Building a Real-Time AI Tutor for 5-Year-Olds Under a 1000ms Budget

Ello (via Hacker News)·medium signal

Ello published an engineering account of building a real-time AI reading tutor for young children, framed entirely around a sub-1000ms response budget — the threshold beyond which a five-year-old disengages. It drew 112 points and an exceptionally high 223 comments. The latency-first framing is the transferable part: it treats time-to-first-token as a hard product constraint that dictates model and architecture selection, rather than a metric to optimize after the fact.