AI Wiki

An AI system that can autonomously plan and execute multi-step tasks, using tools (web search, code execution, API calls) to achieve a goal. Unlike a simple chatbot that answers one question at a time, an agent decides what to do next based on what it's learned so far.

The challenge of making AI systems behave in ways that match human values and intentions. An aligned model does what you mean, not just what you said — and avoids harmful actions even when not explicitly told not to.

A structured way for software to talk to other software. In AI, this usually means sending a request (your prompt) to a provider's server and getting a response (the model's output) back. REST APIs over HTTPS are the standard.

The core mechanism in Transformers that lets a model weigh which parts of the input are most relevant to each other. Instead of reading text left-to-right like older models, attention lets every word "look at" every other word simultaneously to understand context.

A standardized test used to evaluate and compare AI models. Benchmarks measure specific capabilities — reasoning (ARC), math (GSM8K), coding (HumanEval), general knowledge (MMLU) — and produce scores that can be compared across models.

Systematic patterns in AI outputs that reflect or amplify societal prejudices present in training data. Bias can appear in text generation, image creation, hiring tools, and anywhere models make decisions that affect people differently.

A prompting technique where you ask the model to show its reasoning step by step before giving a final answer. Instead of jumping to a conclusion, the model "thinks out loud," which dramatically improves accuracy on complex tasks.

The maximum amount of text (measured in tokens) a model can process in a single conversation. This includes both your input and the model's output. If a model has a 200K context window, that's roughly 150,000 words — about two novels.

The body of text (or other data) used to train a model. A corpus can range from curated collections of books and papers to massive scrapes of the entire internet. The quality and composition of the corpus fundamentally shapes what the model knows and how it behaves.

A type of generative model that creates images (or video, audio) by starting with pure noise and gradually removing it until a coherent output appears. The model learns to reverse the process of adding noise to real data. Stable Diffusion, DALL-E 3, and Midjourney all use variants of this approach.

A way to represent text (or images, or audio) as a list of numbers (a vector) that captures its meaning. Similar concepts end up close together in this number space — "cat" and "kitten" are nearby, while "cat" and "economics" are far apart.

A specific URL where an AI API accepts requests. For example, Anthropic's message endpoint is where you send prompts to Claude. Different endpoints serve different functions: text generation, embeddings, image creation, model listing.

Taking a pre-trained model and training it further on a smaller, specific dataset to specialize its behavior. Like taking a general practitioner and putting them through a surgical residency — same foundational knowledge, new expertise.

A large model trained on broad data that serves as a base for many different tasks. Claude, GPT, Gemini, and Llama are all foundation models. They're "foundational" because they can be adapted to almost anything — writing, coding, analysis, image understanding — without being specifically trained for each task.

A model architecture where two neural networks compete: a generator creates fake data, and a discriminator tries to tell real from fake. Through this adversarial game, the generator gets better at creating realistic outputs. Dominated image generation from 2014 to ~2022.

Originally designed for rendering graphics, GPUs turned out to be perfect for AI because they can do thousands of math operations simultaneously. Training and running AI models is essentially massive matrix multiplication — exactly what GPUs are built for. NVIDIA dominates this market.

Connecting a model's responses to factual, verifiable sources rather than letting it rely solely on its training data. Grounding techniques include RAG, web search integration, and citation requirements. A grounded response says "according to [source]" rather than just asserting facts.

Safety mechanisms that prevent AI models from generating harmful, inappropriate, or off-topic content. Guardrails can be built into the model during training (RLHF), applied through system prompts, or enforced by external filters that check outputs before they reach users.

When an AI model generates information that sounds confident and plausible but is factually wrong or entirely fabricated. The model isn't "lying" — it's pattern-matching its way to fluent text without a concept of truth. Fake citations, invented statistics, and non-existent API methods are common examples.

The process of running a trained model to generate outputs. Training is learning; inference is using what was learned. Every time you send a prompt to Claude or generate an image with Stable Diffusion, that's inference. It's what costs providers GPU hours and what you pay for per token.

The delay between sending a request and getting the first response. In AI, this is often measured as Time to First Token (TTFT) — how long before the model starts streaming its answer. Affected by model size, server load, network distance, and prompt length.

A neural network trained on massive amounts of text to understand and generate human language. "Large" refers to the number of parameters (billions) and the size of the training data (trillions of tokens). Claude, GPT, Gemini, Llama, and Mistral are all LLMs.

A technique that makes fine-tuning dramatically cheaper by only training a small number of additional parameters instead of modifying the entire model. LoRA "adapters" are lightweight add-ons (often just megabytes) that modify a model's behavior without retraining its billions of parameters.

An open protocol (created by Anthropic) that standardizes how AI models connect to external tools and data sources. Think of it as USB-C for AI — one standard interface instead of custom integrations for every tool. MCP servers expose capabilities; MCP clients (like Claude) consume them.

