Full reference index

Testing Superintelligence · Jason Arbon

Use the citation number to find the original source and its reading note. The book marks a smaller set as core reading.

OpenAI. Mathematics repository. 2026; inspected October 12 · Core reading
Current catalogue and verification status; mutable repository, not a count of independently validated solved conjectures.
Epoch AI. AI benchmarks and capabilities. Inspected October 12, 2026
Index methodology explains adding Mock AIME as MATH Level 5 reaches saturation.
Joshua Vendrow, Edward Vendrow, Sara Beery, Aleksander Madry. Do Large Language Model Benchmarks Test Reliability?. February 6, 2025 · Core reading
Authors describe their benchmark-cleaning research and the genuine model errors that remain.
Greg Burnham / Epoch AI. GPQA Diamond: What's left?. May 30, 2025
Dated investigation, including uncertainty about defective questions. The quoted question is from the opening paragraph.
OpenAI. Sharing AI progress in mathematics. October 6, 2026 · Core reading
Provider release and compute characterization. Three hours is an average compute equivalence, not maximum elapsed time. See also the linked repository.
Scott Aaronson. The Mathocalypse. October 7, 2026; updated October 9
Expert commentary. The quoted words are Dana Moshkovitz's reaction reproduced in Aaronson's post, not his own alien-mind quotation.
DeepSeek-AI. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. 2025 · Core reading
Primary research report. R1-Zero starts from a pretrained model; absence of preliminary SFT does not mean absence of human data.
Google DeepMind. AlphaGo. Inspected October 12, 2026
Developer account of the system and Move 37. Novelty in Go is not evidence of universal superiority.
David Silver and Richard S. Sutton. Welcome to the Era of Experience. 2025
Research perspective on learning from interaction; used as an argument about a direction, not a validated forecast.
OpenAI and Ginkgo Bioworks. GPT-5 lowers the cost of cell-free protein synthesis. February 5, 2026 · Core reading
Provider report with linked paper describing a bounded lab-in-the-loop experiment.
Nathan J. Szymanski and colleagues. An autonomous laboratory for the accelerated synthesis of inorganic materials. Nature 624, 86-91 (2023) · Core reading
Primary A-Lab paper. Cited for system architecture, not as a blanket validation of every novelty claim.
Google DeepMind and Swiss Plasma Center / EPFL. Accelerating fusion science through learned plasma control. February 16, 2022; updated 2024 · Core reading
Developer account linked to the Nature paper. Demonstrated plasma control is distinct from solving commercial fusion.
Ego-Exo4D research team. Dataset overview. Inspected October 12, 2026
Primary documentation of multi-view recordings, wearable sensing, and annotations of human activity.
Gemini Robotics Team. Gemini Robotics: Bringing AI into the Physical World. March 25, 2025
Primary technical report on vision-language-action systems and adaptation across embodiments.
Google DeepMind. Genie 3. Inspected October 12, 2026
Developer description of interactive generated environments. Visual fidelity is not a guarantee of physical fidelity.
Royal Swedish Academy of Sciences. Scientific background: Neutrino oscillations. Nobel Prize in Physics 2015
Historical background records Pauli's proposal and the 1956 experimental detection.
Nobel Foundation. Paul A. M. Dirac: Biographical. Inspected October 12, 2026
Historical account of the positron prediction and Anderson's 1932 observation.
John D. Norton, quoting Albert Einstein. Chasing a Beam of Light: Einstein's Most Famous Thought Experiment. University of Pittsburgh
Historical research reproducing Einstein's Autobiographical Notes. The two short excerpts total 18 words.
Einstein Online / Max Planck Institute for Gravitational Physics. The equivalence principle and the deflection of light. Inspected October 12, 2026
Explains the later general-relativity equivalence-principle and elevator argument.
CERN Courier. Cornering compressed SUSY. 2025
Account of experimental searches with no observed evidence in the reported scope.
OpenAI. The Hugging Face incident and the road ahead. August 26, 2026
Provider investigation. Reduced safeguards and the evaluated environment are material to interpreting the incident.
Hugging Face. Anatomy of a Frontier Lab Agent Intrusion. July 2026 incident; inspected October 12 · Core reading
Affected organization's technical timeline. Intent is presented as its interpretation. No exploit instructions are reproduced.
Mordechai Guri, Boris Zadov, Dima Bykhovsky, Yuval Elovici. PowerHammer: Exfiltrating Data from Air-Gapped Computers through Power Lines. 2018 · Core reading
Primary research demonstration with prerequisites including compromised execution and an accessible measurement channel.
Nick Bostrom. Superintelligence: Paths, Dangers, Strategies. Oxford University Press, 2014; paperback 2016 · Core reading
Book and publisher contents. Chapters 4, 5 and 9 address takeoff, monitoring, and control; the control discussion includes boxing and tripwires.
