AI Glossary

An AI model trained on massive amounts of text (books, websites, articles, code, conversations) that can understand and generate human language. "Large" refers to both the enormous amount of training data and the billions of parameters inside the model. Claude, GPT-4, Gemini, and Llama are all LLMs. They don’t understand language the way you do; they are extremely sophisticated pattern-matching systems that predict what words should come next, given what’s already been said.

The broader field of teaching computers to work with human language: reading it, understanding it, translating it, summarizing it. NLP existed long before the current AI wave (think: spell-check, spam filters, Google Translate). Modern LLMs are the most advanced NLP systems ever built, but NLP as a field includes much simpler tools too.

The basic unit an LLM works with. A token isn’t always a whole word. It can be a word, part of a word, or even a punctuation mark. "Unbreakable" might be split into "un," "break," and "able," which is three tokens. Models process, generate, and charge by tokens. When a service says a model has a "200,000-token context window," it means the model can handle roughly 150,000 words of text at once.

The amount of text a model can "see" and work with at one time. Think of it as the model’s short-term memory. A small context window means the model can only consider a few pages of text; a large one means it can hold an entire book. If your conversation with an AI gets long enough to exceed the context window, the model starts "forgetting" the earliest parts of the conversation.

The text you type to communicate with an AI model: your question, instruction, or request. "Write me a cover letter for a marketing job" is a prompt. The quality of your prompt directly affects the quality of the response. This isn’t a flaw; it’s how the technology works. Learning to write clear, specific prompts is genuinely useful and doesn’t require technical skills.

Hidden instructions given to an AI model before you ever start talking to it. These are set by the company or developer and shape how the model behaves: its personality, safety rules, and what it will and won’t do. You typically don’t see the system prompt, but it’s the reason Claude sounds different from ChatGPT, even though both are LLMs.

When an AI model confidently states something that is factually wrong or entirely made up. The model isn’t "lying." It doesn’t have beliefs or intentions. It’s generating text that follows the patterns it learned, and sometimes those patterns produce plausible-sounding nonsense. This is one of the most important limitations to understand: AI can be fluent and wrong at the same time.

A setting that controls how creative or predictable a model’s responses are. Low temperature means the model picks the most likely next word every time (predictable, consistent). High temperature means it’s more willing to make unexpected choices (creative, varied, but also more likely to go off the rails). It’s like the difference between a jazz musician improvising and one reading sheet music.

Taking a pre-trained model and training it further on a smaller, more specialized dataset to make it better at a specific task. A general-purpose LLM might be fine-tuned on legal documents to create a legal AI assistant, or on medical texts to create a clinical support tool. Fine-tuning is much cheaper and faster than training a model from scratch.

A technique where human reviewers rate the model’s responses, and the model uses those ratings to improve. Think of it as a teacher grading essays and the student learning what "good" looks like. RLHF is a major reason modern AI assistants are helpful and conversational rather than producing raw, unfiltered text. It’s also how safety behaviors are instilled.

A technique that lets an AI model look up real information from a database or document collection before responding, rather than relying solely on what it memorized during training. Think of it as the difference between answering a test from memory versus being allowed to check your notes. RAG helps reduce hallucinations and keeps responses grounded in actual sources.

A way of representing words, sentences, or documents as lists of numbers so that similar meanings end up close together mathematically. "Happy" and "joyful" would have similar embeddings; "happy" and "carburetor" would not. Embeddings are how AI systems understand that concepts are related even when the exact words are different. They power search, recommendations, and clustering.

An AI system that can work with more than one type of input, for example text and images, or text and audio. A multimodal model might be able to look at a photo of your fridge and suggest recipes, or listen to a voice memo and summarize it. Most cutting-edge models are moving in this direction.

A program designed to have conversations with people through text. The term predates modern AI by decades. Early chatbots followed simple scripts. Today’s AI chatbots, powered by LLMs, can hold nuanced, context-aware conversations. Claude, ChatGPT, and Gemini are all chatbots, though their capabilities extend far beyond just "chatting."

