AI concepts. One field guide, every lens.

AGI is the idea of a system that generalizes across domains the way a skilled human can — novel problems, physical world, long horizons — not just excelling at benchmarks. No system today is agreed to be AGI. The term anchors timelines, investment, and existential-risk debates more than current product specs.

An agent loops: observe a goal, decide actions, call tools (search, email, code execution, calendars), read results, and continue until done or stopped. Agents can automate workflows but multiply risk — each tool call is a place to go wrong. Serious deployments keep humans in the loop for high-stakes actions.

AI regulation spans safety testing, transparency, sector rules (health, finance, education), export controls on chips, and liability when systems cause harm. Examples include the EU AI Act, U.S. executive orders, and sector guidance from schools or hospitals. Regulation lags technology; compliance is often “best effort against moving targets.”

Alignment research asks how to ensure powerful systems pursue intended goals — helpful, harmless, honest — rather than optimizing the wrong proxy. In practice it includes training techniques, red teaming, policy rules, and monitoring. Perfect alignment is unsolved; “aligned enough for this use case” is an engineering judgment.

An application programming interface lets software send structured requests to a model (text in, text out) and integrate the result into products. When a startup “builds on GPT,” they usually mean API calls plus their own UI and logic. APIs make inference metered and billable per token or request.

A broad label for computer systems that recognize patterns, generate text or images, make predictions, or take actions based on data — without a human executing each step by hand. In casual conversation “AI” might mean a chatbot, a photo filter, a warehouse robot, or a stock-trading model. The term is intentionally wide: it describes a family of techniques, not one single product.

Automation replaces or compresses repetitive human work — invoicing, ticket routing, report generation. AI automation often means models plus rules plus integrations. The design question is not “can it do it?” but “what happens when it misfires, and who is accountable?”

Automation bias is the human tendency to defer to automated recommendations, especially under time pressure. It’s why verification matters in copilot settings: the risk isn’t only model error, it’s humans stopping checks. Training and UI design should keep responsibility visible.

Benchmarks are datasets and scoring rules (MMLU, HumanEval, etc.) that measure model performance on defined tasks. They help compare releases but don’t capture real-world messiness — user prompts, adversarial inputs, or your company’s actual documents. A model that tops a leaderboard can still fail your workflow.

Bias in AI usually means unequal error rates or stereotypes baked in from historical data: hiring tools favoring certain résumés, image models defaulting to narrow demographics, language models mirroring toxic patterns from the web. Fixing bias requires dataset audits, evaluation across groups, and human review — not a single fairness slider.

Closed models run on the provider’s infrastructure; you send prompts and receive outputs under their terms. You trade control and inspectability for convenience, safety filtering, and scale. Most consumer chatbots are closed models behind accounts and rate limits.

Community detection splits a network into clusters with many connections inside and few between: friend groups, fan communities, neighborhoods of related web pages. Popular methods (Louvain, Leiden) keep merging nodes into groups as long as that inside-versus-between score improves. Platforms use clusters to recommend content and to catch coordinated spam accounts — and the same grouping can narrow what people see, reinforcing a filter bubble.

Philosophy of mind asks if machines could be conscious, sentient, or merely simulating conversation without inner experience. Current LLMs show no evidence of subjective experience; they generate text about feelings without necessarily having them. Ethical treatment of AI systems may eventually hinge on this question — today it’s largely theoretical.

Consent questions arise when personal writing, medical records, student work, or biometric data might enter a training or fine-tuning pipeline. Schools and companies often ban pasting identifiable student data into public models for this reason. “Public on the web” is not the same as “consented for AI training” in many ethical frameworks.

The context window is the maximum amount of text (measured in tokens) a model can consider in one request — your prompt, any files you attach, and the answer it generates. If you exceed it, the system truncates or summarizes older material. Bigger windows enable longer documents; they also cost more to run.

Copilot-style features sit inside IDEs, office suites, or CRMs and propose drafts, formulas, or code. The human remains the editor and approver. That posture differs from full automation: speed with supervision, not unsupervised agency.

Copyright debates ask whether scraping books, art, or code to train models infringes creators’ rights, and whether AI-generated works are protectable. Laws differ by country and are still unsettled. For users: assuming “the model said it, so I can publish it” is risky — verify rights on sources and outputs.

Deepfakes use generative models to mimic a person’s appearance or speech. They raise fraud, election, and harassment risks, especially when detection lags creation. Disclosure norms (“this is AI-generated”) and provenance tools are part of the response; skepticism toward unverified media is the user-side defense.

An embedding maps a piece of text (or an image) to a vector — a list of numbers capturing semantic similarity. Similar ideas land near each other in that space, which powers search, recommendations, and retrieval systems. Embeddings are how RAG finds the right paragraph before asking an LLM to answer. Networks can be embedded too: send short random walks through a graph, treat each walk like a sentence, and nodes that keep showing up near each other end up with similar vectors.

Existential risk framing worries that sufficiently capable, misaligned systems could cause irreversible harm — through autonomous weapons, unstable feedback loops, or goals that diverge from human flourishing. Skeptics call it speculative; proponents want safety research prioritized now. Most daily AI use sits far from this debate, but it shapes funding and policy rhetoric.

