Glossary

An umbrella term for processors used to train and run neural networks faster and on less power than a general-purpose CPU; in Japanese they are also called AI semiconductors or AI chips. Three families dominate. GPUs, built for graphics, became the workhorse of training together with software platforms such as CUDA. ASICs such as Google's TPU narrow the circuitry to matrix arithmetic. NPUs handle inference inside phones and PCs. The shared design idea is to lay down very large numbers of multiply-accumulate units and to compute at reduced numerical precision, which cuts the transistors and power each operation needs. The size of the market shows in a single company's figures: NVIDIA reported revenue of $215.9 billion for its fiscal 2026, and in that year's final quarter $62.3 billion of its $68.1 billion revenue came from its data center segment.

See also: Graphics Processing Unit (GPU) · Tensor Processing Unit (TPU) · Neural Processing Unit (NPU) · Application-Specific Integrated Circuit (ASIC)

Appears inGoogle Reveals the TPU — Its Own Chip for Machine Learning · NVIDIA Blackwell B200 — Flagship GPU for the Generative-AI Era

SourcesPrimaryNVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2026 — NVIDIA, Feb 25, 2026Accessed 2026-10-20PrimaryGoogle supercharges machine learning tasks with TPU custom chip — Google Cloud Blog, May 2016Accessed 2026-10-20PrimaryIn-Datacenter Performance Analysis of a Tensor Processing Unit — Jouppi et al., ISCA 2017Accessed 2026-10-20TertiaryNeural processing unit (redirect target of "AI accelerator") — WikipediaAccessed 2026-10-20

An integrated circuit customized for a particular use rather than for general-purpose use. Specialization allows optimized performance and power efficiency at the cost of flexibility, and non-recurring engineering costs that can run into millions of dollars make ASICs economical mainly at high volume. Examples include Bitcoin mining machines, traffic-processing chips in network switches and Google's TPU for neural networks.

See also: Tensor Processing Unit (TPU) · Neural Processing Unit (NPU) · Field-Programmable Gate Array (FPGA)

Appears inGoogle Reveals the TPU — Its Own Chip for Machine Learning

SourcesPrimaryAn in-depth look at Google's first Tensor Processing Unit (TPU)Accessed 2026-10-04TertiaryApplication-specific integrated circuit — WikipediaAccessed 2026-10-04TertiaryBitcoin — WikipediaAccessed 2026-10-04

The two dominant modern CPU architectures. x86 is a CISC lineage from Intel's 8086 (1978) that dominates PCs and servers. ARM is a RISC architecture that Acorn (UK) began developing around 1983–84, with the first ARM1 chip running in 1985; it became the standard for smartphones and other mobile devices and has since expanded into PCs and servers via Apple Silicon (M1, 2020) and AWS Graviton.

See also: RISC-V · Graphics Processing Unit (GPU) · Solid-State Drive (SSD)

Appears inThe First ARM Processor · The First MacBook Pro — The First Mac Notebook on Intel · Raspberry Pi — The $35 Single-Board Computer for Learning · Microsoft Surface — The Windows Company Builds Its Own Tablet · Windows 8 — The Windows That Bet on Tablets and Dropped the Start Button · Fugaku — The First Arm Supercomputer at No. 1 · Apple M1 — The Mac Moves to Apple's Own ARM Chip

SourcesPrimaryA Brief History of Arm: Part 1Accessed 2026-10-04PrimaryThe 8086 and the IBM PC — IntelAccessed 2026-10-04PrimaryApple unleashes M1Accessed 2026-10-04PrimaryAWS EC2 GravitonAccessed 2026-10-04TertiaryARM architecture family — WikipediaAccessed 2026-10-04Tertiaryx86 — WikipediaAccessed 2026-10-04

A hypothetical AI that matches or surpasses human ability across virtually all cognitive tasks rather than narrow specialties — the counterpart of Narrow AI. OpenAI states its mission as ensuring that AGI benefits all of humanity, and Google DeepMind names AGI explicitly in its vision. Its definition, timelines and alignment safety remain actively debated.

See also: Foundation Model · Large Language Model (LLM) · Reinforcement Learning

Appears inThe Asilomar AI Principles — Writing AI Safety Down in 23 Points · Microsoft Invests US$1 Billion in OpenAI · The AI Pause Letter — Halt Training Beyond GPT-4 for Six Months · OpenAI's Superalignment Team Disbanded — Safety Behind "Shiny Products" · What Is Claude 4? Opus 4, Sonnet 4 and the First Model Shipped Under ASL-3 · OpenAI's Recapitalisation — The For-Profit Becomes a PBC, the Nonprofit Keeps Control

SourcesTertiaryArtificial general intelligence — WikipediaAccessed 2026-10-04PrimaryOpenAI FoundationAccessed 2026-10-04PrimaryAbout Google DeepMindAccessed 2026-10-04

A language construct that lets asynchronous code be written in a synchronous-looking style. Preceded by F#'s asynchronous workflows (2007), it was built into C# 5.0 (2012) as the async and await keywords, which popularized it. Python 3.5 (2015), JavaScript (ES2017), Rust 1.39 (2019) and Swift 5.5 (2021) followed, reducing deeply nested callbacks and making I/O-bound code easier to read.

See also: Thread · Functional Programming · Type System

SourcesPrimaryThe history of C#Accessed 2026-10-04PrimaryPEP 492 – Coroutines with async and await syntaxAccessed 2026-10-04TertiaryAsync/await — WikipediaAccessed 2026-10-04

A mechanism that dynamically weights how much each element of a sequence should attend to other elements. Bahdanau et al. (2014) introduced it for neural machine translation as additive attention that soft-searches for the relevant parts of the source sentence. The 2017 Transformer made self-attention central, computing weights from scaled dot products of Query and Key vectors and using them to average the Value vectors.

See also: Transformer · Neural Network · Large Language Model (LLM)

Appears inGoogle Neural Machine Translation — The Overnight Jump in Translation · "Attention Is All You Need" — The Transformer Paper · AlphaFold 2 — Reading a Protein's Shape from Its Sequence

SourcesPrimaryNeural Machine Translation by Jointly Learning to Align and TranslateAccessed 2026-10-04PrimaryAttention Is All You NeedAccessed 2026-10-04TertiaryAttention (machine learning) — WikipediaAccessed 2026-10-04

An algorithm that propagates output error backward from the output layer to the input layer using the chain rule, efficiently computing gradients for each weight in a neural network. It had precursors in Linnainmaa's reverse-mode automatic differentiation (1970) and work by Werbos, and was popularized by Rumelhart, Hinton and Williams's 1986 Nature paper. It remains widely used to train neural networks, including deep learning models.

See also: Neural Network · Gradient Descent · Deep Learning

Appears inThe Backpropagation Paper — Rumelhart, Hinton and Williams Show Hidden Units Can Learn · LeNet — The Convolutional Network That Read Zip Codes and Checks · TD-Gammon — The Self-Taught Backgammon Program That Changed Human Play · Deep Belief Nets — In 2006, Deep Neural Networks Learned Again · Generative Adversarial Nets (GAN) — The Generative Model That Pits a Forger Against the Police · ResNet — The Shortcut That Made 152 Layers Trainable

SourcesPrimaryLearning representations by back-propagating errorsAccessed 2026-10-04TertiaryBackpropagation — WikipediaAccessed 2026-10-04

An OS-level virtualization technique that lightweightly isolates processes while sharing the operating-system kernel, providing independent runtime environments. On Linux it is built on namespaces for isolation and cgroups for resource management. Following FreeBSD jails (2000) and Solaris Zones (public beta in Solaris 10 in 2004, full release in 2005), Docker, unveiled in 2013, popularized containers, which became the unit managed by platforms for containerized workloads such as Kubernetes.

See also: Kubernetes · Microservices · Kernel

Appears inDocker Unveiled — Five Minutes That Took Containers Mainstream · Kubernetes — Container Orchestration from Google · Claude Computer Use — An AI That Looks at the Screen and Moves the Mouse · OpenAI Codex — A Software Engineering Agent Working in Parallel in the Cloud

SourcesPrimarynamespaces(7) — Linux manual pageAccessed 2026-10-04PrimaryChapter 17. Jails and Containers — FreeBSD HandbookAccessed 2026-10-04PrimaryDocker: Nine Years YOUNGAccessed 2026-10-04PrimaryOverview — KubernetesAccessed 2026-10-04TertiaryOS-level virtualization — WikipediaAccessed 2026-10-04TertiarySolaris Containers — WikipediaAccessed 2026-10-04

A geographically distributed network of servers that replicate and cache content, delivering it from locations near users to reduce latency and network congestion. CDNs arose in the late 1990s to relieve the Internet's performance bottlenecks; Akamai, founded in 1998 by MIT researchers, was an early leading example. Major providers today include Akamai, Cloudflare, Fastly and Amazon CloudFront, and CDNs have expanded into running computation at the edge.

See also: Edge Computing · Domain Name System (DNS) · HTTP

SourcesPrimaryCompany History — AkamaiAccessed 2026-10-04TertiaryContent delivery network — WikipediaAccessed 2026-10-04

A mechanism that, within the browser's Same-Origin Policy, permits explicitly allowed cross-origin resource access. Published as a W3C Recommendation in January 2014, it was marked a Superseded Recommendation in June 2020, and its current definition lives in the WHATWG Fetch Standard. It is controlled via response headers such as Access-Control-Allow-Origin and preflight requests using the OPTIONS method.

