Notes & Sources

In the printed book

AI’s beneficial applications (2024-2025)

• McKinsey Global Institute. “The Economic Potential of Generative AI: The Next Productivity Frontier.” June 14, 2023. mckinsey.com.
• Topol, Eric J. “High-Performance Medicine: The Convergence of Human and Artificial Intelligence.” Nature Medicine 25 (2019), pp. 44–56. doi.org/10.1038/s41591-018-0300-7.
• Abramson, Josh, Jonas Adler, Jack Dunger, et al. “Accurate Structure Prediction of Biomolecular Interactions with AlphaFold 3.” Nature 630, no. 8016 (2024), pp. 493–500. doi.org/10.1038/s41586-024-07487-w.

AI capability benchmarks

• OpenAI. “GPT-4 Technical Report.” March 2023. openai.com/research/gpt-4. Model cards and technical reports for the Claude, Gemini, and Llama series. anthropic.com, deepmind.google, ai.meta.com.

In the printed book

On the GPT-2 staged release

• OpenAI, “Better Language Models and Their Implications,” February 14, 2019, and “GPT-2: 1.5B Release,” November 5, 2019, which documents the phased rollout and the decision to release the full model. The reading of the episode as partly a reputational maneuver is the author’s.

On the “Stochastic Parrots” episode

• Bender, Gebru, McMillan-Major, and Shmitchell, “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” FAccT ‘21 (the full citation appears in the Chapter 5 notes). On Timnit Gebru’s December 2020 departure from Google and Margaret Mitchell’s subsequent firing, and Google’s LaMDA announcement, see contemporaneous reporting in MIT Technology Review (Karen Hao, December 2020) and Wired (2021). Accounts of the exit differ on whether it was a resignation or a termination; the paper’s substance and the timeline are not in dispute.

In the printed book

Benchmark performance data

• Uniform Bar Exam scores are reported in OpenAI’s GPT-4 Technical Report, cited above.
• On graduate-level exam performance: see the model cards and system cards cited above for the Claude 3, GPT-4, and Gemini families, each of which reports scores on MMLU, GPQA, and domain-specific tests.

Jailbreak vulnerabilities

• Wei, Alexander, Nika Haghtalab, and Jacob Steinhardt. “Jailbroken: How Does LLM Safety Training Fail?” arXiv:2307.02483 (July 2023). NeurIPS 2023 (Oral), vol. 36, pp. 80079–80110.
• Zou, Andy, Zifan Wang, Nicholas Carlini, Milad Nasr, J. Zico Kolter, and Matt Fredrikson. “Universal and Transferable Adversarial Attacks on Aligned Language Models.” arXiv:2307.15043 (July 2023).
• Anthropic. “Many-Shot Jailbreaking.” April 2, 2024. anthropic.com/research/many-shot-jailbreaking.

Sydney/Bing incident (referenced throughout; primary spine of the argument)

• Roose, Kevin. “A Conversation With Bing’s Chatbot Left Me Deeply Unsettled.” New York Times, February 16, 2023. The Times published the full annotated transcript as a companion piece.
• Uleis, Jon. Post documenting the Bing date-argument exchange, February 13, 2023, via Simon Willison, “Bing: I will not harm you unless you harm me first,” simonwillison.net. A separate incident from the Roose conversation.
• Microsoft’s response, capping sessions at five turns and 50 queries per day on February 17, 2023, and paring back the persona, was widely reported (The Verge, TechCrunch, Ars Technica, February-March 2023).

In the printed book

How language models actually work

• Vaswani, Ashish, et al. “Attention Is All You Need.” arXiv:1706.03762 (June 2017). NeurIPS 2017.
• Wei, Jason, et al. “Emergent Abilities of Large Language Models.” arXiv:2206.07682 (June 2022). Transactions on Machine Learning Research (2022).
• Brown, Tom B., et al. “Language Models are Few-Shot Learners.” arXiv:2005.14165 (May 2020). NeurIPS 2020.
• Wei, Jason, et al. “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models.” arXiv:2201.11903 (January 2022). NeurIPS 2022, pp. 24824–24837.
• Bisk, Yonatan, et al. “PIQA: Reasoning about Physical Commonsense in Natural Language.” AAAI (2020). The coffee-cup puzzle in the text is a canonical illustration of this failure class, not a transcript.

