Small Models Are Eating the Demo
The impressive live demo is increasingly a tiny model doing one narrow thing perfectly. The frontier is overkill for most of it.
Watch the demo reels lately and a pattern shows up: the flashy part is usually a small, specialized model doing one job with scary precision, not a frontier giant. A 27-billion-parameter decider routing support tickets, a tiny transcription model beating Whisper on a CPU, a 2-billion model making a yes/no call for pennies.
This is the right architecture. Frontier models are wonderful generalists and terrible value for a task you run ten million times. The economics flip hard once volume is high: a $0.04-per-million decision model crushes a $2 one on the same label.
For builders it means the smart move is often to start small. Prove the task works with a focused model, then escalate to a frontier model only for the cases that actually need it. Most pipelines are 95% routine and 5% hard, and treating all of it as frontier-class is how budgets blow up.
The demo that wins next year won't be 'look how big our model is.' It'll be 'look how little it costs to do this perfectly.' Small models eating the demo is just cost reality arriving on schedule.