Teaching a child in <1000 ms: the architecture behind a real-time tutor

Building a real-time AI tutor for 5-year-olds

We set out to build the first AI tutor to teach math and reading to kids ages 4-9. For AI to actually teach a five-year-old, pedagogy must be baked into the engineering. A child can't wait for a slow reply, can't read a chat interface, and can't unhear anything a model gets wrong. We wanted to share some of the learnings that shaped our architectural decisions building a real-time AI tutor.

146 points by catalinvoss 104 days ago www.ello.com 394 comments

Summary not yet generated for this article.

Original Text

Hey HN! We've spent the good part of this past year building an AI tutor that teaches kids ages 4-9 reading, math, ESL and more. Getting an AI tutor to effectively teach a child turns out to be a really hard technical challenge, this took getting the underlying architecture right.

Our tutor steers the UX in real-time and makes complex decisions on the fly. Doing both at conversation speed required us to replace the standard tool-use loop. We built our own tutor harness that utilizes a streaming interpreter that executes actions, while an asynchronous planner model reasons ahead of the conversation and makes calls that drive the child's learning. On top of it all, we developed a safety system that checks every turn without it causing an interruption to the activity and conversation flow.

Effective teaching isn't just about answering a child's question quickly, rather making the right move at the right moment. AI is also going to be an integral part shaping how this generation of kids learn to read and think, tackling this responsibly means getting the design right.

Happy to answer questions and curious what you all think, critical feedback included, we've been working on this problem for a long time and love to hear from the HN community.

Comments (394 of ) ⌛

Loading more comments... ⌛

Scroll down to load more comments

No comments available.