Why the launch chart is not a benchmark

Only argues. The caveats state it is a visualization of an interview argument rather than a new measurement, so there is no dataset, harness, or score here to inspect or rerun.

The entry records it as a trust problem. The highlights describe the usual shape of such a chart, with tests as rows and the new model on top of nearly every one, and the description frames that as something you cannot trust as a benchmark.

ML Visualized is restating Diogo Almeida's argument from a Latent Space interview, per the description. So it is commentary on someone else's position, and the full interview is a separate entry worth reading for the original reasoning.

The entry gives no checklist, but the whole clip points one way: a number you cannot reproduce is not evidence. Both the caveats and the description focus on whether a result was measured and checkable, not on who reported it.

ML Visualized published it, based on swyx's Latent Space interview with Diogo Almeida. The entry calls it a visualization of that argument rather than a new measurement.

The usual launch chart, with tests as rows and the new model on top of nearly every one. The description frames that shape as something you cannot trust as a benchmark.

In a separate video, and a separate entry. The entry points there for the original reasoning behind the argument this clip visualizes.