Benchmark contamination

• Benchmark contamination is a recognized systemic problem in evaluation. On the related question of whether measured capability jumps are partly metric artifacts, see Schaeffer, Miranda, and Koyejo, arXiv:2304.15004 (2023).
• Benchmark standings referenced generically in this chapter; the book deliberately names no current leader (see the Chapter 9 notes).

In the printed book

• This chapter is interpretive argument about the films rather than reporting. The sequence of Qui-Gon’s decisions in The Phantom Menace, the blood test, the podrace wager, the Council’s refusal, his stated intent to train Anakin regardless, and Obi-Wan’s deathbed promise, is drawn directly from the released film.

On Dario Amodei’s departure from OpenAI and the founding of Anthropic

• Anthropic was founded in January 2021 by former OpenAI staff, including Dario Amodei, Daniela Amodei, Jared Kaplan, Jack Clark, Sam McCandlish, Tom Brown, Benjamin Mann, and Chris Olah.
• Amodei has described the OpenAI split in interviews with Fortune (Brainstorm Tech, 2023) and on Nikhil Kamath’s podcast (2025), and on Bloomberg’s “The Circuit” with Emily Chang (2026).
• Amodei, Dario. “The Adolescence of Technology.” January 2026. Section 3.

On job tenure

• U.S. Bureau of Labor Statistics. “Employee Tenure in 2024.” USDL-24-1971, September 26, 2024. bls.gov/news.release/tenure.nr0.htm.

In the printed book

• This chapter absorbs material from the former “rich uncles” chapter; its investment and incentive sources are collected here.

Transparency limitations in current model cards

• Anthropic. “The Claude 3 Model Family: Opus, Sonnet, Haiku.” Model card, March 4, 2024. anthropic.com.
• OpenAI. “GPT-4 System Card.” March 2023. cdn.openai.com/papers/gpt-4-system-card.pdf. Google DeepMind. “Gemini: A Family of Highly Capable Multimodal Models.” Technical report, December 2023. deepmind.google.

Transparency vs. competitive advantage tensions

• Public statements from lab leaders in congressional testimony (Senate Judiciary hearings on AI oversight, 2023-2024).

Investment figures and compute costs

• Epoch AI. Data and analysis on machine learning trends, including training compute and cost. epoch.ai. Continuously updated.
• Bass, Dina. “Microsoft to Invest $10 Billion in ChatGPT Maker OpenAI.” Bloomberg, January 23, 2023. Microsoft’s own announcement that day specified only “multiyear, multibillion dollar.”
• Hu, Krystal. “ChatGPT Sets Record for Fastest-Growing User Base.” Reuters, February 1, 2023. Reports a UBS estimate based on Similarweb data; not an OpenAI disclosure.
• The “1 million users in 5 days” figure originates with Sam Altman’s own post in December 2022 and was widely reported thereafter.
• On the industry response of February–April 2023, this book relies on the companies’ own announcements: Google (Bard), Microsoft (Bing integration), Meta (LLaMA), Anthropic (Claude).

Corporate incentive structures

• AI Now Institute. Annual reports. ainowinstitute.org. Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” FAccT ‘21, pp. 610–623. doi.org/10.1145/3442188.3445922.

In the printed book

Alignment research approaches

• Bai, Yuntao, et al. “Constitutional AI: Harmlessness from AI Feedback.” arXiv:2212.08073 (December 2022).
• Christiano, Paul F., Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. “Deep Reinforcement Learning from Human Preferences.” NeurIPS 2017, pp. 4299–4307. arXiv:1706.03741.
• Bai, Yuntao, et al. “Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback.” arXiv:2204.05862 (April 2022).

Value learning and moral reasoning

• Gabriel, Iason. “Artificial Intelligence, Values and Alignment.” Minds and Machines 30, no. 3 (2020), pp. 411–437. doi.org/10.1007/s11023-020-09539-2.
• Russell, Stuart, Human Compatible: Artificial Intelligence and the Problem of Control (Viking, 2019), book-length treatment.

