What the class is for
How they work
What happens between your prompt and the answer, how agents use tools, and how several agents coordinate on one problem.
Using them well
Choosing a model and a workflow for the job, trading quality against speed and cost, and checking what comes back.
Where they fail
Confident wrong answers, brittle tool use and evaluation pitfalls, and how to catch them before they reach your results.
Pushing the frontier
Each of us builds the hardest task we can from our own field, then refines it until the best AI agents fail on the science.
No prior AI expertise is assumed. The emphasis follows your interests, starting with the week 1 survey.
Ten Wednesdays
- Measure, then climb
- Tasks that frontier AI can't solve
Topics are set week by week, shaped by what the class already knows and wants to learn.
This week: measure, then climb
- Page through the slides How a model's ability is trained and measured: RL environments, AI eval, benchmarks, hill climbing and recursive self-improvement, as concepts.
- Read further The same five ideas in more depth, with sources.
- Homework: make your task hold Fix your week 1 task's solvability, then harden it until GPT-5.6 Sol fails on the science. Due before class on Nov 18.
- Missed a class? Recordings of past sessions, and every earlier week's page from the menu.
- Answer the live poll For questions we put on the screen during class.
Also on the web
Practical details
- Discord
- Join the class discussion to ask questions and share ideas between classes.
- When
- Wednesdays, 11:00–11:50 a.m., September 30 to December 2
- Where
- Slichter 3853, UCLA. Joining remotely? Join on Zoom (meeting ID 946 9664 6805, passcode 123456).
- Who
- Registered EPSS 254 students get priority for individual help. Everyone is welcome to audit.
- Led by
- Dr. Zesen Huang, with guest speakers, as part of the EPSS 254 (SPHINX) seminar practicum
- Contact
- Zesen (Jason) Huang, zesenhuang [at] g.ucla.edu