Glossary

An integrated circuit custom-designed for a specific application. ASICs outperform general-purpose CPUs and GPUs in speed and efficiency but carry high design cost and time. Examples include Bitcoin miners, Google TPUs, the coprocessors in Apple Silicon and networking-equipment ASICs — all gaining new attention in the AI era.

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

The two dominant modern CPU architectures. x86 is a CISC lineage from Intel's 8086 (1978) that powers PCs and servers. ARM is a RISC architecture developed by Acorn (UK) in 1985 that dominates 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)

AI capable of performing any intellectual task a human can, across domains rather than narrow specialties — the counterpart of Narrow AI. Stated as a primary goal by OpenAI, DeepMind and Anthropic. Timelines, definitions and alignment safety remain actively debated.

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

A language construct that lets asynchronous code be written in a synchronous-looking style. Pioneered in C# 5.0 (2012), then adopted by Python 3.5 (2015), JavaScript ES2017, Rust, Swift and Kotlin. It eliminated callback hell and transformed the readability of I/O-bound server and UI code.

See also: Thread · Functional Programming · Type System

A mechanism that dynamically weights how much each element of a sequence should attend to other elements. Introduced by Bahdanau et al. (2014) for machine translation, it became central as self-attention in the 2017 Transformer. Relevance is computed via dot products of Query, Key and Value vectors.

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

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. Rediscovered and popularized in 1986 by Rumelhart, Hinton and Williams, it made modern deep learning possible.

See also: Neural Network · Gradient Descent · Deep Learning

A lightweight virtualization technique that isolates processes using Linux kernel namespaces and cgroups to provide independent runtime environments. Following FreeBSD jails (2000) and Solaris Zones (2004), Docker (2013) transformed the developer experience and laid the foundation for the Kubernetes era.

See also: Kubernetes · Microservices · Kernel

A network of edge servers worldwide that replicate and cache content, delivering it from the nearest location to reduce latency and bandwidth cost. Pioneered by Akamai, spun out of MIT in 1998, the market is now led by Cloudflare, Fastly and AWS CloudFront, expanding into edge computing.

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

A mechanism that, within the browser's Same-Origin Policy, permits explicitly allowed cross-origin resource access. Recommended by the W3C in 2014, it is controlled via response headers such as `Access-Control-Allow-Origin` and preflight OPTIONS requests.

See also: HTTP · REST API · HTTPS

Machine learning based on neural networks with many layers. Triggered by Hinton et al.'s 2006 work and AlexNet's 2012 ImageNet victory, it drove the third AI boom and now underpins image recognition, speech, NLP and generative AI such as GPT and diffusion models.

See also: Neural Network · Backpropagation · Transformer

A distributed hierarchical database that translates domain names (example.com) into IP addresses. Designed by Paul Mockapetris in 1983 (RFCs 882/883) to replace ARPANET's centralized hosts.txt, it is core Internet infrastructure, extended over time by DNSSEC and DNS over HTTPS (DoH).

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

A distributed computing model that places compute and processing near where data is produced — user devices, IoT, retail stores — improving latency, bandwidth and privacy. Building on CDNs, products such as Cloudflare Workers, AWS Lambda@Edge and Fastly Compute@Edge are gaining importance with 5G and on-edge AI inference.

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

Representations of words, sentences or images as high-dimensional vectors that preserve semantic distance. Word2Vec (2013) pioneered language embeddings; today models such as OpenAI text-embedding-3, Cohere and Sentence-BERT power RAG pipelines and semantic search.

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

An encryption scheme in which only the sender and recipient hold decryption keys, so intermediaries — including the service provider — cannot read the content. Notable examples include PGP (1991) and the Signal Protocol (2013), implemented in WhatsApp (which made E2EE default in 2016), iMessage and Signal, with ongoing debate over law-enforcement access.

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

Apple's facial recognition system, introduced on the 2017 iPhone X. It projects approximately 30,000 invisible infrared dots onto the user's face and reads their deformation to reconstruct a 3D face mesh, against which it authenticates. Apple's stated false-accept rate is roughly one in a million (Touch ID's was one in fifty thousand).

See also: Touch ID

A logic device whose circuit configuration can be programmed after manufacturing. Xilinx launched the first FPGA, the XC2064, in 1985. AMD (which acquired Xilinx) and Intel (which acquired Altera) are the leading vendors; FPGAs are used in telecom, HFT, data centers and AI inference, and serve as prototyping platforms for ASICs.

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

The process of adapting a pretrained model to a specific task or domain via additional training. Beyond full-parameter updates, parameter-efficient methods such as LoRA (2021) now dominate (PEFT). Preference tuning via RLHF or DPO is also a form of fine-tuning.