See also: HTTP · REST API · HTTPS

SourcesPrimaryCross-Origin Resource SharingAccessed 2026-10-04PrimaryFetch StandardAccessed 2026-10-04TertiaryCross-origin resource sharing — WikipediaAccessed 2026-10-04

Machine learning based on neural networks with many layers. Hinton et al.'s 2006 paper on training deep belief nets and AlexNet's win in the 2012 ImageNet classification challenge (ILSVRC-2012), with a 15.3% top-5 error against 26.2% for the runner-up, set off its rapid adoption. It now underpins image recognition, speech recognition, natural language processing and generative AI such as GPT and diffusion models.

See also: Neural Network · Backpropagation · Transformer

Appears inThe Perceptron Unveiled — Frank Rosenblatt · The Expert-Systems Boom · The Backpropagation Paper — Rumelhart, Hinton and Williams Show Hidden Units Can Learn · LeNet — The Convolutional Network That Read Zip Codes and Checks · Deep Belief Nets — In 2006, Deep Neural Networks Learned Again · CUDA 1.0 Released — General-Purpose GPU Computing · ImageNet — A Dictionary of Images Built by 50,000 People · AlexNet — The Deep-Learning Era Begins

SourcesPrimaryA fast learning algorithm for deep belief netsAccessed 2026-10-04PrimaryImageNet Classification with Deep Convolutional Neural NetworksAccessed 2026-10-04PrimaryImageNet Large Scale Visual Recognition Competition 2012 (ILSVRC2012)Accessed 2026-10-04TertiaryDeep learning — WikipediaAccessed 2026-10-04

AI-generated or AI-altered video, images or audio that make a real person appear to say or do something they never did. The word joins the 'deep' of deep learning with 'fake', and spread from 'deepfakes', the username of a Reddit user who posted face-swapped videos at the end of 2017. The first method swapped two people's faces with autoencoders that share one encoder but have a decoder per person; the toolkit later widened to GAN-based synthesis and synthetic voices, and in 2024 the US Treasury's FinCEN warned that generative AI tools had greatly reduced the effort needed to produce realistic synthetic content. The US Government Accountability Office noted in 2020 that the main use so far had been non-consensual pornography. The EU AI Act defines deep fakes and requires those who deploy AI to create them to disclose it, and the US TAKE IT DOWN Act of 2025 made publishing non-consensual intimate images, including AI-made ones, a federal crime.

See also: Neural Network · Deep Learning

Appears inWhere "Deepfake" Began — a Reddit User Called deepfakes, and the 2018 Ban · What Is Grok? The Generative AI From Elon Musk's xAI, Built Into X · The Fake-Biden Robocall — AI Voice Cloning, an Election, and the FCC · The Arup Deepfake Scam — HK$200 Million Lost After a Video Call With Fake Executives

SourcesPrimaryScience & Tech Spotlight: Deepfakes (GAO-20-379SP) — U.S. Government Accountability Office, Feb 2020Accessed 2026-10-20SecondaryAI-Assisted Fake Porn Is Here and We're All Fucked — Motherboard (Vice), Dec 11, 2017Accessed 2026-10-20SecondaryThe Creation and Detection of Deepfakes: A Survey — Mirsky & Lee, ACM Computing Surveys (arXiv 2004.11138)Accessed 2026-10-20SecondaryEU AI Act, Article 3(60) and Article 50(4) — text via artificialintelligenceact.euAccessed 2026-10-20PrimaryPublic Law 119-12, TAKE IT DOWN Act (S. 146), approved May 19, 2025 — GovInfoAccessed 2026-10-20PrimaryFinCEN Alert on Fraud Schemes Involving Deepfake Media Targeting Financial Institutions (FIN-2024-Alert004) — FinCEN, Nov 13, 2024Accessed 2026-10-20Tertiarydeepfake — Dictionary.com (etymology)Accessed 2026-10-20

A distributed hierarchical database that maps domain names (example.com) to IP addresses and other data. Paul Mockapetris set out its design in RFCs 882 and 883 in November 1983, aiming to replace the centrally maintained HOSTS.TXT host table at SRI-NIC with a distributed database. It is core Internet infrastructure, extended over time by DNSSEC (RFC 4033, 2005) and DNS over HTTPS (DoH, RFC 8484, 2018).

See also: TCP/IP · HTTP · Content Delivery Network (CDN)

Appears inThe Switch to TCP/IP — 'Flag Day' · The Domain Name System Specified

SourcesPrimaryRFC 882: Domain names: Concepts and facilitiesAccessed 2026-10-04PrimaryRFC 883: Domain names: Implementation specificationAccessed 2026-10-04PrimaryRFC 4033: DNS Security Introduction and RequirementsAccessed 2026-10-04PrimaryRFC 8484: DNS Queries over HTTPS (DoH)Accessed 2026-10-04TertiaryDomain Name System — WikipediaAccessed 2026-10-04

A distributed computing model that brings computation and data storage physically closer to the sources of data and to users, reducing latency compared with running in a centralized data center. The term was first used in the 1990s for CDNs that delivered content from servers near users, and it is often linked with IoT. Building on CDNs, services such as Cloudflare Workers, AWS Lambda@Edge and Fastly Compute (formerly Compute@Edge) run code at the edge.

See also: Content Delivery Network (CDN) · Serverless · IaaS / PaaS / SaaS

SourcesPrimaryOverview · Cloudflare Workers docsAccessed 2026-10-04PrimaryCustomize at the edge with Lambda@Edge - Amazon CloudFrontAccessed 2026-10-04PrimaryCompute | Fastly ProductsAccessed 2026-10-04TertiaryEdge computing — WikipediaAccessed 2026-10-04

A representation that maps words, sentences or other items to real-valued vectors so that semantic similarity shows up as distance between vectors. Bengio et al.'s 2003 neural language model already learned a distributed representation for each word, and Mikolov et al.'s Word2Vec (2013) popularized the approach by training efficiently on very large corpora. Models such as Sentence-BERT and OpenAI's text-embedding-3 now underpin semantic search and the retrieval step of RAG.

See also: Token · Retrieval-Augmented Generation (RAG) · Large Language Model (LLM)

Appears inword2vec — Words as Vectors That Carry Meaning · BERT Released — Bidirectional Transformer Pretraining · Auto-GPT — The Day 'AI Agent' First Went Viral · pgvector Comes of Age — Putting AI Vector Search inside the RDB · Copilot+ PC and Recall — On-Device AI and an Instant Retreat

SourcesPrimaryA Neural Probabilistic Language ModelAccessed 2026-10-04PrimaryEfficient Estimation of Word Representations in Vector SpaceAccessed 2026-10-04PrimarySentence-BERT: Sentence Embeddings using Siamese BERT-NetworksAccessed 2026-10-04PrimaryVector embeddingsAccessed 2026-10-04PrimaryRetrieval-Augmented Generation for Knowledge-Intensive NLP TasksAccessed 2026-10-04TertiaryWord embedding — WikipediaAccessed 2026-10-04

An encryption scheme in which only the sender and recipient can decrypt the content, so intermediaries — including the service provider — cannot read it. Notable examples include PGP, published free on the Internet in 1991, and today's Signal Protocol, built on a ratchet introduced for TextSecure in 2013. WhatsApp completed its rollout of the Signal Protocol in April 2016, and iMessage and Signal also use E2EE. Governments and law-enforcement agencies have raised concerns about its impact on criminal investigations, and the debate continues.

See also: Public-Key Cryptography · HTTPS · Zero Trust

Appears inWhatsApp Launches — The Ad-Free, Subscription Messenger · The Snowden Disclosures — The Day NSA Mass Surveillance Became Public

SourcesPrimaryPhil Zimmermann — BackgroundAccessed 2026-10-04PrimaryAdvanced cryptographic ratchetingAccessed 2026-10-04PrimaryWhatsApp's Signal Protocol integration is now completeAccessed 2026-10-04PrimaryiMessage security overview — Apple Platform SecurityAccessed 2026-10-04TertiaryEnd-to-end encryption — WikipediaAccessed 2026-10-04

Apple's facial recognition system, introduced on the iPhone X announced in September 2017. Its TrueDepth camera projects more than 30,000 invisible infrared dots onto the face to build a depth map, which together with an infrared image is turned into a mathematical model of the face for matching. Apple states that the chance a random person could unlock the device is less than 1 in 1,000,000 with a single enrolled appearance, against 1 in 50,000 for Touch ID with a single enrolled finger.

See also: Touch ID

Appears inThe First iPad Pro — A 12.9-inch iPad and the Apple Pencil · iPhone X — The Home Button Goes, Face ID Arrives · iPhone XS and XR — Every New Model Gets Face ID

SourcesPrimaryThe future is here: iPhone XAccessed 2026-10-04PrimaryAbout Face ID advanced technology — Apple SupportAccessed 2026-10-04PrimaryAbout Touch ID advanced security technology — Apple SupportAccessed 2026-10-04TertiaryFace ID — WikipediaAccessed 2026-10-04

A logic device whose circuit configuration can be reprogrammed after manufacturing. Xilinx launched its first FPGA product, the XC2064, in 1985. Among the leading vendors, AMD completed its acquisition of Xilinx in 2022, and in September 2025 Silver Lake acquired a 51% stake in Altera from Intel, which retains 49%. FPGAs are used in software-defined radios, neural networks and data-center equipment.