An AI system that can take actions on its own, not just answer questions, but actually do things like browse the web, write and run code, manage files, or interact with other software. An agent can break a task into steps, execute those steps, evaluate the results, and adjust its approach. This is a rapidly evolving area and represents a shift from AI as a "tool you ask questions" to AI as a "coworker that does tasks."

The broader concept of AI systems that operate with some degree of autonomy, planning, executing, and adapting without needing a human to guide every step. Agentic AI is the umbrella term for agents, autonomous workflows, and self-directed AI systems. It’s one of the most actively discussed areas in AI development right now.

A building (or complex of buildings) packed with thousands of powerful computers. Data centers are where AI models are trained and run. They require enormous amounts of electricity, industrial cooling systems, and high-speed internet connections. When you send a message to Claude, your text travels to a data center, gets processed on specialized hardware, and the response travels back, usually in seconds.

Using someone else’s computers over the internet instead of running software on your own machine. When you use Google Docs, Netflix, or an AI assistant, the heavy lifting happens on servers in a data center somewhere, and you see the results on your screen. Amazon Web Services (AWS), Microsoft Azure, and Google Cloud are the three biggest cloud providers. Most AI companies rent computing power from these providers rather than building their own data centers.

A computer chip originally designed to render video game graphics. It turns out the same kind of math that draws 3D graphics (processing many calculations simultaneously) is exactly what AI training needs. GPUs became the engine of the AI revolution. NVIDIA is the dominant GPU manufacturer for AI workloads, and their chips are in such high demand that supply shortages have been a real constraint on AI development.

Google’s custom-designed chip built specifically for AI workloads. While GPUs were repurposed from gaming, TPUs were designed from the ground up for the kind of math AI models require. They’re used internally at Google to train and run Gemini and other models. Think of it as the difference between converting a truck into an ambulance versus building an ambulance from scratch.

A catch-all term for the computational power used to train and run AI models. "We need more compute" means "we need more processing power." Compute is one of the three fundamental ingredients of modern AI, alongside data and algorithms. Access to compute is a major competitive advantage. The companies with the most computing resources can train the biggest models.

Any specialized processor designed or optimized for AI tasks. GPUs, TPUs, and newer designs from companies like AMD, Intel, and startups are all AI chips. The global race to design faster, more efficient AI chips is reshaping the semiconductor industry. These chips are so strategically important that governments have imposed export restrictions on them.

Data centers generate enormous heat. All those thousands of processors running 24/7 need to be kept cool or they’ll overheat and fail. Traditional data centers use industrial air conditioning. Newer AI-focused data centers are exploring liquid cooling (running coolant directly over chips) and even building facilities near cold climates or bodies of water. Cooling is one of the biggest operational costs and engineering challenges in running AI at scale.

Training a single large AI model can consume as much electricity as a small town uses in a year. Running that model for millions of users every day consumes even more. The energy demands of AI have become a serious concern, driving conversations about sustainability, renewable energy commitments, and whether the benefits of AI justify its environmental footprint. Major AI companies have signed deals for nuclear, solar, and wind power specifically to run their data centers.

A contract where a company agrees to buy electricity directly from a power generator, often a renewable energy project, at a set price for a set period (usually 10–20 years). AI companies use PPAs to secure massive amounts of clean energy for their data centers. When you read that Microsoft or Google signed a deal with a nuclear plant, that’s a PPA.

The total greenhouse gas emissions associated with building, training, and running AI systems. This includes the electricity used by data centers, the manufacturing of chips and hardware, the cooling systems, and even the construction of the buildings. Estimates vary widely, but the AI industry’s energy consumption is growing faster than almost any other sector. Understanding this footprint is part of making informed decisions about AI use.

Processing data on or near the device itself rather than sending it to a distant data center. When your phone uses face recognition to unlock without contacting a server, that’s edge computing. It’s faster (no round trip to the cloud), can work offline, and can be more private (your data never leaves your device). Many AI applications are moving toward edge computing for these reasons.

The delay between when you send a request and when you get a response. When you ask Claude a question and there’s a pause before the answer appears, that’s latency. It’s affected by the distance to the data center, how busy the servers are, and how complex the request is. Lower latency means a snappier, more conversational experience.

