Baseline, intro-level tutorial: key algorithms and techniques that underpin current production AI (especially generative AI), plus a sourced map of major model families (US and non-US) and a few high-impact non-LLM techniques (matching, retrieval, audio). This page is designed as a take-off point for deeper subtopic documents.
Most modern production "AI" is machine learning, and the most capable systems are typically deep neural networks trained using gradient-based optimization (the foundations are covered in standard deep learning texts such as Goodfellow, Bengio, Courville - Deep Learning).
The current production "generative AI" wave is dominated by (a) Transformer-based language/vision models and (b) diffusion-based image generators. GANs are still important historically and remain useful in some niches.
Training is the expensive phase: large datasets, large compute, and iterative optimization to learn model parameters. Inference is the serving phase: you run a trained model to produce outputs (tokens, images, classifications, embeddings).
Base models are typically trained to predict text. Many production chat models are further tuned to follow instructions and avoid obvious unsafe or unhelpful behavior (high-level example: InstructGPT).
Many useful systems are not "just prompting". They are LLM + tools + a loop: search, retrieve documents, call APIs, execute code, and iterate. This is often described as agentic behavior.
For real deployments, you care about throughput, latency, and cost. Common techniques include quantization, parameter-efficient fine-tuning, distillation, and architectural tricks.
See: Efficiency.
"Open weights" typically means you can download model parameters. But open weights alone are not the same as Open Source AI (see OSI - Open Weights).
OSI's Open Source AI Definition 1.0 is published at opensource.org, and the Open Source Definition itself is at opensource.org/osd.
This is intentionally a starting map with primary links (technical reports or model cards). It is not a benchmark report.
Production systems often depend on efficient matching and retrieval. Examples include nearest-neighbor vector search, image feature matching, and audio fingerprinting.
See: ML tooling.