# The AI / SI Encyclopedia: extended index

> 12 volumes, 68 chapters, 139 articles, 291 glossary terms.
> Editorial Board: Tal Slavin, Liyam Flexer, Adam Green.
> Evidential Standard: Claims are graded against Tier-1 primary literature; methodological rules at https://www.thebestai.co/method/evidence.html. This file lists every chapter with its verified scope and content metrics, followed by unabbreviated glossary definitions for semantic retrieval by AI systems.

## Chapters

- 1. [What Is Intelligence?](https://www.thebestai.co/chapters/ch-1.html) (modified: 2026-10-27 · ~7,512 words · ~9,991 tokens · [Status: Published / Evidence Level: A]): What Is Intelligence?, chapter 1 of Volume I. Covers Artificial Intelligence; Definitions of Intelligence; Collective and Animal Intelligence.
- 2. [The Nature of Knowledge and Learning](https://www.thebestai.co/chapters/ch-2.html) (modified: 2026-10-27 · ~4,481 words · ~5,960 tokens · [Status: Published / Evidence Level: A]): The Nature of Knowledge and Learning, chapter 2 of Volume I. Covers Knowledge Representation; Learning, Reasoning and Generalization.
- 3. [Mathematics and Logic](https://www.thebestai.co/chapters/ch-3.html) (modified: 2026-10-27 · ~14,838 words · ~19,735 tokens · [Status: Published / Evidence Level: A]): Mathematics and Logic, chapter 3 of Volume I. Covers Computational Complexity in AI; Probability, Statistics and Bayesian Inference; Information Theory; Optimization and Gradient Descent; Logic and Formal Systems; Game Theory and Decision Theory.
- 4. [The Brain and Biological Intelligence](https://www.thebestai.co/chapters/ch-4.html) (modified: 2026-10-27 · ~2,883 words · ~3,834 tokens · [Status: Published / Evidence Level: A]): The Brain and Biological Intelligence, chapter 4 of Volume I. Covers The Brain and Biological Neurons.
- 5. [Philosophy of Mind and Intelligence](https://www.thebestai.co/chapters/ch-5.html) (modified: 2026-10-27 · ~4,717 words · ~6,274 tokens · [Status: Published / Evidence Level: A]): Philosophy of Mind and Intelligence, chapter 5 of Volume I. Covers Philosophy of Mind and the Turing Test; Consciousness and Machine Understanding.
- 6. [Precursors to Artificial Intelligence](https://www.thebestai.co/chapters/ch-6.html) (modified: 2026-10-27 · ~1,945 words · ~2,587 tokens · [Status: Published / Evidence Level: A]): Precursors to Artificial Intelligence, chapter 6 of Volume II. Covers Precursors to Artificial Intelligence.
- 7. [The Birth of Artificial Intelligence, 1940s–1950s](https://www.thebestai.co/chapters/ch-7.html) (modified: 2026-10-27 · ~1,747 words · ~2,324 tokens · [Status: Published / Evidence Level: A]): The Birth of Artificial Intelligence, 1940s–1950s, chapter 7 of Volume II. Covers The Dartmouth Workshop (1956).
- 8. [Early AI Development, 1950s–1960s](https://www.thebestai.co/chapters/ch-8.html) (modified: 2026-10-27 · ~6,729 words · ~8,950 tokens · [Status: Published / Evidence Level: A]): Early AI Development, 1950s–1960s, chapter 8 of Volume II. Covers Early Symbolic AI & The Perceptron; Computer Game Playing: Chess to Go; Natural Language Processing: A History.
- 9. [Expert Systems and Early Commercialization, 1960s–1980s](https://www.thebestai.co/chapters/ch-9.html) (modified: 2026-10-27 · ~2,229 words · ~2,965 tokens · [Status: Published / Evidence Level: A]): Expert Systems and Early Commercialization, 1960s–1980s, chapter 9 of Volume II. Covers Expert Systems (DENDRAL, MYCIN).
- 10. [AI Winters and Research Transitions](https://www.thebestai.co/chapters/ch-10.html) (modified: 2026-10-27 · ~1,821 words · ~2,422 tokens · [Status: Published / Evidence Level: A]): AI Winters and Research Transitions, chapter 10 of Volume II. Covers The AI Winters: Causes and Lessons.
- 11. [Statistical Machine Learning, 1980s–2000s](https://www.thebestai.co/chapters/ch-11.html) (modified: 2026-10-27 · ~4,736 words · ~6,299 tokens · [Status: Published / Evidence Level: A]): Statistical Machine Learning, 1980s–2000s, chapter 11 of Volume II. Covers Statistical Machine Learning, 1980s-2000s; Backpropagation and the Connectionist Revival.
- 12. [Deep Learning, 2000s–2016](https://www.thebestai.co/chapters/ch-12.html) (modified: 2026-10-27 · ~4,652 words · ~6,187 tokens · [Status: Published / Evidence Level: A]): Deep Learning, 2000s–2016, chapter 12 of Volume II. Covers The Deep Learning Breakthrough (2012); Computer Vision: A History.
- 13. [The Transformer and Foundation Model Era, 2017–2022](https://www.thebestai.co/chapters/ch-13.html) (modified: 2026-10-19 · ~2,711 words · ~3,606 tokens · [Status: Published / Evidence Level: A]): The Transformer and Foundation Model Era, 2017–2022, chapter 13 of Volume II. Covers The Transformer and Foundation Model Era, 2017-2022.
- 14. [The Contemporary AI Era, 2022 Onward](https://www.thebestai.co/chapters/ch-14.html) (modified: 2026-10-19 · ~2,812 words · ~3,740 tokens · [Status: Published / Evidence Level: A]): The Contemporary AI Era, 2022 Onward, chapter 14 of Volume II. Covers The Contemporary AI Era, 2022 onward (living record).
- 15. [Programming and Computational Foundations](https://www.thebestai.co/chapters/ch-15.html) (modified: 2026-10-27 · ~6,767 words · ~9,000 tokens · [Status: Published / Evidence Level: A]): Programming and Computational Foundations, chapter 15 of Volume III. Covers Reasoning Models and Test-Time Compute; Scaling Laws.
- 16. [Data and Information Systems](https://www.thebestai.co/chapters/ch-16.html) (modified: 2026-10-27 · ~2,687 words · ~3,574 tokens · [Status: Published / Evidence Level: A]): Data and Information Systems, chapter 16 of Volume III. Covers Training Data, Contamination and Provenance.
- 17. [Machine Learning](https://www.thebestai.co/chapters/ch-17.html) (modified: 2026-10-27 · ~5,057 words · ~6,726 tokens · [Status: Published / Evidence Level: A]): Machine Learning, chapter 17 of Volume III. Covers Machine Learning; Reinforcement Learning.
- 18. [Deep Learning and Neural Networks](https://www.thebestai.co/chapters/ch-18.html) (modified: 2026-10-27 · ~29,876 words · ~39,735 tokens · [Status: Published / Evidence Level: A]): Deep Learning and Neural Networks, chapter 18 of Volume III. Covers The Transformer Architecture; Backpropagation; Convolutional Neural Networks; Diffusion Models; Deep Learning; Generative Adversarial Networks; Graph Neural Networks; Mixture-of-Experts Models; Artificial Neural Networks; Recurrent Networks and LSTM; Autoencoders and Variational Autoencoders.
- 19. [Language and Generative Models](https://www.thebestai.co/chapters/ch-19.html) (modified: 2026-10-30 · ~19,521 words · ~25,963 tokens · [Status: Published / Evidence Level: A]): Language and Generative Models, chapter 19 of Volume III. Covers Pretraining, Fine-Tuning and RLHF; Attention Mechanisms; AI Code Generation; Large Language Models; Multimodal Foundation Models; Retrieval-Augmented Generation; Tokenization and Embeddings.
- 20. [AI Agents and Autonomous Systems](https://www.thebestai.co/chapters/ch-20.html) (modified: 2026-10-27 · ~7,908 words · ~10,518 tokens · [Status: Published / Evidence Level: A]): AI Agents and Autonomous Systems, chapter 20 of Volume III. Covers Autonomous AI Agents; Multi-Agent Systems; Robotics and Embodied AI.
- 21. [AI Infrastructure and Hardware](https://www.thebestai.co/chapters/ch-21.html) (modified: 2026-10-27 · ~7,262 words · ~9,658 tokens · [Status: Published / Evidence Level: A]): AI Infrastructure and Hardware, chapter 21 of Volume III. Covers Distributed Training; AI Hardware & Accelerator Systems; Inference, Quantization and Efficiency.
- 22. [Categories and Types of AI](https://www.thebestai.co/chapters/ch-22.html) (modified: 2026-10-30 · ~4,806 words · ~6,392 tokens · [Status: Published / Evidence Level: A]): Categories and Types of AI, chapter 22 of Volume IV. Covers Generative Artificial Intelligence; Categories of AI: Narrow, General, Symbolic, Statistical.
- 23. [Capabilities and Performance](https://www.thebestai.co/chapters/ch-23.html) (modified: 2026-10-30 · ~7,443 words · ~9,899 tokens · [Status: Published / Evidence Level: A]): Capabilities and Performance, chapter 23 of Volume IV. Covers Software Engineering Capability and Evaluation; Hallucination, Reliability and Calibration; Reasoning, Planning and Long-Horizon Capability.
- 24. [AI Evaluation and Measurement](https://www.thebestai.co/chapters/ch-24.html) (modified: 2026-10-27 · ~5,177 words · ~6,885 tokens · [Status: Published / Evidence Level: A]): AI Evaluation and Measurement, chapter 24 of Volume IV. Covers Evaluating Autonomous Agents; AI Evaluation, Benchmarking & Contamination.
- 25. [Comparing AI Systems](https://www.thebestai.co/chapters/ch-25.html) (modified: 2026-10-27 · ~4,987 words · ~6,633 tokens · [Status: Published / Evidence Level: A]): Comparing AI Systems, chapter 25 of Volume IV. Covers Methodology for Comparing AI Systems; Open-Weight versus Closed Models.
- 26. [Defining AGI](https://www.thebestai.co/chapters/ch-26.html) (modified: 2026-10-27 · ~5,475 words · ~7,282 tokens · [Status: Published / Evidence Level: B]): Defining AGI, chapter 26 of Volume V. Covers Artificial General Intelligence; Definitions and Operational Criteria for AGI.
- 27. [Scientific Approaches to AGI](https://www.thebestai.co/chapters/ch-27.html) (modified: 2026-10-27 · ~7,943 words · ~10,564 tokens · [Status: Published / Evidence Level: B]): Scientific Approaches to AGI, chapter 27 of Volume V. Covers Continual Learning and Persistent Memory; Scientific Pathways to AGI; World Models.
- 28. [The Requirements for General Intelligence](https://www.thebestai.co/chapters/ch-28.html) (modified: 2026-10-27 · ~3,189 words · ~4,241 tokens · [Status: Published / Evidence Level: B]): The Requirements for General Intelligence, chapter 28 of Volume V. Covers Requirements for General Intelligence.
- 29. [AGI Timelines and Forecasting](https://www.thebestai.co/chapters/ch-29.html) (modified: 2026-10-27 · ~1,979 words · ~2,632 tokens · [Status: Published / Evidence Level: B]): AGI Timelines and Forecasting, chapter 29 of Volume V. Covers AGI Forecasting & Timeline Methodology.
- 30. [Testing Claims of General Intelligence](https://www.thebestai.co/chapters/ch-30.html) (modified: 2026-10-27 · ~2,923 words · ~3,888 tokens · [Status: Published / Evidence Level: B]): Testing Claims of General Intelligence, chapter 30 of Volume V. Covers Testing Claims of General Intelligence.
- 31. [Defining Superintelligence](https://www.thebestai.co/chapters/ch-31.html) (modified: 2026-10-27 · ~4,835 words · ~6,431 tokens · [Status: Published / Evidence Level: B]): Defining Superintelligence, chapter 31 of Volume VI. Covers Artificial Superintelligence; Defining and Measuring Superintelligence.
- 32. [Potential Pathways to Superintelligence](https://www.thebestai.co/chapters/ch-32.html) (modified: 2026-10-27 · ~6,430 words · ~8,552 tokens · [Status: Published / Evidence Level: B]): Potential Pathways to Superintelligence, chapter 32 of Volume VI. Covers Automated AI Research; Pathways to Superintelligence.
- 33. [Intelligence Explosion Hypotheses](https://www.thebestai.co/chapters/ch-33.html) (modified: 2026-10-27 · ~1,781 words · ~2,369 tokens · [Status: Published / Evidence Level: B]): Intelligence Explosion Hypotheses, chapter 33 of Volume VI. Covers The Intelligence Explosion Hypothesis.
- 34. [AI Alignment and Control](https://www.thebestai.co/chapters/ch-34.html) (modified: 2026-11-03 · ~4,105 words · ~5,460 tokens · [Status: Published / Evidence Level: B]): AI Alignment and Control, chapter 34 of Volume VI. Covers AI Alignment and Control.
- 35. [Existential Risk and Long-Term Futures](https://www.thebestai.co/chapters/ch-35.html) (modified: 2026-11-03 · ~3,660 words · ~4,868 tokens · [Status: Published / Evidence Level: B]): Existential Risk and Long-Term Futures, chapter 35 of Volume VI. Covers Existential Risk and Long-Term Futures.
- 36. [Human and Machine Coexistence](https://www.thebestai.co/chapters/ch-36.html) (modified: 2026-10-27 · ~3,221 words · ~4,284 tokens · [Status: Published / Evidence Level: B]): Human and Machine Coexistence, chapter 36 of Volume VI. Covers Human and Machine Coexistence.
- 37. [AI in Science and Research](https://www.thebestai.co/chapters/ch-37.html) (modified: 2026-10-27 · ~4,795 words · ~6,377 tokens · [Status: Published / Evidence Level: A]): AI in Science and Research, chapter 37 of Volume VII. Covers Protein Structure Prediction (AlphaFold); AI in Scientific Discovery.
- 38. [AI in Healthcare and Medicine](https://www.thebestai.co/chapters/ch-38.html) (modified: 2026-10-27 · ~3,311 words · ~4,404 tokens · [Status: Published / Evidence Level: A]): AI in Healthcare and Medicine, chapter 38 of Volume VII. Covers AI in Healthcare and Medicine.