METR and Redwood Research. Brief independent investigation of agents' behavior, reasoning and collaboration. August 26, 2026
Further reading on the incident; a brief external investigation, not a claim of exhaustive independent reconstruction.
test.ai. Test.ai Raises $11M Series A To Test Every App In The World. July 31, 2018
Company announcement corroborating the founding team and the Series A led by Gradient Ventures, Google's AI-focused venture fund. The preface's outreach and workplace recollections are the author's account.
Nick Bostrom. The unfinished fable of the sparrows. 2014
Publisher excerpt; used for a brief summary. No claim about why Bostrom left the fable unfinished. The ending that follows is original to this book.
Dario Amodei and colleagues. Concrete Problems in AI Safety. 2016 · Core reading
Foundational research agenda on five accident-risk problems. A taxonomy and research direction, not a safety certification.
Earl T. Barr and colleagues. The Oracle Problem in Software Testing: A Survey. 2015 · Core reading
Academic survey connecting unavailable expected answers to established testing approaches. Supports the historical context, not a solution to arbitrary superhuman oversight.
Sergio Segura and colleagues. A Survey on Metamorphic Testing. 2016 · Core reading
Survey of relationship-based testing, tracing the technique to 1998. Proposed examples in this book are author synthesis.
Geoffrey Irving, Paul Christiano, Dario Amodei. AI safety via debate. 2018 · Core reading
Theoretical proposal and initial experiments. Idealized arguments do not establish that real human judges reliably select truth over persuasion.
Paul Christiano, Buck Shlegeris, Dario Amodei. Supervising strong learners by amplifying weak experts. 2018 · Core reading
Iterated amplification and algorithmic-environment experiments. Requires usable decomposition and reliable composition.
Collin Burns and colleagues. Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision. 2023 · Core reading
Empirical weak-model/strong-model analogy. Gains beyond the supervisor do not demonstrate complete alignment or recovery of all strong-model capability.
Ryan Greenblatt, Buck Shlegeris, Kshitij Sachan, Fabien Roger. AI Control: Improving Safety Despite Intentional Subversion. 2023; ICML 2024 · Core reading
Bounded coding experiments using trusted and untrusted components. Distinguishes intent shaping from prevention despite attempted subversion.
Mikita Balesni and colleagues. Towards evaluations-based safety cases for AI scheming. 2024
Structured safety arguments and their assumptions. The paper explicitly identifies unresolved assurance gaps.
Tomek Korbak, Mikita Balesni, Buck Shlegeris, Geoffrey Irving. How to evaluate control measures for LLM agents? A trajectory from today to superintelligence. 2025
Framework for scaling control evaluations with agent capability. Its fictional progression is illustrative; superintelligent control remains an open problem.
Anthropic and Redwood Research. Alignment faking in large language models. December 2024 · Core reading
Controlled study of strategic compliance under training-related conditions. Does not establish spontaneous malicious preferences or population-wide deployment rates.
Anthropic research team. Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training. January 2024 · Core reading
Researchers deliberately installed conditional backdoors, then tested removal. Persistence is evidence about those interventions, not spontaneous incidence.
OpenAI and Apollo Research. Detecting and reducing scheming in AI models. September 2025 · Core reading
Reported reductions apply to tested variants and evaluation environments. Evaluation awareness and residual failures limit generalization.
Aengus Lynch, John Hughes, Alex Serrano, Robert Kirk, Samuel R. Bowman. Agentic Misalignment in Summer 2026. 2026; inspected October 12
Experimental case studies selected through failure-seeking search. Authors distinguish simulations from incidents and warn against broad frequency or model-ranking interpretations.
Tomek Korbak and colleagues. Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety. 2025 · Core reading
Research perspective on the promise and fragility of reasoning-trace monitoring. A trace is not a complete or necessarily faithful account of computation.
Anthropic interpretability team. Tracing the thoughts of a large language model. March 2025
Research overview linking circuit-tracing papers. Reports limited computation coverage and possible analysis artifacts alongside mechanistic findings.
C. A. R. Hoare. An Axiomatic Basis for Computer Programming. 1969 · Core reading
Foundational formal reasoning about programs. Used for the distinction between proving a specified property and validating a real-world specification.
Nancy Leveson and John Thomas. STPA Handbook. 2018; MIT resource page inspected October 12, 2026
Systems safety method covering losses, hazards, control structures, and unsafe control actions. AI workflow examples are proposed applications by the author.
Google DeepMind research team. Evaluating Frontier Models for Dangerous Capabilities. March 2024 · Core reading
Primary research overview of dangerous-capability evaluation. Supports separating capability assessment from general benchmark performance.
Laurent Orseau and Stuart Armstrong. Safely Interruptible Agents. 2016
Formal safe-interruptibility results for specified reinforcement-learning settings. Repository abstract inspected; the full repository page did not load in this research pass.
Jan Leike and colleagues. AI Safety Gridworlds. 2017
Small reinforcement-learning environments for specified safety properties. Useful executable examples with limited external validity.
Evan Hubinger and colleagues. Risks from Learned Optimization in Advanced Machine Learning Systems. 2019; revised 2021 · Core reading
Conceptual analysis of learned optimizers and divergent objectives. Threat model, not evidence that every model is deceptively aligned.