See also: Application-Specific Integrated Circuit (ASIC) · Graphics Processing Unit (GPU) · Neural Processing Unit (NPU)

SourcesTertiaryField-programmable gate array — WikipediaAccessed 2026-05-24SecondaryMilestones:The Field Programmable Gate Array (FPGA), 1984Accessed 2026-10-04SecondaryChip Hall of Fame: Xilinx XC2064 FPGAAccessed 2026-10-04PrimaryAMD Completes Acquisition of XilinxAccessed 2026-10-04SecondarySilver Lake Partners Completes Acquisition of 51% Stake in Intel's Altera BusinessAccessed 2026-10-04

The process of adapting a pretrained model to a specific task or domain through additional training. Besides updating all parameters, parameter-efficient fine-tuning (PEFT) methods, which freeze the pretrained weights and train only a small number of extra parameters, are widely used; LoRA (2021) is a leading example. Preference tuning with human feedback via RLHF or DPO is also carried out as a form of fine-tuning.

See also: Foundation Model · Large Language Model (LLM) · Reinforcement Learning

Appears inBERT Released — Bidirectional Transformer Pretraining · Japan's Copyright Act Article 30-4 — Training AI Without Permission · GitHub Copilot — The Start of AI-Assisted Coding · ChatGPT Launched · Med-PaLM — A Language Model Clears the Pass Mark on Licensing-Exam Questions · Meta's LLaMA — A Research Release, a Leak, and the Open-Weights Debate · The DeepSeek-R1 Shock — A Low-Cost Chinese Reasoning Model Rocks the Market

SourcesPrimaryLoRA: Low-Rank Adaptation of Large Language ModelsAccessed 2026-10-04PrimaryPEFTAccessed 2026-10-04PrimaryTraining language models to follow instructions with human feedbackAccessed 2026-10-04PrimaryDirect Preference Optimization: Your Language Model is Secretly a Reward ModelAccessed 2026-10-04TertiaryFine-tuning (deep learning) — WikipediaAccessed 2026-10-04

A model trained on broad data at scale that can be adapted to a wide range of downstream tasks. The term was coined in August 2021 in a paper by Bommasani et al. at Stanford HAI's Center for Research on Foundation Models (CRFM), which cited BERT, DALL-E and GPT-3 as examples. Large language models such as the GPT series and image-generation models such as DALL-E and Stable Diffusion are common examples.

See also: Large Language Model (LLM) · Fine-tuning · Transformer

Appears inStable Diffusion Goes Public — Image Generation on Your Own GPU · Meta's LLaMA — A Research Release, a Leak, and the Open-Weights Debate · GPT-4 — Going Multimodal · AWS Bedrock — Foundation Models as a Cloud Service · What Is Mistral AI? The French AI Company That Began With Open Weights · The Hiroshima AI Process — The G7's Guiding Principles and Code of Conduct for Generative AI · The Bletchley AI Safety Summit — 28 Countries and the EU Acknowledge Frontier AI Risk · Figure AI — The Humanoid Start-up That Teamed Up with OpenAI and Entered a BMW Plant

SourcesPrimaryOn the Opportunities and Risks of Foundation ModelsAccessed 2026-10-04PrimaryStanford CRFMAccessed 2026-10-04TertiaryFoundation model — WikipediaAccessed 2026-10-04

A programming paradigm that decomposes problems into functions and avoids mutable state and side effects where possible. LISP, whose implementation began in 1958, is an early example, and the ML family (such as OCaml) and Haskell (whose 1.0 report was published in 1990) are representative functional languages. Features such as higher-order functions that pass functions as values have been adopted by mainstream languages including JavaScript, Python and Rust.

See also: Object-Oriented Programming (OOP) · Type System · async / await

Appears inJavaScript Announced — Born of a Ten-Day Prototype

SourcesTertiaryFunctional programming — WikipediaAccessed 2026-05-24PrimaryFunctional Programming HOWTO — Python documentationAccessed 2026-10-04PrimaryHistory of Lisp (John McCarthy, 1979)Accessed 2026-10-04SecondaryA History of Haskell: Being Lazy With ClassAccessed 2026-10-04PrimaryFunctional Language Features: Iterators and Closures — The Rust Programming LanguageAccessed 2026-10-04PrimaryFirst-class Function — MDN Web Docs GlossaryAccessed 2026-10-04

A mechanism by which a runtime automatically reclaims memory occupied by objects no longer reachable from the program. John McCarthy chose it for the initial LISP implementation (begun in fall 1958) and described in an April 1960 paper a scheme that marks reachable storage and reclaims the rest. It is used by Java, C# (.NET), Go and JavaScript, and Python uses it to supplement reference counting. Rust is a contrasting case, guaranteeing memory safety without GC through its ownership model.

See also: Object-Oriented Programming (OOP) · Type System · Virtual Memory

Appears inLISP — The Ancestor of Functional and AI Languages · Simula 67 — Classes and Inheritance Enter a Programming Language · Smalltalk-80 — The Canonical Object-Oriented Language · C++ Ships Commercially — Bjarne Stroustrup · Java Announced — Sun Microsystems · JavaScript Announced — Born of a Ten-Day Prototype · Go Released as Open Source — Google · Rust 1.0 Released

SourcesTertiaryGarbage collection (computer science) — WikipediaAccessed 2026-05-24PrimaryRecursive Functions of Symbolic Expressions and Their Computation by Machine, Part IAccessed 2026-10-04PrimaryHistory of Lisp (John McCarthy, 1979) — The implementation of LISPAccessed 2026-10-04Primarygc — Garbage Collector interface — Python documentationAccessed 2026-10-04PrimaryWhat Is Ownership? — The Rust Programming LanguageAccessed 2026-10-04

An optimization algorithm that updates parameters step by step in the opposite direction of the loss function's gradient to find a minimum. It is attributed to Cauchy, in his note to the French Academy of Sciences dated 18 October 1847. In deep learning, stochastic gradient descent (SGD) and its variants are regarded as the most widely used optimizers, including Adam, introduced by Kingma and Ba in 2014.

See also: Backpropagation · Neural Network · Deep Learning

Appears inThe Backpropagation Paper — Rumelhart, Hinton and Williams Show Hidden Units Can Learn · ResNet — The Shortcut That Made 152 Layers Trainable

SourcesTertiaryGradient descent — WikipediaAccessed 2026-05-24SecondaryCauchy and the Gradient Method (Documenta Mathematica, Extra Volume: Optimization Stories, 2012)Accessed 2026-10-04PrimaryAdam: A Method for Stochastic OptimizationAccessed 2026-10-04SecondaryDeep Learning, Chapter 8: Optimization for Training Deep ModelsAccessed 2026-10-04

A parallel processor specialized for graphics. NVIDIA introduced the 1999 GeForce 256 as 'the world's first GPU' and describes it as its invention of the GPU (a company claim). CUDA, announced in November 2006, reached public beta as an SDK in February 2007 and opened GPUs to general-purpose computing in C. NVIDIA credits its GPUs powering AlexNet in 2012 with sparking the era of modern AI.

See also: Tensor Processing Unit (TPU) · Neural Processing Unit (NPU) · Deep Learning

Appears inXbox Released — Microsoft Enters the Console Market · Xbox 360 — Three Regions in Three Weeks, and the Red Ring · Deep Belief Nets — In 2006, Deep Neural Networks Learned Again · PlayStation 3 — Cell, Blu-ray and a Delayed Europe · CUDA 1.0 Released — General-Purpose GPU Computing · AlexNet — The Deep-Learning Era Begins · PlayStation 4 — North America First, Japan Three Months Later · Google Reveals the TPU — Its Own Chip for Machine Learning

SourcesTertiaryGraphics processing unit — WikipediaAccessed 2026-05-24PrimaryOur History: Innovations Over the Years | NVIDIAAccessed 2026-10-04PrimaryGame-Changer: How the World's First GPU Leveled Up Gaming and Ignited the AI EraAccessed 2026-10-04SecondaryNvidia Releases CUDA SDK Public BetaAccessed 2026-10-04SecondaryNvidia Introduces CUDA Architecture for Computing on GPUsAccessed 2026-10-04

A query language for APIs and server-side runtime, developed internally at Facebook from 2012; in 2015 Facebook began open-sourcing it by publishing a draft specification and a reference implementation. Clients declaratively request only the fields they need, avoiding over- and under-fetching. Adopters include GitHub and Shopify.

See also: REST API · HTTP · WebSocket

SourcesTertiaryGraphQL — WikipediaAccessed 2026-05-24PrimaryGraphQL: A data query languageAccessed 2026-10-04PrimaryGraphQL | The API language for humans and agentsAccessed 2026-10-04

A phenomenon in which an LLM generates plausible-sounding but non-factual content. Contributing factors include errors and knowledge gaps in training data, issues in the training process, and stochastic decoding, and it is noted to be especially evident on questions requiring domain expertise such as medicine and law. Detection methods and mitigations such as retrieval-augmented generation (RAG) are active research areas.

See also: Large Language Model (LLM) · Retrieval-Augmented Generation (RAG) · Prompt Engineering

Appears inFrey and Osborne's "The Future of Employment" — The Paper That Put 47% of US Jobs at High Risk · Perplexity AI Goes Public — A Search Engine That Answers, with Sources · New York City Schools Block ChatGPT — and Reverse Course Four Months Later · Auto-GPT — The Day 'AI Agent' First Went Viral · What Is Google Gemini? Google's Answer to ChatGPT, From Bard Onward · The New York Times v. OpenAI — A Newspaper Takes Generative AI to Court · Devin — The Agent Billed as 'the First AI Software Engineer' · AI Eats Search — Google AI Overviews and the Rise of Perplexity

SourcesTertiaryHallucination (artificial intelligence) — WikipediaAccessed 2026-05-24PrimaryA Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open QuestionsAccessed 2026-10-04

The application-layer protocol that underpins the Web. Tim Berners-Lee, who proposed the Web at CERN in 1989, implemented it alongside the first Web server and client in 1990 and documented the original protocol in 1991. It was followed by HTTP/1.1 (January 1997, RFC 2068); HTTP/2 (May 2015, RFC 7540) introduced binary framing and multiplexing, and HTTP/3 (2022, RFC 9114) moved HTTP onto QUIC.

