Introduction to AI Engineering and Foundation Models

Comprehensive vocabulary flashcards covering AI concepts including hallucination, prompt engineering, language model architectures, training methodologies, and computational optimization.

Hallucination

A situation where an AI is given a task but completes it with a mistake or generates incorrect facts based on a wrong assumption.

Prompt Engineering

The process of crafting specific instructions to a model to produce a desired outcome.

AI Engineering

The process of building applications using already existing foundation models, focusing more on model adaptation than training.

Token

The basic unit of a language model used to break down original text into meaningful components.

Tokenization

The process of breaking down original text into smaller units called tokens.

Masked Language Model

A model trained to predict a missing token anywhere in a sequence using context from both before and after the missing area; primarily used for non-generative tasks like text classification.

Autoregressive Language Model

A model used to predict a missing token based specifically on the context of the preceding sequence to generate tokens one after another.

In-context Learning

The method of teaching a model how to perform a task within the prompt itself.

Zero-shot Learning

A scenario where a model is taught to perform a task with no previous examples provided in the prompt.

Few-shot Learning

Teaching a model to learn a task by providing examples within the prompt, usually between $1$ and $50$ examples.

System Prompt

The part of a prompt that provides the task description and shapes the model's persona or overall behavior.

User Prompt

The specific task or query provided by the user within a conversation.

Prompt Breeder

A tool that uses prompt mutation to effectively create prompts that work for different models.

Reverse Prompt Engineering

The process of deducing the original system prompt used for a specific AI application.

PAIR (Prompt Automatic Iterative Refinement)

An automated system that generates, sends, and tunes prompts based on target responses until an objective is achieved.

Generative AI

AI models that can generate open-ended outputs.

Self-Supervision

A training method where a language model creates its own labels from unlabeled data by observing tokens and their context.

Parameter

A variable in a Machine Learning algorithm that is updated by its learning process; the total count determines the model size.

Foundational Models

Models that incorporate multiple data modalities, such as audio and visual data, rather than being limited only to text.

Large Multimodal Model (LMM)

A generative model that includes more than one data modality.

CLIP

A model taught to learn images by receiving $400 \times 10^{6}$ image-text data sets from the internet.

Retrieval Augmented Generation (RAG)

A method to improve a model's ability to construct accurate answers by connecting it to external data sources to fix information-based failures.

MLOps (Machine Learning Operations)

The previous name for AI Engineering, referring to building and maintaining machine learning applications.

Reactive Feature

An application feature that shows a reaction only in response to a user's specific request or action.

Proactive Feature

An application feature that provides a reaction automatically when it identifies an opportunity.

Human-in-the-loop

An application design where humans are required for the final decision-making process.

Pretraining

The initial process of training a model for general text completion tasks.

Fine-tuning

Further training of a pretrained model to adapt it to specific tasks, styles, or behaviors by adjusting its weights.

Dataset Engineering

The process of generating and annotating the data needed for training AI models.

Inference Optimization

The process of making an AI model cheaper and faster during its operational use.

Sampling

The mechanism by which a model selects its final output from a list of possible probabilistic outcomes.

Transformer Architecture

A sequence-to-sequence architecture that uses an encoder to process input and a decoder to process output.

Snowballing Hallucination

A type of self-delusion where a model builds a wrong assumption and then continues to double down with more incorrect facts based on that first error.

Agentic AI

AI systems that can perceive, observe, think, reason, and act autonomously to achieve tasks.

Distillation

The process of training a smaller model to mimic the behavior of a larger model by using data generated by the larger model.

Semantic Parsing

The process of converting natural language into a structured format such as JSON.

PEFT (Parameter-Efficient Fine-Tuning)

A technique to reduce the memory footprint of fine-tuning by updating only a small number of trainable parameters.

Quantization

The process of converting a model from a format with more bits to a format with fewer bits to save memory.

Backpropagation

A model used to train neural networks consisting of a forward pass to compute output and a backward pass to update weights based on the loss.

Gradient

The derivative of the loss with respect to a trainable parameter, representing how much that parameter contributes to the error.

Adam

A common optimizer used for adjusting trainable parameters in transformer-based models.

Inference Memory Formula

The calculation $N \times M \times 1.2$, where $N$ is the parameter count and $M$ is the memory needed for each parameter.

Activation Recomputation

A memory-saving technique where activations are computed each time they are needed instead of being stored.

PTQ (Post-Training Quantization)

The practice of quantizing a model after it has already finished the training process.