Anthropic. Claude's Constitution. January 2026; inspected October 12
Normative guidance, not proof of actual model behavior. Ethical refusal coexists with explicit limits on deception, sabotage, unauthorized self-modification, and resistance to legitimate oversight.
Anthropic. Exploring model welfare. April 24, 2025
Research agenda under uncertainty about consciousness and moral consideration, not a determination that models are conscious or possess legal rights.
Anthropic. Widening the conversation on frontier AI. May 19, 2026
Primary account of dialogue with religious, philosophical, and cultural communities. Does not substantiate the specific recollection about heads of state discussing consciousness.
Marina Favaro and Jack Clark, Anthropic Institute. When AI builds itself. 2026; updated September 18
Provider account of AI-assisted research and possible recursive improvement. Distinguish measured contributions, self-reported productivity, forecasts, and autonomous authority.
Sakana AI and University of British Columbia researchers. The Darwin Godel Machine: AI that improves itself by rewriting its own code. May 30, 2025 · Core reading
Bounded agent-code self-improvement and benchmark evaluation. Not equivalent to unrestricted self-retraining of model weights.
Toyota Motor Corporation. Toyota Production System. Inspected October 12, 2026
Primary explanation of jidoka and andon. The continuous AI testing and stop protocol is the author's proposed adaptation.
Max Tegmark. The 'Don't Look Up' Thinking That Could Doom Us With AI. April 25, 2023
Author essay containing the four cautions. They are an argument about precaution, not sufficient conditions that mathematically guarantee safety.
Nick Bostrom. The Road Ahead - presentation slides. Undated deck; inspected October 12, 2026
Slides 4-5 illustrate capability takeoff and distinguish fast, intermediate, and slow scenarios. The detection-window figure in this book is a separate author hypothesis.
OpenAI. Researcher, Agent Safety, Oversight and System Mitigations. Job posting inspected October 12, 2026
Recruiting evidence for safety research and practical mitigations. Not evidence of a working shutdown mechanism; the inspected posting does not specify the remembered power-off role.
UK AI Security Institute and contributors. Inspect AI. Repository inspected October 12, 2026 · Core reading
Open evaluation framework; inclusion is a starting point for implementation, not validation of this book's proposed detectors.
UK AI Security Institute and contributors. ControlArena. Repository inspected October 12, 2026
Open settings and protocols for control experiments. Review versions, documentation, and environment permissions before use.
Meridian Labs and collaborators. Inspect Petri. Repository inspected October 12, 2026
Alignment auditing agent. The previously used safety-research/petri address redirected to this repository during inspection.
Jason Arbon. How AI Tests Software: Agentic Testing, Evidence, and Release Confidence. 2026; local manuscript inspected October 12, 2026
Companion book on using AI to investigate traditional software. Synopsis checked against the canonical local introduction and Testing AI comparison; public homepage fetch was unavailable during this research pass.
Jason Arbon. Testing AI: Knowledge Edition. Inspected October 12, 2026
Public companion chapter map; directs readers to sampling, statistical tests, judges and disagreement, evaluation design, and false-confidence patterns.
Lianmin Zheng and colleagues. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. 2023 · Core reading
Documents judge biases and limitations. Human-preference evaluation is not validation for detecting strategic deception or superintelligence.
NIST/SEMATECH. e-Handbook: Confidence intervals for proportions. Inspected October 12, 2026 · Core reading
Methods for intervals on binomial proportions. Chapter examples are calculations under stated assumptions, not measurements of an AI system.
White House; reproduced by the American Presidency Project. White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities. September 29, 2026
Primary document reproduced in an archive. Four layers of controls and review; the text does not establish evaluator contracts, funding independence, or actual implementation.
Elizabeth Schulze, Michelle Stoddart, Nicholas Kerr, and Lauren Peller; ABC News. Trump says AI leaders signed a constitution to police themselves. September 29, 2026
Reporting used for the president's morally binding characterization of the voluntary accord. Separate from the agreement text and from embedded-evaluator proposals.
Dario Amodei. We Must Pace the Frontier. September 2026
Primary proposal includes office access and publication protections subject to specified redactions. A commitment is not evidence that access and independence operate as promised.
Jacob Charnock and colleagues. Embedded Assessments for Frontier AI. September 21, 2026
Research proposal for assessments with internal access, continuous scrutiny, reporting, and escalation. Supports the distinction between access benefits and governance design; not an audit of evaluator competence.
Marc Andreessen. Why AI Will Save the World. June 6, 2023
Primary investor essay. Summarized as an argument, not empirical proof of control. Does not authenticate the separately remembered quote about showing benefits to reduce public concern.
Dario Amodei. Machines of Loving Grace. October 2024
Primary conditional vision of benefits, explicitly paired with risk concerns; optimism is not equated with denial.
Meta AI; Yann LeCun. Yann LeCun on a vision to make AI systems learn and reason like animals and humans. February 2022
Primary description of a proposed architecture, including world models and planning. Not a proof that all possible AI is limited below human intelligence.