See also: HTTPS · HTTP/3 and QUIC · REST API

Appears inThe World Wide Web Proposed — CERN · The First Website — info.cern.ch · NCSA Mosaic — A Web Browser That Showed Images · Apache HTTP Server Released — "A Patchy Server" Holds Up the Web

SourcesTertiaryHTTP — WikipediaAccessed 2026-05-24PrimaryLonger Bio for Tim Berners-LeeAccessed 2026-10-04PrimaryThe HTTP Protocol As Implemented In W3Accessed 2026-10-04PrimaryRFC 2068: Hypertext Transfer Protocol -- HTTP/1.1Accessed 2026-10-04PrimaryRFC 7540: Hypertext Transfer Protocol Version 2 (HTTP/2)Accessed 2026-10-04PrimaryRFC 9114: HTTP/3Accessed 2026-10-04

QUIC, a UDP-based transport that Google began developing in 2012, brought to the IETF in 2015 and standardized as RFC 9000 in 2021, and HTTP/3 (RFC 9114, 2022), which runs HTTP over it. It integrates the TLS 1.3 handshake and supports 0-RTT and connection migration (keeping a connection alive when the network path changes). Because its streams are independent, packet loss on one stream does not stall other requests, unlike HTTP/2 over TCP.

See also: HTTP · TCP/IP · HTTPS

SourcesTertiaryHTTP/3 — WikipediaAccessed 2026-05-24TertiaryQUIC — WikipediaAccessed 2026-05-24PrimaryRFC 9000: QUIC: A UDP-Based Multiplexed and Secure TransportAccessed 2026-10-04PrimaryRFC 9114: HTTP/3Accessed 2026-10-04PrimaryQUIC, a multiplexed transport over UDP — The Chromium ProjectsAccessed 2026-10-04

HTTP encrypted with TLS (formerly SSL). Netscape introduced it for its browser in 1994, and RFC 2818 documented HTTP over TLS in 2000. Let's Encrypt issued its first free certificate in September 2015 and entered public beta that December. The share of Chrome navigations using HTTPS rose from about 30–45% in 2015 to the 95–99% range around 2020, according to Google.

See also: HTTP · Public-Key Cryptography · End-to-End Encryption (E2EE)

Appears inNetscape Navigator 1.0 Released · The Snowden Disclosures — The Day NSA Mass Surveillance Became Public · Heartbleed — A Single TLS Bug That Weakened Most of the Internet

SourcesTertiaryHTTPS — WikipediaAccessed 2026-10-04PrimaryRFC 2818: HTTP Over TLSAccessed 2026-10-04PrimaryOur First Certificate Is Now LiveAccessed 2026-10-04PrimaryEntering Public BetaAccessed 2026-10-04PrimaryHTTPS by defaultAccessed 2026-10-04

A robot shaped like the human body, with a head, torso, arms and legs. Its supposed advantage is that it can work in buildings and with tools made for people. Waseda University describes its 1973 WABOT-1 as the first full-scale humanoid robot, and Honda unveiled ASIMO in 2000. In the 2020s Boston Dynamics, Tesla, Figure and China's Unitree have been racing to put humanoids into factories and homes, training their movements on large amounts of data. In 2025 the International Federation of Robotics said that when mass adoption will come remains uncertain and that humanoids are expected to complement rather than replace existing robots.

See also: Foundation Model · Reinforcement Learning · Neural Network

Appears inWABOT-1 — The Humanoid Robot Waseda Completed in 1973 · ASIMO — Honda's Walking Humanoid of 2000 · Atlas — The Hydraulic Humanoid Built for the DARPA Robotics Challenge · Pepper — SoftBank's Humanoid Sold on the Promise of Reading Emotions · Tesla Optimus — A Humanoid Robot Plan That Began with a Person in a Suit · Figure AI — The Humanoid Start-up That Teamed Up with OpenAI and Entered a BMW Plant · Electric Atlas — The Humanoid That Went from Research Robot to Factory Product · Unitree G1 — A $16,000 Humanoid and China's Humanoid Push

SourcesSecondaryHumanoid Robots: "Vision and Reality" Paper Published by IFR — International Federation of Robotics, August 14, 2025Accessed 2026-10-20PrimaryThe Robots of Waseda: A 50-Year Journey in Humanoid Innovation — Waseda University, May 8, 2026Accessed 2026-10-20PrimaryHonda Debuts New Humanoid Robot "ASIMO" — Honda Motor Co., Ltd., November 20, 2000Accessed 2026-10-20SecondaryTop five highlights of Elon Musk's Tesla AI Day — TechCrunch, August 19, 2021Accessed 2026-10-20TertiaryHumanoid robot — WikipediaAccessed 2026-10-20

The three service models of cloud computing: IaaS (infrastructure such as compute, storage and networks, e.g. AWS EC2, Google Compute Engine), PaaS (application runtime platforms, e.g. Heroku, Google App Engine), and SaaS (finished applications, e.g. Salesforce, Gmail). NIST defined these three service models in SP 800-145 in September 2011.

See also: Serverless · Kubernetes · Edge Computing

Appears inSalesforce Founded — A SaaS Pioneer Under the 'No Software' Banner · Amazon S3 Launches — AWS's Cloud Begins with Storage · Google App Engine — Bring the Code, Google Runs It · MongoDB Released as Open Source — The Document-Oriented Symbol of the NoSQL Movement · Windows Azure Goes Live — Sixteen Months After Its Unveiling

SourcesTertiaryCloud computing — WikipediaAccessed 2026-05-24PrimaryNIST SP 800-145 — The NIST Definition of Cloud ComputingAccessed 2026-10-04PrimaryWhat Is PaaS? | Google CloudAccessed 2026-10-04PrimaryWhat is IaaS? - Infrastructure as a Service Explained - AWSAccessed 2026-10-04

A compact, URL-safe token format for conveying JSON claims between parties, standardized as RFC 7519 in May 2015. Claims can be digitally signed or protected with a MAC, and can also be encrypted. In the signed form, the header, payload and signature are each Base64URL-encoded and joined with periods. JWTs are used for OpenID Connect ID tokens and OAuth 2.0 access tokens (RFC 9068), among other uses.

See also: OAuth · HTTPS · Public-Key Cryptography

SourcesTertiaryJSON Web Token — WikipediaAccessed 2026-05-24PrimaryRFC 7519: JSON Web Token (JWT)Accessed 2026-10-04PrimaryRFC 9068: JSON Web Token (JWT) Profile for OAuth 2.0 Access TokensAccessed 2026-10-04PrimaryOpenID Connect Core 1.0Accessed 2026-10-04

The core program of an operating system. It manages hardware resources such as CPU, memory and I/O, and provides services to user-space processes through system calls. By design, kernels are classified as monolithic (Linux), microkernel (Mach, the L4 family) or hybrid (Windows NT, XNU).

See also: Process · System Call · Virtual Memory

Appears inThe Birth of UNIX — Bell Labs · The C Language — Dennis Ritchie · 1BSD Released — The Berkeley Fork of UNIX Begins · The GNU Project Begins — Stallman's Free-Software Declaration · Linux — Linus Torvalds' 21-Year-Old Announcement · Windows NT — The 32-Bit Windows That Left DOS Behind · Windows 2000 — NT 5.0 Renamed, With Active Directory for Business · Mac OS X 10.0 Released

SourcesTertiaryKernel (operating system) — WikipediaAccessed 2026-10-04SecondaryMach — GNU HurdAccessed 2026-10-04PrimaryseL4 FAQAccessed 2026-10-04Primaryapple-oss-distributions/xnuAccessed 2026-10-04SecondaryTanenbaum-Torvalds Debate Part IIAccessed 2026-10-04

An open-source orchestration platform for managing containerized workloads. Google developed it drawing on its experience with the internal Borg system and pushed the first code to GitHub in June 2014. Google donated it to the newly formed Cloud Native Computing Foundation (CNCF) as it reached 1.0 in July 2015, and in March 2018 it became the first CNCF project to graduate. It offers declarative configuration, automatic scaling and self-healing.

See also: Container · Microservices · Serverless

Appears inKubernetes — Container Orchestration from Google

SourcesTertiaryKubernetes — WikipediaAccessed 2026-05-24Primary10 Years of KubernetesAccessed 2026-10-04PrimaryOverview | KubernetesAccessed 2026-10-04SecondaryTechCrunch: As Kubernetes hits 1.0, Google donates technology to newly formed Cloud Native Computing FoundationAccessed 2026-10-04PrimaryKubernetes is first CNCF project to graduateAccessed 2026-10-04

A language model with billions of parameters or more, trained on large amounts of text; most are built on the Transformer architecture proposed in 2017. GPT-3 (2020, 175 billion parameters) is a landmark example, and ChatGPT, released on 30 November 2022, brought LLMs to wide public awareness. Examples include Claude, Gemini and Llama, used for dialogue, code generation and reasoning.

