et al. (Profluent). Scaling Unlocks Broader Generation and Deeper Functional Understanding of Proteins (ProGen3). NeurIPS 2025 (Spotlight) (bioRxiv doi:10.1101/2025.04.15.649055).被引用于：第 4 章（3 处）

[245] Kim K, Kotha S, Liang P, Hashimoto T. Pre-training under infinite compute. arXiv:2509.14786, 2025.被引用于：第 4 章

[246] Schaeffer R, Kazdan J, Hughes J, et al. How Do Large Language Monkeys Get Their Power (Laws)? ICML 2025 (Oral) (arXiv:2502.17578).被引用于：第 4 章 · 第 5 章

[247] Wu Y, Wang Y, Ye Z, Du T, Jegelka S, Wang Y. When More is Less: Understanding Chain-of-Thought Length in LLMs. arXiv:2502.07266, 2025.被引用于：第 4 章

[248] Gema A P, Hägele A, Chen R, et al. Inverse Scaling in Test-Time Compute. TMLR 2025 (arXiv:2507.14417).被引用于：第 4 章 · 第 8 章

[249] DenAdel A, Hughes M, Thoutam A, et al. Evaluating the role of pretraining dataset size and diversity on single-cell foundation model performance. Nature Methods, 2026 (doi:10.1038/s41592-026-03120-y).被引用于：第 4 章

[250] Hansen-Estruch P, Yan D, Chung C-Y, et al. Learnings from Scaling Visual Tokenizers for Reconstruction and Generation. arXiv:2501.09755, 2025.被引用于：第 4 章

[251] Kim Y, Gu K, Park C, et al. Towards a Science of Scaling Agent Systems. arXiv:2512.08296, 2025.被引用于：第 4 章

[252] Kareer S, Pertsch K, Darpinian J, et al. (Physical Intelligence). Emergence of Human to Robot Transfer in Vision-Language-Action Models. arXiv:2512.22414, 2025.被引用于：第 4 章

[253] OpenAI. Introducing OpenAI o3 and o4-mini. OpenAI, 2025-04-16. https://openai.com/index/introducing-o3-and-o4-mini/被引用于：第 11 章

[254] Whitfill P, Wu C, Becker J, Rush N. Many SWE-bench-Passing PRs Would Not Be Merged into Main. METR Notes, 2026-03-10. https://metr.org/notes/2026-03-10-many-swe-bench-passing-prs-would-not-be-merged-into-main/被引用于：第 11 章（2 处）

[255] Shao R, Li S S, Xin R, et al. Spurious Rewards: Rethinking Training Signals in RLVR. arXiv:2506.10947, 2025.被引用于：第 11 章

[256] MiniMax. MiniMax M2.5: Built for Real-World Productivity. MiniMax, 2026-02-12. https://www.minimax.io/news/MiniMax-m25被引用于：第 11 章

[257] Kwa T, Cheng V. How Does Time Horizon Vary Across Domains? METR Blog, 2025-07-14. https://metr.org/blog/2025-07-14-how-does-time-horizon-vary-across-domains/被引用于：第 11 章

[258] Generalist AI. GEN-1: Scaling Embodied Foundation Models to Mastery. Generalist AI Blog, 2026-04-02. https://generalistai.com/blog/gen-1被引用于：第 11 章

[259] Physical Intelligence, Amin A, Aniceto R, et al. π*0.6: a VLA That Learns From Experience. arXiv:2511.14759, 2025.被引用于：第 11 章

[260] Tian Y, Yang Y, Xie Y, et al. InternData-A1: Pioneering High-Fidelity Synthetic Data for Pre-training Generalist Policy. arXiv:2511.16651, 2025.被引用于：第 11 章

[261] Hafner D, Yan W, Lillicrap T. Training Agents Inside of Scalable World Models. arXiv:2509.24527, 2025.被引用于：第 11 章

[262] METR. Task-Completion Time Horizons of Frontier AI Models. METR, 2026-05-08 更新. https://metr.org/time-horizons/ ; METR. Post on Bluesky (Claude Mythos Preview 早期版本的 50% 任务时长测量), 2026-05-08. https://bsky.app/profile/metr.org/post/3mlewafi6pc2u被引用于：第 11 章