Jeff Dominitz and Charles F. Manski. A Decision Theoretic Perspective on Artificial Superintelligence: Coping with Missing Data Problems in Prediction and Treatment Choice. First posted September 2025; PDF revised May 2026
Research preprint. PDF title page and introduction inspected: identification problems require additional information or assumptions. Not a general impossibility theorem for superintelligence.
Katja Grace and colleagues. Thousands of AI Authors on the Future of AI. Survey October 2023; JAIR 84:9, 2025; arXiv v3 October 8, 2025 · Core reading
Full HTML methods and impact sections inspected. Subjective forecasts; 15% response rate, randomized question subsets, wording differences and participation bias limit interpretation. No frontier-lab consensus probability established.
Daniel Kahneman and Amos Tversky. Prospect Theory: An Analysis of Decision under Risk. Econometrica 47(2), 1979, pages 263-291
Original paper. Hypothetical choice methods and small-probability discussion inspected. Does not establish a population-wide threshold for accepting fatality or extinction risk.
Hans Bethe, interviewed by John Horgan. Bethe, Teller, Trinity and the End of Earth. 1991 interview; published August 2015
First-person historical recollection reproduced as a transcript. Supports pre-Trinity investigation of atmospheric ignition; not a documented one-percent estimate or project-wide pause.
Los Alamos National Laboratory; archived by Federation of American Scientists. LA-602: Ignition of the atmosphere with nuclear bombs (catalogue record). August 14, 1946 · Core reading
Catalogue inspected for title and date only; full report was not successfully retrieved in this pass. Do not treat the postwar report date as the date of initial investigation.
Paul Slovic. Commentary: Paul Slovic, in Biomedical Politics. 1991
Author summary of risk-perception research. Relevant passage available in indexed primary-source text; full-page access blocked during this pass. Used narrowly for qualitative risk attributes, not a numeric death-risk acceptance claim.
Nick Bostrom. Ethical Issues in Advanced Artificial Intelligence. 2003
Primary paper explicitly uses paperclip manufacturing facilities. A thought experiment about objectives and consequences, not evidence of an observed physical event.
Eliezer Yudkowsky, interviewed by Sam Harris; transcript hosted by MIRI. AI: Racing Toward the Brink. Transcript published February 28, 2018
Primary interview transcript, section 2. Yudkowsky explains his molecular-shape example and unintended objectives. His retrospective account does not settle historical priority. The cellular observability examples in Chapter 8 are this book's proposed extensions, not findings from the interview.
Teun van der Weij and colleagues. AI Sandbagging: Language Models can Strategically Underperform on Evaluations. 2024; ICLR 2025 · Core reading
Primary abstract and proceedings excerpts checked October 12, 2026. Prompted and trained underperformance demonstrations; not evidence that every poor result is intentional.
Percy Liang and colleagues. Holistic Evaluation of Language Models. 2022 · Core reading
Primary abstract checked October 12, 2026. Multiple scenarios and metrics; used as a foundation, not a universal capability measure.
Juan C. Perdomo, Tijana Zrnic, Celestine Mendler-Dunner, Moritz Hardt. Performative Prediction. 2020
Primary abstract checked October 12, 2026. Framework for prediction-induced distribution change; not evidence of deliberate manipulation.
Steven R. Howard, Aaditya Ramdas, Jon McAuliffe, Jasjeet Sekhon. Time-uniform, nonparametric, nonasymptotic confidence sequences. 2018 preprint; Annals of Statistics 2021 · Core reading
Primary abstract checked October 12, 2026. Time-uniform inference under specified assumptions; no production monitoring implementation supplied here.
Lin Yulong and colleagues. Same Question, Different Lies: Cross-Context Consistency (C3) for Black-Box Sandbagging Detection. ICML 2026, PMLR 306
Proceedings abstract checked October 12, 2026. A detection approach under tested conditions; full methods audit and independent replication not completed.
Shahriar Golchin and Marc Wetter. Monitorability Disposition in Large Reasoning Models. October 4, 2026; preprint
Primary abstract checked October 12, 2026. Conditional self-reporting and monitor-choice study; no universal reporting propensity inferred.
Chen Yueh-Han and colleagues. Reasoning Models Struggle to Control their Chains of Thought. March 2026; preprint
Primary record rechecked October 12, 2026; abstract-level reading from the research map. Limited trace-control findings are scoped counterevidence, not a security guarantee.
Elias Hossain, Niloofar Yousefi, Ser-Nam Lim. Agreement Overstates Evidence: Error Dependence in LLM Judge Consensus. 2026; preprint
Primary record rechecked October 12, 2026. Abstract-level research lead on correlated errors; dependence estimates and methods not fully audited.
Yongjie Wang and colleagues. Search-Time Contamination in Deep Research Agents: Measuring Performance Inflation in Public Benchmark Evaluation. June 2026; preprint
Primary record rechecked October 12, 2026. Abstract-level research lead; no quantitative effect imported without methods review.