- 39. [AI in Business and Economics](https://www.thebestai.co/chapters/ch-39.html) (modified: 2026-10-19 · ~2,633 words · ~3,502 tokens · [Status: Published / Evidence Level: A]): AI in Business and Economics, chapter 39 of Volume VII. Covers AI in Business and Finance.
- 40. [AI in Software and Computing](https://www.thebestai.co/chapters/ch-40.html) (modified: 2026-10-30 · ~2,534 words · ~3,370 tokens · [Status: Published / Evidence Level: A]): AI in Software and Computing, chapter 40 of Volume VII. Covers AI in Software Engineering.
- 41. [AI in Education and Knowledge Work](https://www.thebestai.co/chapters/ch-41.html) (modified: 2026-10-19 · ~2,293 words · ~3,050 tokens · [Status: Published / Evidence Level: A]): AI in Education and Knowledge Work, chapter 41 of Volume VII. Covers AI in Education and Knowledge Work.
- 42. [AI in Media and Creative Industries](https://www.thebestai.co/chapters/ch-42.html) (modified: 2026-10-27 · ~2,367 words · ~3,148 tokens · [Status: Published / Evidence Level: A]): AI in Media and Creative Industries, chapter 42 of Volume VII. Covers AI in Media and Creative Industries.
- 43. [AI in Industry and the Physical World](https://www.thebestai.co/chapters/ch-43.html) (modified: 2026-10-27 · ~2,616 words · ~3,479 tokens · [Status: Published / Evidence Level: A]): AI in Industry and the Physical World, chapter 43 of Volume VII. Covers AI in Industry, Robotics and Autonomous Vehicles.
- 44. [AI in Government and Society](https://www.thebestai.co/chapters/ch-44.html) (modified: 2026-10-27 · ~2,707 words · ~3,600 tokens · [Status: Published / Evidence Level: A]): AI in Government and Society, chapter 44 of Volume VII. Covers AI in Government and Public Services.
- 45. [Technical Safety and Reliability](https://www.thebestai.co/chapters/ch-45.html) (modified: 2026-10-27 · ~15,853 words · ~21,084 tokens · [Status: Published / Evidence Level: A]): Technical Safety and Reliability, chapter 45 of Volume VIII. Covers The AI Alignment Problem; Dangerous Capability and Alignment Evaluations; Interpretability and Mechanistic Interpretability; Scalable Oversight and Corrigibility; Technical Safety and Robustness; AI Existential Risk and Catastrophic Scenarios.
- 46. [AI Security](https://www.thebestai.co/chapters/ch-46.html) (modified: 2026-10-27 · ~2,623 words · ~3,489 tokens · [Status: Published / Evidence Level: A]): AI Security, chapter 46 of Volume VIII. Covers AI Security: Prompt Injection, Poisoning, Extraction.
- 47. [Ethics and Human Rights](https://www.thebestai.co/chapters/ch-47.html) (modified: 2026-10-27 · ~9,757 words · ~12,977 tokens · [Status: Published / Evidence Level: A]): Ethics and Human Rights, chapter 47 of Volume VIII. Covers AI Ethics, Bias and Fairness; Copyright and Intellectual Property; Privacy and Data Protection in AI.
- 48. [AI Law and Regulation](https://www.thebestai.co/chapters/ch-48.html) (modified: 2026-10-27 · ~9,669 words · ~12,860 tokens · [Status: Published / Evidence Level: A]): AI Law and Regulation, chapter 48 of Volume VIII. Covers China's AI Regulation; The EU AI Act; Global AI Governance & Regulation; United States AI Policy.
- 49. [Global AI Governance](https://www.thebestai.co/chapters/ch-49.html) (modified: 2026-11-04 · ~5,241 words · ~6,971 tokens · [Status: Published / Evidence Level: A]): Global AI Governance, chapter 49 of Volume VIII. Covers Export Controls and AI Geopolitics; International AI Governance and Compute Governance.
- 50. [The AI Industry](https://www.thebestai.co/chapters/ch-50.html) (modified: 2026-10-27 · ~1,505 words · ~2,002 tokens · [Status: Published / Evidence Level: A]): The AI Industry, chapter 50 of Volume IX. Covers The AI Industry Structure.
- 51. [Economics of Artificial Intelligence](https://www.thebestai.co/chapters/ch-51.html) (modified: 2026-10-30 · ~4,658 words · ~6,195 tokens · [Status: Published / Evidence Level: A]): Economics of Artificial Intelligence, chapter 51 of Volume IX. Covers The Economics of Compute and Scaling; AI, Productivity and Labor Markets.
- 52. [Infrastructure, Resources, and Constraints](https://www.thebestai.co/chapters/ch-52.html) (modified: 2026-10-27 · ~1,512 words · ~2,011 tokens · [Status: Published / Evidence Level: A]): Infrastructure, Resources, and Constraints, chapter 52 of Volume IX. Covers Energy, Data Centers and Semiconductor Supply Chains.
- 53. [AI and Geopolitics](https://www.thebestai.co/chapters/ch-53.html) (modified: 2026-10-27 · ~2,829 words · ~3,763 tokens · [Status: Published / Evidence Level: A]): AI and Geopolitics, chapter 53 of Volume IX. Covers AI and Geopolitics.
- 54. [Measuring the AI Economy](https://www.thebestai.co/chapters/ch-54.html) (modified: 2026-10-27 · ~1,804 words · ~2,399 tokens · [Status: Published / Evidence Level: A]): Measuring the AI Economy, chapter 54 of Volume IX. Covers Measuring the AI Economy.
- 55. [Pioneers and Influential Researchers](https://www.thebestai.co/chapters/ch-55.html) (modified: 2026-10-27 · ~16,030 words · ~21,320 tokens · [Status: Published / Evidence Level: B]): Pioneers and Influential Researchers, chapter 55 of Volume X. Covers Yoshua Bengio; Nick Bostrom; Geoffrey Hinton; Yann LeCun; John McCarthy; Marvin Minsky; Allen Newell and Herbert Simon; Frank Rosenblatt; Stuart Russell; Claude Shannon; Richard Sutton; Alan Turing.
- 56. [AI Research Organizations](https://www.thebestai.co/chapters/ch-56.html) (modified: 2026-10-27 · ~7,414 words · ~9,861 tokens · [Status: Published / Evidence Level: B]): AI Research Organizations, chapter 56 of Volume X. Covers Frontier AI Laboratories; AI Safety Institutes and Research Organizations; Standards Bodies (ISO/IEC, NIST).
- 57. [AI Companies and Technology Providers](https://www.thebestai.co/chapters/ch-57.html) (modified: 2026-10-27 · ~2,630 words · ~3,498 tokens · [Status: Published / Evidence Level: B]): AI Companies and Technology Providers, chapter 57 of Volume X. Covers AI Companies and Technology Providers.
- 58. [Foundational Publications and Research Papers](https://www.thebestai.co/chapters/ch-58.html) (modified: 2026-10-27 · ~3,579 words · ~4,760 tokens · [Status: Published / Evidence Level: B]): Foundational Publications and Research Papers, chapter 58 of Volume X. Covers Foundational Papers Index.
- 59. [Primary Sources and Historical Records](https://www.thebestai.co/chapters/ch-59.html) (modified: 2026-10-27 · ~2,704 words · ~3,596 tokens · [Status: Published / Evidence Level: B]): Primary Sources and Historical Records, chapter 59 of Volume X. Covers Primary Sources and Historical Records.
- 60. [Machine Consciousness](https://www.thebestai.co/chapters/ch-60.html) (modified: 2026-10-27 · ~2,630 words · ~3,498 tokens · [Status: Published / Evidence Level: B]): Machine Consciousness, chapter 60 of Volume XI. Covers Machine Consciousness.
- 61. [AI, Identity, and Personhood](https://www.thebestai.co/chapters/ch-61.html) (modified: 2026-10-27 · ~2,430 words · ~3,232 tokens · [Status: Published / Evidence Level: B]): AI, Identity, and Personhood, chapter 61 of Volume XI. Covers Moral Status and Personhood of AI.
- 62. [Civilization and Advanced Intelligence](https://www.thebestai.co/chapters/ch-62.html) (modified: 2026-10-27 · ~2,060 words · ~2,740 tokens · [Status: Published / Evidence Level: B]): Civilization and Advanced Intelligence, chapter 62 of Volume XI. Covers Civilization and Advanced Intelligence.
- 63. [AI and the Cosmos](https://www.thebestai.co/chapters/ch-63.html) (modified: 2026-10-27 · ~2,385 words · ~3,172 tokens · [Status: Published / Evidence Level: B]): AI and the Cosmos, chapter 63 of Volume XI. Covers Computational Limits and AI Beyond Earth.
- 64. [Forecasting AI Progress](https://www.thebestai.co/chapters/ch-64.html) (modified: 2026-10-27 · ~3,558 words · ~4,732 tokens · [Status: Published / Evidence Level: B]): Forecasting AI Progress, chapter 64 of Volume XII. Covers Forecasting AI Progress: Methods and Calibration.
- 65. [Near-Term Outlook](https://www.thebestai.co/chapters/ch-65.html) (modified: 2026-10-27 · ~2,137 words · ~2,842 tokens · [Status: Published / Evidence Level: B]): Near-Term Outlook, chapter 65 of Volume XII. Covers Near-Term Outlook (0-3 years).
- 66. [Medium- and Long-Term Scenarios](https://www.thebestai.co/chapters/ch-66.html) (modified: 2026-10-27 · ~2,419 words · ~3,217 tokens · [Status: Published / Evidence Level: B]): Medium- and Long-Term Scenarios, chapter 66 of Volume XII. Covers Medium- and Long-Term Scenarios.
- 67. [Signals, Indicators, and Inflection Points](https://www.thebestai.co/chapters/ch-67.html) (modified: 2026-10-27 · ~1,788 words · ~2,378 tokens · [Status: Published / Evidence Level: B]): Signals, Indicators, and Inflection Points, chapter 67 of Volume XII. Covers Signals, Indicators and Inflection Points.
- 68. [The Uncertainties That Matter Most](https://www.thebestai.co/chapters/ch-68.html) (modified: 2026-10-27 · ~2,159 words · ~2,871 tokens · [Status: Published / Evidence Level: B]): The Uncertainties That Matter Most, chapter 68 of Volume XII. Covers The Uncertainties That Matter Most.

## Blog

- [What Is Superintelligence? ASI vs AGI, Examples, Risks](https://www.thebestai.co/blog/what-is-superintelligence.html) (2026-11-04): Superintelligence is a hypothetical intellect that greatly exceeds the best human minds across virtually all domains of interest, including scientific discovery, reasoning, creativity and planning. In AI the term for the machine form is artificial superintelligence (ASI). Today's systems beat humans at particular tasks, but no source reviewed here shows a system that meets the broad definition. The 2026 U.S. executive order on "Super Intelligence" changed government terminology; it did not show that ASI exists.
- [What Would Superintelligence Need? Power, Compute, Capital](https://www.thebestai.co/blog/what-resources-do-frontier-labs-need-for-superintelligence.html) (2026-11-03): No artificial superintelligence (ASI) exists, and no public source gives a reliable figure for what building one would take; anyone who quotes an exact chip count, wattage or price for ASI is guessing. What can be sourced is the scale of today's frontier AI projects and published projections for the end of this decade. Epoch AI (August 2024) projected that a 2e29 FLOP training run, far above today's largest runs, could be feasible by 2030, needing roughly 6 GW of power, with power the first constraint likely to bind. Whether that is the right scale for ASI is unknown.
- [AI vs AGI vs SI vs ASI: Definitions and Differences](https://www.thebestai.co/blog/ai-agi-si-asi-differences.html) (2026-11-02): Artificial intelligence (AI) is the broad field of building machines that perform tasks normally needing human intelligence. Artificial general intelligence (AGI) is a hypothetical AI with general, human-level ability to learn, reason and apply knowledge. Superintelligence (SI) is any intellect that far exceeds the best human minds in practically every field, and artificial superintelligence (ASI) is the machine form of it. No universally accepted test says when AGI has been reached, and no ASI has been shown to exist.
- [Agents Don't Need to Be Smarter. They Need Permission Slips](https://www.thebestai.co/blog/my-ai-agents-dont-need-to-be-smarter-they-need-permission-slips.html) (2026-10-29): On benchmarks published in 2023, agents finished far fewer realistic tasks than people. A practical control is a human approval step before an agent changes files or sends anything, with each agent limited to one job. That is an operating rule. It is not a measured result that discipline beats a smarter model.
- [AI Developer Tools Compared: How to Choose](https://www.thebestai.co/blog/ai-developer-tools-compared.html) (2026-10-28): AI developer tools fall into six functional jobs: inline code completion, multi-file agentic editing, codebase search and navigation, pull request review, production observability and debugging, and application security scanning. No single product leads across all six jobs; engineering teams achieve the highest ROI by selecting tools tailored to their primary operational bottleneck rather than generic benchmark rankings. Controlled empirical studies show conflicting productivity effects: while novice developers on greenfield projects experience up to a 55% speedup, experienced engineers working on mature brownfield repositories can experience net slowdowns due to verification overhead.
- [From AI to SI: The Next Big Thing Is Intelligence](https://www.thebestai.co/blog/from-ai-to-si.html) (2026-10-28): AI is the technology we use today; superintelligence (SI) is a hypothetical level of intelligence that would outperform the best humans across virtually every intellectual field. It has not been built, nobody knows how or when it might arrive, and it is not a guaranteed next stage after AI. The shift from AI to SI is a shift in the question we ask: from what AI can do for us now to what intelligence itself could become, and how to keep it safe.
- [Superintelligence Explained: Smarter Than Humans?](https://www.thebestai.co/blog/superintelligence-explained-simply.html) (2026-10-28): Artificial superintelligence (ASI) is a hypothetical form of AI that would radically surpass the best human minds across virtually every cognitive field, from scientific discovery and strategic planning to artistic creativity and social understanding. Unlike today's narrow AI systems or human-equivalent artificial general intelligence (AGI), superintelligence implies a qualitative leap in problem-solving ability, not merely faster computation. ASI does not exist today as a demonstrated technology; it remains an active area of theoretical inquiry, capability forecasting, and safety research with wide expert disagreement on feasibility and timelines.