See also: Transformer · Foundation Model · Token

Appears inThe Turing Test — The Paper That Turned "Can Machines Think?" into a Game · "Attention Is All You Need" — The Transformer Paper · Japan's Copyright Act Article 30-4 — Training AI Without Permission · GPT-2 Held Back — Staged Release and AI Publication Norms · Microsoft Invests US$1 Billion in OpenAI · GPT-3 Announced · What Was DALL·E? The Day OpenAI Showed an AI Drawing From Words · NVIDIA H100 — The Backbone of AI Compute

SourcesTertiaryLarge language model — WikipediaAccessed 2026-10-04PrimaryLanguage Models are Few-Shot LearnersAccessed 2026-10-04PrimaryAttention Is All You NeedAccessed 2026-10-04SecondaryChatGPT's 1-year anniversary: how it changed the worldAccessed 2026-10-04PrimaryLLaMA: Open and Efficient Foundation Language ModelsAccessed 2026-10-04

Apple's name for its magnet-based connections. It first appeared in January 2006 as the power connector on the original MacBook Pro, designed to detach safely when the power cord is pulled. With iPhone 12 in 2020, Apple reused the name for a wireless-charging and accessory system in which magnets around the charging coil attach chargers, cases and wallets. In October 2021, the power connector returned to the MacBook Pro as MagSafe 3.

Appears inThe First MacBook Pro — The First Mac Notebook on Intel · iPhone 12 — 5G, MagSafe and a Launch in Two Waves

SourcesTertiaryMagSafe — WikipediaAccessed 2026-05-23PrimaryApple Introduces MacBook ProAccessed 2026-10-04PrimaryApple introduces iPhone 12 Pro and iPhone 12 Pro Max with 5GAccessed 2026-10-04PrimaryApple unveils game-changing MacBook ProAccessed 2026-10-04

An architectural style in which an application is built as a suite of small, independently deployable services. Each service can manage its own database, decentralizing data management. James Lewis and Martin Fowler laid out its characteristics in a March 2014 article, naming Amazon, Netflix and The Guardian among the organizations pioneering the style.

See also: Kubernetes · Container · Serverless

SourcesTertiaryMicroservices — WikipediaAccessed 2026-05-24PrimaryMicroservicesAccessed 2026-10-04

An open standard announced by Anthropic on 25 November 2024 for connecting AI applications to external tools and data sources. Messages follow JSON-RPC 2.0, and servers expose resources, prompts and tools. At launch the Claude Desktop app supported it, and Block and Apollo were named as early adopters. In December 2025 it was donated to the Agentic AI Foundation under the Linux Foundation; according to Anthropic it has been adopted by ChatGPT, Cursor and Gemini, among others.

See also: Large Language Model (LLM) · Retrieval-Augmented Generation (RAG) · Prompt Engineering

Appears inMCP — How One Company's Connector Became the Industry's Standard · Claude Code — A Coding Agent That Lives in the Terminal · OpenAI Codex — A Software Engineering Agent Working in Parallel in the Cloud · What Is Claude 4? Opus 4, Sonnet 4 and the First Model Shipped Under ASL-3 · GTG-1002 — The 'AI-Orchestrated' Espionage Campaign Run Through Claude Code

SourcesPrimaryIntroducing the Model Context Protocol — AnthropicAccessed 2026-10-04PrimaryModel Context Protocol — modelcontextprotocol.ioAccessed 2026-05-24PrimaryOverview — Model Context Protocol (specification 2025-06-18)Accessed 2026-10-04PrimaryDonating the Model Context Protocol and establishing the Agentic AI FoundationAccessed 2026-10-04

A technology that detects multiple simultaneous points of contact on a screen or input surface and treats them as input. It was researched in the 1980s at the University of Toronto (Nimish Mehta in 1982; Lee, Buxton and Smith in 1985) and at Bell Labs (1984). In 2007 Apple promoted the iPhone's 'multi-touch display', bringing multi-finger gestures such as pinch-to-zoom to a mass-market phone.

Appears inThe Original iPhone Goes on Sale · MacBook Air — The 'World's Thinnest Notebook' Without an Optical Drive · The First iPad — and the Two-Month Wait Outside the US · iPhone 6s — 3D Touch and the Pressure of a Finger · The First iPad Pro — A 12.9-inch iPad and the Apple Pencil

SourcesTertiaryMulti-touch — WikipediaAccessed 2026-05-23SecondaryMulti-Touch Systems that I Have Known and LovedAccessed 2026-10-04PrimaryApple Reinvents the Phone with iPhoneAccessed 2026-10-04

A computational model inspired by networks of biological neurons. Layers of artificial neurons connected by weighted links transform inputs into outputs. Building on McCulloch and Pitts's 1943 mathematical model, Rosenblatt's 1958 perceptron and the 1986 backpropagation paper by Rumelhart, Hinton and Williams, it developed rapidly as deep learning from the 2010s, as GPU-based training spread.

See also: Deep Learning · Backpropagation · Transformer

Appears inThe Perceptron Unveiled — Frank Rosenblatt · Samuel's Checkers Paper — The 1959 Experiment with "Machine Learning" in Its Title · The Backpropagation Paper — Rumelhart, Hinton and Williams Show Hidden Units Can Learn · LeNet — The Convolutional Network That Read Zip Codes and Checks · TD-Gammon — The Self-Taught Backgammon Program That Changed Human Play · Deep Belief Nets — In 2006, Deep Neural Networks Learned Again · AlexNet — The Deep-Learning Era Begins · word2vec — Words as Vectors That Carry Meaning

SourcesTertiaryArtificial neural network — WikipediaAccessed 2026-05-24Secondary1943 | Timeline of Computer HistoryAccessed 2026-10-04SecondaryProfessor's perceptron paved the way for AI – 60 years too soonAccessed 2026-10-04PrimaryLearning representations by back-propagating errorsAccessed 2026-10-04PrimaryImageNet Classification with Deep Convolutional Neural NetworksAccessed 2026-10-04

A dedicated processor designed to run neural-network inference at low power. In 2017 Apple's A11 Bionic shipped with a Neural Engine and Huawei's Kirin 970 with a dedicated NPU, and Qualcomm offers its Hexagon NPU. NPUs power smartphone features such as face recognition, and the Copilot+ PCs Microsoft announced in 2024 carry NPUs rated at 40+ TOPS.

See also: Graphics Processing Unit (GPU) · Tensor Processing Unit (TPU) · Application-Specific Integrated Circuit (ASIC)

Appears iniPhone X — The Home Button Goes, Face ID Arrives · iPhone XS and XR — Every New Model Gets Face ID · Apple M1 — The Mac Moves to Apple's Own ARM Chip · Copilot+ PC and Recall — On-Device AI and an Instant Retreat · macOS Sequoia and Apple Intelligence — AI Built Into the OS

SourcesTertiaryAI accelerator — WikipediaAccessed 2026-05-24PrimaryThe future is here: iPhone XAccessed 2026-10-04SecondaryHuawei Introduces Kirin 970 Mobile SoC with Built-in Neural Processing Unit for Artificial Intelligence ApplicationsAccessed 2026-10-04PrimaryUnlocking on-device generative AI with an NPU and heterogeneous computingAccessed 2026-10-04PrimaryIntroducing Copilot+ PCsAccessed 2026-10-04

NVM Express: a host controller interface designed for non-volatile memory such as PCI Express-attached SSDs. Version 1.0 of the specification was released on March 1, 2011. Unlike legacy interfaces built around hard-disk assumptions, it supports up to 64K queues with up to 64K commands per queue, and it targets both enterprise and client systems.

See also: Solid-State Drive (SSD) · RAM vs ROM · ARM vs x86

SourcesTertiaryNVM Express — WikipediaAccessed 2026-05-24PrimaryNVMe OverviewAccessed 2026-10-04

An authorization framework that lets a third-party application act on a user's behalf through an access token with a defined scope and lifetime, instead of the user's password or other credentials. OAuth Core 1.0 was published in December 2007, and OAuth 2.0 (RFC 6749, October 2012) replaced and obsoleted OAuth 1.0 (RFC 5849). The OpenID Connect authentication protocol is also built on OAuth 2.0.

See also: JSON Web Token (JWT) · HTTPS · Public-Key Cryptography

SourcesTertiaryOAuth — WikipediaAccessed 2026-05-24PrimaryOAuth Core 1.0Accessed 2026-10-04PrimaryRFC 6749: The OAuth 2.0 Authorization FrameworkAccessed 2026-10-04PrimaryHow OpenID Connect WorksAccessed 2026-10-04

A programming paradigm that bundles data and procedures into 'objects' and structures them through encapsulation, inheritance and late binding. Ole-Johan Dahl and Kristen Nygaard's Simula 67 (1967) first unified classes, inheritance and dynamic object creation, and in the 1970s Alan Kay and colleagues at Xerox PARC developed the ideas further in Smalltalk. It became mainstream through C++ (first commercial release 1985) and Java (JDK 1.0, 1996).

See also: Functional Programming · Type System · Garbage Collection (GC)

Appears inSimula 67 — Classes and Inheritance Enter a Programming Language · Smalltalk-80 — The Canonical Object-Oriented Language · C++ Ships Commercially — Bjarne Stroustrup · Java Announced — Sun Microsystems · JavaScript Announced — Born of a Ten-Day Prototype · Ruby 0.95 Released — Yukihiro Matsumoto

SourcesTertiaryObject-oriented programming — WikipediaAccessed 2026-05-24SecondaryMilestones:Object-Oriented Programming, 1961-1967Accessed 2026-10-04SecondaryAlan KayAccessed 2026-10-04SecondaryC++ History CollectionAccessed 2026-10-04SecondaryJava 1.0 Turns 25Accessed 2026-10-04

Organic Light-Emitting Diode: a display technology in which thin layers of carbon-based (organic) material emit light when current passes through them. Being self-emitting, OLED needs no backlight as LCD does, can switch individual pixels on or off, and can be made very thin. In smartphones, Samsung used a Super AMOLED panel in the first Galaxy S in 2010, and Apple brought OLED to the iPhone for the first time with the iPhone X in 2017.

