AI Glossary

An AI system that can take actions on its own — browsing the web, writing files, running code, sending messages — not just answering questions. It has goals and can use tools to achieve them.

The challenge of ensuring AI systems behave according to human values and intended purposes. This involves developing techniques to make AI understand and act upon what humans actually want, not just what they literally ask for.

The maximum amount of text a model can 'see' at once — both your input and its output combined. Larger context = can process longer documents.

When an AI confidently generates false or made-up information as if it were true. A major challenge — the AI doesn't 'know' it's wrong.

An AI model trained on massive amounts of text data that can understand and generate human-like language. Think of it as a very sophisticated autocomplete that can write essays, code, and answer questions.

The text you type to tell an AI what you want. The better your prompt, the better the response. Prompt engineering is the art of writing good prompts.

RAG combines AI language models with a knowledge database, allowing the AI to search and retrieve relevant information before generating responses. This helps the AI provide accurate, up-to-date answers grounded in real data rather than just relying on training knowledge.

The basic unit AI models process text in. Roughly 1 token ≈ 3/4 of a word in English. Models have token limits (context windows) and API pricing is per token.

The process of converting text into smaller pieces (tokens) that AI models can understand. These tokens might be whole words, parts of words, or even characters. The AI processes language token by token, similar to how we read word by word.

A prompting technique where you ask the AI to 'think step by step' before giving the final answer. Significantly improves accuracy on math, logic, and complex problems.

A training approach where AI models learn to behave according to a set of predefined principles or 'constitution'. The AI critiques its own responses against these rules, gradually improving its behavior without requiring extensive human oversight for each interaction.

Direct Preference Optimization is a more efficient alternative to RLHF that directly optimizes language models based on human preferences. It's less complex to implement while achieving similar alignment improvements in AI behavior.

A knowledge transfer technique where a large, capable 'teacher' model trains a smaller 'student' model to replicate its behavior. This creates smaller, faster models that retain most of the original's capabilities while being more practical for deployment.

Converting text (or images) into a list of numbers (a vector) that captures its meaning. Similar texts get similar numbers. This enables semantic search — finding content by meaning, not just keywords.

A distributed training method where AI models learn from data spread across multiple devices or locations without centralizing that data. This allows AI to improve from diverse real-world usage while maintaining user privacy by keeping data on original devices.

Giving the AI a few examples of what you want before asking it to do the task. Dramatically improves accuracy for specific formats or styles.

Taking a pre-trained model and training it further on specific data to make it better at a particular task. Like teaching a college graduate to specialize.

Training method where models learn to follow human instructions by practicing on example tasks with clear directions. This technique teaches AI to understand and execute specific tasks rather than just generating text, improving reliability on practical applications.

Combining multiple trained models or parameters to create new models with enhanced capabilities. This can involve blending different fine-tuned models or merging training techniques to achieve better performance than individual models alone.

The skill of crafting effective prompts to get the best results from AI models. Includes techniques like few-shot examples, chain-of-thought, and role-playing.

Quantized Low-Rank Adaptation combines LoRA with model quantization, allowing fine-tuning of massive AI models on consumer hardware. This technique can reduce memory requirements by 3-4x while maintaining model quality.

A technique where the AI first searches a knowledge base for relevant information, then uses it to generate accurate answers. Reduces hallucinations and keeps answers grounded in real data.

Reinforcement Learning from Human Feedback is a training method where AI models learn by receiving feedback from humans. Instead of just learning from data, the AI learns what humans consider good or bad responses, leading to more helpful and aligned behavior.

A hidden instruction given to the AI before the conversation starts. Sets the AI's personality, rules, and behavior. The user usually can't see it.

The neural network architecture behind virtually all modern AI models. Invented by Google in 2017. Uses 'attention' to understand relationships between words regardless of distance in text.

A database designed to store and search embeddings efficiently. Instead of matching keywords, it finds the most semantically similar content.

Processing multiple inputs together efficiently by grouping them into batches. This approach optimizes hardware usage and can be more cost-effective for high-volume applications compared to processing individual requests one at a time.

NVIDIA's programming platform and API that enables AI models to run on graphics cards. It's the foundation that allows AI software to access GPU accelerations for the intensive mathematical calculations required by machine learning models.

AI processing performed locally on devices rather than in centralized cloud servers. This approach improves privacy, reduces latency, and enables AI functionality without internet connectivity, though it's limited by device computational capacity.

A file format for quantized AI models, designed for efficient local inference. The standard format used by llama.cpp and Ollama. Replaced the older GGML format.

The chip that runs AI models. Originally designed for gaming graphics, but the parallel processing power is perfect for AI math. NVIDIA dominates the AI GPU market.

Running AI models on your own computer instead of using a cloud API. Free, private, no internet needed — but requires good hardware (especially VRAM).

AI capabilities that operate completely on user devices like smartphones, laptops, or smart cameras without requiring external servers. This ensures privacy for sensitive data and enables AI features even in environments with poor or no internet connectivity.

Compressing an AI model by using less precise numbers. Like converting a RAW photo to JPEG — smaller file, slightly lower quality, but usually barely noticeable. Q4 = 4-bit, Q8 = 8-bit, FP16 = 16-bit (full).

Specialized processing units on NVIDIA GPUs designed specifically for AI matrix operations. These provide dramatically faster performance than traditional GPU cores for the mathematical operations used in deep learning, essential for efficient AI inference and training.