[263] Alphabet. 2026 Q2 Earnings Call (2026 年资本开支指引 1950–2050 亿美元). 2026-07-22. https://abc.xyz/investor/events/event-details/2026/2026-Q2-Earnings-Call-2026-GgTAq7Is0z/default.aspx ; Microsoft. Fiscal Year 2026 Fourth Quarter Earnings Conference Call (CY2026 约 1750 亿美元，因租约分类调整). 2026-07-29. https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4 ; Meta. Meta Reports Second Quarter 2026 Results (1300–1450 亿美元). 2026-07-29. https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Second-Quarter-2026-Results/default.aspx ; Amazon.com. Q2 2026 Earnings Call (约 2200 亿美元). 2026-07-30. https://ir.aboutamazon.com/quarterly-results/default.aspx ; Oracle. Oracle Announces Q1 Results Driven by Triple Digit Growth in Cloud Infrastructure Revenues (Q1 FY2027；FY2027 资本开支 900–950 亿美元、净现金支出不超过 700 亿). 2026-09-10. https://investor.oracle.com/investor-news/news-details/2026/Oracle-Announces-Q1-Results-Driven-by-Triple-Digit-Growth-in-Cloud-Infrastructure-Revenues/default.aspx被引用于：第 10 章

[264] Reuters. US to allow Nvidia H200 chip shipments to China, Trump says (引 Trump 在 Truth Social 的原帖). Reuters, 2025-12-08. https://www.reuters.com/world/china/us-open-up-exports-nvidia-h200-chips-china-semafor-reports-2025-12-08/被引用于：第 10 章

[265] James L. First Nvidia H200 shipments reach China, ByteDance and Tencent take deliveries as Beijing loosens its import block (引 Financial Times 报道). Tom's Hardware, 2026-08-19. https://www.tomshardware.com/pc-components/gpus/first-nvidia-h200-shipments-reach-bytedance-and-tencent-as-beijing-loosens-its-import-block被引用于：第 10 章

[266] Nie S, Zhu F, Du C, et al. Scaling up Masked Diffusion Models on Text. ICLR 2025 (arXiv:2410.18514).被引用于：第 6 章（2 处）

[267] Zhang S, Roller S, Goyal N, et al. OPT: Open Pre-trained Transformer Language Models. arXiv:2205.01068, 2022.被引用于：第 3 章

[268] BigScience Workshop. BLOOM: A 176B-Parameter Open-Access Multilingual Language Model. arXiv:2211.05100, 2022.被引用于：第 3 章

[269] Gemma Team, Google DeepMind. Gemma 2: Improving Open Language Models at a Practical Size. arXiv:2408.00118, 2024.被引用于：第 3 章

[270] Meta AI. Introducing Meta Llama 3: The most capable openly available LLM to date. Meta AI Blog, 2024-04-18. https://ai.meta.com/blog/meta-llama-3/被引用于：第 3 章（2 处）

[271] Kimi Team (Moonshot AI). Kimi K3: Open Frontier Intelligence. arXiv:2607.24653, 2026. 权重与模型卡：https://github.com/MoonshotAI/Kimi-K3被引用于：附录 A（3 处）

[272] Z.ai. GLM-5.3-Flash: Frontier Intelligence, Flash Cost. Z.ai Blog, 2026-08-26. https://z.ai/blog/glm-5.3-flash被引用于：附录 A（3 处）

[273] ByteDance Seed. Seed2.1 正式发布，深入 AI 生产力. 字节跳动 Seed 博客, 2026-06-23. https://seed.bytedance.com/zh/blog/seed2-1-officially-released-advancing-ai-productivity被引用于：附录 A

[274] Meta. Meta Model API. Meta for Developers, 2026 (2026-09-28 访问). https://dev.meta.ai/products/meta-model-api被引用于：附录 A

[275] Pearl M. xAI Is Dead. Long Live SpaceXAI (引 @SpaceXAI 2026-07-06 帖). Gizmodo, 2026-07-06. https://gizmodo.com/xai-is-dead-long-live-spacexai-2000782034被引用于：附录 A

[276] Kandpal N, Deng H, Roberts A, Wallace E, Raffel C. Large Language Models Struggle to Learn Long-Tail Knowledge. ICML 2023 (arXiv:2211.08411).被引用于：第 8 章

[277] Romera-Paredes B, Barekatain M, Novikov A, et al. Mathematical Discoveries from Program Search with Large Language Models. Nature 625:468–475, 2024 (online 2023-12-14; doi:10.1038/s41586-023-06924-6).被引用于：第 8 章

[278] Hubert T, Mehta R, Sartran L, et al. Olympiad-Level Formal Mathematical Reasoning with Reinforcement Learning. Nature 651:607–613, 2026 (online 2025-11-12; doi:10.1038/s41586-025-09833-y).被引用于：第 8 章

[279] Pichai S, Hassabis D, Kavukcuoglu K. A new era of intelligence with Gemini 3. Google (The Keyword), 2025-11-18. https://blog.google/products-and-platforms/products/gemini/gemini-3/被引用于：附录 A