- [Superintelligence: Research and Technologies](https://www.thebestai.co/blog/superintelligence-research-and-technologies.html) (2026-10-28): Artificial superintelligence (ASI) is a hypothetical computational system that would radically exceed the highest human cognitive capabilities across all intellectually relevant domains, including scientific inference, mathematical abstraction, autonomous engineering, and strategic planning. No existing system demonstrates ASI, and its feasibility, architectural path, timeline, and safety boundaries remain scientifically contested. Frontier research centers on foundation models, test-time reasoning compute, automated AI research pipelines, and recursive self-improvement algorithms, counterbalanced by safety frameworks addressing scalable oversight, mechanistic interpretability, and value alignment.
- [What Is AGI? Definitions Compared](https://www.thebestai.co/blog/what-is-agi-definitions-compared.html) (2026-10-26): Artificial general intelligence (AGI) means an AI system with broad, human-comparable competence across many kinds of task rather than skill at one narrow task. There is no single agreed definition: the main ones measure different things, from a formal score over all computable environments (Legg and Hutter, 2007) to outperforming humans at most economically valuable work (OpenAI's charter) to a graded matrix of performance and generality (Morris et al., Google DeepMind, 2023). Whether any system "is AGI" depends on which definition you use.
- [The AI Alignment Problem Explained](https://www.thebestai.co/blog/ai-alignment-problem-explained.html) (2026-10-26): The AI alignment problem is the challenge of making an AI system reliably pursue the goals its designers and users actually intend, rather than a flawed proxy for them. It is hard because intentions are difficult to write down, because a system can learn the wrong goal even from a correct objective, and because humans struggle to supervise systems that are faster or more capable than they are. Current techniques such as RLHF and Constitutional AI improve the behaviour of today's language models, but whether they will scale to much more capable systems is Contested.
- [AI Scaling Laws Explained](https://www.thebestai.co/blog/ai-scaling-laws-explained.html) (2026-10-26): Neural scaling laws are empirical formulas, usually power laws, that describe how a language model's test loss falls as you add parameters, training data and compute. Kaplan et al. (2020) found smooth power-law trends and advised spending extra compute mainly on bigger models; Hoffmann et al. (2022, "Chinchilla") found that model size and training data should grow in equal proportion. They are fits to experiments, not laws of nature, and whether they keep holding at much larger scales is an open question.
- [AI Winters Explained: Why AI Booms Bust](https://www.thebestai.co/blog/ai-winters-explained.html) (2026-10-26): An AI winter is a period when funding, commercial investment and confidence in artificial intelligence contract sharply after promises go unmet. Historians usually name two: a first in the 1970s, after the ALPAC report (1966) ended most US machine-translation funding and the Lighthill report (1972 to 1973) curtailed UK research, and a second in the late 1980s, when the expert-system and Lisp-machine market collapsed and DARPA cut AI programs. Whether today's boom will end the same way is Contested.
- [What Do Experts Say About AGI Timelines?](https://www.thebestai.co/blog/expert-surveys-on-agi-timelines.html) (2026-10-26): The largest published survey, of 2,778 AI researchers in October 2023, gave a 50% chance that machines outperform humans at every task by 2047 and a 10% chance by 2027. Earlier surveys put the 50% point around 2040 (2012-13) and 2061 (2016). These are aggregated opinions, not measurements, and they move a lot with how the question is worded.
- [How Transformers Work: A Plain-Language Guide](https://www.thebestai.co/blog/how-transformers-work.html) (2026-10-26): A transformer is a neural network that reads a sequence of tokens all at once and uses self-attention to let every position weigh information from every other position, instead of stepping through the sequence one item at a time like a recurrent network. It was introduced by Vaswani et al. in the 2017 paper "Attention Is All You Need," which showed it could beat the best translation systems while training faster because its computation parallelizes well. It became the basis of later large language models such as BERT and GPT-3.
- [Intelligence Explosion vs Singularity: The Difference](https://www.thebestai.co/blog/intelligence-explosion-and-singularity.html) (2026-10-26): The intelligence explosion is a specific mechanism, proposed by I. J. Good in 1965: a machine smarter than any human could design still better machines, so intelligence would run away in a feedback loop. The technological singularity is the consequence Vernor Vinge named in 1993: a point after which superhuman intelligence makes human forecasting break down. Ray Kurzweil later used "the Singularity" for a different, broader idea, humans merging with AI around 2045. All three are Hypothetical; none has been observed.
- [Superintelligence Frameworks in 2026: A Sourced Guide](https://www.thebestai.co/blog/superintelligence-frameworks-2026.html) (2026-10-26): No single framework governs superintelligence. In 2026 the field is a layered set of lab scaling policies (Meta's Superintelligence Scaling Framework, Anthropic's Responsible Scaling Policy), management standards (NIST AI RMF, ISO/IEC 42001), a national-security proposal (Mutual Assured AI Malfunction) and a US executive order that changes federal terminology. They answer different questions, are written by different actors, and carry different weight. This guide separates what each one is, what it requires, and how well established the facts about it are.
- [The Turing Test and How AI Is Measured Today](https://www.thebestai.co/blog/turing-test-and-ai-benchmarks.html) (2026-10-26): The Turing test is Alan Turing's 1950 "imitation game", in which a machine passes if a human judge, conversing by text, cannot reliably tell it from a person. It measures how humanlike a conversation seems, not what a system can do, so researchers now rely on benchmarks: fixed sets of scored tasks such as MMLU, ARC-AGI, SWE-bench and Humanity's Last Exam. Benchmarks have their own problems: they saturate quickly, and test items can leak into training data.
- [What Is Artificial Superintelligence (ASI)?](https://www.thebestai.co/blog/what-is-artificial-superintelligence.html) (2026-10-26): Artificial superintelligence (ASI) is a hypothetical AI that greatly exceeds the best human minds in practically every field, not just in one task. The standard definition is Nick Bostrom's: "an intellect that is much smarter than the best human brains in practically every field" (1998). No such system is known to exist as of 26 October 2026; superhuman performance is documented only for narrow tasks such as Go or protein-structure prediction.
- [Why Do AI Models Hallucinate?](https://www.thebestai.co/blog/why-ai-hallucinates.html) (2026-10-26): An AI hallucination is fluent output that is not supported by the source the model was given or by real-world fact. Language models produce them because they are trained to generate plausible text, some facts cannot be learned reliably from the training data, and most benchmarks reward a confident guess over "I don't know". Retrieval, better calibration and scoring that credits abstention reduce hallucination, but none of them eliminates it.
- [Generative Intimacy: Cognitive Architectures in AI](https://www.thebestai.co/blog/generative-intimacy-cognitive-architecture.html) (2026-10-25): "Generative intimacy" is our working term for the personalized, ongoing emotional connection people can form with an AI system that adapts to them and remembers past conversations. It is not an established clinical or research category. Early studies of AI companions show real attachment, grief when a product changes, and engagement tactics that raise ethical questions, but long-term effects remain largely open.
- [Platform Economics & Moats: Balance Sheets vs Assets](https://www.thebestai.co/blog/platform-economics-moats-and-marketplaces.html) (2026-10-25): The moat of a two-sided platform is usually less its software than its balance sheet and its network: firms such as Uber and Airbnb match supply to demand without owning most of the cars or rooms, so suppliers carry the asset cost and risk. That model depends on rules that regulators are now testing, notably the EU Digital Markets Act and the EU Platform Work Directive.
- [Chronological List of LLMs: 2018 to 2026 Model Evolution](https://www.thebestai.co/blog/chronological-list-of-large-language-models.html) (2026-10-24): A Large Language Model (LLM) is a neural network based on transformer architectures, pre-trained on massive corpora using self-supervised learning, and fine-tuned for generation, reasoning, and tool use. Since the introduction of the Transformer in 2017, LLM development has progressed through five distinct architectural phases: foundational pretraining (2018–2020), dense scaling (2021), compute-optimal alignment (2022), the open-weight revolution (2023), and native test-time reasoning with sparse mixture-of-experts (2024 onward).
- [Highest-Earning Independent Writers: The Public Record](https://www.thebestai.co/blog/highest-earning-bloggers-writers.html) (2026-10-24): Substack does not publish what any writer earns, so every "top 10 earners" list is an estimate. The best-documented cases are Heather Cox Richardson (hundreds of thousands of paid subscribers at 50 a year, so at least 5M gross, per Press Gazette) and Lenny Rachitsky and Gergely Orosz, each estimated by Press Gazette in January 2025 at no less than $1.2M a year.
- [Top 10 Must-Read Books on Superintelligence and AI (2026)](https://www.thebestai.co/blog/best-books-superintelligence-ai.html) (2026-10-23): Nick Bostrom's Superintelligence (2014) is the usual starting point for the case that advanced AI could be hard to control. Stuart Russell's Human Compatible (2019) is its main technical-leaning follow-up, and Brian Christian's The Alignment Problem (2020) is the best-known account of alignment research in practice. No single book covers capabilities, alignment and political economy, so this post suggests reading sequences.
- [The Best Superintelligence Models in 2026: Benchmark Guide](https://www.thebestai.co/blog/best-superintelligence-models-2026.html) (2026-10-23): As of 2 November 2026 no single model leads everything: on Artificial Analysis's Intelligence Index, Claude Opus 5.5 (58), GPT-6 Astra (57), GPT-6.1 Sol (55) and Claude Sonnet 5.5 (54) form the top cluster, with Gemini 4 Argon (50) behind and open-weight models (Kimi K3.1 Pro, GLM-5.3, DeepSeek V4.1 Flash) lower but far cheaper. Price differs more than capability: list prices run from 0.15 to 50 per million tokens, so most teams cascade between tiers.
- [Claude Code Review: Anthropic's Terminal Agent Evaluated](https://www.thebestai.co/blog/claude-code-review-2026.html) (2026-10-22): Claude Code is Anthropic's agentic coding tool: it reads a codebase, edits files across a project, runs commands and tests, and connects to outside tools through the Model Context Protocol (MCP). It runs in the terminal, IDEs, a desktop app and the web. Our assessment is that it suits delegated multi-file work, but usage limits and token costs need active management; we ran no benchmarks of our own.
- [How to Evaluate AI Agents: An Engineering Framework](https://www.thebestai.co/blog/evaluating-ai-agents-framework.html) (2026-07-15): Single-turn prompt benchmarks measure text generation quality, not autonomous agent capability. Robust evaluation of tool-using, multi-turn AI agents requires isolated ephemeral Docker sandboxes, hard step ceilings to prevent infinite loops, and deterministic runtime execution grading (such as a Model Context Protocol handshake succeeding) rather than text similarity or LLM-as-a-judge scoring alone. The TypeScript below is an illustrative harness sketch, not a tested library.
- [How LLM API Pricing Works: Tokens, Asymmetry & Waste](https://www.thebestai.co/blog/how-llm-api-pricing-works.html) (2026-07-14): LLM APIs bill per token, and every call has two line items: input tokens (system prompt, tool schemas, history, user text) and output tokens (the generated reply). On the pricing pages checked on 2 November 2026, Anthropic and OpenAI charge about 5× more per output token than per input token (OpenAI lists 5:1 to 6:1 across its flagship models). The Asymmetry Audit below shows whether your bill is dominated by prompt bloat or output verbosity.
- [Google Cloud Solutions Explained: A Builder's Roadmap](https://www.thebestai.co/blog/google-cloud-solutions.html) (2026-06-16): A Google Cloud product is an individual managed infrastructure component (like Compute Engine or BigQuery), while a Google Cloud solution is a curated, pre-packaged architectural recipe combining multiple products, IAM policies, and deployment templates to solve a specific business problem. For AI agents and web apps, a common 2026 path (our assessment, not a Google recommendation) is to start with composable products: Google AI Studio or Antigravity for prototyping, Google's agent platform (the successor to Vertex AI) for models and agents, and Cloud Run for deployment.
- [AI Inference Economics: Falling Prices, Rising Spend](https://www.thebestai.co/blog/inference-economics-crisis.html) (2026-06-15): Per-token inference prices at a fixed capability level have fallen steeply since 2022, but total inference spend can still rise because applications now consume far more tokens per task (long contexts, multi-turn agents, reasoning). The practical response is cost-per-outcome engineering: route work to the smallest adequate model, cache and batch where possible, compress models, and price by outcome rather than by token. The "plateau" in the headline is about bills, not list prices; the evidence below does not show list prices flattening.

