What happens in ninety minutes.
The same nine chapters every time, so listeners know where to jump and authors know what to prepare. The hardest chapter is the one in the middle: what the reviewers asked.
| Chapter | Minutes | What it is for |
|---|---|---|
| Cold open | 3 | The one result, in the author's words |
| The paper in five minutes | 5 | Claim, method, evidence |
| The path | 20 | Where the idea came from |
| The experiment that nearly failed | 12 | What did not work, and why |
| What the reviewers asked | 10 | The hardest questions, answered |
| Limits | 10 | Where the result stops |
| What comes next | 20 | Open problems and the next paper |
| Advice | 8 | For students and for teams |
| Sixty seconds | 2 | Why it matters, plainly |
Before, during and after
Preparation call, 30 minutes
The author, the host and an MTRI researcher who has read the paper. We agree the outline, what is out of scope (unpublished or confidential work) and the figures to show.
Recording, two hours booked
In the Singapore studio, at the conference or remotely. Three cameras in the studio; the author can stop and restate anything.
Edit and fact review
The author sees the edit and checks technical accuracy. Corrections to what they said about their own work are made; fair questions and their answers stay.
Release
Video, audio, transcript, Brief and a Notes item, all linked to the paper.
Corrections
If a finding is misstated after release, we correct the transcript and show notes and log the change.
The transcript
Free transcripts are full and timestamped, with links to the paper. Meridian Pro transcripts add the labels below to every substantive line, link each finding to its table, figure or section, and let you filter by label.
We expected the smaller model to collapse on the longer problems. It didn't, and that is the part I still find surprising.
Author's viewOn the held-out set, the method improves accuracy over the strongest baseline in four of the five task families; the fifth is within noise.
Paper finding · Table 2, §4.1If that holds at larger scale, it would make the technique cheap enough for most teams to try. The paper does not test that.
Meridian interpretationReviewer two asked whether the gains came from the data rather than the method. That was the right question.
Author's viewThe ablation that trains on the same data without the method recovers less than half of the gain.
Paper finding · Table 4, §5