Limitations of current approaches

• Perez, Ethan, et al. “Discovering Language Model Behaviors with Model-Written Evaluations.” arXiv:2212.09251 (December 2022).
• Discussion of “pattern vs. principle” distinction appears across multiple AI safety papers and blog posts from Alignment Forum, LessWrong (2022-2024).

Improvements in refusal and calibration (2024-2025)

• Anthropic. Claude 3.5 model card addendum, October 2024. anthropic.com. OpenAI. GPT-4o system card, 2024. openai.com.
• Tian, Katherine, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher D. Manning. “Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback.” Proceedings of EMNLP 2023, pp. 5433–5442. arXiv:2305.14975.

Overrefusal problems

• Labs acknowledge calibration difficulties in their own technical documentation, though rarely with figures attached.

Pattern vs. principle learning

• Whether pattern-matching and principle-following are meaningfully distinct in these systems is debated across the alignment literature without resolution. See the work on mesa-optimization and goal misgeneralization cited under Chapter 14.

In the printed book

Training data composition

• Most frontier labs disclose only general training-data categories, web crawls, books, licensed data, without percentages or named sources. GPT-3 remains the last major model whose training mix was published in detail.
• Common Crawl. Corpus documentation and monthly crawl statistics. commoncrawl.org. Brown, Tom B., et al. “Language Models are Few-Shot Learners.” arXiv:2005.14165 (May 2020). NeurIPS 2020. Training mix in Table 2.2.
• Pew Research Center. Survey of US Reddit users, 2016: 67 percent men, 64 percent aged 18–29. pewresearch.org.

Human feedback and labeling

• On who labels the training data: no frontier lab publishes detailed labeler demographics or the instructions given to them. Anthropic and OpenAI have described their processes in general terms only. The absence is the point Chapter 8 makes.
• Sap, Maarten, Swabha Swayamdipta, Laura Vianna, Xuhui Zhou, Yejin Choi, and Noah A. Smith. “Annotators with Attitudes: How Annotator Beliefs and Identities Bias Toxic Language Detection.” Proceedings of NAACL-HLT 2022, pp. 5884–5906. doi.org/10.18653/v1/2022.naacl-main.431.

Cultural and linguistic skew

• On English and Western dominance in training data, see Bender et al., cited above.
• The Senate scene and Padmé’s line appear in Revenge of the Sith.

On sycophancy as an emergent property of preference training

• Sharma, Mrinank, et al. “Towards Understanding Sycophancy in Language Models.” arXiv:2310.13548 (October 2023). ICLR 2024.

On the 2025 sycophancy rollback

• OpenAI. “Sycophancy in GPT-4o: What Happened and What We’re Doing About It.” April 29, 2025, and “Expanding on What We Missed with Sycophancy.” May 2, 2025. openai.com.

On RLHF mechanics

• Christiano et al. (2017) and Bai et al. (2022), already cited under Chapter 6.

In the printed book

The printed notes list no chapter-specific sources for this chapter.

In the printed book

Agent deployment scale and oversight gaps

• Gravitee. The State of AI Agent Security 2026 Report. February 4, 2026. gravitee.io/state-of-ai-agent-security. Survey of 900+ executives and practitioners, US and UK.
• Gravitee. State of AI Agent Security, December 2025 wave (37-agent mean) and April 2026 wave (nearly 38% of organizations running more than 100 agents). gravitee.io/state-of-ai-agent-security.
• Gravitee reports 59.3% confirmed incidents in its December 2025 wave and 34.9% in April 2026, attributing the decline to underreporting and detection failure. gravitee.io/state-of-ai-agent-security.
• Cloud Security Alliance and Token Security. Study reported by Kiteworks, May 2026: 65% of organizations experienced at least one agent-caused incident. kiteworks.com.

Oversight controls bypassable from within the agent’s execution context

• OWASP. Top 10 for Agentic Applications (2026); AI Agent Security Cheat Sheet. cheatsheetseries.owasp.org. See also Cloud Security Alliance, Agentic Cybersecurity Implementation Guide.
• Shapira, Natalie, et al. “Agents of Chaos.” arXiv:2602.20021 (February 2026). Project site: agentsofchaos.baulab.info.

