Talking Out of School: Why is a “small-language model” better for K12?

You’ve probably heard about AI’s large language models. But what is a “small language model” and why might this fine-tuned version of artificial intelligence work better in K12? 

In the latest episode of District Administration‘s “Talking Out of School” podcast, Merlyn Mind’s Levi Belnap fills listeners in on the emerging technology. The more well-known large language model, such as ChatGPT, is “what makes a computer able to understand our human language in ways that, previously, would have been shocking,” Belnap says.

“Its most current versions is, you take all the world’s information, you stick it in there and now [your computer] can kind of understand anything you say, and it can respond in ways that seem really coherent,” he adds. “It’s like, ‘Oh, my gosh! The computer understands me.'”

Merlyn Mind’s founder, Satya Nitta, has described using such a powerful in K12 learning as “putting a Ferrari engine in a lawnmower,” Belnap notes.

“In a classroom, we need to take into account privacy, safety and security needs,” he adds. “We need to take into account costs and, very importantly, we have to take into account ‘How do you help students learn?’ And that may not be doing everything that the big models can do, but doing very focused, specific things that you want to do in the classroom.”

Merlyn has built a small-language “appropriateness” model that reduces the risk of returning inappropriate answers to students’ questions. The company’s technology is also designed to increase student-teacher interaction after witnessing the struggles of AI tutors to fully engage students.

“You gotta make sure that what’s in those models is safe for those students to see,” he explains.

To learn more, listen to the podcast below, or on Apple, Podbean or Spotify.