This week emphasizes that prompting and sampling guide behavior without changing the underlying model. By briefly revisiting earlier concepts such as transformers and multimodal generative architectures, learners place prompting within the broader AI landscape while staying focused on practical control. The week closes by raising a key question: if users can shape behavior so effectively at inference time, how does the model learn what “good” behavior is in the first place? That question leads directly into the next week’s exploration of training, reinforcement learning, and fine-tuning.

The week centers on reinforcement learning from human feedback (RLHF) and evaluator models as mechanisms for encoding preferences, alignment, and style into generative systems. Learners examine how feedback shapes model behavior, why RLHF has been so effective, and why it can also contribute to issues such as hallucinations and reward misalignment. The week closes by shifting attention back to inference time, asking how structured prompting and additional compute can enable models to reason, revise, and refine outputs without retraining, setting the stage for the study of reasoning scaffolds and chain-of-thought in the following week.

This week focuses on inference-time compute and reasoning scaffolds such as chain-of-thought and step-by-step prompting, highlighting how large context windows allow models to “think on the page.” Rather than producing a single answer, models can fill the context window with intermediate steps, enabling feedback into their own reasoning and improving accuracy on complex tasks. The week emphasizes that this process does not involve learning or parameter updates. Instead, reasoning emerges from structure, additional context, and the ability to revisit earlier steps within the same prompt. Learners explore how self-critique, revision, and iterative prompting take advantage of large context windows to refine outputs without retraining. The week closes by shifting from individual reasoning strategies to broader comparison, preparing learners to examine how different models reason, specialize, and perform across tasks, which leads directly into the study of open-source models, mixture-of-experts architectures, and systematic evaluation in the following week.

This week introduces the open-source model ecosystem and Mixture of Experts architectures, using models such as Mistral to illustrate how specialization and routing can improve performance without relying on a single monolithic model. Learners connect these ideas to earlier discussions of fine-tuning, seeing how different approaches shape behavior and capability in complementary ways. The week then shifts to evaluation and benchmarking as essential practices for understanding model strengths, limitations, and tradeoffs. Learners examine the history of benchmarking to see how rapidly frontier models have advanced, from outperforming grade-school benchmarks to surpassing expert-level performance when paired with tools. Concepts such as alignment, alignment drift, and reward hacking are introduced through examples, including Goodhart’s Law, to show why evaluation must evolve alongside model capability. The week closes by highlighting practical considerations around data ownership, IP boundaries, and deployment constraints—particularly in open-source settings—setting up the next week’s focus on tool use, memory, and systems that operate across time.

This week explores tool use, showing how models invoke calculators, search, APIs, and retrieval-augmented generation systems to access external capabilities. Tool use marks a key transition from passive reasoning to action-oriented behavior, where models no longer operate solely within their training data or context window. The week also introduces memory and context persistence, examining how short-term context, long-term storage, and summarization enable systems to operate across multiple interactions rather than isolated prompts. Learners explore basic evaluation heuristics that help monitor reliability as systems grow more complex. Together, tools and memory allow AI systems to maintain continuity over time, setting the stage for real-world applications and ethical considerations in the following week.

This week surveys modern applications across domains such as code generation, business workflows, accessibility enhancements, and creative media (music, speech, image, video), emphasizing how AI systems function as productivity multipliers rather than replacements. The week also introduces ethical reflection at the individual level, focusing on what it means to work alongside AI systems in daily practice. Learners consider tradeoffs related to cognition, autonomy, and reliance and discussions of the AI productivity paradox. “Every time you interact with an AI, realize you’re giving something up in exchange.” The course closes by inviting learners to reflect on what AI can do in their chosen field today, forming the basis for the Course 2 capstone and setting up Course 3, where attention shifts from current capabilities to the emergence of agents and the implications of more autonomous systems.