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

Large general-purpose models pretrained self-supervised on broad data and adaptable to many downstream tasks. The term was coined in 2021 by Stanford HAI. It covers GPT, Claude, Gemini, Llama and Stable Diffusion, and forms the backbone of today's AI industry.

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

A paradigm that treats computation as the evaluation of mathematical functions, minimizing state and side effects. From LISP (1958), the lineage runs through Haskell (1990), OCaml, Erlang, Scala and Clojure. Higher-order functions and immutability are now common in mainstream languages such as JavaScript, Python and Rust.

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

A mechanism by which a runtime automatically reclaims memory occupied by objects no longer reachable from the program. Invented by John McCarthy for LISP in 1959, it is used in Java, C#, Go, Python and JavaScript. Rust contrasts by achieving memory safety without GC through its ownership model.

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

An optimization algorithm that updates weights step by step in the opposite direction of the loss function's gradient to find a minimum. Originating with Cauchy in 1847, modern deep learning uses variants such as Stochastic Gradient Descent (SGD) and Adam (2014).

See also: Backpropagation · Neural Network · Deep Learning

A parallel processor specialized for graphics. NVIDIA popularized the term 'GPU' with the GeForce 256 in 1999. CUDA (2007) opened the GPU to general-purpose computing (GPGPU), and from the late 2010s GPUs became the central engine for deep learning and LLM training, anchoring the AI silicon market.

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

A query language and runtime for APIs developed internally at Facebook in 2012 and open-sourced in 2015. Clients declaratively request only the fields they need, addressing REST's over- and under-fetching problems. Adopted at scale by GitHub, Shopify and Netflix.

See also: REST API · HTTP · WebSocket

A phenomenon in which LLMs generate plausible-sounding but factually incorrect or fabricated information. Caused by gaps in training data and the probabilistic nature of generation, it poses risks in medical and legal domains. Mitigations include RAG, citation grounding and factuality-focused reinforcement learning.

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

The application-layer protocol that underpins the Web. Tim Berners-Lee designed the original at CERN in 1991. HTTP/1.1 (1997, RFC 2068) saw long use; HTTP/2 (2015) introduced binary multiplexing, and HTTP/3 (2022, RFC 9114) moved to a QUIC-based transport.

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

QUIC, a UDP-based transport developed by Google from 2012 and standardized as RFC 9000 by the IETF in 2021, and HTTP/3 (RFC 9114, 2022) which runs over it. With integrated TLS 1.3, 0-RTT and connection migration, it outperforms TCP/TLS on mobile networks.

See also: HTTP · TCP/IP · HTTPS

HTTP encrypted with TLS (formerly SSL). Introduced by Netscape alongside SSL in 1994, it became the de facto default after Let's Encrypt began issuing free certificates in 2015 and Chrome pushed HTTPS-everywhere; today it covers over 95% of Web traffic.

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

The three-layer model of cloud services: IaaS (infrastructure like VMs and storage, e.g. AWS EC2), PaaS (runtime environments, e.g. Heroku, Google App Engine), and SaaS (finished applications, e.g. Salesforce, Gmail). NIST formalized these definitions in SP 800-145 in 2011.

See also: Serverless · Kubernetes · Edge Computing

A compact, signed (and optionally encrypted) token format that carries JSON claims, standardized as RFC 7519 in 2015. Encoded as Base64URL header.payload.signature, JWTs are widely used in OAuth 2.0, OpenID Connect and stateless API authentication.

See also: OAuth · HTTPS · Public-Key Cryptography

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

See also: Process · System Call · Virtual Memory

An open-source container-orchestration platform released by Google in 2014, derived from its internal Borg system. As the flagship CNCF project, with declarative APIs, autoscaling and self-healing, it has become the de facto standard for modern cloud-native infrastructure.

See also: Container · Microservices · Serverless

Transformer-based language models with billions to trillions of parameters. The term gained traction after GPT-3 (2020, 175B parameters), and ChatGPT (2022) brought LLMs to mainstream awareness. Examples include Claude, Gemini and Llama, used for dialogue, code generation and reasoning.

See also: Transformer · Foundation Model · Token

Apple's magnet-based connection standard. Originally the name (2006–2019) of a MacBook power connector designed to detach safely when pulled. Apple reused the name with iPhone 12 in 2020 for a new wireless-charging-and-accessory-attachment standard: magnets embedded in the phone's back snap onto cases, card holders, and chargers.

An architecture in which an application is built as a suite of small independently deployable services, each with its own data and API. Martin Fowler and James Lewis formalized the definition in 2014. Pioneered in practice by Netflix and Amazon, microservices became the standard pattern of the Kubernetes era.