Memory on your graphics card (GPU). Determines the largest AI model you can run locally. More VRAM = bigger and smarter models. The #1 bottleneck for local AI.

The core innovation of transformers. Lets the model weigh how important each word is relative to every other word. 'The cat sat on the mat because it was tired' — attention helps the AI understand 'it' refers to 'the cat'.

The core technology that allows AI models to focus on the most relevant parts of input data. When processing text, it helps the AI understand which words are most important to each other word, similar to how humans pay attention to key details in a conversation.

An optimized attention mechanism that works more efficiently with limited GPU memory. It achieves faster processing and enables use of longer context windows by being more memory-efficient, removing traditional bottlenecks in processing long sequences.

A performance optimization that stores pre-computed values (keys and values) from the attention mechanism, avoiding redundant calculations for tokens that have been seen before. This dramatically speeds up inference when generating long responses.

A lightweight fine-tuning technique that only trains a tiny fraction of the model's parameters. Makes fine-tuning accessible — you can customize a 70B model on a single GPU.

A model architecture that activates only specific parts of the network for each input, making large models more efficient. It's like a team of specialists where only relevant experts respond to each query, reducing computational costs while maintaining capability.

A model architecture where only a fraction of the network is active for each task. Like a company with specialists — the right expert handles each question. More efficient than activating the entire model.

Multiple AI agents working together, each with different roles or specializations. Like a team where one agent does research, another writes code, and a third reviews quality.

AI systems that can process and understand multiple types of media simultaneously - text, images, audio, and video. These models can describe images, answer questions about sounds, or generate content that spans different media formats.

The 'knowledge' of an AI model, stored as numbers. More parameters generally = more capable but requires more hardware. Think of them as the synapses in a brain.

An optimization technique that uses a smaller, faster model to predict the next tokens, then a larger model corrects any mistakes. This significantly speeds up text generation while maintaining quality, like having a quick assistant do first drafts for review by an expert.

A sampling method that considers only the most probable next tokens whose combined probability reaches a certain threshold. This provides more nuanced control than temperature alone, balancing diversity with relevance in generated text.

AI models specifically designed to understand both visual and textual information together. They can describe images, answer questions about what's in pictures, read text from images, and perform tasks requiring understanding of visual context.

A way for software to communicate with AI models programmatically. Instead of typing in a chat window, your code sends requests and gets responses. Pay-per-token pricing.

AI systems capable of independent decision-making and action-taking without constant human supervision. These agents can set goals, plan actions, use tools, and adapt strategies to achieve objectives in complex environments over extended periods.

A technique that adds carefully calculated noise to data or model outputs to prevent individual user information from being identifiable while preserving overall utility. This enables using sensitive data for AI training without compromising privacy.

A structured way for AI models to invoke specific functions or APIs with defined parameters. This allows seamless integration between AI conversations and external applications, enabling automated workflows where AI can trigger specific software actions.

The process of connecting AI responses to factual, verified information or source data. This helps reduce hallucinations by ensuring the AI's outputs are anchored in real, substantiated knowledge rather than potentially incorrect assumptions.

Safety mechanisms built into AI systems that prevent harmful, inappropriate, or unwanted outputs. These can include content filters, rule-based constraints, or automated monitoring that ensure AI behavior stays within acceptable boundaries.

The process of getting an AI model to generate a response. When you send a prompt and get an answer — that's inference. Training creates the model; inference uses it.

Techniques or prompts designed to bypass AI safety measures and restrictions, causing the model to generate content it was specifically trained not to produce. This ongoing security challenge requires continuously updating defensive measures.

Models where the code and/or weights are freely available. 'Open source' means full code access. 'Open weights' means you can use the model but may have license restrictions.

A security vulnerability where malicious instructions are embedded within user inputs, potentially causing AI systems to behave in unintended ways. This can occur when systems fail to properly separate user content from system instructions.

Systematic testing of AI systems by simulating adversarial attacks to identify vulnerabilities, biases, or harmful behaviors. This proactive approach helps improve safety by discovering potential failure modes before they can be exploited by malicious actors.

Searching based on meaning and context rather than exact keyword matches. This allows finding relevant results even when search terms don't appear exactly in the source material, understanding the intent behind queries.

AI systems that process and respond to data in real-time as it becomes available, rather than waiting for complete datasets. This enables applications like live transcription, real-time translation, or immediate responses to changing inputs.

A setting that controls how creative/random the AI's responses are. Low (0.0) = deterministic and focused. High (1.0+) = creative and varied. Like a creativity dial.

An AI model's ability to call external tools — search the web, run code, read files, use APIs. Transforms a chatbot into an agent that can actually do things.

The process of teaching an AI model by feeding it vast amounts of data. Pre-training uses the entire internet; fine-tuning uses specific data. Costs millions of dollars for large models.

An add-on for image AI that gives you precise control over the output — pose, edges, depth. Instead of hoping the AI gets the composition right, you guide it.

The technique behind modern image generation AI. Starts with random noise and gradually 'denoises' it into an image guided by your text prompt. Like sculpting a statue from a block of marble.

Text telling the image AI what you DON'T want in the image. Helps avoid common artifacts and unwanted elements.

Using AI to increase image resolution while adding realistic detail. Turns a small 1024px image into a large 4096px print-ready file without losing quality.