The presenters describe Ace, a live voice tutor whose main constraint is response delay rather than model benchmark performance. Their design separates lesson planning, progress tracking and the choice of the next activity from the model that speaks to the student.
A state machine coordinates the lesson and passes a structured brief to a smaller language model on each turn. The demonstration contrasts a slower reasoning-led response with a faster scaffolded response, but it is a specific example rather than a controlled claim about every frontier model.
They acknowledge that small models can drift when asked to manage long, structured tasks without support. The architectural work is therefore essential, not a free benefit: build explicit control and reasoning around the fastest model that satisfies the interaction's latency requirements.
Watch on YouTube
