The openai/whisper repo and paper referenced by the model card seem to be saying it's transformer based.

It’s an encoder-decoder transformer trained on audio (language?) and transcription.

Seems kinda weird for it not to meet the definition in a tautological way even if it’s not the typical sense or doesn’t tend to be used for autoregressive token generation?

I think part of getting by with a lower PPD is the IRL pixels are fixed and have hard boundaries that OS affordances have co-evolved with.

(pixel alignment via lots of rectangular things - windows, buttons; text rendering w/ that in mind; "pixel perfect" historical design philosophy)

The VR PPD is in arbitrary orientations which will lead to more aliasing. MacOS kinda killed their low-dpi experience via bad aliasing as they moved to the hi-dpi regime. Now we have svg-like rendering instead of screen-pixel-aligned baked rasterized UIs.

macOS exposes a lot of affordances to code/xrpc/services/etc that Shortcuts (and previously automator) used. They let you do basically anything you'd want on macOS programmatically, without going through accessibility frameworks, code signing and sand-boxing issues. iOS as well to some extent.

Presumably if OpenAI is dog-walked/locked out of these by Apple at some point, they would be stuck in the Chrome/Chromebook feature jail. My guess is this gives OpenAI a team to put in charge to give them a chance to wedge themselves into the OS before Apple changes their mind or puts scare-box dialogs everywhere.

Either that or there's nothing so complicated and OpenAI just wants to re-build this stack inside ChatGPT as quickly and well as they can.

I assume this comment in relation to the starting from a stop being unpleasant?

If it's w.r.t. effect of low max power on low cumulative generation, I agree it does seem like a little silly to arbitrage your power generation this way. But maybe the tradeoff is worth it in some circumstances in their view?

Or maybe it's just a low cost addition as other commenters say.