I wouldn't call it wasted effort (neither would I call space exploration and lunar or martian colonization as wasted effort) - it's not the kind of bullshit "prompt engineering" that tries to find magic phrases and sentence structures, and gets invalidated every time the model is updated. If I understood the article correctly, this language implements its operators by stepping the language model directly, applying the logical operations to token probability distributions. This actually makes sense as a generic and reliable method for these kinds of models, as it works below the prompt level, directly on the underlying machinery.

We agree! But now I understand the work itself better thanks to your summary, so thank you.