Artificial intelligence, from first principles.
A chronological and practical library covering the long history of computation, classical AI and machine learning, neural networks, transformers and LLMs, current model families, AI in business and design, robotics, asset tokenization, governance and the documented directions shaping AI in 2026.
History & Foundations
15 source-linked pagesFrom the Abacus to Artificial Intelligence
AI sits at the end of a much longer history of calculation, logic, programmability, electronics, statistics and learning systems. The abacus itself is not AI; its relevance is that it is an early tool for externalizing arithmetic procedures.
↗Pascal's Calculator and Mechanical Arithmetic
Pascal's mechanical calculator used geared wheels to automate arithmetic. It belongs to the pre-electronic history of computing because it showed that a physical mechanism could carry out a formal numerical procedure.
↗Leibniz and the Stepped Reckoner
Leibniz's stepped reckoner was a mechanical calculating machine intended to automate addition, subtraction, multiplication and division, extending earlier mechanical arithmetic devices.
↗The Jacquard Loom and Programmable Patterns
The Jacquard loom used punched cards to control textile patterns. It was not a general-purpose computer, but it demonstrated that external encoded instructions could control a machine's behavior.
↗Charles Babbage and the Analytical Engine
Babbage's Analytical Engine was a nineteenth-century design for a general-purpose mechanical computer with concepts resembling a processor, memory and program control.
↗Ada Lovelace and the First Published Algorithm
Ada Lovelace's notes on Babbage's Analytical Engine included a procedure for calculating Bernoulli numbers and are widely described as containing the first published algorithm intended for machine execution.
↗Boolean Algebra and Machine Logic
Boolean algebra formalizes logical operations with values such as true and false. Its application to switching circuits made it foundational to digital logic.
↗Punched Cards and Automated Data Processing
Punched cards encode information through the presence or absence of holes. They were widely used for automated tabulation, data input and early computer programming.
↗Alan Turing and the Universal Machine
Turing's 1936 work defined an abstract computing machine and the concept of a universal machine capable of simulating other machines, becoming foundational to theoretical computer science.
↗ENIAC and Electronic General-Purpose Computing
ENIAC was an early large-scale electronic digital computer completed in the 1940s. It used vacuum tubes and helped demonstrate the practicality of electronic general-purpose computation.
↗Stored-Program Computers and the Von Neumann Architecture
Stored-program computing places instructions in memory so they can be fetched and executed rather than requiring a machine to be rewired for every task.
↗Cybernetics and Feedback Systems
Cybernetics studied control and communication in animals and machines, especially feedback loops in which systems sense state, compare it with a goal and adjust behavior.
↗The Turing Test and the 1950 Question of Machine Intelligence
In 1950 Alan Turing published “Computing Machinery and Intelligence” and proposed the imitation game as a behavioral way to discuss machine intelligence rather than trying to define the word “thinking” directly.
↗The Dartmouth Workshop and the Birth of Artificial Intelligence
The 1956 Dartmouth Summer Research Project on Artificial Intelligence is widely treated as a founding event of AI as a research field. The proposal used the term “artificial intelligence” and brought together researchers working on machine reasoning, learning and symbolic problem solving.
↗The Perceptron and Early Machine Learning
The perceptron is an early trainable model associated with Frank Rosenblatt. It learns a linear decision boundary from examples and became historically important in the development of neural-network research.
↗Classical AI & Machine Learning
15 source-linked pagesSymbolic AI
Symbolic AI represents knowledge with explicit symbols, rules and structures, then manipulates those representations with logic, search or planning procedures.
↗Search Algorithms in Artificial Intelligence
AI search algorithms explore possible states or actions in order to find paths, plans or solutions. Classical examples include breadth-first search, depth-first search and A*.
↗Heuristics and Problem Solving
A heuristic is an estimate or rule used to guide search toward promising states without exhaustively checking every alternative.
↗Knowledge Representation
Knowledge representation studies how facts, relations, rules and concepts can be encoded so a computer system can reason over them.
↗Expert Systems
Expert systems are narrow rule-based programs that separate a knowledge base from an inference mechanism to reproduce parts of specialist decision-making.
↗ELIZA and Early Conversational Programs
ELIZA, created by Joseph Weizenbaum in the 1960s, used pattern matching and scripted text transformations to simulate conversation without modern statistical language understanding.
↗SHRDLU and Language in a Blocks World
SHRDLU, developed by Terry Winograd around 1968–1970, accepted natural-language instructions about a small simulated blocks world and combined language parsing with a symbolic world model.