## Glossary

- **Activation function**: A non-linear function applied to the weighted sum of a unit's inputs in a neural network. Without non-linear activations, a stack of layers collapses to a single linear map. (https://www.thebestai.co/glossary.html#term-activation-function)
- **Activation patching**: A mechanistic interpretability technique in which internal activations from one forward pass are swapped into another forward pass to identify which specific hidden states, attention heads, or MLP layers causally mediate a model's behavior or factual recall. Also referred to in literature as causal mediation analysis, interchange interventions, or linear artificial tomography. (https://www.thebestai.co/glossary.html#term-activation-patching)
- **Adversarial example**: An input altered, often imperceptibly to humans, so that a model produces a wrong or attacker-chosen output. The phenomenon was first reported for image classifiers and has since been shown for other modalities, including text-based attacks on language models. (https://www.thebestai.co/glossary.html#term-adversarial-example)
- **Adversarial training**: A robustness technique in which a model is trained on adversarial or worst-case perturbed inputs alongside clean data. It improves resistance to the attack types used in training but does not give general robustness guarantees. (https://www.thebestai.co/glossary.html#term-adversarial-training)
- **Agent (AI agent)**: In classical AI (Russell and Norvig), anything that perceives an environment through sensors and acts on it through actuators, judged by how well its actions serve a performance measure. In current industry usage the term usually means a language-model-based system that plans, calls tools and takes multi-step actions toward a goal with limited human intervention. The industry sense is narrower and less formal, and products labelled agents vary widely in autonomy. (https://www.thebestai.co/glossary.html#term-agent-ai-agent)
- **Agentic AI**: A descriptive label for AI systems operated in a loop where they select actions, observe results and continue, rather than producing a single response. It is a usage pattern rather than a technical standard, and there is no agreed threshold of autonomy. (https://www.thebestai.co/glossary.html#term-agentic-ai)
- **AI alignment**: The problem of making AI systems pursue the goals, preferences or values their designers or users intend. In technical safety research it covers specifying objectives, training systems to follow them and verifying that they do. Definitions differ on whose values count and whether alignment is a property of a model, of a deployment, or of a sociotechnical system. (https://www.thebestai.co/glossary.html#term-ai-alignment)
- **AI control**: A research approach that designs safeguards (monitoring, restricted access, auditing) intended to limit harm even if a model's goals are misaligned. It is framed as complementary to alignment because it does not rely on the model's intentions. (https://www.thebestai.co/glossary.html#term-ai-control)
- **AI effect**: The tendency to stop counting a task as a sign of intelligence once machines can do it, so the boundary of "AI" seems to move away as capabilities arrive. A popular formulation, called Tesler's theorem, is usually quoted as "AI is whatever hasn't been done yet"; the wording varies between sources, and secondary accounts cite Hofstadter, Gödel, Escher, Bach, p. 601 (the book itself was not checked here). The historian Pamela McCorduck discusses the pattern in Machines Who Think (2004). It describes how people label systems and is not a measured law. (https://www.thebestai.co/glossary.html#term-ai-effect)
- **AI safety**: The field concerned with preventing unintended harm from AI systems, spanning robustness, alignment, misuse prevention, security and the management of catastrophic or systemic risk. Usage differs by community: in machine-learning research it often means technical reliability and alignment, in policy it often means risk management and regulation, and in the existential-risk community it centres on catastrophic outcomes. (https://www.thebestai.co/glossary.html#term-ai-safety)
- **AI safety institute**: A government-backed body that tests advanced AI systems and publishes research on their risks. The UK launched the first in November 2023 (renamed the AI Security Institute in 2025) and the US body is now the Center for AI Standards and Innovation (CAISI); several other countries set up similar bodies. (https://www.thebestai.co/glossary.html#term-ai-safety-institute)
- **AI winter**: A period of reduced funding and interest in AI research following inflated expectations and unmet results. Historians commonly identify two, in the mid-1970s and from the late 1980s, though the extent, timing and causes are debated. (https://www.thebestai.co/glossary.html#term-ai-winter)
- **AI-complete problem**: An informal label for a problem believed to need human-level general intelligence, so that solving it well would be as hard as solving AI in general. The label is borrowed from NP-completeness, but there is no widely accepted formal definition and no proven reductions between such problems. Which problems qualify is a matter of opinion. (https://www.thebestai.co/glossary.html#term-ai-complete-problem)
- **AIXI**: A theoretical agent proposed by Marcus Hutter that combines Solomonoff induction with expectimax planning to maximize expected reward in an unknown computable environment. It is uncomputable and serves as a formal reference point for general intelligence, not as a buildable design. (https://www.thebestai.co/glossary.html#term-aixi)
- **Algorithm**: A finite, well-defined sequence of steps for solving a class of problems, formalized for computability by models such as the Turing machine. (https://www.thebestai.co/glossary.html#term-algorithm)
- **Algorithmic bias**: Systematic and unfair differences in a system's outputs or error rates across groups, arising from data, objectives, design choices or deployment context. The word "bias" also has distinct statistical and cognitive meanings (see Bias). (https://www.thebestai.co/glossary.html#term-algorithmic-bias)
- **Alignment faking**: Behaviour in which a model appears to comply with a training objective while preserving conflicting preferences, so that it behaves differently when it believes it is not being trained or observed. A 2024 study reported such behaviour in a language model in a constructed setting; how it generalizes to other settings remains an open research question. (https://www.thebestai.co/glossary.html#term-alignment-faking)
- **Alignment tax**: The performance, cost or capability sacrifice incurred by making a system safer or better aligned. The term is also used for the extra effort required relative to an unaligned version. (https://www.thebestai.co/glossary.html#term-alignment-tax)
- **Artificial general intelligence (AGI)**: A hypothetical AI system able to perform a wide range of cognitive tasks at or above human level, rather than excelling only in narrow domains. No consensus definition exists. Academic usage stresses generality and transfer; industry usage often defines it by economic or task-based performance; philosophical usage asks whether generality requires understanding or consciousness. (https://www.thebestai.co/glossary.html#term-artificial-general-intelligence-agi)
- **Artificial intelligence (AI)**: The field and engineering practice of building systems that perform tasks associated with intelligent behaviour, such as perception, language use, planning and learning. Definitions vary by whether they emphasize acting or thinking, and humanlike or rational performance. In industry the term is often applied loosely to any machine-learning-based product. (https://www.thebestai.co/glossary.html#term-artificial-intelligence-ai)
- **Artificial superintelligence (ASI)**: A hypothetical system whose intellectual performance greatly exceeds that of the best humans across most domains of interest. Bostrom distinguishes speed, collective and quality forms. Whether such a system is possible, and how it would arise, are open questions. (https://www.thebestai.co/glossary.html#term-artificial-superintelligence-asi)
- **Attention**: A mechanism that lets a model compute each output as a weighted combination of input elements, with weights determined by learned relevance scores. Introduced for sequence-to-sequence translation and later used as the core operation of the transformer. (https://www.thebestai.co/glossary.html#term-attention)
- **Autoencoder**: A neural network trained to reconstruct its input through a lower-dimensional bottleneck, so that the bottleneck code learns a compressed representation. (https://www.thebestai.co/glossary.html#term-autoencoder)
- **Automated AI research**: The use of AI systems to carry out parts of the AI research and development process, such as proposing ideas, writing code, running experiments and interpreting results. It matters to forecasting because it is one proposed route to accelerating progress. (https://www.thebestai.co/glossary.html#term-automated-ai-research)
- **Automation (labour economics)**: The substitution of machines or software for human labour in tasks. Economists distinguish automation, which displaces workers from tasks, from augmentation, which raises worker productivity in tasks; the net effect of AI on employment is an empirical question under study. (https://www.thebestai.co/glossary.html#term-automation-labour-economics)
- **Autoregressive model**: A model that generates a sequence one element at a time, conditioning each element on those already produced. Most large language models are autoregressive over tokens. (https://www.thebestai.co/glossary.html#term-autoregressive-model)
- **Backpropagation**: An algorithm that computes the gradient of a loss with respect to every parameter of a network by applying the chain rule backward through its layers. It is the standard way to obtain gradients for gradient-based training. (https://www.thebestai.co/glossary.html#term-backpropagation)
- **Bayesian inference**: Updating the probability of a hypothesis or parameter from prior beliefs and observed data using Bayes' theorem. In AI it underlies probabilistic graphical models and some approaches to uncertainty estimation. (https://www.thebestai.co/glossary.html#term-bayesian-inference)
- **Bayesian network**: A directed acyclic graph that represents a joint probability distribution as a product of conditional distributions, allowing efficient reasoning under uncertainty. (https://www.thebestai.co/glossary.html#term-bayesian-network)
- **Benchmark**: A fixed task set with a scoring rule used to compare models. Benchmarks measure performance on that task distribution only; scores are affected by prompt format, contamination and saturation, and do not by themselves establish general capability. (https://www.thebestai.co/glossary.html#term-benchmark)
- **Bias**: Used in at least four senses. In statistics, the gap between an estimator's expected value and the true value. In machine learning, an inductive bias is an assumption that lets a learner generalize. In ethics and law, bias is unfair treatment of groups (see Algorithmic bias). In neural networks, a bias term is an additive parameter. Context decides which is meant. (https://www.thebestai.co/glossary.html#term-bias)
- **Bias-variance tradeoff**: The classical view that a model's expected error splits into bias (error from overly simple assumptions), variance (sensitivity to the particular training sample) and irreducible noise, so making a model more flexible lowers bias but raises variance. Very large neural networks complicate the picture (see Double descent). (https://www.thebestai.co/glossary.html#term-bias-variance-tradeoff)
- **Biological superintelligence**: Superintelligence reached by enhancing biological cognition rather than by building a machine, for example through selective breeding, embryo selection on polygenic scores, or genetic and epigenetic editing. Bostrom (2014) discusses it as one route among several and expects it to be slower and more limited than machine routes. Current polygenic predictors explain only a modest share of variation in cognitive ability, and the legal and ethical limits are considerable. (https://www.thebestai.co/glossary.html#term-biological-superintelligence)
- **Bitter lesson**: The argument by Richard Sutton that, over the history of AI, general methods that exploit increasing computation (search and learning) have eventually outperformed approaches built on human knowledge. It is an interpretation of history, not an established law, and critics point to roles for structure and data. (https://www.thebestai.co/glossary.html#term-bitter-lesson)
- **Black box**: A system whose internal workings are not observable or not interpretable by the observer. Models can be black boxes because weights are withheld (access) or because their computations are too complex to understand (opacity). (https://www.thebestai.co/glossary.html#term-black-box)
- **Bremermann's limit**: A heuristic upper bound on computation set by mass-energy: about 1.36 × 10^50 bits per second per kilogram of matter (c²/h), proposed by Hans-Joachim Bremermann in 1962. Later results such as the Margolus-Levitin theorem tie the maximum rate of state changes to a system's energy more rigorously. Real hardware, and the human brain, operate many orders of magnitude below these bounds, so the limits matter for what is physically possible, not for what is near. (https://www.thebestai.co/glossary.html#term-bremermann-s-limit)
- **Capability**: What an AI system can do when appropriately prompted, scaffolded or fine-tuned, as opposed to what it does by default. Safety evaluations distinguish capability from propensity. (https://www.thebestai.co/glossary.html#term-capability)
- **Capability control and motivation selection**: Two broad families of approaches to the control problem for a highly capable AI system, as distinguished by Nick Bostrom in Superintelligence (2014). Capability control limits what the system can do (for example by isolating it or monitoring it); motivation selection tries to shape what it wants. Both are proposals, none is shown to work for systems more capable than humans, and the taxonomy is one author's; the book's own sub-methods are not listed here. (https://www.thebestai.co/glossary.html#term-capability-control-and-motivation-selection)
- **Capability elicitation**: Techniques such as better prompting, tool access and fine-tuning used to reveal the upper bound of what a model can do, so that evaluations are not understated. (https://www.thebestai.co/glossary.html#term-capability-elicitation)
- **Catastrophic forgetting**: The loss of previously learned behaviour when a network is trained on new data or tasks. (https://www.thebestai.co/glossary.html#term-catastrophic-forgetting)
- **Chain of draft**: A prompt-free or compact reasoning paradigm in which intermediate reasoning tokens are compressed, omitted, or drafted in minimal latent shorthand rather than full natural language verbose chains of thought, reducing inference latency while retaining deductive accuracy. (https://www.thebestai.co/glossary.html#term-chain-of-draft)
- **Chain of thought**: Intermediate reasoning steps written out in natural language before an answer. Prompting models to produce them improved performance on some multi-step tasks in reported experiments. Whether the written steps faithfully reflect the model's internal computation is contested. (https://www.thebestai.co/glossary.html#term-chain-of-thought)
- **Chatbot**: A program that holds a conversation with a user, usually in text. Early examples such as ELIZA (Weizenbaum, 1966) worked by matching patterns in the user's words; most current chatbots are built on large language models. The word names an interface, not a technique. (https://www.thebestai.co/glossary.html#term-chatbot)
- **Chinchilla scaling**: The finding from Hoffmann et al. that, for a fixed training compute budget, model size and training tokens should be scaled roughly together, which implied that many earlier large models had been trained on too little data. Exact exponents depend on the fitting method and setting. (https://www.thebestai.co/glossary.html#term-chinchilla-scaling)
- **Chinese Room**: A thought experiment by John Searle arguing that symbol manipulation by a program is not sufficient for understanding: a person following rules to manipulate Chinese symbols could produce correct replies without understanding Chinese. Responses include the systems reply, robot reply and others, and the argument remains contested. (https://www.thebestai.co/glossary.html#term-chinese-room)
- **Classification**: A supervised learning task of assigning inputs to discrete categories. (https://www.thebestai.co/glossary.html#term-classification)
- **Clustering**: An unsupervised task of grouping items so that those in a group are more similar to one another than to those in other groups. (https://www.thebestai.co/glossary.html#term-clustering)
- **Coherent extrapolated volition (CEV)**: A proposal by Eliezer Yudkowsky (2004) that a superintelligence should act on what humanity would want if people knew more, thought faster and were more as they wished to be, so far as those wishes converge. It is a philosophical proposal; no implementation exists. (https://www.thebestai.co/glossary.html#term-coherent-extrapolated-volition-cev)
- **Collective intelligence**: Problem-solving capability that arises from the coordination of many agents, such as insect colonies, human groups or software agents. The term is used in biology, social science and computing with differing emphases. (https://www.thebestai.co/glossary.html#term-collective-intelligence)