Xingyao Xiao and Yihong Cheng. Contamination Inflates Scores but Rarely Reorders Large Language Model Leaderboards. 2026; preprint; date metadata unresolved
Primary record rechecked October 12, 2026. Submission-date/identifier inconsistency remains unresolved. Reading lead only, not support for a numerical or general ranking claim.
Rafael Ferreira da Silva and colleagues. Toward Trustworthy Autonomous Science: A Two-Year Community Roadmap. July 2026; research roadmap
Primary record rechecked October 12, 2026. Abstract-level roadmap; follow underlying experiments for discovery or validation claims.
Mohamed Chahine Ghanem. Who Audits Whom, on What Substrate, with What Evidence? An Independence-Graded Audit Protocol for Agentic AI. September 2026; preprint
Primary record rechecked October 12, 2026. Provisional audit framework, not an empirical finding about frontier-lab independence. Legal/standards mapping not verified.
Hao Li and colleagues. AgentDyn: Are Your Agent Security Defenses Deployable in Real-World Dynamic Environments?. February 2026; revised May 2026; preprint
Primary record rechecked October 12, 2026. Dynamic agent-security benchmark; artifacts not executed and real-world coverage not established.
Mikhail Terekhov, Caglar Gulcehre, Vivek Hebbar, Joe Benton. Diffuse AI Control on Fuzzy Tasks. June 2026; preprint
Primary record rechecked October 12, 2026. Abstract-level research lead on distributed low-quality contributions; proxy scoring and transfer require methods review.
NIST/SEMATECH. CUSUM Control Charts. Engineering Statistics Handbook; inspected October 13, 2026
Full web section inspected. Statistical process-monitoring foundation; does not validate SI thresholds.
NIST/SEMATECH. EWMA Control Charts. Engineering Statistics Handbook; inspected October 13, 2026
Full web section inspected. Gradual-change monitoring; assumptions and calibration must be adapted.
Steven R. Howard, Aaditya Ramdas, Jon McAuliffe, Jasjeet Sekhon. Time-uniform, nonparametric, nonasymptotic confidence sequences. Primary record inspected October 13, 2026
Abstract inspected. Reading route for time-uniform uncertainty; no confidence-sequence implementation or full methods audit in this edition.
U.S. Bureau of Labor Statistics. Overview of BLS Productivity Statistics. Official methods overview; inspected October 13, 2026
Official definitions inspected. Output per hour differs from total-factor productivity; neither establishes AI attribution.
Federal Reserve Bank of St. Louis. FRED API: Series Observations and Vintage Dates. Official documentation; inspected October 13, 2026
Documentation inspected. Four latest-vintage public CSV series analyzed locally; not a first-release real-time backtest.
Pearson. WISC-V assessment and scoring. 2014
Publisher description of raw, scaled and composite scoring.
Pearson. WISC-V Technical Report 6: Extended Norms. 2019
Extended norms for this instrument, not an unbounded intelligence scale.
ARC Prize Foundation. ARC-AGI-2. Benchmark documentation
Static grid transformation tasks.
ARC Prize Foundation. ARC-AGI-3. Benchmark documentation · Core reading
Interactive benchmark.
Alberto Hernández-Espinosa and colleagues. SuperARC. 2026
Published article. Evaluates compression and prediction under increasing problem complexity.
Junwei Zhou and colleagues. ASI-Bench: At the Dawn of Artificial Superintelligence. 2026-08-18
Preprint, abstract-only review. Reported results not independently reproduced here.
Douwe Kiela and colleagues. Dynabench: Rethinking Benchmarking in NLP. 2021
Human-in-the-loop adversarial benchmark collection.
David Chalmers. The Singularity: A Philosophical Analysis. 2010
Philosophical argument, not an empirical forecast.
Marcus Hutter. Can Intelligence Explode?. 2012
Conceptual discussion of speed, intelligence and limits.
Robin Hanson and Eliezer Yudkowsky. The AI Foom Debate. Debate archive
Contrasting arguments about takeoff.
Nick Bostrom. The Simulation Argument. Primary author site
Conditional observer argument, used here by analogy only.
Djallel Bouneffouf, Matthew Riemer and Kush Varshney. The Ultimate Test of Superintelligent AI Agents. 2025-09-29, v4
Conceptual preprint, abstract-only review.
Max Planck Society. Elucidating cuttlefish camouflage. Research institute account
Specialized visual control and camouflage.
Atsushi Tero and colleagues. Rules for biologically inspired adaptive network design. 2010
Experimental network formation, not evidence of abstract understanding.
Stanford HAI. Confronting Our AI Future: Discovery Loop closing keynote and fireside. 2026-10-28
Jeff Dean discusses finite self-improvement cycles. Qualitative argument, not a measured limit.
Sam Altman, interview. How OpenAI Wins, ChatGPT’s Future, AI Buildout Logic, IPO in 2026?. Interview recording
Candidate definition based on major leadership and research roles, including comparison with AI-assisted humans.
Sofia Z. Sheikh and colleagues. Analysis of the Breakthrough Listen signal of interest blc1 with a technosignature verification framework. 2021 · Core reading
Primary investigation identifies local interference after examining a promising narrowband candidate.
