"32 bit is totally lossless unlike 16-bit or 8-bit" In about minute 4.
It's not lossless if edits are applied to rendered data. Now parametric edits produce RGB data and one could suggest it's lossless but that's got little if anything to do with bit depth since all raw data is 'high bit' (more than 8-bits per color). The raw's capture bit depth is what it is no matter what a software product may call it (15+1, "32"). The pie is still 3lbs no matter how sliced.
"It is a raw layer"
I have no idea what he's suggesting. If it's raw, it's raw. As a 'layer' if so (I suspect it isn't) or as a Smart Object. Now with an SO, you edit this parametrically in ACR. If it's indeed rendered as a layer in RGB, no matter the bit depth, edits will apply rounding errors. It isn't lossless. But because it's high bit (16 or 32-bit), it's moot; the edits do not apply damage that is ever visible at that bit depth or when printed.
Layers are not truly lossless anyway. Sure, while you are in Photoshop, you can always revert an edit (you can do a Save As too and the original did not undergo data loss). But at some point, outside of Photoshop, layers must be flattened and the data loss, moot or not, will exists. IF you print the image from Photoshop, PS has to flatten all the data to make the print. So same condition.
This talk of 'lossless' is mostly marketing hype and misunderstanding of what, how and where data loss occurs. Outside parametric edits, any edits to any RGB data in any bit depth will undergo 'some' data loss. The edits and if it's visible is the key here.
He seems somewhat confused by the ACR Histogram too. The fact is, ACR has highlight recovery IF (bit if), one or two channels is clipped. But the only way know how much or how little is to examine a raw Histogram which ACR, Photoshop, LR etc cannot provide. Highlight clipping in raw is based upon Exposure and that means the amount of light striking the sensor (or film), nothing else (including ISO). The way to view exposure of raw is by viewing a raw Histogram in a product like RawDigger. So there's some minor Highlight recovery in the ACR engine. There is its Histogram which is always based not on the exposure or the actual raw data, but based on the current rendering settings applied.
"So here is a 32 bit raw file". 5:15 in to the video
No, the raw file's bit depth is what it is.
"Not rasterized yet"
Well mostly yes and a little no. It hasn't been fully rendered from the original data, there is a preview that's rasterized so you can see what you've done so far. But then that's true too of the SO from raw; not rasterized yet, maybe never until the SO and other layers are flattened.
Look you can take ONE raw file, in its native bit depth and render it multiple times to blend it as you desire. There's nothing new in what this guy's showing here. You can't recover highlight data that's fully clipped due to true over exposure of the raw and visible with a raw Histogram. He's taking a very old concept (multiple rendering and blending of one raw) and adding some complexity by speaking of bit depth etc.
IF you have a 12 bit raw capture with clipped highlights, that's what you have, no matter how ACR or Photoshop slices up that pie. You gain no more highlight data or recovery using more bits; it's there or it isn't.
Now IF you bracketed the shot, then used multiple raws, that's a totally different technique and process but again, the recoverability isn't due to the bit depth per se, it's due to the data actually being in one or more of the raws.
All this talk of "Raw Layers" and "32-bit" in the video kind of reminds me of an old saying by the late, great Bruce Fraser:
"You can do all sorts of things that are fiendishly clever, then fall
in love with them because they're fiendishly clever, while
overlooking the fact that they take a great deal more work to obtain
results that stupid people get in half the time. As someone who has
created a lot of fiendishly clever but ultimately useless techniques
in his day, I'd say this sounds like an example."
Bruce Fraser