Your Complete Guide to Artificial Intelligence

Architecture using self-attention to process sequences in parallel — powers virtually all modern LLMs.

Allows models to focus on relevant input parts — enables long-range dependency capture.

Breaking text into tokens that a model processes numerically. Affects pricing and context limits.

Dense vector where semantic similarity maps to proximity. Foundation of semantic search.

Adapting a pre-trained model to a specific task using a smaller task-specific dataset.

Retrieval-Augmented Generation — combines LLMs with retrieved documents to reduce hallucinations.

When an LLM generates plausible-sounding but factually wrong or fabricated output.

Max tokens a model processes at once — affects memory and reasoning over long content.

Reducing model precision (float32 → int4) to shrink model size with minimal quality loss.

Low-Rank Adaptation — trains small adapter matrices for efficient fine-tuning. Saves GPU memory.

Reinforcement Learning from Human Feedback — trains models to align with human preferences.

Anthropic's approach: train AI with guiding principles to be helpful, harmless, and honest.

Mixture of Experts — routes inputs to specialized sub-networks for efficient large-scale models.

Learns to reverse a noise-addition process to generate high-quality images or other content.

Running a trained model to generate output — distinct from training which updates model weights.

Artificial General Intelligence — hypothetical AI matching human-level ability across all domains.