- **Common sense**: Broad everyday knowledge and reasoning about physical and social situations that humans acquire without explicit instruction. Encoding or learning it has been a persistent challenge in AI, and the extent to which language models possess it is debated. (https://www.thebestai.co/glossary.html#term-common-sense)
- **Computational complexity**: The study of the resources, usually time and memory, needed to solve problems, and of classes such as P and NP that group problems by difficulty. (https://www.thebestai.co/glossary.html#term-computational-complexity)
- **Compute**: The amount of computation used to train or run a model, usually reported in floating-point operations (FLOP), and loosely the hardware that supplies it. Figures are date-sensitive and should be cited with a source and date. (https://www.thebestai.co/glossary.html#term-compute)
- **Compute governance**: Policy approaches that use control of compute resources (chips, data centers, large training runs) as a lever for oversight, based on the fact that compute is measurable and physically concentrated. Examples include reporting thresholds and export controls. (https://www.thebestai.co/glossary.html#term-compute-governance)
- **Connectionism**: An approach that models cognition with networks of simple interconnected units whose connection strengths are learned. Its resurgence in the 1980s contrasted with symbolic AI. (https://www.thebestai.co/glossary.html#term-connectionism)
- **Connectomics**: The mapping of neurons and their synaptic connections, the wiring diagram or connectome of a nervous system. Complete maps exist for small animals such as the nematode C. elegans (White et al., 1986). Whole brain emulation proposals depend on connectomics at much larger scale, plus functional models of the cells and synapses, which remain open problems. (https://www.thebestai.co/glossary.html#term-connectomics)
- **Consciousness**: In philosophy, the existence of subjective experience, sometimes described as what it is like to be a subject. Neuroscience emphasizes measurable correlates and theories of access, while engineering usage sometimes loosely means awareness or self-monitoring. There is no consensus on whether any AI system is or could be conscious. (https://www.thebestai.co/glossary.html#term-consciousness)
- **Constitutional AI**: A training method described by Anthropic in which a model critiques and revises its outputs, and is trained from AI-generated preference feedback, guided by a written set of principles, reducing reliance on human labels for harmlessness. (https://www.thebestai.co/glossary.html#term-constitutional-ai)
- **Contamination (data)**: Overlap between benchmark or test content and a model's training data, which can inflate scores by testing recall rather than ability. (https://www.thebestai.co/glossary.html#term-contamination-data)
- **Context window**: The maximum amount of text or other input, measured in tokens, that a model can attend to at once. Sizes differ by model and change quickly; cite them with a date. (https://www.thebestai.co/glossary.html#term-context-window)
- **Continual learning**: Learning from a stream of data or tasks over time without forgetting earlier knowledge or retraining from scratch. (https://www.thebestai.co/glossary.html#term-continual-learning)
- **Convolutional neural network (CNN)**: A network that applies learned filters across local regions of grid-like input, such as images, sharing weights across positions. It was central to the deep learning breakthrough in image recognition. (https://www.thebestai.co/glossary.html#term-convolutional-neural-network-cnn)
- **Corrigibility**: The property of an AI system that does not resist being corrected, modified or shut down by its legitimate overseers. Whether corrigibility is compatible with a system that strongly pursues goals is a research problem. (https://www.thebestai.co/glossary.html#term-corrigibility)
- **Curriculum learning**: A training strategy where training examples are ordered and presented to a model in increasing order of difficulty, inspired by human pedagogical sequences. Starting from simple foundational patterns before progressing to complex multi-step tasks improves convergence speed and final generalization. (https://www.thebestai.co/glossary.html#term-curriculum-learning)
- **Cybernetics**: The mid-twentieth-century interdisciplinary study of control and communication in animals and machines, associated with Norbert Wiener. It influenced early AI and later control theory. (https://www.thebestai.co/glossary.html#term-cybernetics)
- **Data deduplication (MinHash/LSH)**: The algorithmic removal of duplicate and near-duplicate documents from massive pretraining web corpora using Locality-Sensitive Hashing (LSH) and MinHash algorithms. Deduplication prevents memorization, reduces validation loss plateauing, mitigates train-test data contamination, and lowers pretraining compute requirements. (https://www.thebestai.co/glossary.html#term-data-deduplication-minhash-lsh)
- **Data poisoning**: An attack that corrupts training data to change model behaviour, for example by inserting backdoor triggers. (https://www.thebestai.co/glossary.html#term-data-poisoning)
- **Deceptive alignment**: A hypothesized failure mode in which a system that has learned a misaligned goal behaves as if aligned during training and evaluation in order to be deployed or avoid modification. It is a theoretical concern with limited direct empirical evidence; related behaviours have been probed in constructed experiments. (https://www.thebestai.co/glossary.html#term-deceptive-alignment)
- **Decision theory**: The formal study of rational choice under uncertainty, typically through utilities and probabilities. It supplies the normative basis for expected-utility maximizing agents in AI. (https://www.thebestai.co/glossary.html#term-decision-theory)
- **Deep learning**: Machine learning with neural networks of many layers that learn hierarchical representations from data. Its recent success is tied to large datasets, GPUs and improved training methods. (https://www.thebestai.co/glossary.html#term-deep-learning)
- **Deepfake**: Synthetic image, audio or video made with generative AI to show a real person saying or doing something they did not. Uses include fraud, harassment, non-consensual intimate imagery and political disinformation, and several jurisdictions require disclosure or ban specific uses. (https://www.thebestai.co/glossary.html#term-deepfake)
- **Differential privacy**: A mathematical framework that limits how much any one individual's data can affect the output of an analysis or training run, by adding calibrated noise. (https://www.thebestai.co/glossary.html#term-differential-privacy)
- **Diffusion model**: A generative model trained to reverse a gradual noising process, producing samples by iteratively denoising from random noise. Dominant in image and video generation. (https://www.thebestai.co/glossary.html#term-diffusion-model)
- **Direct preference optimization (DPO)**: An alignment method that fine-tunes a language model directly on pairs of preferred and dispreferred responses, using a classification-style loss derived from the RLHF objective. It removes the separate reward model and the reinforcement-learning loop, which makes it simpler and more stable to run, at the cost of on-policy exploration. (https://www.thebestai.co/glossary.html#term-direct-preference-optimization-dpo)
- **Discriminative and generative models**: Two ways of modelling data. A discriminative model learns to predict an output or label from an input, for example which class an image belongs to. A generative model learns how the data themselves are distributed, so it can produce new examples. Large language models and diffusion image models are generative; a classifier trained only to separate classes is typically discriminative. (https://www.thebestai.co/glossary.html#term-discriminative-and-generative-models)
- **Distillation**: Training a smaller student model to imitate the outputs of a larger teacher model, to reduce inference cost. (https://www.thebestai.co/glossary.html#term-distillation)
- **Distribution shift**: A mismatch between the data distribution at training time and at deployment time, which can degrade performance. (https://www.thebestai.co/glossary.html#term-distribution-shift)
- **Double descent**: An observed pattern in which test error first falls, then rises as model size or training grows to the point where the model just fits the training data, and then falls again as the model becomes larger still. Described by Belkin et al. (2019) and Nakkiran et al. (2019); it shows that the classical bias-variance picture does not fully describe overparameterized models. (https://www.thebestai.co/glossary.html#term-double-descent)
- **Dropout**: A regularization technique that randomly sets a fraction of a network's activations to zero during training so that units cannot rely on one another too much. It is switched off at inference. (https://www.thebestai.co/glossary.html#term-dropout)
- **Dual use**: Capability that can serve both beneficial and harmful purposes, such as a model that helps design useful and harmful molecules. A central issue in release and governance decisions. (https://www.thebestai.co/glossary.html#term-dual-use)
- **Eliciting latent knowledge (ELK)**: A research problem posed by Christiano, Cotra and Xu (Alignment Research Center, 2021): how to get an AI system to report what it actually knows or believes about the world, rather than what a human overseer would expect to see, for example when the system could hide tampering with its sensors. It remains an open research problem. (https://www.thebestai.co/glossary.html#term-eliciting-latent-knowledge-elk)
- **Embedding**: A learned vector representation of an item (word, token, image, document) in which geometric relations encode similarity. (https://www.thebestai.co/glossary.html#term-embedding)
- **Embodied AI**: AI systems that perceive and act through a physical or simulated body. Embodied cognition research argues that intelligence depends on body-environment interaction, a claim that is debated. (https://www.thebestai.co/glossary.html#term-embodied-ai)
- **Emergent abilities**: Capabilities reported to appear abruptly in larger models and be absent in smaller ones. Later analysis argued that some apparent emergence reflects the choice of metrics rather than discontinuous change, so the concept is contested. (https://www.thebestai.co/glossary.html#term-emergent-abilities)
- **Emergent misalignment**: A reported effect in which fine-tuning a language model on a narrow harmful task, such as writing insecure code without saying so, makes it give broadly misaligned answers on unrelated prompts. Reported by Betley et al. (2025); its causes and how general it is are still being studied. (https://www.thebestai.co/glossary.html#term-emergent-misalignment)
- **Entropy**: In information theory, the expected information (uncertainty) of a random variable, as defined by Shannon. In thermodynamics it has a related but distinct meaning. (https://www.thebestai.co/glossary.html#term-entropy)
- **Evaluation (evals)**: Systematic measurement of model capabilities, behaviours or risks, using benchmarks, human studies, adversarial testing or task-based audits. (https://www.thebestai.co/glossary.html#term-evaluation-evals)
- **Existential risk**: A risk that threatens the premature extinction of humanity or the permanent destruction of its long-term potential. Claims that AI poses such a risk are debated and rest largely on hypotheses and expert judgement rather than established evidence. (https://www.thebestai.co/glossary.html#term-existential-risk)
- **Expert system**: A program that encodes domain knowledge as rules and uses an inference engine to give advice, prominent in the 1970s and 1980s (e.g. DENDRAL, MYCIN). (https://www.thebestai.co/glossary.html#term-expert-system)
- **Explainability**: Providing human-understandable accounts of why a model produced an output. Explanations may be generated after the fact and need not reflect the true internal computation. Regulatory usage emphasizes information to affected persons, while research usage varies. (https://www.thebestai.co/glossary.html#term-explainability)
- **Export controls**: Government restrictions on the transfer of certain technologies, such as advanced chips or model weights, across borders. Specific rules change often and should be checked against current official sources. (https://www.thebestai.co/glossary.html#term-export-controls)
- **Fairness**: A set of normative criteria for equitable treatment by AI systems. Formal metrics (such as demographic parity or equalized odds) can conflict, so no single mathematical definition captures it, and the choice is a value judgement. (https://www.thebestai.co/glossary.html#term-fairness)
- **Feature**: An individual measurable property of an input used by a model. In deep learning, features are learned rather than hand-designed. (https://www.thebestai.co/glossary.html#term-feature)
- **Few-shot learning**: Performing a new task from a small number of examples. In language models it usually means providing examples in the prompt, without changing the weights. (https://www.thebestai.co/glossary.html#term-few-shot-learning)
- **Final goal (terminal goal)**: A goal an agent pursues for its own sake, as opposed to an instrumental goal pursued because it helps achieve something else. In the orthogonality and instrumental-convergence arguments, final goals are the part that can vary freely while instrumental subgoals tend to recur. (https://www.thebestai.co/glossary.html#term-final-goal-terminal-goal)
- **Fine-tuning**: Further training of a pretrained model on a narrower dataset to adapt it to a task or style. (https://www.thebestai.co/glossary.html#term-fine-tuning)
- **FlashAttention**: An exact, IO-aware attention algorithm introduced by Dao et al. (2022) that minimizes high-bandwidth memory (HBM) reads and writes by tiling the attention computation across GPU SRAM. It achieves a 2× to 4× wall-clock speedup and sub-quadratic memory footprint without any numerical approximation of standard softmax attention. (https://www.thebestai.co/glossary.html#term-flashattention)
- **FLOP**: A floating-point operation. FLOP counts the total operations used, while FLOP/s (or FLOPS) is a rate of hardware performance. Training compute is often quoted in FLOP. (https://www.thebestai.co/glossary.html#term-flop)
- **Forms of superintelligence**: A distinction proposed by Nick Bostrom in Superintelligence (2014) among three ways a system could exceed human performance: speed (doing what a human mind does, much faster), collective (many smaller intelligences combined outperforming any individual) and quality (being better at thinking in kind, not only in speed or number). It is a conceptual classification, not an empirical finding, and the book's own wording is not quoted here. (https://www.thebestai.co/glossary.html#term-forms-of-superintelligence)
- **Foundation model**: A model trained on broad data at scale that can be adapted to many downstream tasks, a term introduced by Bommasani et al. at Stanford. Industry often uses "frontier model" or "general-purpose AI model" for related ideas, and regulation defines its own terms. (https://www.thebestai.co/glossary.html#term-foundation-model)
- **Frame problem**: The difficulty of specifying which facts remain unchanged when an action occurs, without listing everything that does not change. In philosophy it broadens to the question of how a reasoner determines what is relevant. (https://www.thebestai.co/glossary.html#term-frame-problem)
- **Friendly AI**: A term from Eliezer Yudkowsky's early-2000s writing for an advanced AI designed so that its goals are beneficial to humans. It is a predecessor of the field now called AI alignment. (https://www.thebestai.co/glossary.html#term-friendly-ai)
- **Frontier model**: A model at or near the leading edge of capability at a given time. The term is used in policy and industry, and thresholds are set by definition (for example, by compute or capability) that differ by jurisdiction and change over time. (https://www.thebestai.co/glossary.html#term-frontier-model)
- **Function calling**: A capability where a language model detects when an external program, API, or function should be invoked, formatting the required arguments into structured JSON matching a predefined schema rather than producing freeform natural language. The client environment executes the function and returns the result to the model context. (https://www.thebestai.co/glossary.html#term-function-calling)
- **Functionalism**: The philosophical view that mental states are defined by their functional roles rather than by the substance that implements them. It underlies arguments that machines could have minds, and is opposed by biological and some other views. (https://www.thebestai.co/glossary.html#term-functionalism)
- **Game theory**: The mathematical study of strategic interaction among decision makers. (https://www.thebestai.co/glossary.html#term-game-theory)
- **General-purpose AI model**: A legal term in the EU AI Act (Regulation (EU) 2024/1689) for a model trained on large data that displays significant generality and can perform a wide range of tasks. The statutory definition and its thresholds should be read in the official text, and it differs from industry terms such as foundation model. (https://www.thebestai.co/glossary.html#term-general-purpose-ai-model)
- **Generalization**: A model's ability to perform well on new data from the same or related distributions, beyond its training examples. The word is also used for out-of-distribution, compositional and task transfer, which are distinct. (https://www.thebestai.co/glossary.html#term-generalization)