↗The AI Winters
AI winters were periods in which enthusiasm and funding fell after systems failed to meet expectations or became economically unattractive. The term is especially associated with downturns in the 1970s and again in the late 1980s and early 1990s.
↗Bayesian Networks
A Bayesian network is a directed acyclic graph representing probabilistic relationships between variables using graph structure and conditional probability distributions.
↗Fuzzy Logic
Fuzzy logic allows degrees of membership between 0 and 1 rather than forcing every proposition into a strict true-or-false category.
↗Genetic Algorithms
Genetic algorithms are optimization methods inspired by evolutionary selection, iteratively varying and selecting candidate solutions according to a fitness function.
↗Reinforcement Learning
Reinforcement learning trains an agent by interaction: actions change an environment, rewards provide feedback and the agent learns a policy intended to maximize cumulative reward.
↗Supervised Learning
Supervised learning trains a model from examples paired with known labels or target values. Classification and regression are common supervised tasks.
↗Unsupervised Learning
Supervised learning trains a model from examples paired with known labels or target values. Classification and regression are common supervised tasks.
↗Artificial Intelligence vs Machine Learning
Artificial intelligence is the broader field concerned with systems performing tasks associated with intelligent behavior; machine learning is a major subfield in which systems learn patterns from data.
↗Neural Networks & Deep Learning
15 source-linked pagesArtificial Neural Networks
Artificial neural networks are parameterized models built from layers of connected computational units. Training adjusts the parameters to reduce an objective function.
↗Backpropagation
Backpropagation efficiently computes gradients through differentiable neural networks by applying the chain rule from outputs back toward earlier layers.
↗Convolutional Neural Networks
Artificial neural networks are parameterized models built from layers of connected computational units. Training adjusts the parameters to reduce an objective function.
↗Recurrent Neural Networks
Artificial neural networks are parameterized models built from layers of connected computational units. Training adjusts the parameters to reduce an objective function.
↗Long Short-Term Memory Networks
LSTM networks are recurrent neural networks with gated memory mechanisms designed to preserve and control information across longer sequences.
↗Word Embeddings
Word embeddings represent words or tokens as dense numerical vectors whose geometry captures statistical relationships learned from data.
↗ImageNet and AlexNet
AlexNet's strong result in the 2012 ImageNet competition helped demonstrate the effectiveness of deep convolutional neural networks trained on GPUs and large image datasets.
↗GPUs and the Rise of Deep Learning
GPUs perform many numerical operations in parallel, which fits the matrix and tensor workloads used in neural-network training and inference.
↗Deep Learning
Deep learning uses neural networks with multiple layers of learned representations. Its modern rise combined larger datasets, better training methods and parallel hardware.
↗Attention Mechanisms
Attention computes context-dependent weighted combinations of representations, allowing a model to emphasize information relevant to the current prediction.
↗The Transformer Architecture
The Transformer was introduced in the 2017 paper “Attention Is All You Need.” It replaced recurrent sequence processing with attention-based layers, enabled highly parallel training and became the architectural basis of most modern large language models.
↗Encoder-Decoder Models
Encoder-decoder architectures turn an input into internal representations and then generate an output sequence or structure conditioned on those representations.
↗Self-Supervised Learning
Supervised learning trains a model from examples paired with known labels or target values. Classification and regression are common supervised tasks.
↗Foundation Models
Foundation models are trained on broad data at scale and then adapted or prompted for many downstream tasks.
↗Multimodal AI
Multimodal AI processes or generates multiple data types such as text, images, audio or video in one coordinated system.
↗Large Language Models
15 source-linked pagesWhat Is a Large Language Model?
A large language model is a statistical model trained on large collections of text or multimodal data to predict and generate token sequences. Modern LLMs are usually transformer-based and are adapted to conversation, coding, tool use and other tasks through post-training and application-level systems.
↗LLM Pretraining
LLM pretraining exposes a model to large datasets and optimizes a predictive objective, commonly next-token prediction, before later post-training or adaptation.
↗Fine-Tuning Language Models
Fine-tuning continues training a pretrained model on a narrower dataset or objective so it performs better for a particular domain, task or behavior.
↗RLHF and Preference Training
Reinforcement learning from human feedback uses preference information from people to train or guide models toward preferred behavior; modern post-training also uses related preference-optimization techniques.
↗Tokens in AI
Tokens are the discrete units processed by a language model. Depending on the tokenizer, a token can represent a word, part of a word, punctuation, whitespace patterns or another text fragment.
↗Tokenization
Tokenization converts text into token identifiers a model can process. Model families use different tokenizers, so the same text can consume different numbers of tokens in different systems.