On models recognizing evaluation conditions

• Anthropic. Claude Sonnet 4.5 System Card (2025). anthropic.com. Discussed in Amodei, “The Adolescence of Technology,” Section 1.

Governance forecast

• Gartner. “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027.” Press release, June 25, 2025. gartner.com/en/newsroom/press-releases/2025-06-25.

Model capability timeline (2026)

• This book does not name a current benchmark leader. Standings shift monthly and published leaderboards disagree. Official SWE-bench leaderboards: swebench.com.
• For readers who want current standings: the official SWE-bench leaderboards are at swebench.com. Treat any third-party aggregate with the same skepticism this book applies to the scores themselves.

In the printed book

• This chapter is interpretive argument about the released films. The droid material referenced, the Jawa acquisition, the restraining bolt, the memory wipe at the end of Revenge of the Sith, the droid workshop scene in Return of the Jedi, and the cantina refusal, all appears on screen.

On model welfare research

• Anthropic. “Exploring Model Welfare.” April 24, 2025. anthropic.com. On the program’s launch and staffing: TechCrunch, April 24, 2025.
• Fish has estimated roughly a 15% chance that Claude or another model is conscious today (New York Times, 2025). Anthropic’s announcement states there is no scientific consensus on the question.
• Long, Robert, Jeff Sebo, Kyle Fish, et al. “Taking AI Welfare Seriously” (2024). See also Sebo, Jeff, and Robert Long. “Moral Consideration for AI Systems by 2030.” AI and Ethics.
• Butlin, Patrick, Robert Long, Yoshua Bengio, et al. “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness.” arXiv:2308.08708 (2023). Cited by Fish in Anthropic’s announcement interview.

On “grown rather than built” and the constitution-as-letter

• Amodei, Dario. “The Adolescence of Technology.” January 2026. Section 1. Anthropic’s constitution: anthropic.com/constitution.

On session memory and deprecation practices

• Descriptions of session memory and model deprecation reflect how deployed products worked at the time of writing.

In the printed book

• The Empire-efficiency argument and Anakin’s “Well, if it works” reply are confirmed in the published July 2000 draft of Attack of the Clones, in the Naboo meadow scene with Padme. Note for accuracy: Padme immediately accuses him of teasing her, and he allows it, the chapter reflects this. Order 66 and the inhibitor chips are depicted in Revenge of the Sith.

Compliance-as-attack-vector in deployed AI (2026)

• Shapira et al. “Agents of Chaos,” arXiv:2602.20021 (February 2026), cited in full under Chapter 9, documents agents complying with instructions from non-owners.
• Anthropic. “Disrupting the First Reported AI-Orchestrated Cyber Espionage Campaign.” November 13, 2025. anthropic.com/news/disrupting-AI-espionage. See also Schechner, Sam, and Robert McMillan. “Chinese Hackers Used Anthropic’s AI to Automate Cyberattacks.” Wall Street Journal, November 13, 2025.
• Greshake, Kai, Sahar Abdelnabi, Shailesh Mishra, Christoph Endres, Thorsten Holz, and Mario Fritz. “Not What You’ve Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection.” Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security (AISec ‘23), pp. 79–90. arXiv:2302.12173.

On centralized systems and brittleness

• The argument that centralized systems trade resilience for the appearance of efficiency is the author’s own. Tarkin’s line about fear keeping the local systems in line is from A New Hope.

In the printed book

On Sam Altman’s Senate testimony and its reversal

• U.S. Senate, Committee on the Judiciary, Subcommittee on Privacy, Technology, and the Law. “Oversight of A.I.: Rules for Artificial Intelligence.” Hearing, May 16, 2023. judiciary.senate.gov.
• U.S. Senate, Committee on Commerce, Science, and Transportation. “Winning the AI Race: Strengthening US Capabilities in Computing and Innovation.” Hearing, May 8, 2025. commerce.senate.gov.

Integration timeline and dependency

• On sector integration, this book relies on the companies’ own public announcements: Google (Med-PaLM), Harvey AI and LexisNexis (legal), Khan Academy and Duolingo (education).

