Concepts

Learn how LLMs predict tokens at scale, why the training pipeline has three distinct stages, and how to choose the right model for your system.

Learn how tokenization converts text to tokens LLMs actually process, why it drives pricing and multilingual costs, and how to budget your context window.

Learn how embeddings encode meaning as vectors, why they power semantic search and RAG, and how to choose the right model for production.

Understand how attention lets transformers focus on relevant context, why it replaced RNNs, and what query-key-value matrices mean at an intuitive level for engineers building AI systems.

Learn how LLMs train on unlabeled text by predicting masked or next tokens, why this makes labeled data unnecessary at scale, and what it means for how models generalize.

Understand how the transformer's encoder-decoder structure, positional encoding, and residual connections work together, and why this architecture has dominated AI since 2017.

Learn how to construct the context window to get the best results from LLMs, why 'context engineering' has replaced prompt engineering as the key skill, and what belongs in a production system prompt.

Understand how few-shot examples guide LLM behavior, when they outperform fine-tuning, and how to select and order examples to maximize response quality.

Learn when fine-tuning outperforms prompting, how LoRA makes it affordable, and how to decide between full fine-tuning, LoRA, QLoRA, and instruction tuning for your use case.

Learn how function calling lets LLMs trigger real APIs and return structured data, why it's essential for production AI systems, and how to design tool schemas that work reliably.

Learn how RAG grounds LLM responses in your data, how the ingestion and retrieval pipelines work, and how to diagnose the most common failure modes in production RAG systems.

Learn how vector databases power RAG and semantic search, how HNSW and IVF indexes work, why metadata filtering is the most common production failure point, and how to choose between pgvector and dedicated solutions.

Understand Anthropic's Model Context Protocol (MCP), how it standardizes tool and resource access for AI agents, and when to build an MCP server instead of raw function calling.

Learn how chain-of-thought prompting makes LLMs show their reasoning steps, why it dramatically improves accuracy on complex tasks, and when to use zero-shot vs few-shot CoT.

Understand how RLHF adapts a capable base LLM toward preferred behavior, what reward models and PPO do, and why DPO is often a simpler engineering choice.

Learn how reasoning models such as o1, o3, and DeepSeek R1 use additional inference-time computation on complex tasks, and how to evaluate when the extra cost is justified.

Learn how multimodal models process images, audio, and video alongside text, what CLIP-based architectures look like, and how to use vision LLMs effectively in production systems.

Learn when small language models (1B-14B parameters) outperform large ones, how Phi-4, Gemma 3, and Llama 3.2 are closing the quality gap, and how to choose between cloud APIs and self-hosted deployment.

Learn how knowledge distillation transfers capability from large teacher models to smaller student models, when it beats fine-tuning, and how it powers DeepSeek and Phi.

Learn how quantization reduces LLM memory footprint, what INT4 and GGUF mean in practice, and how to evaluate lower-precision models on constrained hardware.

Understand what AI agents are, how the ReAct loop works, what memory and tool primitives look like, and why production agents fail so often at tasks that demos make look easy.

Learn how KV caching, continuous batching, and speculative decoding cut LLM serving costs, what TTFT and TBT mean for UX, and how vLLM and TGI handle production throughput.

Learn how to measure LLM application quality with assertion-based tests and LLM-as-judge, why evals come before architecture, and how to build an evaluation pipeline that gates production deploys.

Learn how to instrument LLM applications with traces, logs, and metrics to debug failures, detect prompt drift, and link production issues back to specific prompts and model versions.

Learn how constrained decoding forces LLMs to produce valid JSON every time, why regex-guided generation works, and how to choose between JSON mode, tool-use schemas, and libraries like Outlines.

Learn how diffusion models generate images by iteratively denoising random noise, why they replaced GANs as the dominant generative architecture, and how latent diffusion powers Stable Diffusion and DALL-E 3.

Learn how Mixture of Experts (MoE) architectures activate only a fraction of parameters per token, why GPT-4 and Mixtral use sparse routing, and how MoE trades memory for throughput in production serving.

Learn how prompt injection, jailbreaks, and adversarial attacks exploit LLMs in production, how red teaming identifies vulnerabilities before attackers do, and how to build defense-in-depth for AI systems.

Learn how LLM routers pick the cheapest model that can handle each query, why cascading from small to large models cuts costs 60-80%, and how to build a routing layer for production AI systems.

Understand how LLMs are priced by token, calculate real costs for agent systems, and apply the 60/30/10 model tiering rule, batch API strategies, and context compression to build cost-efficient AI at scale.