Tiny AI Matches a Bigger Model at Meeting Notes, Using Less Power

Meeting AI usually works like this: you ask "what did we decide about the budget?" and the software digs through the transcript to answer. The problem is that most systems swallow the whole conversation, which is expensive and slow. In a new paper, researcher Edward Xi Yang (Ertas AI) tested a shortcut: pull out only the roughly 2,000 words that matter for your question, then have the AI summarize those. He also re-scored 15 existing systems the same way, because the standard meeting dataset (QMSum, a public set of real transcripts) had no consistent scoring, making comparisons unreliable.

The shortcut worked surprisingly well. Feeding the AI relevant excerpts instead of the first 4,500 words of raw transcript improved results, and extra training on that style added a large jump in quality. The payoff: a 406-million-setting model scored 36.33 versus 35.41 for a model three times bigger — a statistical tie — while using about one-third the total size and less than half the peak memory. In plain terms, similar quality for a fraction of the computing cost, which means it could run on cheaper hardware, maybe even a good laptop.

There was a second, flashier result. Under one short instruction and a word-overlap scorer, the small specialized model beat five paid, hosted AI services by at least 6.2 points. That suggests a narrowly trained small model can outperform giant general-purpose cloud AI on one specific job — a pattern worth watching for anyone paying per-query AI bills.

The catch is real. All of this was measured by automatic metrics that reward matching words in a reference summary, not truth, usefulness, or brevity. Output length was not controlled, and there were no human raters and no factuality checks. So the ranking over paid services is suggestive, not settled, and the conclusions apply only to this one dataset. Treat it as a promising signal about cost and size, not proof your next meeting bot will be flawless.