Modern AI Techniques

State-of-the-art methods driving the current AI wave. These build on everything in the previous sections.

Generative Models

• Diffusion Models — how Stable Diffusion, DALL-E, and Midjourney work
• Generative Adversarial Networks — generator vs discriminator

Scaling & Architecture

• Mixture of Experts — sparse models that scale efficiently (GPT-4, Mixtral)
• State Space Models — Mamba, linear-time alternatives to attention
• Multimodal Models — vision + language in a single model (GPT-4V, LLaVA)
• Scaling Laws — power-law relationships between compute, data, and performance
• Vision Transformers — ViT, DeiT, Swin — transformers replace CNNs for vision

Alignment & Training

• RLHF and Alignment — how LLMs learn to be helpful, harmless, honest
• Constitutional AI — rule-based self-improvement
• Instruction Tuning — teaching models to follow instructions

Inference & Efficiency

• Quantization — shrink models from 16-bit to 4-bit with minimal loss
• Knowledge Distillation — train a small model to mimic a large one
• Speculative Decoding — speed up inference by predicting ahead
• LoRA and PEFT — fine-tune large models with tiny parameter budgets

Agents & Applications

• AI Agents — LLMs with tools, memory, and planning
• Tool Use and Function Calling — models that use APIs and code
• Retrieval Augmented Generation — grounding with external knowledge

Links

• Key Papers — the research behind these techniques
• Deep Learning Roadmap — foundations you need first
• Training Projects — hands-on practice