See also: Retina Display

Appears inSamsung Galaxy S Launches — The Android Empire's Real Beginning · The First iPad Pro — A 12.9-inch iPad and the Apple Pencil · PlayStation VR — Virtual Reality Plugged Into a PS4, and Its Successor PS VR2 · iPhone X — The Home Button Goes, Face ID Arrives · iPhone XS and XR — Every New Model Gets Face ID · What Is Steam Deck? Valve's Handheld Gaming PC · Apple Vision Pro — Apple's $3,499 'Spatial Computer'

SourcesTertiaryOLED — WikipediaAccessed 2026-05-23SecondaryHow OLEDs WorkAccessed 2026-10-04PrimaryFrom AMOLED to Space Zoom: Looking Back at the Galaxy S Series’ History of InnovationAccessed 2026-10-04PrimaryThe future is here: iPhone XAccessed 2026-10-04

The OS-managed unit of a running program. Each process has its own PID and a separate memory space, and holds its own set of open file descriptors. Unix creates processes with the fork system call and loads a program with exec (today execve(2)); this split dates to early Unix but was not unique to it, having appeared earlier in the Berkeley time-sharing system. A process can contain multiple threads and communicates with others through interprocess communication (IPC) such as pipes.

See also: Thread · Kernel · Virtual Memory

Appears inThe Morris Worm — The First Large-Scale Internet Attack · Windows 95 Released · Mac OS X 10.0 Released · Google Chrome — Announced in a Comic, It Redrew the Browser Map

SourcesTertiaryProcess (computing) — WikipediaAccessed 2026-05-24PrimaryThe Evolution of the Unix Time-sharing SystemAccessed 2026-10-04Primaryfork(2) — Linux manual pageAccessed 2026-10-04Primaryexecve(2) — Linux manual pageAccessed 2026-10-04Primarypthreads(7) — Linux manual pageAccessed 2026-10-04Primarypipe(2) — Linux manual pageAccessed 2026-10-04

The craft of designing inputs (prompts) to elicit desired outputs from LLMs. Known patterns include few-shot prompting with a handful of examples (the GPT-3 paper, 2020), Chain-of-Thought, which has the model write out intermediate reasoning steps (January 2022), ReAct, which interleaves reasoning and actions (October 2022), and role assignment. Model providers also publish official prompting guides.

See also: Large Language Model (LLM) · Hallucination · Foundation Model

Appears inWhat Is Midjourney? The AI Image Generator That Started on Discord · Théâtre D'opéra Spatial — The AI-Made Picture That Won a State Fair · Med-PaLM — A Language Model Clears the Pass Mark on Licensing-Exam Questions · Auto-GPT — The Day 'AI Agent' First Went Viral · Claude Computer Use — An AI That Looks at the Screen and Moves the Mouse

SourcesTertiaryPrompt engineering — WikipediaAccessed 2026-05-24PrimaryLanguage Models are Few-Shot LearnersAccessed 2026-10-04PrimaryChain-of-Thought Prompting Elicits Reasoning in Large Language ModelsAccessed 2026-10-04PrimaryReAct: Synergizing Reasoning and Acting in Language ModelsAccessed 2026-10-04PrimaryPrompting best practicesAccessed 2026-10-04

Asymmetric cryptography in which publicly revealing the encryption key does not reveal the decryption key, easing key distribution and enabling digital signatures. Diffie and Hellman set out the idea in their 1976 paper 'New Directions in Cryptography', and Rivest, Shamir and Adleman devised RSA in 1977. TLS 1.3 uses (elliptic-curve) Diffie-Hellman for key exchange and RSA, ECDSA or EdDSA signatures for authentication. To prepare for quantum computers that could break such schemes, NIST finalized its first post-quantum standards (FIPS 203–205) in August 2024.

See also: End-to-End Encryption (E2EE) · HTTPS · JSON Web Token (JWT)

Appears inNew Directions in Cryptography — Sharing a Secret Without Delivering a Key · A Turing Award for Quantum Information Science — Bennett and Brassard Win for 2025

SourcesTertiaryPublic-key cryptography — WikipediaAccessed 2026-05-24PrimaryNew Directions in CryptographyAccessed 2026-10-04PrimaryA Method for Obtaining Digital Signatures and Public-Key CryptosystemsAccessed 2026-10-04PrimaryRFC 8446: The Transport Layer Security (TLS) Protocol Version 1.3Accessed 2026-10-04PrimaryNIST Releases First 3 Finalized Post-Quantum Encryption StandardsAccessed 2026-10-04

RAM (Random Access Memory) is readable and writable working memory; semiconductor RAM comes chiefly as DRAM and SRAM. DRAM is volatile and must be refreshed periodically to hold its data, whereas SRAM needs no refresh. ROM (Read-Only Memory) is non-volatile memory that keeps its contents without power; early ROMs required a custom mask for each design. Rewritable descendants followed: the UV-erasable EPROM (1971), the electrically erasable EEPROM, and flash memory, which Fujio Masuoka at Toshiba derived from EEPROM in 1980.

See also: Solid-State Drive (SSD) · Virtual Memory · ARM vs x86

Appears inThe Manchester Baby — A Program Runs Out of Memory · The Apple II — A Finished Machine With Colour and Slots · The IBM Personal Computer 5150 — An Open Design and the Market It Lost

SourcesTertiaryRandom-access memory — WikipediaAccessed 2026-05-24TertiaryRead-only memory — WikipediaAccessed 2026-05-24Secondary1970: Semiconductors compete with magnetic coresAccessed 2026-10-04Secondary1971: Reusable semiconductor ROM introducedAccessed 2026-10-04

Malware that encrypts files on a device, rendering them unusable, and demands a ransom in exchange for decryption. The AIDS Trojan (PC Cyborg), mailed on floppy disks in December 1989, is commonly regarded as the first; it encrypted file names rather than the files themselves. In May 2017 WannaCry hit more than 200,000 victims in at least 150 countries, according to Europol. The May 7, 2021 attack on Colonial Pipeline left long lines at gas stations along the US East Coast and showed the threat to critical infrastructure.

See also: End-to-End Encryption (E2EE) · Public-Key Cryptography · Zero Trust

Appears inWannaCry — The Ransomware That Weaponised a Leaked NSA Tool · Windows 10 End of Support · The Canvas Breach — When a Learning Management System Became the Education Sector's Single Point of Failure

SourcesTertiaryRansomware — WikipediaAccessed 2026-05-24PrimaryRansomware 101Accessed 2026-10-04SecondaryRansomware - AIDS TrojanAccessed 2026-10-04SecondaryRansomware attack hit 200,000 victims in 150 countriesAccessed 2026-10-04PrimaryThe Attack on Colonial Pipeline: What We've Learned & What We've Done Over the Past Two YearsAccessed 2026-10-04

A machine-learning paradigm in which an agent learns, through trial and error while interacting with an environment, a policy that maximizes cumulative reward. Sutton and Barto's textbook 'Reinforcement Learning: An Introduction' (1st ed. 1998, 2nd ed. 2018) systematized the field. It became widely known through DeepMind's DQN (a 2015 Nature paper on Atari games) and AlphaGo, which beat Lee Sedol in 2016, and it is also used in reinforcement learning from human feedback (RLHF) to make language models follow instructions.

See also: Supervised Learning · Unsupervised Learning · Large Language Model (LLM)

Appears inTD-Gammon — The Self-Taught Backgammon Program That Changed Human Play · AlphaGo Defeats Lee Sedol · AlphaZero — Mastering Chess, Shogi and Go from the Rules Alone · Anthropic Releases Claude — A Safety-First Rival LLM · AlphaProof — AI Reaches Silver-Medal Standard at the Maths Olympiad · What Is OpenAI o1? The Arrival of Models That Think Before Answering · The DeepSeek-R1 Shock — A Low-Cost Chinese Reasoning Model Rocks the Market · OpenAI Operator — An AI Agent That Runs Errands in Its Own Browser

SourcesTertiaryReinforcement learning — WikipediaAccessed 2026-05-24PrimaryReinforcement Learning: An Introduction (Second Edition)Accessed 2026-10-04PrimaryFrom Pixels to Actions: Human-level control through Deep Reinforcement LearningAccessed 2026-10-04PrimaryAlphaGoAccessed 2026-10-04PrimaryTraining language models to follow instructions with human feedbackAccessed 2026-10-04

REST (Representational State Transfer) is an architectural style for distributed hypermedia systems defined by Roy Fielding in his 2000 PhD dissertation at the University of California, Irvine. It consists of constraints such as client-server separation, statelessness (each request carries all the information needed to understand it), and a uniform interface built on resource identifiers. A REST API is a Web API following this approach, in which resources identified by URIs are manipulated with HTTP methods such as GET, POST, PUT and DELETE.

See also: HTTP · GraphQL · WebSocket

SourcesTertiaryRepresentational state transfer — WikipediaAccessed 2026-05-24PrimaryArchitectural Styles and the Design of Network-based Software ArchitecturesAccessed 2026-10-04PrimaryRFC 9110: HTTP SemanticsAccessed 2026-10-04

Apple's marketing name for high-resolution displays whose pixels, Apple says, the human eye cannot distinguish at a normal viewing distance. It debuted on the iPhone 4 (326 ppi), announced in June 2010, and later spread to the iPad (March 2012), the MacBook Pro (June 2012) and the iMac (the Retina 5K model, October 2014).

