explained

A foundation model is any broadly capable model trained on massive data. An LLM is a specific kind — foundation models also include vision, audio, and multimodal.

Parameters are learned by the model during training. Hyperparameters are set by humans before training. Mixing them up causes confused debugging.

Parameters is the umbrella term for every learnable number in a model — weights plus biases. In practice, people use the two words interchangeably.

An algorithm is a procedure. A model is the trained result of running that procedure on data. Same algorithm can produce many models.

Narrow AI excels at one task. General AI (AGI) would match humans at any intellectual task. All deployed AI today is narrow.

Generative AI creates new content. Predictive AI forecasts outcomes from patterns. Different goals, different models, often used together.

AI makes decisions. Automation executes repetitive tasks. Robotics is the physical body. They combine in many modern systems but are distinct concepts.

A jailbreak bypasses an AI's safety training. Prompt injection hijacks the AI's task. Different goals, overlapping techniques.

Prompt injection is when an attacker hides instructions in user input or external content, hijacking the AI to do something it should not.

RAG retrieves facts at query time. Fine-tuning bakes behavior into model weights. Use RAG for facts; fine-tune for style or narrow tasks.

Supervised learning uses labeled data. Unsupervised learning finds patterns in unlabeled data. Most production systems use both plus self-supervised learning.

Training is how a model learns from data. Inference is how it applies what it learned to new inputs. Different costs, hardware, and time scales.

Three ways to get an AI model to do a task: ask it (zero-shot), show examples (few-shot), or retrain it (fine-tuning). Each has different costs and trade-offs.

Grounding is the practice of tying AI output to verifiable external sources so answers are factual and citable.

Hallucination is when an AI model generates confident but false information. It is the biggest risk in production LLM applications.

A vector database stores embeddings and finds the most similar ones fast. It powers semantic search, RAG, and recommendations.

An embedding is a list of numbers that represents the meaning of text, images, or other data. Similar meanings produce similar numbers.

The context window is the maximum number of tokens an AI model can read and write at once. Bigger windows let the model handle longer documents and conversations.

A token is the basic unit of text an AI model reads and writes. One token is roughly 3-4 English characters, not a full word.

Temperature is a parameter that controls how random or focused an AI model's output is. Lower values produce predictable text; higher values produce creative text.