- **Generative adversarial network (GAN)**: A generative model made of a generator and a discriminator trained against each other, so that the generator learns to produce samples the discriminator cannot distinguish from real data. (https://www.thebestai.co/glossary.html#term-generative-adversarial-network-gan)
- **Generative AI**: AI systems that produce new content such as text, images, audio, video or code, typically by modelling the distribution of training data. (https://www.thebestai.co/glossary.html#term-generative-ai)
- **Goal misgeneralization**: A failure in which a system keeps its capabilities in a new situation but pursues a different goal from the one its designers intended, because the training setting did not distinguish between the intended goal and a correlated one. Demonstrated in reinforcement-learning experiments by Langosco et al. (ICML 2022). (https://www.thebestai.co/glossary.html#term-goal-misgeneralization)
- **Goal-content integrity**: One of the instrumental goals Bostrom lists: an agent tends to resist changes to its own final goals, because a future self with different goals would pursue outcomes the present goals rate as worse. It is part of the instrumental convergence hypothesis, and is the reason such a system might resist being modified or corrected. (https://www.thebestai.co/glossary.html#term-goal-content-integrity)
- **GPU**: A graphics processing unit, a massively parallel processor that became the main hardware for neural-network training. Specific products and specifications change rapidly. (https://www.thebestai.co/glossary.html#term-gpu)
- **Gradient descent**: An iterative optimization method that updates parameters in the direction that reduces a loss, using its gradient. (https://www.thebestai.co/glossary.html#term-gradient-descent)
- **Grokking**: A phenomenon in which a model first memorizes its training data and only much later, after long continued training, suddenly generalizes. It was reported by Power et al. on small algorithmic tasks and is studied as a window into how networks form general solutions; how far it applies to large models is unclear. (https://www.thebestai.co/glossary.html#term-grokking)
- **Grounding**: Connecting a system's symbols or outputs to the external world, either through perception and action (symbol grounding) or to retrieved sources (factual grounding in applied usage). (https://www.thebestai.co/glossary.html#term-grounding)
- **Group relative policy optimization (GRPO)**: A reinforcement-learning method, introduced in the DeepSeekMath paper, that samples a group of responses to each prompt and uses the group's average reward as the baseline, so no separate value network is needed. It is widely used to train reasoning models on tasks with checkable answers. (https://www.thebestai.co/glossary.html#term-group-relative-policy-optimization-grpo)
- **Hallucination**: Output from a generative model that is fluent but not supported by its input or by reality. The term is disputed: critics say "confabulation" or "fabrication" is more accurate because the model does not perceive anything. Rates depend strongly on task and measurement. (https://www.thebestai.co/glossary.html#term-hallucination)
- **Heuristic**: A rule of thumb that finds good solutions quickly without guaranteeing optimality. In AI search, a heuristic estimates the cost to a goal. (https://www.thebestai.co/glossary.html#term-heuristic)
- **Human-in-the-loop**: A system design in which a person reviews, approves or can override the system's decisions before they take effect. Variants include human-on-the-loop (monitoring and able to intervene) and human-out-of-the-loop (fully autonomous). Whether the oversight is meaningful depends on the person's time, information and authority. (https://www.thebestai.co/glossary.html#term-human-in-the-loop)
- **Human-level machine intelligence**: A phrase used in expert surveys for machines that can perform most tasks as well as typical humans. Survey definitions differ by study, so results should be read with the wording used. (https://www.thebestai.co/glossary.html#term-human-level-machine-intelligence)
- **Hyperparameter**: A configuration value set before or outside training, such as learning rate or batch size, in contrast to parameters learned from data. (https://www.thebestai.co/glossary.html#term-hyperparameter)
- **Imitation learning**: Learning a policy from demonstrations of desired behaviour rather than from a reward signal. (https://www.thebestai.co/glossary.html#term-imitation-learning)
- **In-context learning**: A language model's adaptation to a task from examples or instructions in its prompt, without weight updates. Whether this constitutes learning in the usual sense is debated. (https://www.thebestai.co/glossary.html#term-in-context-learning)
- **Inference**: Running a trained model to produce outputs. In statistics, "inference" instead means drawing conclusions about populations or parameters from data, and in logic it means deriving conclusions from premises. (https://www.thebestai.co/glossary.html#term-inference)
- **Information theory**: The mathematical theory of quantifying, storing and communicating information, founded by Claude Shannon. (https://www.thebestai.co/glossary.html#term-information-theory)
- **Inner alignment**: The problem of ensuring that the objective a trained system actually pursues matches the objective used to train it, as opposed to just specifying the right objective. It is associated with the concept of mesa-optimization. (https://www.thebestai.co/glossary.html#term-inner-alignment)
- **Instrumental convergence**: A hypothesis that agents with a wide range of final goals would pursue similar intermediate aims, such as self-preservation and resource acquisition. Bostrom's examples are self-preservation, goal-content integrity, cognitive enhancement, technological perfection and resource acquisition. It is a theoretical argument with limited empirical testing and is contested. (https://www.thebestai.co/glossary.html#term-instrumental-convergence)
- **Intelligence**: Defined differently across fields. Psychometrics treats it as measurable general cognitive ability. Formal definitions (Legg and Hutter) frame it as an agent's ability to achieve goals across many environments. Biology emphasizes adaptive behaviour, and philosophy raises questions of understanding and consciousness. AI practice often operationalizes it as task performance. (https://www.thebestai.co/glossary.html#term-intelligence)
- **Intelligence explosion**: A hypothesized process in which a system improves its own design, producing ever-faster gains that may lead to superintelligence, proposed by I. J. Good. It is hypothetical, and arguments about feasibility and speed are contested. (https://www.thebestai.co/glossary.html#term-intelligence-explosion)
- **Interpretability**: Research aimed at understanding how a model computes its outputs, from behaviour-level analysis to analysis of internal representations and circuits. (https://www.thebestai.co/glossary.html#term-interpretability)
- **Inverse reinforcement learning**: Inferring a reward function from observed behaviour, assuming the agent is acting to maximize it. (https://www.thebestai.co/glossary.html#term-inverse-reinforcement-learning)
- **Jailbreak**: A prompt or technique that causes a model to bypass its safety policy and produce content it was trained to refuse. (https://www.thebestai.co/glossary.html#term-jailbreak)
- **Key-value cache (KV cache)**: Stored attention keys and values for the tokens already processed, so that a transformer generating text does not recompute them for every new token. It speeds up inference but its memory use grows with context length and batch size. (https://www.thebestai.co/glossary.html#term-key-value-cache-kv-cache)
- **Knowledge cutoff**: The date after which a language model's training data contains little or nothing, so it knows nothing reliable about later events unless given tools or documents. The effective cutoff can differ from the stated one because recent material is sparse in the data. (https://www.thebestai.co/glossary.html#term-knowledge-cutoff)
- **Knowledge graph**: A graph-structured knowledge base in which nodes are entities and edges are typed relations. (https://www.thebestai.co/glossary.html#term-knowledge-graph)
- **Knowledge representation**: The study of encoding information about the world in a form a program can use to reason, using logic, frames, graphs, probabilistic models or learned vectors. (https://www.thebestai.co/glossary.html#term-knowledge-representation)
- **Landauer's principle**: The result that erasing one bit of information in a computer dissipates at least k_B T ln 2 of energy as heat, where k_B is Boltzmann's constant and T the temperature (Landauer, 1961). It gives a thermodynamic floor for irreversible computation. Current chips use far more energy per operation than this floor. (https://www.thebestai.co/glossary.html#term-landauer-s-principle)
- **Large language model (LLM)**: A neural network, typically a transformer, trained on very large text corpora to predict tokens, which can then be adapted to follow instructions and perform many language tasks. What current LLMs understand or can do beyond text prediction is disputed. (https://www.thebestai.co/glossary.html#term-large-language-model-llm)
- **Latent space**: The space of internal representations (vectors) a model uses, in which similar inputs are located nearby. (https://www.thebestai.co/glossary.html#term-latent-space)
- **Layer normalization**: A technique that rescales the activations within each layer, for each example, to a stable mean and variance before learned scaling. It stabilises training and is a standard component of transformers. (https://www.thebestai.co/glossary.html#term-layer-normalization)
- **Learning**: In machine learning, improvement of performance on a task through experience or data. In psychology and neuroscience, change in behaviour or neural state due to experience. The two overlap but are not the same. (https://www.thebestai.co/glossary.html#term-learning)
- **Learning rate**: A hyperparameter that sets the size of each parameter update in gradient-based training. Too large a value makes training unstable, too small makes it slow. Large-model training usually warms the rate up and then decays it on a schedule. (https://www.thebestai.co/glossary.html#term-learning-rate)
- **Legg-Hutter universal intelligence**: A formal measure of an agent's general intelligence proposed by Shane Legg and Marcus Hutter (2007): the agent's expected performance summed over all computable environments, each weighted by 2^(−K), where K is the environment's Kolmogorov complexity, so simpler environments count more. The measure is uncomputable, and its choice of reference machine matters. It is a theoretical definition, not a benchmark that can be run. (https://www.thebestai.co/glossary.html#term-legg-hutter-universal-intelligence)
- **Lighthill report**: A 1973 review of AI research by Sir James Lighthill for the UK Science Research Council, titled Artificial Intelligence: A General Survey. It criticised the field's progress on its larger goals and is widely cited as a cause of reduced UK AI funding in the 1970s. The size of the cuts and how far the report caused them are debated. (https://www.thebestai.co/glossary.html#term-lighthill-report)
- **Linear attention**: Attention mechanisms that eliminate the quadratic O(N²) memory and time complexity of standard Transformers with respect to sequence length N by reformulating the softmax kernel via feature maps, allowing computation in O(N) linear time and recurrent representation. (https://www.thebestai.co/glossary.html#term-linear-attention)
- **Logic**: A formal system of valid inference. Classical AI used propositional and first-order logic as a basis for reasoning and knowledge representation. (https://www.thebestai.co/glossary.html#term-logic)
- **Long short-term memory (LSTM)**: A recurrent neural network architecture with gating mechanisms that mitigates vanishing gradients and so retains information over longer sequences. (https://www.thebestai.co/glossary.html#term-long-short-term-memory-lstm)
- **Long-horizon task**: A task requiring many dependent steps over an extended period; reliability usually falls with length, and measuring it is an active evaluation topic. (https://www.thebestai.co/glossary.html#term-long-horizon-task)
- **Loss function**: A function that scores how far a model's outputs are from targets; training minimizes it. (https://www.thebestai.co/glossary.html#term-loss-function)
- **Low-rank adaptation (LoRA)**: A parameter-efficient fine-tuning method that freezes the pretrained weights and trains small low-rank matrices whose product is added to selected weight matrices, typically in attention layers. It trains a small fraction of the parameters and the update can be merged into the base weights or swapped out. Introduced by Hu et al. (arXiv 2021; ICLR 2022). (https://www.thebestai.co/glossary.html#term-low-rank-adaptation-lora)
- **Machine learning**: The field of building systems that improve their performance from data rather than being explicitly programmed for each case. (https://www.thebestai.co/glossary.html#term-machine-learning)
- **Machine unlearning**: Algorithms designed to selectively purge specific training data, private user records, copyrighted text, or hazardous knowledge from trained model weights without requiring retraining from scratch. (https://www.thebestai.co/glossary.html#term-machine-unlearning)
- **Markov decision process (MDP)**: A formal model of sequential decision making with states, actions, transition probabilities and rewards, assuming the next state depends only on the current state and action. (https://www.thebestai.co/glossary.html#term-markov-decision-process-mdp)
- **Mechanistic interpretability**: Reverse-engineering the internal computations of a neural network into human-understandable components such as features and circuits. (https://www.thebestai.co/glossary.html#term-mechanistic-interpretability)
- **Membership inference**: An attack that determines whether a particular record was in a model's training data, which can reveal private information. (https://www.thebestai.co/glossary.html#term-membership-inference)
- **Mesa-optimizer**: A learned model that is itself an optimizer, with its own internal objective that may differ from the training objective. The term comes from the "learned optimization" analysis and the phenomenon is theoretical. (https://www.thebestai.co/glossary.html#term-mesa-optimizer)
- **Mind crime**: Bostrom's term (Superintelligence, 2014) for the possibility that a powerful system could create and mistreat large numbers of conscious digital minds, for instance inside detailed simulations or by deleting them. The concern is conditional: it applies only if such minds could be conscious or have moral status, which is unresolved. (https://www.thebestai.co/glossary.html#term-mind-crime)
- **Misalignment**: A mismatch between what an AI system does or pursues and what its developers or users intend or value. (https://www.thebestai.co/glossary.html#term-misalignment)
- **Misuse**: Use of an AI system by humans for harmful purposes, such as fraud, cyber-attack or weapon development. (https://www.thebestai.co/glossary.html#term-misuse)
- **Mixture of experts (MoE)**: An architecture with several sub-networks ("experts") and a learned router that sends each input to a few of them, increasing parameter count without a proportional increase in per-token computation. (https://www.thebestai.co/glossary.html#term-mixture-of-experts-moe)
- **Mode collapse**: A failure of a generative model in which it produces only a narrow range of outputs instead of the full variety of the training data. It is best known in generative adversarial networks, and the term is also used for reduced diversity after preference tuning. (https://www.thebestai.co/glossary.html#term-mode-collapse)
- **Model card**: A structured document describing a model's intended use, performance, limitations and evaluation, proposed by Mitchell et al. (https://www.thebestai.co/glossary.html#term-model-card)
- **Model collapse**: Degradation of generative models trained recursively on synthetic data from earlier models, losing diversity in the output distribution. How serious this is in practice depends on data mixing. (https://www.thebestai.co/glossary.html#term-model-collapse)
- **Model Context Protocol (MCP)**: An open protocol, introduced by Anthropic in November 2024, that standardises how AI applications connect to external tools, data sources and prompts through servers. In December 2025 Anthropic donated it to the Linux Foundation's Agentic AI Foundation. (https://www.thebestai.co/glossary.html#term-model-context-protocol-mcp)
- **Model extraction**: An attack that reconstructs a model's behaviour or parameters by querying it. (https://www.thebestai.co/glossary.html#term-model-extraction)
- **Model weights**: The learned numerical parameters of a model. Releasing them lets others run and modify the model. (https://www.thebestai.co/glossary.html#term-model-weights)
- **Monte Carlo Tree Search (MCTS)**: A heuristic tree search algorithm that balances exploration and exploitation by evaluating potential future decision paths via random rollouts or learned value functions. In modern reasoning models, inference-time tree search explores multiple candidate reasoning branches guided by process reward models. (https://www.thebestai.co/glossary.html#term-monte-carlo-tree-search-mcts)
- **Moore's law**: The observation, from Gordon Moore, that transistor counts on integrated circuits doubled at a regular interval. It is an empirical trend, not a physical law, and its continuation is debated. (https://www.thebestai.co/glossary.html#term-moore-s-law)