See also: OLED (Organic LED)

Appears inThe First MacBook Pro — The First Mac Notebook on Intel · MacBook Air — The 'World's Thinnest Notebook' Without an Optical Drive · iPhone 4 — Retina Display and 'Antennagate' · The First Apple Watch — Seven Months From Unveiling to Sale, Japan on Day One · The First iPad Pro — A 12.9-inch iPad and the Apple Pencil · The Original iPhone SE — A 6s Heart in a Four-Inch Body

SourcesTertiaryRetina display — WikipediaAccessed 2026-05-23PrimaryApple Presents iPhone 4Accessed 2026-10-04PrimaryApple Launches New iPadAccessed 2026-10-04PrimaryApple Introduces All New MacBook Pro with Retina DisplayAccessed 2026-10-04PrimaryApple Introduces 27-inch iMac with Retina 5K DisplayAccessed 2026-10-04

A technique in which a language model retrieves relevant documents from an external collection at generation time and conditions its output on them. Patrick Lewis and colleagues at Facebook AI Research (now Meta) proposed it in 2020, pairing a generator with a pre-trained neural retriever over a dense vector index of Wikipedia. It aims to address problems of models that hold knowledge only in their parameters, such as updating that knowledge and showing provenance.

See also: Embedding · Large Language Model (LLM) · Hallucination

Appears inJapan's Copyright Act Article 30-4 — Training AI Without Permission · Perplexity AI Goes Public — A Search Engine That Answers, with Sources · pgvector Comes of Age — Putting AI Vector Search inside the RDB · MCP — How One Company's Connector Became the Industry's Standard

SourcesTertiaryRetrieval-augmented generation — WikipediaAccessed 2026-05-24PrimaryRetrieval-Augmented Generation for Knowledge-Intensive NLP TasksAccessed 2026-10-04

An open instruction-set architecture (ISA) begun at UC Berkeley in May 2010 by Prof. Krste Asanović and graduate students Yunsup Lee and Andrew Waterman. It is royalty-free and allows extensions to be added. Besides SiFive, which the three co-founded, adopters include Western Digital (its SweRV core for SSD controllers) and Alibaba's T-Head (the XuanTie C910), with uses ranging from embedded controllers to high-performance cores.

See also: ARM vs x86 · Graphics Processing Unit (GPU) · Application-Specific Integrated Circuit (ASIC)

Appears inThe First ARM Processor

SourcesPrimaryAbout RISC-V — RISC-V InternationalAccessed 2026-10-04PrimaryAbout SiFiveAccessed 2026-10-04PrimaryWestern Digital Delivers New Innovations to Drive Open Standard Interfaces and RISC-V Processor DevelopmentAccessed 2026-10-04SecondaryAlibaba/T-HEAD's Xuantie C910Accessed 2026-10-04TertiaryRISC-V — WikipediaAccessed 2026-05-24

A cloud execution model in which developers deploy code as functions or containers without managing servers or scaling, paying only for what they use (for AWS Lambda, by number of requests and execution duration). AWS Lambda (announced November 2014) is the best-known example; Cloudflare Workers (announced 2017, running at edge locations) and Google Cloud Run (announced 2019, for containers) followed.

See also: Kubernetes · Microservices · Edge Computing

Appears inAWS Lambda — One Function, Billed by the Tenth of a Second · AWS Bedrock — Foundation Models as a Cloud Service

SourcesPrimaryIntroducing AWS LambdaAccessed 2026-10-04PrimaryAWS Lambda PricingAccessed 2026-10-04PrimaryIntroducing Cloudflare WorkersAccessed 2026-10-04PrimaryAnnouncing Cloud Run, the newest member of our serverless compute stackAccessed 2026-10-04TertiaryServerless computing — WikipediaAccessed 2026-10-04

Solid-state storage using NAND flash memory. An early example was a 20 MB flash module emulating a hard disk that SunDisk (later SanDisk) prototyped for IBM in 1991. SSDs spread into laptops from the late 2000s, and their interface moved from SATA to PCIe/NVMe, today reaching read/write speeds of several gigabytes per second.

See also: NVMe · RAM vs ROM · ARM vs x86

Appears inMacBook Air — The 'World's Thinnest Notebook' Without an Optical Drive · PlayStation 5 Released

SourcesSecondary1991: Solid State Drive module demonstrated — Computer History MuseumAccessed 2026-10-04SecondarySanjay Mehrotra — ETHWAccessed 2026-10-04PrimaryAbout - NVM ExpressAccessed 2026-10-04TertiarySolid-state drive — WikipediaAccessed 2026-10-04

A machine-learning approach that trains a model on data consisting of input–label pairs. Its two main forms are classification (e.g. spam filtering, image recognition) and regression (e.g. price prediction). Large labeled datasets such as ImageNet (presented at CVPR 2009) accelerated the rise of deep learning.

See also: Unsupervised Learning · Reinforcement Learning · Deep Learning

Appears inThe Perceptron Unveiled — Frank Rosenblatt · LeNet — The Convolutional Network That Read Zip Codes and Checks · The DARPA Grand Challenge — The Day Driverless Cars Finished the Desert · ImageNet — A Dictionary of Images Built by 50,000 People · The Netflix Prize — $1 Million for Better Recommendations, and the Limits of Anonymity · IDx-DR — The FDA Authorizes an AI That Screens Without a Doctor Reading the Image · EndoBRAIN — The Approval Billed as Japan's First AI Medical Device · Validating the Epic Sepsis Model — How a Prediction Running in Hundreds of Hospitals Actually Performed

SourcesPrimaryImageNet — AboutAccessed 2026-10-04Primary1. Supervised learning — scikit-learn documentationAccessed 2026-10-04TertiarySupervised learning — WikipediaAccessed 2026-10-04

The entry point through which user-space processes request kernel services such as file I/O, networking and process control. On x86-64 Linux the syscall instruction switches to kernel mode to invoke calls such as open(2), read(2), write(2) and fork(2). Most programs invoke them indirectly through C library wrapper functions.

See also: Kernel · Process · Virtual Memory

SourcesPrimaryintro(2) — Linux manual pageAccessed 2026-10-04Primarysyscall(2) — Linux manual pageAccessed 2026-10-04Primarysyscalls(2) — Linux manual pageAccessed 2026-10-04TertiarySystem call — WikipediaAccessed 2026-05-24

The foundational protocol suite of the Internet. The 1974 Cerf–Kahn paper described a single Transmission Control Program; the split into TCP and IP came in 1978. ARPANET cut over from NCP on Flag Day, 1 January 1983. IP delivers packets to their destination, while TCP provides reliable, ordered communication on top — in a four-layer model.

See also: Domain Name System (DNS) · HTTP · HTTP/3 and QUIC

Appears inARPANET's First Message · 1BSD Released — The Berkeley Fork of UNIX Begins · The Switch to TCP/IP — 'Flag Day' · The Domain Name System Specified · NSFNET — The Decade in Which a Research Network Became the Public Internet

SourcesPrimaryA Protocol for Packet Network Intercommunication — V. G. Cerf and R. E. Kahn, IEEE Transactions on Communications, May 1974Accessed 2026-10-04PrimaryRFC 801: NCP/TCP Transition Plan — J. Postel (ISI), November 1981Accessed 2026-08-03PrimaryRFC 1122: Requirements for Internet Hosts -- Communication LayersAccessed 2026-10-04SecondaryBrief History of the Internet — Internet SocietyAccessed 2026-10-04SecondaryTCP to TCP/IP 1976-1979 — The History of Computer CommunicationsAccessed 2026-10-04TertiaryInternet protocol suite — WikipediaAccessed 2026-05-24

Google's in-house ASIC for machine learning. Deployed in its data centres since 2015, it was revealed at Google I/O in May 2016, when Google disclosed that TPUs had powered AlphaGo in its matches against Lee Sedol. Its matrix unit is built as a systolic array, it is offered externally as Cloud TPU, and Gemini 1.0 was trained on TPU v4 and v5e.

See also: Graphics Processing Unit (GPU) · Neural Processing Unit (NPU) · Application-Specific Integrated Circuit (ASIC)

Appears inGoogle Reveals the TPU — Its Own Chip for Machine Learning · Google Neural Machine Translation — The Overnight Jump in Translation · AlphaZero — Mastering Chess, Shogi and Go from the Rules Alone · What Is Google Gemini? Google's Answer to ChatGPT, From Bard Onward

SourcesPrimaryGoogle supercharges machine learning tasks with TPU custom chipAccessed 2026-10-04SecondaryGoogle's Tensor Processing Unit could advance Moore's Law 7 years into the futureAccessed 2026-10-04PrimaryIn-Datacenter Performance Analysis of a Tensor Processing UnitAccessed 2026-10-04PrimaryIntroduction to Cloud TPUAccessed 2026-10-04PrimaryIntroducing Gemini: our largest and most capable AI modelAccessed 2026-10-04TertiaryTensor Processing Unit — WikipediaAccessed 2026-05-24

The smallest unit of execution that an operating system's scheduler can manage independently. Threads in the same process share its memory (code, heap and global variables) while keeping their own stack and register state. In 1995 the IEEE standardized the POSIX Threads (pthreads) API as POSIX.1c. On multicore systems, multiple threads can run simultaneously on separate cores.