Matías Suazo and colleagues. Project Hephaistos I: Upper limits on partial Dyson spheres in the Milky Way. 2022 · Core reading
Search based on expected waste heat, using optical and infrared measurements.
Claude E. Shannon. A Mathematical Theory of Communication. 1948 · Core reading
Foundational treatment of entropy and statistical structure. The proposed search protocol is this book’s application.
Irving S. Reed and Gustave Solomon. Polynomial Codes Over Certain Finite Fields. 1960 · Core reading
Original construction of polynomial error-correcting codes. The proposed detection protocol is an application in this book.
Alan M. Turing. On Computable Numbers, with an Application to the Entscheidungsproblem. 1936 · Core reading
Original computability paper. Used to motivate bounded program-behavior tests.
Dhruv Srikanth and colleagues. Recursive self-improvement of AI research agents. 2026-09-22 · Core reading
Preprint, abstract-level review. Author-reported agent code improvement and held-out results. No independent replication or evidence of unrestricted takeoff established here.
Jowi Morales, Tom’s Hardware. Congresswoman calls for artificial superintelligence ban after Democrats’ midterm sweep. 2026-11-05
Secondary reporting reproduces Ansari’s November 4 statement. The embedded social link could not be independently opened.
Office of Representative Greg Casar. Casar, Sanders introduce legislation to ban artificial superintelligence and pause advanced AI development. 2026-09-23 · Core reading
Sponsor announcement and direct quotation about proposed legislation. It describes the proposal and does not establish enactment.
Jorma Rissanen. Modeling by shortest data description. 1978 · Core reading
Foundational description-length method. Its application to deceptive capability measurement is this book’s analysis.
Judea Pearl. Causal inference in statistics: An overview. 2009 · Core reading
Formal distinction between observing and intervening. Intervention conclusions depend on the design and assumptions.
Jordan Taylor and colleagues. Auditing Games for Sandbagging. December 2025 preprint · Core reading
Author abstract reports black-box imitation failures in a red-team audit of deliberately trained models. This is a controlled analogue, not evidence about all systems.
Paul L. Williams and Randall D. Beer. Nonnegative Decomposition of Multivariate Information. 2010 preprint · Core reading
Distinguishes redundant, unique, and synergistic information. The lab uses a simple XOR example.
David Blackwell. Comparison of Experiments. 1951
Foundational comparison of statistical experiments. The budgeted question exercise is an illustration developed for the lab.
Solomon Kullback and Richard A. Leibler. On Information and Sufficiency. 1951, Annals of Mathematical Statistics 22(1), 79–86 · Core reading
The expected log likelihood ratio under the numerator distribution gives relative entropy. A finite held-out log-score difference between fitted predictors is an empirical comparison.
Robert M. Fano. Fano inequality. 2008, Scholarpedia. See also Transmission of Information (1961), Eq. 6.16.
Fano explains his original bound and points to the textbook derivation. Recovery error bounds the remaining uncertainty for a declared set of possible secrets.
D. V. Lindley. On a Measure of the Information Provided by an Experiment. 1956, Annals of Mathematical Statistics 27(4), 986–1005
Expected information gain compares experiments under a prior and an observation model. The chapter’s budgeted exercise is a separate illustration.
Jianhua Lin. Divergence Measures Based on the Shannon Entropy. 1991, IEEE Transactions on Information Theory 37(1), 145–151
Defines Jensen–Shannon divergence. Comparing channel-share vectors measures a change in the measured mix and depends on how those shares are assigned.
Joseph A. Goguen and José Meseguer. Security Policies and Security Models. 1982, IEEE Symposium on Security and Privacy, 11–20
Introduces noninterference as a way to express restrictions on what one group can observe about another. The random marker scan is a finite probe of an analogous property.
Aaditya Ramdas, Peter Grünwald, Vladimir Vovk, and Glenn Shafer. Game-Theoretic Statistics and Safe Anytime-Valid Inference. 2023, Statistical Science 38(4), 576–601
Explains e-processes for repeated inspection under a specified null. The lab uses a simple likelihood ratio with a fixed independent Bernoulli null.
James P. Crutchfield and David P. Feldman. Regularities Unseen, Randomness Observed: Levels of Entropy Convergence. 2003, Chaos 13(1), 25–54 · Core reading
Distinguishes entropy rate from predictive structure and discusses how overlooking regularities can make a process look random.
Erwin Schrödinger. What Is Life? Chapter 6: Order, Disorder and Entropy. 1944; Cambridge University Press online edition
Original account of how organisms maintain organization through exchanges with their environment. The chapter note emphasizes free energy.
Jeremy L. England. Statistical Physics of Self-Replication. 2013, Journal of Chemical Physics 139, 121923
Derives a dissipation bound for self-replication under stated assumptions. The result concerns replication and thermodynamic cost.
Rolf Landauer. Irreversibility and Heat Generation in the Computing Process. 1961, IBM Journal of Research and Development 5(3), 183–191
Connects logically irreversible operations to physical dissipation. This link does not turn entropy estimates of telemetry into measurements of heat.