See also: Kubernetes · Container · Serverless

An open standard protocol announced by Anthropic in November 2024 for connecting LLMs to external tools and data sources. Built on JSON-RPC, it unifies resources, tools and prompts under one interface. Early adopters include Claude Desktop and Cursor, making MCP a de facto standard for AI-agent integrations.

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

A technology that detects and treats multiple simultaneous points of contact on a screen as input. Researched by Bill Buxton and others in the 1980s, it was not adopted at mass-consumer scale until the iPhone in 2007. It enables pinch, zoom, rotation, and other gestures performed with multiple fingers.

A computational model inspired by biological neurons in the brain. Layers of artificial neurons connected by weighted edges transform inputs into outputs. Starting with the 1958 Perceptron and revived by backpropagation (rediscovered 1986) and GPU-accelerated training, it became today's Deep Learning from the 2010s.

See also: Deep Learning · Backpropagation · Transformer

Dedicated processors optimized for neural-network inference. Examples include the Apple Neural Engine (A11 Bionic, 2017), Huawei Kirin and Qualcomm Hexagon. NPUs run on-device AI — face recognition, image generation, speech — on smartphones and laptops at low power.

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

A protocol designed for communicating with SSDs over the PCIe bus. NVMe 1.0 (2011) removed the legacy constraints of SATA/AHCI and unlocked SSDs' real performance through massive command parallelism (up to 64K queues x 64K commands). It is now standard from data centers to consumer NVMe SSDs.

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

An authorization framework that delegates access to third-party applications without sharing the user's password. OAuth 1.0 was drafted in 2007, and OAuth 2.0 (RFC 6749, 2012) is the current standard. It powers 'Sign in with' flows for Google, Facebook and GitHub, and mobile-app API integrations.

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

A programming paradigm that bundles data and behavior into 'objects' structured by encapsulation, inheritance and polymorphism. Established by Alan Kay and colleagues with Smalltalk in the 1970s, it became mainstream through C++ (1985) and Java (1995); modern code commonly blends it with functional styles.

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

Organic Light-Emitting Diode—a display technology in which organic compounds emit light directly when current is applied. Unlike LCD, OLED needs no backlight and can switch individual pixels on or off, yielding true blacks, effectively infinite contrast, and thinner or curved form factors. Adopted early in smartphones by Samsung; Apple introduced OLED on the iPhone X in 2017.

See also: Retina Display

The OS-managed unit of a running program. Each process owns a private virtual address space, file descriptors and a PID, isolated from others. The classic model dates to Unix's fork(2)/exec(3) of the 1970s. A process can host multiple threads and communicate with others via IPC.

See also: Thread · Kernel · Virtual Memory

The craft of designing inputs (prompts) to elicit desired outputs from LLMs. Patterns include few-shot, Chain-of-Thought (2022), ReAct and role assignment. It rose as a discipline after ChatGPT's release, though its emphasis is shifting as newer LLMs follow instructions more reliably.

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

Asymmetric cryptography in which data encrypted with a public key can only be decrypted with the corresponding private key. Diffie and Hellman proposed the concept in 1976, and Rivest, Shamir and Adleman published RSA in 1977. It underpins TLS, SSH, PGP and blockchains; today elliptic-curve cryptography (ECC) and post-quantum cryptography (PQC) are central.

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

RAM (Random Access Memory) is fast volatile memory that retains data only while powered — DRAM and SRAM are common variants. ROM (Read-Only Memory) is non-volatile and read-only, retaining data without power. Modern Flash and EEPROM are rewritable derivatives of ROM.

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

Malware that encrypts a victim's files and demands a ransom — typically in cryptocurrency — in exchange for the decryption key. The prototype was the 1989 AIDS Trojan. WannaCry infected ~300,000 machines across 150 countries in 2017, bringing the threat to wide public awareness, and attacks like Colonial Pipeline (2021) made ransomware a critical-infrastructure risk.

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

A machine-learning paradigm where an agent learns a policy by interacting with an environment and maximizing cumulative reward through trial and error. Formalized by Sutton & Barto and popularized by DeepMind's DQN (2015) and AlphaGo (2016), it also underpins RLHF used to align LLMs.

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

A Web API design style proposed by Roy Fielding in his 2000 PhD dissertation, where resources are identified by URIs and manipulated via HTTP methods such as GET, POST, PUT and DELETE. Widely adopted in the late 2000s by Twitter, Amazon S3 and others, it became the de facto standard for Web APIs.

See also: HTTP · GraphQL · WebSocket

Apple's marketing term for high-resolution displays whose pixels are claimed to be indistinguishable to the human retina at typical viewing distance. Introduced on the 2010 iPhone 4 (326 ppi) and later extended to the iPad, MacBook, and Mac mini lines.