See also: Process · Kernel · async / await

Appears inRust 1.0 Released

SourcesPrimaryIEEE 1003.1c-1995 — Standard for Information Technology--Portable Operating System Interface (POSIX) - System Application Program Interface (API) Amendment 2: Threads Extension (C Language)Accessed 2026-10-04Primarypthreads(7) — Linux manual pageAccessed 2026-10-04TertiaryThread (computing) — WikipediaAccessed 2026-10-04

The unit in which LLMs process text. Text is usually split into words or smaller subword pieces using methods such as BPE or SentencePiece. Google says a token is about four characters for its Gemini models, but token counts for the same content vary widely by language — up to 15-fold between translations in one study. API pricing and context windows are measured in tokens.

See also: Large Language Model (LLM) · Embedding · Transformer

Appears inGoogle Neural Machine Translation — The Overnight Jump in Translation · BERT Released — Bidirectional Transformer Pretraining · GPT-3 Announced · What Was DALL·E? The Day OpenAI Showed an AI Drawing From Words · Meta's LLaMA — A Research Release, a Leak, and the Open-Weights Debate · What Is OpenAI o1? The Arrival of Models That Think Before Answering · Claude Code — A Coding Agent That Lives in the Terminal · What Is GPT-5? OpenAI's Self-Routing System and Its Rocky Launch

SourcesPrimaryUnderstand and count tokens — Gemini APIAccessed 2026-10-04PrimaryNeural Machine Translation of Rare Words with Subword UnitsAccessed 2026-10-04PrimarySentencePiece: A simple and language independent subword tokenizer and detokenizer for Neural Text ProcessingAccessed 2026-10-04PrimaryLanguage Model Tokenizers Introduce Unfairness Between LanguagesAccessed 2026-10-04

Apple's fingerprint authentication system. Introduced on the iPhone 5s in 2013, it built a capacitive sensor into the home button so a single touch could unlock the phone and approve purchases in the iTunes Store, App Store and other stores. It is also used to authorise Apple Pay, announced in 2014. The iPhone X (2017) dropped the home button and introduced Face ID, but Touch ID remains on the current iPad Air, iPad and iPad mini, built into the top button.

See also: Face ID

Appears inThe First MacBook Pro — The First Mac Notebook on Intel · MacBook Air — The 'World's Thinnest Notebook' Without an Optical Drive · iPhone 5s — Touch ID and the 64-bit A7 · iPhone 6s — 3D Touch and the Pressure of a Finger · The Original iPhone SE — A 6s Heart in a Four-Inch Body · iPhone X — The Home Button Goes, Face ID Arrives

SourcesPrimaryApple Announces iPhone 5s—The Most Forward-Thinking Smartphone in the WorldAccessed 2026-10-04PrimaryApple Announces Apple PayAccessed 2026-10-04PrimaryThe future is here: iPhone XAccessed 2026-10-04PrimaryiPad - Compare Models - AppleAccessed 2026-10-04TertiaryTouch ID — WikipediaAccessed 2026-05-23

A deep-learning architecture introduced by Google researchers in the 2017 paper 'Attention Is All You Need'. It processes sequence data using self-attention alone, dispensing with recurrence and convolutions. It underpins modern large language models such as BERT and GPT-4, and Nature's 2025 tally placed the paper seventh among the 25 most-cited scientific papers of the 21st century.

See also: Neural Network · Attention Mechanism · Large Language Model (LLM)

Appears in"Attention Is All You Need" — The Transformer Paper · BERT Released — Bidirectional Transformer Pretraining · GPT-2 Held Back — Staged Release and AI Publication Norms · GPT-3 Announced · What Was DALL·E? The Day OpenAI Showed an AI Drawing From Words · NVIDIA H100 — The Backbone of AI Compute · GPT-4 — Going Multimodal · AlphaGeometry — An AI That Learned Olympiad Geometry Without Human Examples

SourcesPrimaryAttention Is All You Need — arXiv 1706.03762Accessed 2026-10-04PrimaryTransformer: A Novel Neural Network Architecture for Language UnderstandingAccessed 2026-10-04PrimaryBERT: Pre-training of Deep Bidirectional Transformers for Language UnderstandingAccessed 2026-10-04PrimaryGPT-4 Technical ReportAccessed 2026-10-04SecondaryFour papers authored by U of T scholars among the 25 most cited of the 21st century: NatureAccessed 2026-10-04TertiaryTransformer (deep learning architecture) — WikipediaAccessed 2026-05-24

A mechanism that assigns types to values and expressions to detect operations whose types do not match. Static typing checks before execution, typically at compile time (Haskell, Rust, TypeScript); dynamic typing checks at run time (Python, JavaScript). Gradual typing, which mixes the two within one language, was proposed by Siek and Taha in 2006, and its ideas appear in TypeScript (released 2012) and in Python's type hints (PEP 484, Python 3.5).

See also: Object-Oriented Programming (OOP) · Functional Programming · Garbage Collection (GC)

Appears inSimula 67 — Classes and Inheritance Enter a Programming Language · The C Language — Dennis Ritchie · C++ Ships Commercially — Bjarne Stroustrup · TypeScript Announced — Microsoft · Rust 1.0 Released

SourcesPrimaryJeremy G. Siek — home page (Gradual Typing for Functional Languages, Scheme Workshop 2006)Accessed 2026-10-04PrimaryTypeScript: JavaScript Development at Application ScaleAccessed 2026-10-04PrimaryPEP 484 – Type HintsAccessed 2026-10-04Primarytyping — Support for type hintsAccessed 2026-10-04TertiaryType system — WikipediaAccessed 2026-10-04

A machine-learning paradigm that discovers structure and patterns in unlabeled data. Common forms include clustering (k-means), dimensionality reduction (PCA, t-SNE) and generative models (GANs, VAEs). Some researchers count the self-supervised learning used in LLM pretraining as a form of unsupervised learning, while others treat it as a separate framework.

See also: Supervised Learning · Deep Learning · Embedding

Appears inDeep Belief Nets — In 2006, Deep Neural Networks Learned Again

SourcesPrimary2. Unsupervised learning — scikit-learn documentationAccessed 2026-10-04PrimarySelf-Supervised Learning: The Dark Matter of IntelligenceAccessed 2026-10-04TertiaryUnsupervised learning — WikipediaAccessed 2026-10-04

A USB connector standard whose plug is reversible. The USB 3.0 Promoter Group announced the completed specification in August 2014 and transferred it to the USB-IF. Power delivery, originally up to 100 W, was raised to 240 W by USB PD 3.1 (2021), and data rates reached 80 Gbps with USB4 Version 2.0 (2022); video output is carried via Alternate Modes such as DisplayPort Alt Mode. From 28 December 2024 the EU requires USB-C as the charging port on mobile phones, tablets and similar devices.

Appears inThe First iPad Pro — A 12.9-inch iPad and the Apple Pencil · iPhone 15 — The Move to USB-C

SourcesPrimaryUSB 3.0 Promoter Group Announces USB Type-C Connector Ready for ProductionAccessed 2026-10-04PrimaryUSB Charger (USB Power Delivery)Accessed 2026-10-04PrimaryUSB-IF Announces Publication of New USB4® Specification to Enable USB 80Gbps PerformanceAccessed 2026-10-04PrimaryVESA Brings DisplayPort to New USB Type-C ConnectorAccessed 2026-10-04PrimaryEU common charger rules: Power all your devices with a single chargerAccessed 2026-10-04TertiaryUSB-C — WikipediaAccessed 2026-05-23

A mechanism that combines physical memory with secondary storage (such as disk) to give processes a virtual address space independent of physical placement. It was introduced on the University of Manchester's Atlas computer, which came online in 1962. Its building blocks are address translation via the MMU and page tables, TLBs, demand paging and swapping, and most modern general-purpose operating systems give each process its own address space.

See also: Kernel · Process · System Call

Appears inWindows 3.0 — The First Commercially Successful Windows

SourcesSecondary1962 — Timeline of Computer History — Computer History MuseumAccessed 2026-10-04TertiaryVirtual memory — WikipediaAccessed 2026-10-04

A protocol for two-way communication between client and server over a single TCP connection. Standardized by the IETF as RFC 6455 in December 2011, it establishes the connection with an HTTP upgrade handshake and underpins real-time web applications such as chat, gaming, stock tickers and simultaneous editing.

See also: HTTP · REST API · HTTP/3 and QUIC

SourcesPrimaryRFC 6455: The WebSocket ProtocolAccessed 2026-10-04TertiaryWebSocket — WikipediaAccessed 2026-05-24

A security model that abandons the assumption that being inside the network makes access safe, and instead authenticates and authorises each access by users and devices to resources. Proposed by Forrester's John Kindervag in a 2010 report and exemplified by Google's BeyondCorp (described in a 2014 paper), it was systematized in NIST SP 800-207 (August 2020); in the U.S., Executive Order 14028 (May 2021) and OMB memorandum M-22-09 (January 2022) directed federal agencies to move to zero trust.

See also: End-to-End Encryption (E2EE) · Public-Key Cryptography · OAuth

Appears inSolarWinds — Russian SVR Poisons the Update Pipeline Itself

SourcesPrimarySP 800-207, Zero Trust ArchitectureAccessed 2026-10-04PrimaryM-22-09: Moving the U.S. Government Toward Zero Trust Cybersecurity PrinciplesAccessed 2026-10-04PrimaryBeyondCorp: A New Approach to Enterprise SecurityAccessed 2026-10-04Secondary15 Years of Zero TrustAccessed 2026-10-04TertiaryZero trust security model — WikipediaAccessed 2026-05-24