A practical knowledge base for AI protocols, agents, core concepts, frameworks, and tools.
A diffusion model is trained by gradually adding noise to real data until it becomes pure static, then learning to run that process backwards. To generate, it starts from random noise and denoises step by step according to a prompt, arriving at an image, video, or audio clip
Embodied AI is AI with a physical body: it senses a real environment through sensors, changes it through actuators, and learns from the results. Compared with pure software models it has one extra thing — a real world with no undo — which is why data, real-time control, and safety are its three hard limits
GraphRAG layers a knowledge graph on top of ordinary RAG: entities and relationships are extracted from documents into a graph, and retrieval traverses relationships rather than only matching semantically similar passages. It excels at questions requiring multi-document chaining or a global summary
LoRA is a parameter-efficient fine-tuning method: freeze the original model weights and train only a small pair of low-rank matrices alongside them. Training cost drops to a fraction of full fine-tuning, and the resulting adapter file is usually only tens of megabytes, so it can be attached and swapped at will
Open weights means a model's trained parameters are published for anyone to download, self-host, and fine-tune. People call these "open source models," but most don't meet the full open source definition — training data and training code usually stay private, and licenses may restrict use
Quantization compresses model weights from 16-bit floats down to 8-bit, 4-bit, or lower integer representations, trading a controlled amount of precision for smaller memory footprint and faster inference. It is the standard way to run large models locally and to cut inference cost
Scaling laws are empirical regularities: a model's prediction error falls as a predictable power law in parameters, data, and compute. They turned "how good a model does this budget buy" into something you can estimate in advance, and they are the direct justification for the compute build-out of recent years
Shadow AI is employees using AI tools on company data without IT or security approval. What separates it from shadow IT is that the data itself leaves, not just a system entry point — and bans rarely work, so the viable path is providing a compliant alternative
A small language model is one small enough to run on ordinary consumer hardware and fast enough to serve a single user. In agent systems, most model calls are repetitive, narrow, fixed-format subtasks — and handling those with a small model is often cheaper and more stable
Spec-driven development treats versioned specifications as the single source of truth, with code generated and maintained against them by humans and AI agents. It targets three characteristic failures of coding agents: intent drift, context decay, and outputs no one can judge for lack of acceptance criteria
Synthetic data is training data generated by a model or a program rather than collected from the real world. It has become a necessity as high-quality real data runs short and in domains like robotics where collection is prohibitively expensive — but used carelessly it amplifies bias and degrades models
The Transformer is a neural network architecture introduced in 2017. It dropped the word-by-word recurrent structure of its predecessors and used attention to read an entire input at once. Today it is the skeleton of nearly every large language model, multimodal model, and mainstream image or video generator
Showing 12 of 12 entries
Every entry grouped by topic — one click straight to the explanation.