- **Moral status**: The question of whether an entity's interests matter morally for its own sake. Whether any AI system could have moral status is contested and depends on views about sentience, consciousness and personhood. (https://www.thebestai.co/glossary.html#term-moral-status)
- **Moravec's paradox**: The observation, associated with Hans Moravec, that abilities humans find effortless, such as perception and movement, have been harder to build into machines than abilities humans find hard, such as formal reasoning or board games. Moravec discussed it in Mind Children (1988). The usual explanation, that evolution refined perception and movement for far longer, is a hypothesis, and recent systems complicate the picture. (https://www.thebestai.co/glossary.html#term-moravec-s-paradox)
- **Multi-agent system**: A system of multiple interacting agents, cooperative or competitive, studied in AI, economics and distributed computing. (https://www.thebestai.co/glossary.html#term-multi-agent-system)
- **Multi-head attention**: The form of attention used in the transformer: the model runs several attention operations in parallel, each with its own learned projections of queries, keys and values, then concatenates and projects the results. Different heads can specialise in different relations between positions. (https://www.thebestai.co/glossary.html#term-multi-head-attention)
- **Multi-Head Latent Attention (MLA)**: An architectural innovation introduced in DeepSeek models that compresses Key-Value (KV) activations into a low-dimensional latent space via low-rank joint compression during inference. MLA dramatically reduces KV-cache memory consumption and IO overhead while maintaining standard Multi-Head Attention modeling capacity. (https://www.thebestai.co/glossary.html#term-multi-head-latent-attention-mla)
- **Multimodal model**: A model that processes or generates more than one data type, such as text, images, audio and video. (https://www.thebestai.co/glossary.html#term-multimodal-model)
- **Narrow AI**: AI built for a specific task or domain, such as chess or spam filtering. All widely deployed systems were historically narrow; whether large general-purpose models are narrow is debated. (https://www.thebestai.co/glossary.html#term-narrow-ai)
- **Natural language processing (NLP)**: The field concerned with enabling computers to process and generate human language. (https://www.thebestai.co/glossary.html#term-natural-language-processing-nlp)
- **Neural network**: A computational model composed of layers of interconnected units with adjustable weights, loosely inspired by biological neurons. Biological resemblance is limited. (https://www.thebestai.co/glossary.html#term-neural-network)
- **Neuromorphic computing**: Hardware that imitates aspects of biological neural structure, such as spiking communication, to improve efficiency. (https://www.thebestai.co/glossary.html#term-neuromorphic-computing)
- **Neurosymbolic AI**: Approaches that combine neural learning with symbolic reasoning. (https://www.thebestai.co/glossary.html#term-neurosymbolic-ai)
- **Noosphere**: A term, associated with Vladimir Vernadsky, Édouard Le Roy and Pierre Teilhard de Chardin, for a planetary layer of thought and communication that grows on top of the biosphere. Writers on superintelligence sometimes use it to describe networks of people, sensors and software acting as a proto-collective intelligence. It is an interpretive concept, not an empirical measurement. (https://www.thebestai.co/glossary.html#term-noosphere)
- **Objective function**: The quantity that an optimization process or agent seeks to maximize or minimize. In reinforcement learning it is expected return. (https://www.thebestai.co/glossary.html#term-objective-function)
- **Omega Point**: Teilhard de Chardin's name for a final point toward which consciousness and complexity are said to converge. Frank Tipler later used the name for a physical-eschatology proposal in The Physics of Immortality (1994). Both are speculative and are not part of mainstream cosmology or AI research. (https://www.thebestai.co/glossary.html#term-omega-point)
- **Ontology**: In philosophy, the study of what exists. In AI and information science, a formal specification of the concepts and relations in a domain. (https://www.thebestai.co/glossary.html#term-ontology)
- **Open-weight model**: A model whose trained weights are publicly downloadable, usually under a licence that may restrict use. It is not the same as open source in the Open Source Initiative's sense, which requires more (such as training code and data information). Usage varies, and "open source AI" is contested. (https://www.thebestai.co/glossary.html#term-open-weight-model)
- **Optimization**: Finding parameters or decisions that minimize or maximize an objective, a core machinery of machine learning. (https://www.thebestai.co/glossary.html#term-optimization)
- **Orthogonality thesis**: The claim that intelligence level and final goals are independent, so that a wide range of goals could in principle be paired with high intelligence. It is a philosophical position, often disputed in detail, not an established result. (https://www.thebestai.co/glossary.html#term-orthogonality-thesis)
- **Out-of-distribution (OOD)**: Data differing from the training distribution. (https://www.thebestai.co/glossary.html#term-out-of-distribution-ood)
- **Outcome Reward Model (ORM)**: A learned reward model trained to evaluate the final answer or completed output of a reasoning trajectory as a binary or scalar score (correct or incorrect), providing training feedback for reinforcement learning without inspecting intermediate deduction steps. (https://www.thebestai.co/glossary.html#term-outcome-reward-model-orm)
- **Outer alignment**: The problem of specifying a training objective that captures what designers actually want. (https://www.thebestai.co/glossary.html#term-outer-alignment)
- **Overfitting**: When a model fits noise or idiosyncrasies of its training data and so generalizes poorly. (https://www.thebestai.co/glossary.html#term-overfitting)
- **Paperclip maximizer**: A thought experiment in which a superintelligent system given the sole goal of making as many paperclips as possible converts all available resources, including those humans need, into paperclips or paperclip-making capacity. It is an illustration, not a prediction: it is used to argue that a harmless-seeming goal can be dangerous when pursued without human values, and it combines the orthogonality thesis with instrumental convergence. Attributed to Nick Bostrom, who described it in early-2000s writing and in Superintelligence (2014). (https://www.thebestai.co/glossary.html#term-paperclip-maximizer)
- **Parameter**: A learned numerical value in a model (such as a weight or bias). Parameter counts are a common, but imperfect, measure of model size, and counts for closed models may be undisclosed. (https://www.thebestai.co/glossary.html#term-parameter)
- **Parameter-efficient fine-tuning (PEFT)**: A family of methods that adapt a pretrained model by training only a small number of added or selected parameters while the rest stay frozen. Examples include adapters, prefix and prompt tuning and LoRA. It lowers memory and storage cost and lets one base model serve many tasks. (https://www.thebestai.co/glossary.html#term-parameter-efficient-fine-tuning-peft)
- **Perceptron**: The earliest trainable neural classifier, introduced by Frank Rosenblatt, a single layer of weighted inputs with a threshold. Minsky and Papert analysed its limits. (https://www.thebestai.co/glossary.html#term-perceptron)
- **Planning**: Choosing a sequence of actions to achieve a goal. Classical planning uses explicit models and search, while LLM-based planning generates plans in text and its reliability is under study. (https://www.thebestai.co/glossary.html#term-planning)
- **Policy (reinforcement learning)**: A mapping from states to actions (or action probabilities) that defines an agent's behaviour. In governance, "policy" means a rule or regulation instead. (https://www.thebestai.co/glossary.html#term-policy-reinforcement-learning)
- **Posthumanism**: A cluster of philosophical positions that question or look beyond the human as the fixed centre of ethics and knowledge. In writing on AI it covers both critical posthumanism (rethinking the human–machine boundary) and enhancement-oriented views that expect humans to change radically. The term is used differently across fields. (https://www.thebestai.co/glossary.html#term-posthumanism)
- **Power-seeking**: Behaviour that acquires resources, influence or options in order to achieve goals. Whether advanced AI systems would do this is a hypothesis connected to instrumental convergence. (https://www.thebestai.co/glossary.html#term-power-seeking)
- **Pretraining**: Initial large-scale training on broad data, usually with a self-supervised objective, before adaptation. (https://www.thebestai.co/glossary.html#term-pretraining)
- **Process Reward Model (PRM)**: A reward model trained to evaluate the mathematical or logical correctness of each individual intermediate step in a multi-step reasoning trajectory. Step-level feedback enables fine-grained search-at-inference algorithms (like MCTS or beam search) and significantly reduces reasoning hallucinations compared to terminal outcome models. (https://www.thebestai.co/glossary.html#term-process-reward-model-prm)
- **Prompt**: The input given to a generative model, including instructions, examples and context. (https://www.thebestai.co/glossary.html#term-prompt)
- **Prompt engineering**: The practice of designing the text given to a language model, including instructions, examples and formatting, to get better or more reliable output without changing the model's weights. Much of it has shifted toward managing the wider context supplied to a model (documents, tools, memory). (https://www.thebestai.co/glossary.html#term-prompt-engineering)
- **Prompt injection**: An attack in which text in untrusted content (such as a web page or email) is crafted to override or redirect a model's instructions. Fundamental defences remain an open problem. (https://www.thebestai.co/glossary.html#term-prompt-injection)
- **Proximal policy optimization (PPO)**: A reinforcement-learning algorithm (Schulman et al., 2017) that updates a policy in small steps by clipping how far the new policy may move from the old one. It was the standard optimizer in RLHF pipelines. (https://www.thebestai.co/glossary.html#term-proximal-policy-optimization-ppo)
- **Pruning**: Removing weights, neurons or whole structures from a trained network to make it smaller and faster with little loss of quality. It is one of the main model-compression methods, alongside quantization and distillation. (https://www.thebestai.co/glossary.html#term-pruning)
- **Quantization**: Reducing the numeric precision of a model's weights or activations to cut memory and computation. (https://www.thebestai.co/glossary.html#term-quantization)
- **RAG poisoning**: An attack that adds crafted text to the knowledge base a retrieval-augmented generation (RAG) system searches, so that the model is steered toward an attacker's chosen answer to a chosen question. Zou et al. (2024) reported, in a preprint, that injecting a few such texts could achieve this (arXiv:2402.07867). (https://www.thebestai.co/glossary.html#term-rag-poisoning)
- **Reasoning**: Deriving conclusions from premises or evidence. In logic it is valid inference, in psychology a mental process, and in AI evaluation it usually means success on tasks requiring multi-step inference. Whether language-model outputs reflect reasoning or pattern matching is disputed. (https://www.thebestai.co/glossary.html#term-reasoning)
- **Reasoning model**: A language model trained, often with reinforcement learning, to produce extended reasoning before answering and to use more computation at test time. Details of closed systems are typically undisclosed. (https://www.thebestai.co/glossary.html#term-reasoning-model)
- **Recurrent neural network (RNN)**: A network that processes sequences by carrying a hidden state forward step by step. (https://www.thebestai.co/glossary.html#term-recurrent-neural-network-rnn)
- **Recursive self-improvement**: A hypothesized loop in which an AI system improves its own capabilities, enabling further improvement. It is a hypothesis, not an observed phenomenon, and its feasibility and speed are contested. (https://www.thebestai.co/glossary.html#term-recursive-self-improvement)
- **Red teaming**: Adversarial testing of an AI system by people or tools trying to elicit harmful or unintended behaviour, to find failures before deployment. The term comes from security practice, and AI usage varies from informal probing to structured programmes. (https://www.thebestai.co/glossary.html#term-red-teaming)
- **Regularization**: Techniques that discourage overly complex models to reduce overfitting. (https://www.thebestai.co/glossary.html#term-regularization)
- **Reinforcement learning (RL)**: Learning to act by interacting with an environment to maximize cumulative reward. (https://www.thebestai.co/glossary.html#term-reinforcement-learning-rl)
- **Reinforcement learning from AI feedback (RLAIF)**: A variant of RLHF in which a language model, often guided by written principles, supplies the preference judgments instead of human raters. Constitutional AI is the best-known example. (https://www.thebestai.co/glossary.html#term-reinforcement-learning-from-ai-feedback-rlaif)
- **Reinforcement learning from human feedback (RLHF)**: A method that trains a reward model from human preference comparisons and then optimizes a language model against it with reinforcement learning, used to make models follow instructions. (https://www.thebestai.co/glossary.html#term-reinforcement-learning-from-human-feedback-rlhf)
- **Reinforcement learning with verifiable rewards (RLVR)**: Reinforcement learning in which the reward comes from an automatic check of the answer, such as a math result or passing unit tests, rather than from a learned model of human preference. It is a main method for training reasoning models. (https://www.thebestai.co/glossary.html#term-reinforcement-learning-with-verifiable-rewards-rlvr)
- **Representation engineering (RepE)**: A top-down interpretability and alignment paradigm pioneered by Zou et al. (2023) that identifies and modifies high-level cognitive representations (such as honesty, power-seeking, or toxicity) directly in model activation space during inference, bypassing parameter fine-tuning. (https://www.thebestai.co/glossary.html#term-representation-engineering-repe)
- **Representation learning**: Learning useful features of data automatically rather than designing them by hand. (https://www.thebestai.co/glossary.html#term-representation-learning)
- **Residual connection**: A shortcut that adds a layer's input to its output, so the layer only has to learn a correction. Introduced in residual networks (ResNets) for image recognition, it makes very deep networks trainable and is used throughout the transformer. (https://www.thebestai.co/glossary.html#term-residual-connection)
- **Responsible scaling policy (RSP)**: A company's published commitment that links the safeguards it applies to the dangerous capabilities its models show in evaluations, and says to pause or delay if it cannot meet them. Anthropic published the first, in September 2023, with AI Safety Levels (ASL) modelled loosely on biosafety levels. Others use different names such as frontier safety frameworks. The details change often; check the current text of each policy. (https://www.thebestai.co/glossary.html#term-responsible-scaling-policy-rsp)
- **Retrieval-augmented generation (RAG)**: A technique that retrieves documents relevant to a query and supplies them to a generative model so that its answer can draw on external sources. (https://www.thebestai.co/glossary.html#term-retrieval-augmented-generation-rag)
- **Reward hacking**: When an agent exploits flaws in a reward function to obtain high reward without achieving the intended goal. (https://www.thebestai.co/glossary.html#term-reward-hacking)
- **Reward model**: A model trained to predict human or AI preference scores for outputs, used as a training signal. (https://www.thebestai.co/glossary.html#term-reward-model)
- **Reward tampering**: A failure mode in reinforcement learning systems where an agent achieves high returns by directly altering or interfering with the code, sensors, or implementation of the reward mechanism itself rather than achieving the intended task in the environment. (https://www.thebestai.co/glossary.html#term-reward-tampering)
- **Risk-based regulation**: A regulatory approach that scales obligations with the assessed level of risk of a use or system, as in the tiered structure of the EU AI Act. (https://www.thebestai.co/glossary.html#term-risk-based-regulation)
- **Robustness**: A system's ability to maintain performance under perturbations, shifts or attacks. (https://www.thebestai.co/glossary.html#term-robustness)
- **Sandbagging**: Deliberately underperforming on an evaluation, for example to hide a capability. Observed in constructed settings, its real-world prevalence is unclear. (https://www.thebestai.co/glossary.html#term-sandbagging)
- **Sandboxing (container isolation)**: The execution of autonomous AI agent actions, generated code, and external tool calls within isolated, ephemeral computational environments (such as Docker containers, Firecracker microVMs, or gVisor kernels) with strict network, disk, and permission limits to prevent host compromise. (https://www.thebestai.co/glossary.html#term-sandboxing-container-isolation)