See also: OLED (Organic LED)

A technique that retrieves relevant documents from an external knowledge base via vector search at inference time and injects them into the LLM's context to improve accuracy and reduce hallucinations. Proposed by Lewis et al. at Meta in 2020, it is now a standard pattern for enterprise QA and fresh-information access.

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

An open instruction-set architecture (ISA) initiated at UC Berkeley in 2010. Royalty-free and freely extensible, it is adopted by SiFive, Western Digital and Alibaba T-Head, and is rapidly spreading from embedded systems to HPC and AI workloads — drawing attention as an alternative to ARM and x86.

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

A cloud execution model where developers deploy code as functions without managing servers or scaling, paying only for execution time and invocations. AWS Lambda (2014) pioneered the category; Cloudflare Workers, Vercel and Google Cloud Run followed, increasingly converging with edge computing.

See also: Kubernetes · Microservices · Edge Computing

Solid-state storage using NAND flash memory. SanDisk's 20MB SSD in 1991 was an early commercial example. SSDs spread into laptops in the late 2000s, displacing HDDs, and evolved from SATA to PCIe/NVMe — today delivering read/write speeds in the gigabytes-per-second range.

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

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

See also: Unsupervised Learning · Reinforcement Learning · Deep Learning

The interface by 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 privileged mode and invokes calls like open(2), read(2), write(2) and fork(2). It is the OS's API boundary.

See also: Kernel · Process · Virtual Memory

The foundational protocol suite of the Internet. Proposed by Vint Cerf and Bob Kahn in 1974 and adopted Internet-wide on ARPANET's Flag Day, January 1, 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

Google's machine-learning-specific ASIC. The first generation was unveiled at Google I/O 2016 and used in AlphaGo's match operations. Built around a systolic array for fast matrix multiplication, TPUs are offered externally via Google Cloud TPU and power training of Google's large models such as Gemini.

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

The smallest unit of concurrent execution within a process. Threads share the process's virtual memory while maintaining their own stack and register state. Standardized in the 1990s as POSIX Threads (pthreads), they became the fundamental unit of parallelism in the multicore era.

See also: Process · Kernel · async / await

The smallest unit LLMs use to process text. Text is split into subwords using algorithms such as BPE or SentencePiece. As a rule of thumb, ~4 English characters equal one token; one Japanese character is 1–2 tokens. API pricing and context windows are measured in tokens.

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

Apple's fingerprint authentication system. Introduced on the iPhone 5s in 2013, it embedded a capacitive fingerprint sensor in the home button, replacing passcodes for unlock and authorising App Store and Apple Pay transactions with a single touch. Replaced on the iPhone X and later flagships by Face ID, but still in use on current iPads and other Apple devices.

See also: Face ID

A deep-learning architecture published by Google in 2017. Its self-attention mechanism captures long-range dependencies in sequence data efficiently, underpinning modern LLMs such as BERT, GPT and Claude. The foundational paper 'Attention Is All You Need' is one of the most cited works in AI history.

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

A mechanism that assigns types to values and expressions to detect invalid operations, typically at compile time. Systems are broadly classified as static (Haskell, Rust, TypeScript) or dynamic (Python, JavaScript). Since TypeScript (2012), gradual typing has spread to JavaScript and Python.

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

A machine-learning paradigm that discovers structure in unlabeled data. Common forms include clustering (k-means), dimensionality reduction (PCA, t-SNE) and generative models (GANs, VAEs). LLM pretraining, as self-supervised learning, is broadly considered part of this family.

See also: Supervised Learning · Deep Learning · Embedding

A reversible USB connector standard published by the USB Implementers Forum in 2014. A single cable carries up to 100 W of power, up to 40 Gbps of data (USB4 / Thunderbolt), and video output via DisplayPort or HDMI Alternate Modes. Through the 2020s it became the de facto universal connector across smartphones, PCs, and accessories.

A mechanism that combines physical memory with secondary storage to give each process a contiguous virtual address space. Pioneered by the Atlas computer in 1962, it relies on paging via the MMU, TLBs, demand paging and swapping, and is essential to all modern operating systems.

See also: Kernel · Process · System Call

A protocol providing full-duplex real-time communication over a single TCP connection. Standardized as RFC 6455 by the IETF in 2011, it upgrades from HTTP to establish full-duplex channels and underpins real-time Web use cases such as chat, gaming, market data and collaborative editing.

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

A security model that abandons implicit trust based on network location and continuously verifies every access. Coined by Forrester's John Kindervag in 2010 and exemplified by Google BeyondCorp (2014), it was systematized in NIST SP 800-207 (2020) and made the U.S. federal government's standard policy.

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