This paper makes the points that it's the saddles and not local minima that are the problem: https://arxiv.org/abs/1406.2572 It was the basis for adding 'momentum' to optimizers - so that you could skate across the saddles.

...hmm, that was counter to my understanding (limited though it may be...) which was partially formed by this paper: https://arxiv.org/abs/1712.09913

TLDR - loss landscapes are nasty, but you can tame them with skip connections.

These two papers are not necessarily contradicting each other, but perhaps my description was a bit sloppy.

Sagun et al. (and derivative works) only focus on the Hessian on the trajectory followed by gradient descent, while Li et al. give a broader look at the loss surface as a whole.

...sounds like it costs much more than the $10 pi0...

I imagine the OP's use case of IoT did not require huge amounts of bandwidth.

"That's not to say it won't function well enough in a pinch but I wouldn't seek this out as a permanent answer if I were buying new."

Like I said this functions fine, as is evident by it having been posted, but the question was why it wasn't a very good repeater not was it a functional use of an existing pi 0 laying around.

It's also not the best you could do buying new if that is what you had to do as ~$16 will get you a 2x2 802.11n repeater that functions much better, has external antennas, and has an ethernet port too. Not $10 but it's also enclosed and doesn't need a power supply or uSD. ~$30 will get you decent AC repeaters you could use outside of a singular IoT project. Less for either if you're willing to go used. The ideal 4x4 ax double phy wireless repeater is indeed going to be more than a pi 0 but it was an example of what makes a repeater good (to the question) not what should have been purchased for this project.

Again nothing wrong with the solution used here, that's not where my comment was pointed.

Ahem...trebuchets are superior to catapults in every way.

Okay, but is this viable?