↗Context Windows
A context window is the amount of tokenized information a model can consider in one working sequence or request. Larger context increases capacity for long inputs but does not guarantee perfect retrieval or reasoning over every detail.
↗Embeddings in Modern AI
Embeddings are dense numerical vectors used to represent text, images or other objects in spaces where distance or similarity can be computed.
↗Inference
Inference is the phase in which a trained model is run to make predictions or generate outputs. LLM inference usually produces tokens sequentially.
↗Temperature and Sampling
Temperature changes the probability distribution used during sampling. Lower values concentrate probability more strongly; higher values make less likely alternatives easier to sample.
↗AI Hallucinations
A hallucination is generated content that is false, unsupported or inconsistent with the evidence available to the system. Language generation optimizes likely sequences, not guaranteed factual truth, so high-stakes factual work needs grounding and verification.
↗Retrieval-Augmented Generation
Retrieval-augmented generation combines a generative model with retrieval from an external collection such as documents, databases or search results. Retrieved material is inserted into the model's context so the answer can be grounded in information outside the model's parameters.
↗Function Calling and Tool Use
Function calling and tool use let a model request structured actions such as querying a database, calling an API or using search, with the surrounding application executing the action.
↗AI Agents
An AI agent is a system in which a model can select or plan actions over multiple steps, usually with tools, memory and feedback from an environment. The language model is only one part of the agent; orchestration, permissions, tools and stopping conditions are equally important.
↗Mixture-of-Experts Models
A mixture-of-experts model contains multiple expert subnetworks and routes each token or input through only a subset of them. This can increase total parameter capacity without activating every parameter on every forward pass.
↗Modern Model Families
15 source-linked pagesThe GPT Model Family
GPT is OpenAI's family of generative pretrained transformer models, evolving from research language models into multimodal, reasoning and tool-using systems.
↗GPT-5.4
OpenAI released GPT-5.4 in March 2026 for professional work, with GPT-5.4 Thinking in ChatGPT and GPT-5.4 plus GPT-5.4 Pro in the API. OpenAI later released GPT-5.4 mini and nano as faster, lower-cost members of the same generation.
↗The Gemini Model Family
Gemini is Google's multimodal model family developed by Google DeepMind and integrated into Google products and developer platforms.
↗Gemini 3.5
Google introduced Gemini 3.5 in May 2026, beginning with Gemini 3.5 Flash, which Google positions for fast agentic workflows. Google also announced Gemini 3.5 Pro as part of the same family.
↗The Claude Model Family
Claude is Anthropic's family of language and multimodal models, with product lines including higher-capability Opus models and efficiency-oriented Sonnet models.
↗Claude Opus 4.6
Claude Opus 4.6 is a 2026 Anthropic frontier model. Anthropic's system card documents extensive evaluation across coding, computer use, safety, alignment and other capability areas.
↗The DeepSeek Model Family
DeepSeek is a Chinese AI model family known internationally for open-weight language and reasoning systems such as DeepSeek-V3 and R1.
↗DeepSeek V4
DeepSeek's April 2026 API changelog lists V4-Pro and V4-Flash. It also states that the older deepseek-chat and deepseek-reasoner names would be discontinued after temporarily mapping to modes of V4-Flash.
↗The Llama Model Family
Llama is Meta's open-weight language-model family, beginning with a 2023 research release and expanding into developer-focused multimodal models.
↗Llama 4 Scout and Maverick
Meta released Llama 4 Scout and Maverick in April 2025 as natively multimodal mixture-of-experts models. Meta describes Scout as having 17 billion active parameters and 16 experts, and Maverick as having 17 billion active parameters and 128 experts.
↗The Grok Model Family
Grok is the model family developed by xAI/SpaceXAI, spanning general reasoning, coding, voice and generative-media systems.
↗Grok 4.5
SpaceXAI introduced Grok 4.5 in July 2026 for coding, agentic tasks and knowledge work, then expanded its availability through the company's own products and integrations including GitHub Copilot.
↗The Mistral Model Family
Mistral AI develops commercial and open-weight language models for general generation, coding, multimodal and specialist workloads.
↗Cohere Command Models
Cohere's Command family is designed heavily around enterprise generation, retrieval, tool use and multilingual applications. Cohere's 2026 documentation lists Command A+ as a mixture-of-experts model with text and image input support.
↗Open, Open-Weight and Closed AI Models
AI models differ in what users can inspect and control. Open-weight models make trained parameters available under a license; closed models are typically accessed through hosted products or APIs.
↗Business, Design & Work
15 source-linked pagesAI Adoption in Companies in 2026
Stanford's 2026 AI Index reports organizational AI adoption at 88%. That figure shows how widely AI has entered organizations, but it does not imply that most deployments are fully autonomous or equally mature.
