TEACHING & LEARNING
Making tools understandable, along with the thinking behind research.
Course materials on human–AI collaboration and AI-assisted research, covering how to ask questions, build context, use research skills, manage agents, and organize research workflows.
Turning vague needs into questions with clear conditions, goals, and room for further inquiry.
Organizing literature, materials, research history, and sources that can be checked.
Understanding how skills, agent management, and execution workflows support research.
Checking outputs while explaining the methods, their limits, and why those checks matter.
Research tools should help people develop their own understanding. I design courses around clear explanations, further inquiry, and feedback, connecting technical practice with the ideas behind it. This page introduces my course materials and areas of knowledge sharing.
Get in touch about courses & sharingThese are prepared slides, speaker notes, and lesson plans. Dates record materials preparation.
For students entering research, I organize a map of post-training: supervised fine-tuning, preference optimization, reinforcement learning with verifiable rewards, agent training, and the role of rewards and evaluation. The materials move from what the field does to what a student might investigate, bringing open questions, compute and evaluation requirements into the choice of a research direction.
These materials compare the learning signals in SFT, offline DPO and online GRPO, then connect the feedback structure of language models to visual generation. Group-advantage calculations, a reward-scaling counterexample and a two-dimensional point-cloud demonstration distinguish overall diversity, task-relevant variation and actual training gains. The lesson ends by turning an intuition into a research question with controls, measurements and a possible failure.
A two-dimensional demonstration gives two outputs the same mean, total variance and response singular values, but different variation along the task direction. Switching the reward from the horizontal to the vertical axis reverses that ordering. This is a constructed teaching counterexample, rather than a measured SFT or GRPO result.
The introduction begins with a cup-pushing scenario, connecting state, observation, action and an imagined feedback loop. It then compares generative, representation-based and decision-oriented world models through architectures including World Models, Dreamer, MuZero and JEPA. Interaction, counterfactuals, persistence and uncertainty lead to a central question: does the simulation help a decision, beyond producing a convincing image?
From the same state, would a different action change the prediction and the next decision? Counterfactual branches in the cup scenario help distinguish coherent imagery from a model that actually uses action information.