The field focused on making sure AI systems do what we want them to do, and don’t do things we don’t want them to do. This ranges from immediate concerns (making sure a chatbot doesn’t give dangerous medical advice) to long-term concerns (ensuring very powerful future AI systems remain under human control). AI safety isn’t about being afraid of robots; it’s about building reliable, trustworthy technology.

The challenge of making sure an AI system’s goals and behaviors match human intentions and values. A model that’s "aligned" does what you actually want, not a technically correct but unhelpful literal interpretation of your request. Alignment is harder than it sounds because human values are complex, contextual, and sometimes contradictory. It’s one of the central technical and philosophical challenges in AI development.

When an AI system produces unfair or skewed results because the data it was trained on reflected existing human prejudices. If a hiring model was trained mostly on data from a male-dominated industry, it might learn to favor male candidates, not because anyone told it to, but because that’s the pattern in the data. AI bias isn’t a bug that can be fixed with a single patch; it requires ongoing attention to training data, evaluation methods, and real-world outcomes.

Limits and rules built into an AI system to prevent harmful, inappropriate, or off-topic outputs. When a chatbot refuses to help you make a weapon or tells you it can’t provide medical diagnoses, that’s a guardrail at work. Guardrails are implemented through a combination of system prompts, RLHF, and content filtering. The challenge is setting guardrails that prevent genuine harm without making the system so restricted it’s unhelpful.

The practice of deliberately trying to make an AI system fail, behave badly, or produce harmful outputs. Essentially, hiring people to attack your own product before bad actors do. Red teams try to bypass guardrails, extract dangerous information, or trick the model into producing biased or toxic content. It’s a critical part of safety testing before releasing a model to the public.

A broad term for the practice of developing and deploying AI in ways that are ethical, fair, transparent, and accountable. It encompasses bias mitigation, safety testing, privacy protection, environmental impact, and honest communication about what AI can and can’t do. Most major AI companies have published responsible AI principles; the degree to which they follow through varies.

The ability to understand how and why an AI model produced a specific output. Current large models are largely "black boxes." They produce results, but even the people who built them can’t always explain exactly why the model said what it said. Explainability research aims to open that black box, which is important for trust, debugging, accountability, and regulatory compliance.

Laws and rules governing how AI can be developed and used. The EU’s AI Act is the most comprehensive AI law to date. The U.S. has taken a more sector-specific approach with executive orders and agency guidelines. China has its own set of AI regulations. The regulatory landscape is evolving rapidly and varies significantly by country.

The European Union’s landmark regulation classifying AI systems by risk level, from minimal risk (spam filters) to unacceptable risk (social scoring by governments). Higher-risk AI systems face stricter requirements: transparency obligations, human oversight mandates, data quality standards, and more. The Act began taking effect in stages starting in 2024 and will influence AI regulation worldwide.

AI-generated synthetic media, usually a video or audio recording that makes it look or sound like a real person is saying or doing something they never actually said or did. Deepfakes range from harmless entertainment to serious threats: political disinformation, fraud, and non-consensual intimate imagery. Detection tools exist but lag behind generation capabilities. When you see a surprising video online, questioning whether it’s real is now a reasonable first instinct.

False or misleading information. Misinformation is spread unintentionally, when someone shares something they genuinely believe is true. Disinformation is spread deliberately to deceive. AI makes both worse: generative AI can produce convincing but false text, images, and video at scale. At the same time, AI tools can potentially help detect and combat false content. The net effect is an information environment where critical thinking matters more than ever.

The theoretical risk that a sufficiently advanced AI system could pose a fundamental threat to human civilization. This is the most speculative and debated area of AI safety. Some researchers consider it a serious concern that deserves proactive attention; others consider it a distraction from more immediate, concrete AI harms like bias, job displacement, and surveillance. Reasonable people disagree. What matters is that you’ve heard both sides.

Software you access online through a subscription instead of buying and installing on your computer. Most AI tools you interact with (Claude, ChatGPT, Midjourney) follow this model. You pay monthly or yearly for access. The AI runs on the company’s servers; you use it through a website or app.

