AI at Berkeley

The San Francisco Chronicle runs a story about the use of AI in university courses on various Bay Area campuses. For UC-Berkeley, there is this:

...New and different course offerings and increased interest in them aren’t the only changes being wrought by the technology at the university level. UC Berkeley professor Narges Norouzi started encouraging her computer science students this fall to take advantage of OpenAI’s GPT-4, albeit in a limited way, to solve knotty coding challenges. The technology’s AI-generated coding hints can help students push through tough problems and get faster results.

The effort is still in the testing phase, mainly in her introduction to coding class, Norouzi said, and does not give students access to the full-on chatbot. Instead, a program she fashioned using the technology allows students to click a button that then prompts GPT-4 for a hint based on the material they are working on and, increasingly, what the program knows about the class and student learning patterns. Students have been critiquing the coding hints, which Norouzi said will be incorporated into future versions of the tool through a process called reinforcement learning from human feedback. Over time, it will be able to increasingly understand how students learn and make better prompts.

Another program still in the testing phase that Norouzi hopes to incorporate into classes next year will feed GPT-4 with information from course texts and an online student Q&A portal, called Edison. The goal is for the program to learn from that material and be able to speedily answer many of the thousands of questions students pose to Edison every semester. One data course has roughly 1,300 students, Norouzi said, and teachers’ assistants can be heavily burdened with answering every question that pops up on Edison.

“We’ve been working on how to give the (model) the right context,” Norouzi said. That means training it to know whether to reference a course text or a homework description or the syllabus to see if a question has already been asked and answered.

She said a related tool that she plans to make available to students next semester can perform similar context queries of course materials, but with an eye toward shortening the long waits that can crop up when students sign up for office hours for one-on-one time with professors. “The queues get as long as four hours wait time,” Norouzi said. So she has created a chatbot based on GPT-4 versed in the course material that students can converse with, hopefully solving their problem without having to wait in line.

“Hopefully this can help us move faster in the queue and help serve more students,” especially in larger courses, she said...

Full story at https://www.sfchronicle.com/bayarea/article/ai-artificial-intelligence-college-18550411.php.

The times, they are a-changing.