Show Notes
This episode addresses how Reinforcement Learning from Human Feedback (RLHF) adds the final layer of alignment after supervised fine-tuning, shifting the training signal from “right vs wrong” to “better vs worse.” We explore how preference rankings create a reward signal (reward models plus PPO) and the newer shortcut (DPO) that learns preferences directly, then connect RLHF to safety through the Helpful, Honest, Harmless goal. We also unpack the “alignment tax,” the trade-off between being safe and being genuinely useful, and close by setting up the next module on running models at scale, starting with GPU memory limits, plus a personal reflection on starting later without being behind.
More from The AI Concepts Podcast
In this episode, we bring everything together by following one LLM application from the user's first request to the final response. We see retrieval, context, tools, model calls, state, memory and orchestration working as one system, revealing the bigger lesson of the module: building with LLMs is not just about what the model can do, but deciding what the model should do and what the rest of the application should handle.
In this episode, we map the modern LLM application stack and explore where frameworks, SDKs, runtimes and protocols actually fit. We look at how tools like LangChain, LangGraph, LlamaIndex, provider SDKs, durable workflows and MCP solve different problems, and why understanding the architecture matters more than memorizing product names.
This episode separates two concepts that are often confused: state, which keeps track of what is happening now, and memory, which allows information from the past to become useful later. We explore how applications create continuity around a model, and why good memory is not about remembering everything, but remembering the right things at the right time.
This episode breaks down what orchestration actually means, from sequencing and routing to retries, parallel execution and human approvals, and explores the different ways those flows can be managed. Most importantly, we separate orchestration from the model itself and show why it is really about controlling how work moves through an application.
Who actually decides what happens next inside an LLM application? This episode explores the difference between decisions made by code and decisions made by the model, and why real applications often use both. We follow the loop that emerges when a model can request information or actions, receive the results and decide what to do next, revealing what developers mean when they talk about “owning the loop” and setting the stage for orchestration.