Domain 1: AI Foundations — Learning Objectives

Throughout this program, guardrails refers to the technical controls — input validators, output filters, classifiers, circuit breakers, monitoring systems, and other mechanisms — that operate around an AI model to prevent it from producing harmful, incorrect, or unauthorized outputs. Guardrail engineering is the discipline of designing, building, and validating these controls.

After completing this module, you will be able to:

  • Explain how neural networks learn by describing the training loop (forward pass, loss computation, backpropagation, weight update) and why knowledge is distributed across billions of parameters rather than stored in discrete, inspectable locations.

  • Describe the transformer architecture including self-attention, positional encoding, and the distinction between encoder-decoder and decoder-only designs, and explain why transformers replaced recurrent architectures for language tasks.

  • Trace the full path of a single inference request from raw text through tokenization, embedding lookup, positional encoding, multi-head masked self-attention, feed-forward layers, and the output probability distribution over the vocabulary.

  • Explain autoregressive text generation including the generation loop, sampling strategies (temperature, top-k, top-p), and why each token choice is irreversible and shapes all subsequent output.

  • Distinguish each stage of the LLM training pipeline — pre-training, instruction tuning, and alignment (RLHF/DPO) — and identify what capabilities each stage creates, what can go wrong at each stage, and what the application developer can versus cannot control.

  • Identify common production AI system architectures — simple chat, RAG, agentic, and multi-model pipelines — and describe the data flow through each pattern.

  • Map guardrail placement points within a production AI pipeline (pre-model input, post-model output, system-prompt level, retrieval level, tool-call level) and explain the trade-offs of each placement.

  • Explain why LLMs cannot be made safe by training alone, connecting concepts like distributed knowledge, stochastic generation, and instruction hierarchy as learned preference to the need for external guardrail systems.

  • Describe embedding models and vector representations, explaining how they enable semantic search and retrieval-augmented generation, and where guardrails must be applied in retrieval pipelines.

  • Articulate the guardrail implications of each foundational concept — translating technical understanding of how AI systems work into reasoning about where and why they fail and how to defend against those failures.

← PreviousAI Foundations