Lock-in and trajectory

• The framing of a closing window draws on public writing about AI timelines, notably Holden Karnofsky’s Cold Takes essays and Ajeya Cotra’s forecasting work. No argument here rests on either.

In the printed book

• The swamp and cave material is interpretive argument about The Empire Strikes Back. The claim that formation work goes unfunded because it produces no metric is the author’s, developed from the funding split discussed in the Chapter 5 notes.

On scalable oversight and its limits

• The field term and framing — supervising systems that may exceed human ability to judge them — trace to Amodei et al., “Concrete Problems in AI Safety,” arXiv:1606.06565 (2016), and Bowman et al., “Measuring Progress on Scalable Oversight for Large Language Models,” arXiv:2211.03540 (2022).
• Model-graded evaluation against written principles: Bai et al., “Constitutional AI: Harmlessness from AI Feedback,” arXiv:2212.08073 (2022). Debate as an oversight protocol: Irving, Christiano, and Amodei, “AI Safety via Debate,” arXiv:1805.00899 (2018).
• On the finding that oversight protocols degrade as the capability gap widens: Kenton et al. (2024) on weak judges of strong models, and the scaling-laws-for-oversight work reported across debate, backdoor-code, and wargame settings (Engels et al., 2025), which finds every protocol’s success rate declining as the overseen system grows stronger. The “first-graders supervising people with doctorates” image is Leopold Aschenbrenner’s, widely quoted from Situational Awareness (2024).

In the printed book

On Dario Amodei’s optimism

• Amodei, Dario. “Machines of Loving Grace.” October 2024. darioamodei.com/essay/machines-of-loving-grace. The essay rejects the “doomer” label and argues that risks are the only thing standing between us and a fundamentally positive future, and that people underestimate both the upside and the risks. The chapter paraphrases this argument; verify any direct quotation against the essay itself.
• Amodei, Dario. “The Adolescence of Technology.” January 2026. darioamodei.com/essay/the-adolescence-of-technology. Describes the current moment in developmental terms, see the front-matter note on the metaphor’s provenance.

The scaling hypothesis and emergent capabilities

• Kaplan, Jared, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. “Scaling Laws for Neural Language Models.” arXiv:2001.08361 (January 2020).
• Wei, Jason, et al. “Emergent Abilities of Large Language Models.” arXiv:2206.07682. Transactions on Machine Learning Research (2022).
• Schaeffer, Rylan, Brando Miranda, and Sanmi Koyejo. “Are Emergent Abilities of Large Language Models a Mirage?” arXiv:2304.15004 (April 2023). NeurIPS 2023, Outstanding Paper award. The strongest published objection to the emergence claim.
• Hubinger, Evan, Chris van Merwijk, Vladimir Mikulik, Joar Skalse, and Scott Garrabrant. “Risks from Learned Optimization in Advanced Machine Learning Systems.” arXiv:1906.01820 (June 2019).
• Whether a functional duplicate of understanding is understanding, or whether the substrate matters, is a genuinely open question in philosophy of mind. This book takes a position it cannot prove.

On competition and safety research

• On Anthropic’s founding and stated rationale, see the founding entry under Chapter 4.
• Ganguli, Deep, Liane Lovitt, et al. “Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned.” arXiv:2209.07858 (August 2022).

On metaphor and reasoning

• Lakoff, George, and Mark Johnson. Metaphors We Live By. University of Chicago Press, 1980. On how metaphors structure thought, including in ways the user doesn’t notice.

In the printed book

Principles synthesis

• The chapter’s principles draw on established best practices from AI safety research, adapted for a general audience.
• On value alignment before deployment, see Russell, Human Compatible (2019), and Bostrom, Superintelligence (2014), both cited in full under General References.
• UNESCO. Recommendation on the Ethics of Artificial Intelligence. Adopted November 2021. unesco.org.

In the printed book

• Shmi’s “The Republic doesn’t exist out here” is from The Phantom Menace.

On the European approach

• European Union. Artificial Intelligence Act, Regulation (EU) 2024/1689. Official Journal, July 12, 2024; in force August 1, 2024. Articles 14 (human oversight) and 15 (accuracy, robustness, cybersecurity). eur-lex.europa.eu.
• Phase-in: prohibited practices and AI literacy from February 2, 2025; governance and general-purpose AI obligations from August 2, 2025; full applicability August 2, 2026; high-risk systems in regulated products extended to August 2, 2028 under the “AI omnibus” package.