Explainability (and interpretability) asks for reasons behind predictions — critical in lending, medicine, and discipline decisions. Deep neural networks are often black boxes; post-hoc explanations can be approximate or misleading. Regulated domains may require human-readable justification even when the model can’t truly “explain” itself.

Fine-tuning starts from a general model and continues training on a smaller, targeted corpus — your company’s support tickets, a medical style guide, a classroom rubric. It’s cheaper than training from scratch but still requires clean data and evaluation. It does not automatically fix hallucinations or safety issues.

Generative systems produce novel outputs from prompts: essays, logos, music, Python scripts. That’s different from traditional ML that mostly labels or ranks (spam vs. not spam, which photo contains a cat). Most public excitement since 2022 is about generative models, especially LLMs and image generators.

Graphics processing units (and similar accelerators) train and run large models. Limited supply concentrates power among cloud providers and well-funded labs. Compute access is a geopolitical and economic variable — export controls on chips are part of the AI policy toolkit.

In computing, a graph isn’t a bar chart — it’s a set of things (nodes) and the connections between them (edges). Friends on a social app, web pages linking to each other, roads between towns, and atoms in a molecule are all graphs. Edges can point one way (a link from one page to another) and can carry labels (“is the parent of”). Some databases store a graph as nodes with properties attached, others as long lists of subject–relationship–object facts; either way, the connections are the data.

A graph neural network learns from data shaped like a network. Each node starts with its own information, then repeatedly blends in what its neighbors know — called message passing — so after a few rounds it reflects its surroundings, not just itself. GNNs are used to predict properties of molecules, flag fraud rings, estimate travel times on road networks, and recommend items. The variants differ mostly in how they weigh neighbors; some learn which connections matter most.

GraphRAG is a family of retrieval techniques that pull from a graph instead of (or alongside) a pile of similar-sounding text. The system might follow links between people, places, and things to answer multi-step questions (“which of our suppliers depend on this one factory?”), translate a question into a database query, or summarize clusters of related documents to answer big-picture questions. It makes answers easier to trace, but the model can still misread what it retrieved.

A hallucination is a fluent, plausible-sounding output that isn’t grounded in fact — fake citations, invented dates, wrong math presented smoothly. The model isn’t “lying” in a human sense; it’s completing patterns without a built-in truth checker. The fix is verification habits, not hoping the next model version eliminates the behavior entirely.

Human-in-the-loop designs require approval for high-stakes steps: sending email, grading, medical suggestions, financial trades. It trades speed for accountability. The opposite — fully autonomous loops — is where agent hype meets governance anxiety.

If training is studying, inference is taking the test: you send inputs, the model runs forward, and you get outputs. Every chat message, image generation, or autocomplete suggestion is inference. Inference cost (compute per query) is what shapes pricing, speed limits, and why “free unlimited AI” rarely stays free at scale.

Inference cost is compute per query: GPUs, electricity, datacenter overhead. It drives API pricing, context limits, and why “run your own 70B model” is expensive. Training is a giant one-time bill; inference is the recurring meter that scales with users.

In AI, “intelligence” often means measurable task success: exam scores, coding puzzles, game playing. In human terms it includes judgment, values, social context, and learning from few examples. Conflating the two leads to both overhype (“it’s thinking”) and overpanic (“it’s already smarter than us”). Being precise about which kind you mean clears a lot of debates.

A knowledge graph stores facts as things linked by named relationships: Paris → capital of → France, Marie Curie → won → Nobel Prize. Search engines use them for the fact boxes beside results, and public ones like Wikidata are built by volunteers. Because each fact is an explicit link, you can trace where an answer came from — but the graph is only as accurate and complete as whoever entered the facts.

Economic debates distinguish automation that replaces tasks from tools that augment workers. Historical pattern: technology shifts job mix rather than eliminating work entirely, but transitions are uneven and painful. Policy arguments focus on retraining, wages, and which roles get deskilled vs. upskilled.

An LLM is trained on enormous amounts of text and learns to predict what token comes next given everything before it. That simple objective, at scale, produces fluent answers, code, and summaries — but fluency is not the same as understanding or truth. Most consumer “AI chat” today is an LLM plus safety layers, tools, and a user interface.

In business debates, a moat is defensibility: proprietary data, distribution, brand, regulatory approval, workflow lock-in, or fine-tuned models on exclusive corpora. Base models trend toward commodity; applications fight on trust, integration, and outcomes.

A model is the actual learned system: weights, architecture, and training history packaged so it can take inputs and produce outputs. When someone says “GPT” or “Claude,” they usually mean a model (or a family of models) behind an interface. The app, login screen, and buttons are not the model.

A multimodal model can take combinations of media in one workflow: describe this photo, transcribe this clip, read a PDF with diagrams. “Modal” refers to mode of data. Multimodal systems blur the line between separate tools (OCR, speech-to-text, chat) into one interface — with the same verification problems as text-only models.

