Compute is only interesting for what comes out of it. These are the projects that have run on Navon hardware, written up with the scores, the costs and the parts that are still weak.
Most vendor case studies are an advert with a logo on it. These rules are what stop ours becoming that, and they are the reason there are few of them.
No anonymised “a leading university”. If the customer will not be named, the study does not run. That is also why every one of these is signed off in writing before it goes up.
Scores, costs, hardware and dates, as measured. Where a figure came from a third party we say which one. Where it is ours, we say that too.
Every real project has a part that failed or is still weak. It stays in. A study that only reports wins tells you nothing about what to expect from yours.
The model, the corpus, the hyperparameters, the machine. A reader at another university should be able to start from what is on the page.
We put serious compute in the hands of Kenyan research teams and get out of the way. If you are at a university with work that is held up by hardware, tell us what you are trying to train. You keep the model, the corpus and the credit.