↗How to Build an AI Strategy Roadmap
An AI strategy roadmap connects business objectives with candidate workflows, data, model choices, evaluations, controls, owners and staged deployment.
↗How Companies Choose AI Use Cases
Companies choose AI use cases by matching model capabilities to tasks with sufficient data, measurable outcomes, acceptable error costs and clear human or software ownership.
↗AI in Customer Support
AI in customer support is used for search, summarization, classification, routing, response drafting and increasingly tool-using assistants that can perform defined service actions.
↗Virtual AI Assistants and Chatbots
Modern virtual assistants combine a language model with retrieval, business rules, APIs, identity controls and conversation state. The surrounding system determines what the assistant can know and do.
↗AI in Sales and Marketing
AI supports sales and marketing through research, segmentation, drafting, variation, lead triage, forecasting assistance and workflow automation, while claims and customer data still require governance.
↗AI in Finance and Accounting
Finance and accounting teams use AI for document extraction, reconciliation support, anomaly detection, research, forecasting assistance and conversational access to enterprise data.
↗AI in Legal Services
Legal AI systems are used for search, document review, summarization, drafting assistance and contract analysis; jurisdiction-specific and consequential outputs require source grounding and professional review.
↗AI in Healthcare Administration
Healthcare organizations use AI in administration for documentation support, scheduling, coding assistance, call-center workflows and information retrieval; clinical uses require stronger validation and governance.
↗AI in Manufacturing
Manufacturing AI includes predictive maintenance, visual inspection, process optimization, planning and robotics, usually tied to operational data and measurable production outcomes.
↗AI in Retail and Ecommerce
Retail and ecommerce use AI for search, recommendations, catalog enrichment, customer support, forecasting, fraud detection and merchandising assistance.
↗AI for Small Businesses
Small businesses can use hosted AI for drafting, research, customer-service support, document processing and automation without training foundation models themselves.
↗AI and Web Design
AI is increasingly used in web design for research, ideation, content variants, asset generation, code assistance, accessibility support and design-to-code workflows. Figma's 2026 research reports substantial growth in cross-functional overlap between design and development as AI tools spread.
↗AI and Graphic Design
AI has become part of image generation, editing, layout exploration and asset production. The designer still owns art direction, brand consistency, rights and provenance decisions, accessibility and the judgment of what should actually be created.
↗AI in Software Development
AI coding systems support code generation, explanation, testing, refactoring, repository search and increasingly agentic software tasks that operate across tools.
↗Robotics, Finance & Governance
10 source-linked pagesAI and Robotics
AI and robotics combine perception, planning, control and learning with physical machines. Unlike a text-only application, a robot must operate under sensor noise, timing constraints, changing environments and physical safety requirements.
↗Physical AI
Physical AI refers to AI systems that perceive and act in the physical world through robots, autonomous machines or embodied devices. Current development increasingly combines foundation models, simulation, accelerated computing and on-device inference.
↗Autonomous Vehicles and AI
Autonomous-vehicle systems combine sensors, perception, prediction, planning and control. Their safety is a property of the complete system, not a single neural network.
↗Industrial Robots and AI
Industrial robots traditionally excel at repeatable structured tasks; AI expands their ability to perceive variable objects, adapt to changing conditions and plan more flexible work.
↗AI in Financial Services
Financial institutions use AI for fraud detection, risk analysis, customer service, document processing, research and operational automation, with strong model-risk and governance requirements.
↗Asset Tokenization
Asset tokenization represents claims on financial or physical assets as digital tokens on programmable platforms. BIS work treats tokenization as a change in representation, transfer and settlement infrastructure—not as a form of artificial intelligence.
↗AI and Tokenized Financial Assets
AI and tokenization are distinct technologies that can interact. AI can support monitoring, compliance, analysis or operational automation around tokenized markets; tokenization itself is a ledger and market-structure technology.
↗The NIST AI Risk Management Framework
The NIST AI Risk Management Framework is a voluntary framework organized around governing, mapping, measuring and managing AI risk. NIST's Generative AI Profile extends that framework with risks and suggested actions specific to generative systems.
↗AI Governance and Regulation
AI governance is the set of policies, ownership, controls, documentation, evaluation and monitoring used to manage AI systems; regulation is one external component of that broader governance system.
↗AI Roadmap 2026: Agents, Multimodality, Efficiency and Robotics
As of 2026, first-party releases and independent measurement show several major directions: more tool-using and agentic systems, stronger multimodality, smaller efficient models, longer context, model routing, open-weight competition and closer integration between foundation models and robotics.
↗