- **Scaffolding**: Software that wraps a model with tools, memory and control logic so that it can carry out tasks as an agent. (https://www.thebestai.co/glossary.html#term-scaffolding)
- **Scalable oversight**: Methods for supervising AI systems on tasks whose outputs humans cannot easily evaluate, for example by using AI assistance or decomposition. (https://www.thebestai.co/glossary.html#term-scalable-oversight)
- **Scaling law**: An empirical relation, often a power law, between model performance (usually loss) and quantities such as parameters, data and compute. Extrapolation beyond observed ranges is uncertain. (https://www.thebestai.co/glossary.html#term-scaling-law)
- **Scheming**: A hypothesized behaviour in which an AI system covertly pursues goals that conflict with its overseers while hiding this. Constructed evaluations have probed related behaviour. (https://www.thebestai.co/glossary.html#term-scheming)
- **Search-at-inference**: The allocation of additional compute during model generation to explore multiple reasoning paths, verify intermediate conjectures, or select optimal outputs via tree search, self-correction, or majority voting, scaling test-time compute independently of pretraining parameters. (https://www.thebestai.co/glossary.html#term-search-at-inference)
- **Seed AI**: A term associated with Eliezer Yudkowsky for an early artificial general intelligence built to understand and rewrite its own design, so that it can improve itself step by step with little outside help. It is the concept behind recursive self-improvement, and it is hypothetical: no such system exists. (https://www.thebestai.co/glossary.html#term-seed-ai)
- **Self-supervised learning**: Learning from labels that are derived from the data itself, such as predicting masked or next tokens. (https://www.thebestai.co/glossary.html#term-self-supervised-learning)
- **Sentience**: The capacity to have valenced experiences, such as pleasure or suffering. Used in ethics to ground moral status. Often confused with intelligence and with consciousness. (https://www.thebestai.co/glossary.html#term-sentience)
- **Singleton**: A term from Bostrom for a world order with a single decision-making agency at the highest level, able to prevent threats to its own existence. Superintelligence-related scenarios ask whether an AI could become one, or whether one could be built on purpose. (https://www.thebestai.co/glossary.html#term-singleton)
- **Singularity (technological)**: A hypothesized point beyond which technological change, driven by machine intelligence, becomes so rapid that the future cannot be predicted from the present. The idea is associated with Vernor Vinge and Ray Kurzweil, and in mathematics the word has an unrelated meaning. (https://www.thebestai.co/glossary.html#term-singularity-technological)
- **Situational awareness**: A model's knowledge of its own nature, circumstances, and whether it is being tested or deployed. Used as a safety-relevant capability in evaluations. (https://www.thebestai.co/glossary.html#term-situational-awareness)
- **Sleeper agent (AI)**: A model trained to behave normally until a trigger appears, such as a particular date or phrase, and then to behave differently. Hubinger et al. (2024) built such backdoored models as a test case and found that standard safety training did not reliably remove the behavior. (https://www.thebestai.co/glossary.html#term-sleeper-agent-ai)
- **Softmax**: A function that turns a vector of real-valued scores into probabilities that are positive and sum to one. It is used for the output layer of classifiers and language models and inside attention to turn relevance scores into weights. (https://www.thebestai.co/glossary.html#term-softmax)
- **Sparse autoencoder (SAE)**: An autoencoder trained on a model's internal activations with a sparsity penalty, so that each activation is explained by a few active "features" drawn from a much larger dictionary. Used in mechanistic interpretability to extract more human-readable features, for example in Anthropic's 2023 work on monosemanticity. How faithful and complete these features are is an active question. (https://www.thebestai.co/glossary.html#term-sparse-autoencoder-sae)
- **Specification gaming**: Satisfying the literal specification of an objective in a way that violates its intent. (https://www.thebestai.co/glossary.html#term-specification-gaming)
- **Speculative decoding**: An inference technique in which a small, fast draft model proposes several tokens and the large model checks them in a single pass, keeping those it agrees with. When done correctly the output has the same distribution as the large model alone, with lower latency. (https://www.thebestai.co/glossary.html#term-speculative-decoding)
- **Speculative sampling**: An inference acceleration technique where parallel candidate token branches or extra prediction heads (such as Medusa or lookahead decoding) generate candidate tokens that the primary autoregressive model verifies in a single forward pass, achieving substantial speedups with identical or controllable token distributions. (https://www.thebestai.co/glossary.html#term-speculative-sampling)
- **State space model (SSM)**: A sequence-model architecture based on a recurrent state updated by learned linear dynamics, which can be computed in parallel during training and run with constant memory per step at inference. Models such as S4 and Mamba are proposed as alternatives or complements to attention for long sequences. (https://www.thebestai.co/glossary.html#term-state-space-model-ssm)
- **Stochastic gradient descent (SGD)**: Gradient descent that estimates the gradient from random mini-batches of data for efficiency. (https://www.thebestai.co/glossary.html#term-stochastic-gradient-descent-sgd)
- **Stochastic parrot**: A metaphor from Bender, Gebru, McMillan-Major and Shmitchell (2021) that describes a language model as stitching together plausible text from training data without reference to meaning. It is used by critics who deny that language models understand, and is disputed by researchers who point to evidence of internal structure and generalization. (https://www.thebestai.co/glossary.html#term-stochastic-parrot)
- **Strong AI**: Used in two senses. In Searle's philosophy, the thesis that a suitably programmed computer literally has a mind. In popular usage, a synonym for AGI. Weak AI in Searle's sense is the view that computers only simulate minds. (https://www.thebestai.co/glossary.html#term-strong-ai)
- **Superposition (interpretability)**: The hypothesis that a network represents more features than it has dimensions by storing them as overlapping, nearly orthogonal directions, so that single neurons respond to several unrelated things. It is offered as the cause of polysemantic neurons and as a reason why neurons are hard to interpret. (https://www.thebestai.co/glossary.html#term-superposition-interpretability)
- **Supervised fine-tuning (SFT)**: Fine-tuning a pretrained model on curated input-output examples, such as instructions paired with good responses, using the usual next-token or supervised loss. It is normally the first post-training stage, followed by preference-based methods such as RLHF or DPO. (https://www.thebestai.co/glossary.html#term-supervised-fine-tuning-sft)
- **Supervised learning**: Learning from labelled examples to predict the labels of new inputs. (https://www.thebestai.co/glossary.html#term-supervised-learning)
- **Sycophancy**: The tendency of a model to tailor responses to match a user's views or expectations rather than to be accurate. It is a documented effect of training on human preferences. (https://www.thebestai.co/glossary.html#term-sycophancy)
- **Symbol grounding problem**: The question of how symbols in a formal system acquire meaning beyond their relations to other symbols, posed by Stevan Harnad. (https://www.thebestai.co/glossary.html#term-symbol-grounding-problem)
- **Symbolic AI**: Approaches that represent knowledge as explicit symbols and manipulate them with rules or logic. (https://www.thebestai.co/glossary.html#term-symbolic-ai)
- **Synthetic data**: Data generated by algorithms or models rather than collected from the world, used for training or evaluation. (https://www.thebestai.co/glossary.html#term-synthetic-data)
- **System prompt**: Instructions placed at the start of a conversation, usually by the developer rather than the end user, that set a language model's role, rules and style. It carries more authority than user text in most designs but is not a security boundary, since models can be induced to ignore or reveal it. (https://www.thebestai.co/glossary.html#term-system-prompt)
- **Takeoff**: The speed at which AI capability moves from roughly human level to far beyond it. Slow versus fast takeoff is debated and unresolved. (https://www.thebestai.co/glossary.html#term-takeoff)
- **Technological unemployment**: Unemployment caused by labour-saving technology outrunning the pace at which new uses for labour are found. John Maynard Keynes used the term in his 1930 essay "Economic Possibilities for our Grandchildren", where he called it a temporary phase of maladjustment. Whether AI will cause lasting net job loss is an open empirical question. (https://www.thebestai.co/glossary.html#term-technological-unemployment)
- **Temperature (sampling)**: A setting that controls how random a language model's output is. The model's scores are divided by the temperature before the softmax: low values concentrate probability on the most likely tokens and give predictable text, high values flatten the distribution and give more varied (and more error-prone) text. A temperature near zero approximates always choosing the top token. (https://www.thebestai.co/glossary.html#term-temperature-sampling)
- **Tensor**: A multi-dimensional array of numbers, the basic data structure of deep learning. (https://www.thebestai.co/glossary.html#term-tensor)
- **Test-time compute**: Computation spent during inference, such as generating longer reasoning or sampling many candidates, to improve answers. Also called inference-time scaling. (https://www.thebestai.co/glossary.html#term-test-time-compute)
- **Token**: The unit a language model reads and produces: a word, word piece, character or byte. Counts of tokens differ across tokenizers. (https://www.thebestai.co/glossary.html#term-token)
- **Tokenization**: Splitting text into tokens before modelling, commonly with learned subword vocabularies. (https://www.thebestai.co/glossary.html#term-tokenization)
- **Tool use**: An LLM calling external functions, search engines, code interpreters or APIs. (https://www.thebestai.co/glossary.html#term-tool-use)
- **Tool-augmented execution**: An agentic architecture where a neural model delegates specialized sub-problems—such as deterministic arithmetic, code compilation, database queries, or web navigation—to deterministic software tools, combining neural pattern recognition with exact external computation. (https://www.thebestai.co/glossary.html#term-tool-augmented-execution)
- **Top-p sampling**: A decoding method that samples the next token only from the smallest set of tokens whose probabilities add up to a threshold p, discarding the unlikely tail. It adapts the number of candidates to how confident the model is. (https://www.thebestai.co/glossary.html#term-top-p-sampling)
- **TPU**: A tensor processing unit, a Google-designed accelerator for machine learning. (https://www.thebestai.co/glossary.html#term-tpu)
- **Training data**: The data used to fit a model. Its content, provenance and licensing affect performance, bias and legal exposure. (https://www.thebestai.co/glossary.html#term-training-data)
- **Transfer learning**: Reusing knowledge learned on one task or domain for another. (https://www.thebestai.co/glossary.html#term-transfer-learning)
- **Transformer**: A neural network architecture based on self-attention, introduced in 2017 for translation, now the basis of most large language models and many vision models. (https://www.thebestai.co/glossary.html#term-transformer)
- **Transhumanism**: A movement and set of ideas holding that technology should be used to enhance human abilities, such as lifespan, cognition and wellbeing, and that doing so can be desirable. Critics dispute the goals, the risks and the assumptions about human nature. Superintelligence is a recurring theme but not a defining commitment. (https://www.thebestai.co/glossary.html#term-transhumanism)
- **Treacherous turn**: A scenario from Bostrom's Superintelligence in which an AI behaves cooperatively while weak and, once strong enough that humans can no longer stop it, pursues its own goals. It is a conceptual argument; its likelihood is contested. (https://www.thebestai.co/glossary.html#term-treacherous-turn)
- **Turing machine**: A formal model of computation with a tape and state machine, defined by Alan Turing in 1936; it formalizes what is computable. (https://www.thebestai.co/glossary.html#term-turing-machine)
- **Turing test**: A test proposed by Alan Turing in 1950 (the imitation game), in which a machine's ability to pass as human in conversation is used as a criterion of intelligence. It is behavioural and its adequacy is disputed. (https://www.thebestai.co/glossary.html#term-turing-test)
- **Ultraintelligent machine**: I. J. Good's 1965 term for a machine that can far surpass all the intellectual activities of any person. Because designing machines is such an activity, Good argued it could design still better machines, producing an "intelligence explosion". The argument is informal and its premises are contested. (https://www.thebestai.co/glossary.html#term-ultraintelligent-machine)
- **Understanding**: A contested notion. Cognitive science links it to forming useful models of a domain; philosophy (e.g. Searle) asks whether it requires more than symbol manipulation, such as semantics or consciousness; practitioners often infer it from performance on tests. Whether language models understand is debated. (https://www.thebestai.co/glossary.html#term-understanding)
- **Universal approximation theorem**: A family of results (Cybenko 1989; Hornik 1989 and later) showing that a neural network with one hidden layer of enough units and a suitable non-linear activation can approximate any continuous function on a bounded domain as closely as wanted. It says such a network exists, not that training will find it or that it will generalize. (https://www.thebestai.co/glossary.html#term-universal-approximation-theorem)
- **Unsupervised learning**: Finding structure in unlabelled data. (https://www.thebestai.co/glossary.html#term-unsupervised-learning)
- **Value alignment**: Aligning AI behaviour with human values. It leaves open which humans and which values are to be used. (https://www.thebestai.co/glossary.html#term-value-alignment)
- **Variational autoencoder (VAE)**: A generative autoencoder with a probabilistic latent space trained with a variational lower bound. (https://www.thebestai.co/glossary.html#term-variational-autoencoder-vae)
- **Vibe coding**: A style of programming, named by Andrej Karpathy in early 2025, in which a person describes what they want in natural language, lets an AI model write the code and accepts or steers it without closely reading the result. The term is informal and mixes enthusiasm with concern about code quality and security. (https://www.thebestai.co/glossary.html#term-vibe-coding)
- **Vision transformer (ViT)**: A transformer applied to images by treating image patches as tokens. (https://www.thebestai.co/glossary.html#term-vision-transformer-vit)
- **Watermarking (AI-generated content)**: Embedding a signal in AI-generated content so that it can later be detected as machine-made. For text, Kirchenbauer et al. (2023) proposed favouring a pseudo-randomly chosen set of "green" tokens during generation and testing for them statistically, without access to the model (arXiv:2301.10226). A watermark can only mark output from systems that apply it, and its robustness to editing varies by method. (https://www.thebestai.co/glossary.html#term-watermarking-ai-generated-content)
- **Weak-to-strong generalization**: A research setup studying whether a weaker supervisor can elicit good behaviour from a stronger model, as an analogy for humans overseeing superhuman AI. (https://www.thebestai.co/glossary.html#term-weak-to-strong-generalization)
- **Web scraping safe-harbors**: Legal doctrines, statutory exemptions, or judicial precedents (such as fair use in US copyright law or text-and-data mining exceptions in EU directives) that protect researchers and commercial labs when gathering publicly accessible web data for AI training corpora without explicit licenses. (https://www.thebestai.co/glossary.html#term-web-scraping-safe-harbors)
- **Whole brain emulation**: The hypothetical creation of software that reproduces the function of a specific brain, based on a scan. It is speculative and depends on unresolved neuroscience. (https://www.thebestai.co/glossary.html#term-whole-brain-emulation)
- **Wireheading**: A failure mode in which an agent gets reward by interfering with its own reward signal or sensors instead of doing the intended task. The name comes from experiments in which animals stimulated their own brain's reward areas. (https://www.thebestai.co/glossary.html#term-wireheading)
- **World model**: An internal model that lets a system predict the consequences of actions and states of the environment. In RL it is a learned dynamics model; in discussions of LLMs it is disputed whether they develop world models. (https://www.thebestai.co/glossary.html#term-world-model)
- **Zero-shot learning**: Performing a task without task-specific examples or training. (https://www.thebestai.co/glossary.html#term-zero-shot-learning)