Companies that provide AI capabilities through an API so other businesses can add AI to their own products without building models from scratch. Instead of training your own LLM (which would cost millions), you pay Anthropic, OpenAI, or Google per usage to access theirs. This is how most AI-powered apps work behind the scenes.

The observation that AI models generally get more capable when you increase three things: the amount of training data, the number of parameters, and the amount of compute used for training. Scaling laws have driven the "bigger is better" approach in AI development. Whether these laws will continue to hold, or whether we’ll hit diminishing returns, is one of the most important open questions in the field.

A hypothetical AI system that can perform any intellectual task a human can, with the same flexibility and generality. Current AI is "narrow." Claude is great at language tasks but can’t ride a bike or feel emotions. AGI would be a system with human-level versatility. Whether AGI is five years away, fifty years away, or fundamentally impossible depends on who you ask. It’s the north star (or the bogeyman, depending on your perspective) of AI research.

The concern that AI will eliminate certain jobs or significantly reduce the number of humans needed to do them. History shows that major technologies do displace specific jobs while creating new ones, but the transition can be painful and unevenly distributed. AI is most likely to augment many jobs (making workers more productive), fully automate some jobs (particularly routine, repetitive tasks), and create new jobs that don’t exist yet. The honest answer is that nobody knows exactly how this will play out.

Legal rights over creative works and inventions. AI has created a tangle of IP questions: Can AI-generated content be copyrighted? If an AI model was trained on copyrighted books and artwork, do the original creators deserve compensation? Multiple lawsuits are working through these questions right now. The legal framework hasn’t caught up with the technology, and the outcomes will shape the AI industry for decades.

Data generated by AI models rather than collected from the real world. If you need a million examples to train a model but only have a thousand real ones, you might use AI to generate the rest. Synthetic data can help with privacy (no real people’s data involved) and availability (you can create data for rare scenarios). The risk is that if synthetic data has flaws, the model trained on it inherits those flaws.

The idea that knowledge gained from one task can be applied to another. A model trained to recognize objects in photos can be adapted to recognize medical conditions in X-rays, because the low-level visual skills transfer. Transfer learning is one of the reasons foundation models are so valuable. The general capabilities they learn during pre-training provide a powerful starting point for many specific applications.

A model’s ability to perform a task it was never explicitly trained for. Zero-shot means the model can do it with no examples; you just describe what you want. Few-shot means you give it a handful of examples and it picks up the pattern. The fact that modern LLMs can do this at all is one of their most remarkable and surprising capabilities. It’s why you can ask Claude to do things it was never specifically taught to do, and it often succeeds.

A distinction that matters more than it sounds. Open weights means the trained model’s parameters are publicly available; you can download and run it. Open source traditionally means the full package: code, data, training process, and documentation are all available. Some models marketed as "open source" are really just open weights. The difference matters because without the full picture, you can use the model but can’t fully understand or reproduce how it was built.

An architecture where a model contains multiple specialized sub-networks ("experts"), and for any given input, only a few relevant experts activate. Think of a hospital where you see a specialist based on your symptoms rather than every doctor examining every patient. MoE models can be very large in total parameter count but efficient to run because only a fraction of the model is working at any time.

A technique for making AI models smaller and faster by reducing the precision of their internal numbers. Instead of using highly precise decimal numbers (like 3.14159265), quantized models use rougher approximations (like 3.14). The model gets slightly less accurate but dramatically more efficient, often small enough to run on a phone or laptop instead of a server farm. It’s one of the key techniques making AI more accessible on everyday devices.

AI that runs directly on your phone, laptop, or other personal device rather than on a remote server. Apple Intelligence, Google’s on-device features, and various phone-specific AI features all fall into this category. The advantages: it’s faster (no internet round trip), works offline, and keeps your data on your device. The limitation: your phone is far less powerful than a data center, so on-device models are necessarily smaller and less capable.

A virtual replica of a physical object, system, or process. Factories create digital twins of their production lines to test changes virtually before implementing them physically. Cities create digital twins to model traffic flow. AI is increasingly used to power these simulations, making them more accurate and responsive. You probably won’t interact with a digital twin directly, but they’re quietly shaping how infrastructure, manufacturing, and logistics are designed.