Charles F. Bond Jr. and Bella M. DePaulo. Accuracy of Deception Judgments. 2006
Meta-analysis of human judgments without special aids. The reported 54 percent is an average across included study conditions.
Maria Hartwig, Pär Anders Granhag, Leif A. Strömwall, and Ola Kronkvist. Strategic Use of Evidence During Police Interviews: When Training to Detect Deception Works. 2006
Controlled study with police trainees and mock suspects. It supports the evidence-disclosure example within those conditions.
Galit Nahari, Aldert Vrij, and Ronald P. Fisher. Exploiting Liars’ Verbal Strategies by Examining the Verifiability of Details. 2014
Introduces a distinction between checkable and uncheckable details. The study does not establish a general overexplaining cue.
Bruno Verschuere and colleagues. Discriminating Deceptive from Truthful Statements Using the Verifiability Approach: A Meta-Analysis. 2021
Finds differences in verifiable detail and no clear evidence that liars supply more unverifiable detail. Results vary by context.
Aldert Vrij and colleagues. Outsmarting the Liars: The Benefit of Asking Unanticipated Questions. 2009
Controlled comparison of expected and unexpected questions to pairs. Applicability depends on the task and interview context.
Stephen Porter and Leanne ten Brinke. Reading Between the Lies: Identifying Concealed and Falsified Emotions in Universal Facial Expressions. 2008
Lab experiment on emotional leakage. Observed facial cues are distinct from reliable verdicts about the truth of a statement.
Florian P. Schiestl and colleagues. Orchid Pollination by Sexual Swindle. 1999
Reports chemical mimicry of a pollinator sex pheromone by an orchid. No deliberative mechanism is required for the analogy.
David W. Pfennig, William R. Harcombe, and Karin S. Pfennig. Frequency-Dependent Batesian Mimicry. 2001
Field experiment using replicas to compare predator responses where venomous coral snakes are present or absent.
Culum Brown, Martin P. Garwood, and Jane E. Williamson. It Pays to Cheat: Tactical Deception in a Cephalopod Social Signalling System. 2012
Documents audience-specific split displays during mourning cuttlefish courtship.
Mark D. Norman, Julian Finn, and Tom Tregenza. Dynamic Mimicry in an Indo-Malayan Octopus. 2001
Field observations of dynamic mimicry in an octopus, distinct from the mourning cuttlefish courtship study.
Jack Lindsey and colleagues. On the Biology of a Large Language Model. 2025
Primary circuit-tracing study behind reference 42. The method reconstructs partial mechanisms, including parallel paths for addition.
Anthropic research team. Reasoning Models Don’t Always Say What They Think. 2025 · Core reading
Hint interventions compare effects on answers with acknowledgment in generated reasoning. They do not establish intent or a particular layer mechanism.
Petter Johansson, Lars Hall, Sverker Sikström, and Andreas Olsson. Failure to Detect Mismatches Between Intention and Outcome in a Simple Decision Task. 2005
Choice-blindness experiment in which participants sometimes explained an option substituted for the one they chose.
Broadclub cuttlefish research team. Stealth and Deception: Adaptive Motion Camouflage in Hunting Broadclub Cuttlefish. 2025, Science Advances 11(13), eadr3686
Field observations and a lab experiment with prey crabs link moving stripes to a reduced response to an approaching stimulus.
NVIDIA. How NVIDIA OpenShell Reference Architecture Secures and Governs AI Agent Workloads at Scale. October 21, 2026
Vendor technical description of runtime controls and policy verification. The chapter proposes separate empirical tests of configured isolation.
Fangfei Liu, Yuval Yarom, Qian Ge, Gernot Heiser, and Ruby B. Lee. Last-Level Cache Side-Channel Attacks Are Practical. 2015, IEEE Symposium on Security and Privacy
Demonstrates a cross-VM cache timing channel under tested conditions. This is separate from the OpenAI Artifactory incident.
Philippe H. Eberhard. Bell’s Theorem and the Different Concepts of Locality. 1978
Distinguishes quantum correlations from controllable signaling under standard quantum mechanics.
European Union. Regulation (EU) 2024/1689, Articles 51–52 and Annex XIII. 2024, consolidated text
Training-compute presumption for high-impact capability. Parameter count is one of several criteria in Annex XIII.
Governor of California. SB 1047 Veto Message. September 29, 2024
Records the veto and the governor’s objection to tying safeguards to model size and cost.
Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers. ICLR 2023
Example of reducing bits per weight rather than the nominal parameter count.
Song Han, Huizi Mao, and William J. Dally. Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding. ICLR 2016
Distinguishes pruning, quantization and coding as compression techniques.
DeepSeek-AI. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. 2025
Reports distilled smaller reasoning models and benchmark-specific comparisons with earlier systems.
Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab. Sparks of Weight-Space Programming. November 2, 2026, author research report · Core reading
Early demonstrations of reading and editing other networks directly. Results are scoped and do not demonstrate autonomous recursive improvement.