A neural network is a stack of simple units (neurons) that transform inputs through weighted connections. Training adjusts those weights so the network maps inputs to useful outputs. Deep learning means many layers. You don’t need to implement one to use AI, but the term explains why “parameters” and “training data” keep coming up.

Open-weight models release the actual trained network (often with a license governing use). That enables local running, inspection, and fine-tuning without calling a vendor API. “Open” varies — some licenses restrict commercial use or require sharing derivatives. Open weights are not the same as open training data.

PageRank was Google’s original idea for ranking web pages, published in 1998: a page is important if important pages link to it. Picture a surfer who clicks random links and now and then jumps to a random page — the pages where they spend the most time rank highest. Modern search mixes in many other signals. More broadly, “centrality” is the family of ways to score importance in a network, from the most-connected node to the one that sits on the most shortest paths between others (a bridge).

Parameters are the learned numbers inside a model that determine its behavior. More parameters generally mean more capacity to memorize patterns, but also more compute to train and run. “70B model” means roughly seventy billion parameters — a shorthand for size and cost, not quality by itself.

Privacy in AI use means knowing whether your prompts are retained, reviewed by humans, used for training, or shared with subprocessors. Enterprise tiers often offer “no training on customer data”; consumer free tiers may not. Treat prompts like email to a third party: don’t send secrets you wouldn’t want leaked.

A prompt is whatever you send in: a question, a document to summarize, a role (“act as a tutor”), constraints (“under 200 words”), or examples of the format you want. Better prompts are specific about goal, audience, and what “good” looks like. Vague prompts reliably produce vague answers.

Retrieval-augmented generation combines search with generation: the system retrieves snippets from a knowledge base, injects them into the prompt, and asks the LLM to respond using that material. RAG reduces some hallucinations about your own docs but doesn’t eliminate errors — the model can still misread or overreach beyond the retrieved text.

Red teams probe for jailbreaks, bias, data leaks, and unsafe advice — adversarial testing before launch. It’s standard in serious AI products and increasingly expected by regulators. Red teaming finds known failure modes; it doesn’t prove safety in all future situations.

Reinforcement learning from human feedback ranks model outputs (helpful vs. harmful, accurate vs. sloppy) and nudges the model toward preferred behavior. It’s a major reason chatbots feel “assistant-like.” RLHF shapes tone and refusal patterns; it doesn’t guarantee factual correctness on its own.

Sensors convert real-world signals into data: cameras, microphones, lidar, thermometers, accelerometers. In robotics and edge AI, sensors ground the system in something outside the training corpus. A language model alone has no sensors — it only knows what you type.

A social graph is the network of people and their connections on a service: friends, follows, contacts, who messages whom. Apps use it to suggest “people you may know,” rank feeds, and spot fake accounts. It also lets a service guess things you never told it — researchers have shown that details a person kept private, like their school or class year, can often be inferred from what their friends share publicly. Contact uploads can even put people who never signed up into the graph.

Most generative models sample from probability distributions — they’re stochastic. Temperature and seed settings control randomness. That’s why “ask twice” is a legitimate test. Nondeterminism complicates testing, grading, and compliance workflows that assume fixed outputs.

Surveillance AI includes facial recognition, behavioral scoring in workplaces or schools, and mass metadata analysis. It raises civil-liberty questions separate from chatbot hype: power asymmetry, false positives, and chilling effects on speech. Not every camera has AI; not every AI deployment is surveillance — but the overlap is politically charged.

Models don’t see whole words the way humans do — they break text into tokens (often pieces of words or common short words). Longer prompts and longer answers mean more tokens, which means more compute, time, and often more cost. You don’t need to count tokens to use AI well, but the term explains why “just paste the whole book” hits limits.

Training is the expensive upfront phase where a model adjusts its internal parameters to fit patterns in data — predicting the next word, classifying images, etc. Training from scratch takes enormous compute and curated datasets. Most people and companies instead use already-trained models and adapt them lightly.

Training data is the text, images, or logs used during training. Quality, consent, copyright, and representation in that data shape model behavior. When people ask “where did it learn that?”, they’re asking about training data and fine-tuning — usually opaque for closed models.

Transformers are a neural-network design that uses “attention” to weigh which parts of an input matter for each output token. Introduced in 2017, they scaled better than earlier approaches and enabled today’s large language models. When someone says “GPT-style model,” they usually mean a transformer trained to predict text.

A vector database indexes embeddings so you can quickly retrieve the closest matches to a query. It’s infrastructure behind many enterprise AI assistants, support bots, and internal search. The quality of RAG depends as much on what you indexed and how you chunk documents as on the LLM itself. Some databases also store vectors inside the nodes of a graph, so one query can search by meaning and follow connections.

Verification means treating model output as a draft: cross-check claims, run code, read sources, ask what would falsify the answer. It’s the skill that separates tool use from blind trust. Kids and adults need the same loop — models don’t get a pass because they sound confident.

Vertical SaaS serves one domain deeply — legal, dentistry, HVAC — and now embeds models for drafting, scheduling, or diagnostics inside that workflow. The moat is often data and integrations, not the raw model. Generic chatbots compete on breadth; vertical tools compete on “it fits how we already work.”