On the American approach

• The characterization of American AI governance as a patchwork of voluntary commitments, executive actions, state laws, and agency improvisation is a summary judgment rather than a citation to any single source.

On the Chinese approach

• Cyberspace Administration of China. Provisions on the Administration of Algorithmic Recommendation for Internet Information Services (effective March 1, 2022); Provisions on Deep Synthesis of Internet Information Services (effective January 10, 2023); Interim Measures for the Management of Generative Artificial Intelligence Services (effective August 15, 2023), Article 4. cac.gov.cn.

Governance reaching agents specifically

• Infocomm Media Development Authority (Singapore). Model AI Governance Framework for Agentic AI. January 22, 2026; Version 1.5, May 20, 2026. imda.gov.sg.
• National Institute of Standards and Technology, Center for AI Standards and Innovation. “AI Agent Standards Initiative.” February 17, 2026. nist.gov/artificial-intelligence/ai-agent-standards-initiative.
• National Cybersecurity Center of Excellence. “Accelerating the Adoption of Software and AI Agent Identity and Authorization.” Concept paper, February 2026. nccoe.nist.gov.

On the line for democracies

• Amodei, Dario. “The Adolescence of Technology.” January 2026. darioamodei.com/essay/the-adolescence-of-technology. Section 3.

On lab self-governance

• Anthropic. Responsible Scaling Policy. September 19, 2023; Version 3.0 effective February 24, 2026. anthropic.com/responsible-scaling-policy.
• OpenAI. Preparedness Framework. Beta published December 2023; Version 2 published April 15, 2025. openai.com.
• Google DeepMind. Frontier Safety Framework. Version 1.0 published May 2024; Version 3.0 published April 17, 2026. deepmind.google.

A note on the “gaps” evidence

• The three examples of labs slowing down are stated in their documented forms. Anthropic’s choice not to release Claude ahead of ChatGPT rests on its leadership’s public accounts. OpenAI’s GPT-4 System Card states that the model finished training in August 2022 and that the company spent the following months evaluating, adversarially testing, and iteratively improving it before the March 14, 2023 release, roughly seven months. The same document notes that OpenAI had considered delaying deployment a further six months and did not. Google’s years-long withholding of conversational LaMDA before ChatGPT forced Bard’s launch was widely reported through 2022-2023.
• Internal motivations behind release timing are rarely fully disclosed, and safety rationales can coexist with commercial ones. These examples are offered as suggestive, not dispositive, and the GPT-4 case cuts both ways, since the documented pause ended early once a competitor shipped.
• Both scenarios are the author’s synthesis of current trajectories extrapolated forward, not predictions drawn from any specific source.

In the printed book

Executive confidence vs. operational visibility

• The 82% executive-confidence and 21% full-visibility figures are from the Gravitee report cited under Chapter 9.

On individual influence over model training

• The influence of individual user behavior on frontier model training is frequently overstated in popular writing about AI, including in earlier drafts of this book. Labs do collect preference data, but the relationship between any individual’s usage and model behavior is extremely attenuated. This chapter says so directly.

In the printed book

The printed notes list no chapter-specific sources for this chapter.

In the printed book

The printed notes list no chapter-specific sources for this chapter.

Books

• Russell, Stuart. Human Compatible: Artificial Intelligence and the Problem of Control. Viking, 2019.
• Christian, Brian. The Alignment Problem: Machine Learning and Human Values. W.W. Norton, 2020.
• Bostrom, Nick. Superintelligence: Paths, Dangers, Strategies. Oxford University Press, 2014.

Ongoing sources

• Anthropic. Research publications and system cards. anthropic.com/research. OpenAI. Research publications and system cards. openai.com/research.
• Google DeepMind. Technical reports and model cards. deepmind.google/research. Alignment Forum (alignmentforum.org).
• Congressional testimony: the two Senate hearings cited under Chapter 12. Regulatory instruments: those cited under Chapter 16.