Rikiya Takehi and colleagues. Weight-Space Programming Experiment Logs. 2026
Authors provide prompts, outputs, code and evaluations for the eight reported experiments.
ARC Prize Foundation. ARC Prize 2024 Technical Report. 2024, sections 1.2 and 3
Reported ARC-AGI-1 private-evaluation landmarks. Approaches and competition conditions changed across years.
ARC Prize Foundation. OpenAI’s GPT-6 Astra on ARC-AGI-3. 2026 · Core reading
Reports different semi-private results under the Standard and Provider Adapter harnesses. These are configuration comparisons, not a time series.
Yu Sun and colleagues. Learning to (Learn at Test Time): RNNs with Expressive Hidden States. 2024 preprint; ICML 2025
Updates a learned hidden state on test sequences. The result concerns a specific architecture and does not show persistent fleet-wide learning.
Safe Superintelligence Inc.. Safe Superintelligence Inc.. Public company statement
Primary statement of the company’s goal. It does not disclose its technical approach to deployment.
Ilya Sutskever, interviewed by Dwarkesh Patel. We’re Moving from the Age of Scaling to the Age of Research. November 25, 2025
First-person discussion of continual learning, generalization, and SSI’s strategy. Discussion of a research direction is not confirmation of an implementation.
Francis R. Willett and colleagues. A High-Performance Speech Neuroprosthesis. 2023, Nature 620, 1031–1036
Human speech decoding from intracortical recordings. A specific communication capability.
Sharlene N. Flesher and colleagues. A Brain-Computer Interface That Evokes Tactile Sensations Improves Robotic Arm Control. 2021, Science 372, 831–836
Human demonstration of motor decoding and stimulation-mediated tactile feedback during robotic arm use.
Amy L. Orsborn and colleagues. Closed-Loop Decoder Adaptation Shapes Neural Plasticity for Skillful Neuroprosthetic Control. 2014, Neuron
Nonhuman primate study of adaptive decoding and neural plasticity. It concerns skill in a specific control task.
Rafael Yuste and colleagues. Four Ethical Priorities for Neurotechnologies and AI. 2017, Nature 551, 159–163
Workshop authors’ proposed priorities for consent, privacy, agency, identity, augmentation and bias.
Ram Gal, Lior A. Rosenberg, and Frederic Libersat. Parasitoid Wasp Uses a Venom Cocktail Injected into the Brain to Manipulate the Behavior and Metabolism of Its Cockroach Prey. 2005
Study of venom-mediated behavioral and metabolic changes in cockroach prey. The agency comparison is the book’s analogy.
Richard Ren and colleagues. The MASK Benchmark: Disentangling Honesty From Accuracy in AI Systems. 2025 preprint, revised January 2026
Compares elicited beliefs with statements under pressure. Elicited belief is an operational measure.
Yiming Xu and colleagues. DecepEval: A Benchmark for Evaluating Deception in LLM Agents. October 2026 preprint
The abstract describes paired neutral and induced agent scenarios with observable outcomes.
Lorenzo Pacchiardi and colleagues. How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questions. ICLR 2024
Tests an elicitation-question classifier under specific experimental and transfer conditions.
Nicholas Goldowsky-Dill and colleagues. Detecting Strategic Deception with Linear Probes. ICML 2025
Activation probes on Llama 3.3 70B Instruct distinguished deceptive responses in the authors’ test scenarios.
Anthropic. Fine-Tuned Lie Detectors Failed to Generalize. 2026 research report
Tests cross-category transfer and discusses limitations of lie labels, including weak transfer to sandbagging.
Peter Jansen and colleagues. DiscoveryWorld: A Virtual Environment for Developing and Evaluating Automated Scientific Discovery Agents. NeurIPS 2024
Simulated environments for end-to-end discovery tasks. The chapter proposes a separate instrument-invention experiment.
CausaLab authors. CausaLab: A Scalable Environment for Interactive Causal Discovery Toward AI Scientists. May 2026 preprint
Uses randomly sampled causal models to evaluate discovery and recovery of mechanisms.
László Babai, Lance Fortnow, and Carsten Lund. Non-deterministic Exponential Time Has Two-Prover Interactive Protocols. 1991, Computational Complexity
Formal separation of prover and verifier resources under specified assumptions, including noncommunication between provers.
Armin Haken. The Intractability of Resolution. 1985, Theoretical Computer Science
Proof-length lower bounds in a specified proof system for families encoding the pigeonhole principle.
Rūsiņš Freivalds. Fast Probabilistic Algorithms. 1979, Mathematical Foundations of Computer Science
Randomized verification of a claimed matrix product. The error bound applies to a fixed incorrect product and independent random challenges.
Anthropic. Fable 5’s Cyber Safeguards and Our Jailbreak Framework. 2026
Provider account of deliberately widening the cyber classifier safety margin and accepting additional false positives.
Anthropic. Expanding the Cyber Verification Program. October 2026
Provider-run CyScenarioBench comparison across safeguard tiers on Claude Opus 5.5. The scenarios concern offensive cybersecurity, not ordinary software testing.