An architecture where the model contains multiple "expert" sub-networks, but only activates a few of them for each input. A router network decides which experts are relevant for a given token. This means a model can have 100B+ total parameters but only use 20B for any single forward pass.

A model that can understand and/or generate multiple types of data: text, images, audio, video, code. Claude can read images and text; some models can also produce images or speech. "Multimodal" contrasts with "unimodal" models that only handle one type.

A computing system loosely inspired by biological brains, made of layers of interconnected "neurons" (mathematical functions) that learn patterns from data. Information flows through layers, getting progressively transformed until the network produces an output. Every modern AI model is a neural network of some kind.

When a company releases a model's trained parameters for anyone to download and run. "Open weights" is more accurate than "open source" because most released models don't include training data or training code — you get the finished model but not the recipe. Llama, Mistral, and Qwen are open-weights models.

When a model memorizes its training data too well and loses the ability to generalize to new inputs. Like a student who memorizes answers to practice tests but can't solve new problems. The model performs great on training data but poorly on anything it hasn't seen before.

The initial, massive training phase where a model learns language (or other modalities) from a huge corpus. This is the expensive part — thousands of GPUs running for weeks or months, costing millions of dollars. The result is a foundation model that understands language but hasn't been specialized for any task yet.

The practice of crafting inputs to get better outputs from AI models. This ranges from simple techniques (being specific, providing examples) to advanced methods (chain of thought, few-shot prompting, role assignment). Despite the fancy name, it's fundamentally about communicating clearly with a statistical system.

Reducing a model's precision to make it smaller and faster. A model trained in 32-bit floating point can be quantized to 8-bit, 4-bit, or even lower — shrinking its size by 4-8x with surprisingly small quality loss. GGUF is the popular format for local inference via llama.cpp.

A technique that gives AI models access to external knowledge by retrieving relevant documents before generating a response. Instead of relying only on what the model learned during training, RAG searches a knowledge base, finds relevant chunks, and includes them in the prompt as context.

Restrictions on how many API requests you can make per minute/hour/day. Providers impose rate limits to prevent server overload and ensure fair access. Limits typically apply per API key and can restrict requests per minute (RPM) and tokens per minute (TPM).

The practice of deliberately trying to make an AI model fail, misbehave, or produce harmful outputs. Red teams probe for vulnerabilities: jailbreaks, bias, misinformation generation, privacy leaks. Named after military wargaming where a "red team" plays the adversary.

A training technique where human evaluators rank model outputs by quality, and this feedback is used to train a reward model that guides the AI toward better responses. It's what turns a raw pre-trained model (which just predicts next words) into a helpful, harmless assistant.

An alternative to Transformers that processes sequences by maintaining a compressed "state" instead of using attention over all tokens. Mamba is the most well-known SSM architecture. SSMs scale linearly with sequence length (vs. quadratic for attention), making them potentially much more efficient for very long contexts.

A special instruction given to a model at the start of a conversation that sets its behavior, personality, and rules. Unlike user messages, the system prompt is meant to be persistent and authoritative — it defines who the model is for this session. "You are a helpful coding assistant. Always use TypeScript."

A parameter that controls how random or deterministic a model's output is. Temperature 0 makes the model always pick the most probable next token (deterministic, focused). Temperature 1+ makes it more willing to pick less probable tokens (creative, unpredictable). Most APIs default to around 0.7.

The basic unit of text that AI models process. A token is typically a word or word fragment — "understanding" might be one token, while "un" + "der" + "standing" could be three. On average, one token is roughly 3/4 of a word in English. Models read, think, and charge in tokens.

The ability of an AI model to call external functions or tools during a conversation. Instead of just generating text, the model can decide to search the web, run code, query a database, or call an API — then incorporate the results into its response. The model outputs a structured "tool call" that the host application executes.

The neural network architecture behind virtually all modern LLMs and many image/audio models. Introduced by Google in the 2017 paper "Attention Is All You Need," Transformers use self-attention to process all parts of an input simultaneously rather than sequentially, enabling massive parallelism during training.

A database optimized for storing and searching embeddings (vectors). Instead of matching exact keywords like a traditional database, vector databases find the most semantically similar items. You ask "how to fix a memory leak" and it returns documents about "debugging RAM consumption" because the embeddings are close.

The memory on a GPU, separate from system RAM. AI models must fit in VRAM to run on a GPU. A 7B parameter model in 16-bit precision needs ~14GB of VRAM. Consumer GPUs have 8-24GB; datacenter GPUs (A100, H100) have 40-80GB. VRAM is almost always the bottleneck for local AI.

Zero-shot means asking a model to do a task with no examples — just the instruction. Few-shot means providing a handful of input-output examples in the prompt before the actual request. "Here are 3 examples of how to format this data... now do this one." The model learns the pattern from context alone, no training required.