Oxide and Friends AI in Computer Science Education
49 snips
May 10, 2026 
Shriram Krishnamurthi, a Brown CS professor who designs programming courses, and Kathi Fisler, a Brown CS educator focused on curriculum, discuss an experimental intro course using agentic programming. They cover why to redesign CS now, crafting assignments that reveal LLM brittleness, Tetris as a reveal of strengths and flaws, testing at scale, types and peer crits, and curricular models for teaching trustworthy AI practices.
AI Snips
Chapters
Transcript
Episode notes
Design Assignments That Force LLM Failure
- Design assignments the LLM will likely 'vibe code' poorly so students must recognize failures.
- Follow failures with targeted instruction: testing, types, constraint solvers, or databases as remediation and future-course pointers.
Scaling Data Revealed Naive LLM Solutions
- Data analysis task: small CSV then a 10,000-row set to break naive LLM-generated code.
- Students learned testing limits and had to move from quick prompts to designing scalable data workflows.
Types Are High-Leverage Constraints For LLMs
- Use TypeScript/types as a moderate, meaningful review target to constrain agents and guide correct code.
- Shriram iterates on types first, reads and corrects them, then lets the agent produce code that follows the refined types.