Talk-ready slides

Twelve minutes, twelve slides

One claim per slide, and the title states the claim: “Materialised attention moves 268 MB per head at 4096” rather than “Attention”. The first slide carries the one number the talk exists for. The last carries the next lever.

The five moves

Problem: one diagram of the pipeline and one sentence about what was slow. Demo: a recorded sixty seconds, because the venue network is not yours and a live demo is a second project. Numbers: one table, three rows, a band, and the rig line in small type. Lessons: three bullets, one of which is about the code review. Next: one line.

Rehearse against a timer

Run it once at speaking speed and cut a fifth. The cut is usually the second example in each section, the material added because it was interesting rather than because it carried the argument. Run it again and check that the numbers on the slides are the numbers in the writeup, with the same shapes and the same band.

The backup section

Every question you can predict gets a slide after the end: the roofline, the tile sizes that failed, the upstream diff, the GPU comparison with its caveats. Answering from a slide is faster than reconstructing a number from memory, and it keeps the claim attached to its rig line.

Things that break

A live demo on a conference network. Forty slides at twelve minutes, which is a reading rather than a talk. A title slide that says “MLIR-based AI compiler” and no number, so the audience spends the first minute working out what you did. And a GPU comparison presented as a contest between two machines that were never comparable.