Seems horrendously brittle. I'd trust badly written bash over this.

When they say modern, which standard to they mean?

Literally in the title of the linked page:

> Modern C++ Programming Course (C++11/14/17/20)

It's described in the course, but canonically this means at least c++11 (and these days, likely 14/17 in practice - there is broad toolchain support)

I think it's kind of disingenuous maybe to claim such improvements in training efficiency when they rely on:

- Existing models for data pseudo-labelling

- ImageNet pretraining

- A frozen text encoder

- A frozen image encoder