Since a lot of people will come to the comments without reading the article itself:
It is embedding the Youtube video, like a lot of websites and apps do - if you ever watched a youtube video outside of youtube.com or official apps, odds are it is using the same API as that app.
It still shows ads and uses the Youtube website directly for certain parts, but with some custom CSS and JS akin to how browser extensions would work.
From the article (right under "Can I give feedback"):
I’ve only been able to develop this in the simulator, which obviously has its limitations, so once I get my hands on a device this Friday I’ll probably have a lot of thoughts on things I want to improve as well.
the 300/month is for the 5$ plan? possible "fair use"-like hidden limits aside, the 10$ sounds unlimited
from their front page they claim that "We do not log or associate searches with an account", and their privacy page is fairly detailed
the story is written from the perspective of a literal toaster, not a human
(the primal urges / bodily needs being just normal hunger for food)
Might want to clarify: The "model" in this case is not a full model like Stable Diffusion, but rather something used like a patch, more comparable to something like LoRA
I don't think that anyone would misunderstand anyway, but better safe than sorry
It is not "in the whole fediverse", it is out of approximately 325,000 posts analyzed over a two day period.
And that is just for known images that matched the hash.
Quoting the entire paragraph:
Out of approximately 325,000 posts analyzed over a two day period, we detected
112 instances of known CSAM, as well as 554 instances of content identified as
sexually explicit with highest confidence by Google SafeSearch in posts that also
matched hashtags or keywords commonly used by child exploitation communities.
We also found 713 uses of the top 20 CSAM-related hashtags on the Fediverse
on posts containing media, as well as 1,217 posts containing no media (the text
content of which primarily related to off-site CSAM trading or grooming of minors).
From post metadata, we observed the presence of emerging content categories
including Computer-Generated CSAM (CG-CSAM) as well as Self-Generated CSAM
(SG-CSAM).
Here's a link to the report: https://stacks.stanford.edu/file/druid:vb515nd6874/20230724-fediverse-csam-report.pdf
It is from 2023-07-24, so there's a considerable chance it is not the one you were thinking about?
They could live in the Southern hemisphere Subtropics like somewhere in Latin America
Their arguments still hold up pretty well as far as I can tell. If anything "improved" since then, you could argue that what the biggest platforms decided to use (Mastodon, Lemmy) became the de-facto dialect in use, but you still have to explicitly refer to how certain projects do things if you want to implement ActivityPub, which can be pretty demotivating for developers, and doesn't makes the user experience any better.
...and nothing prevents new apps such as Threads from using ActivityPub differently, being incompatible with existing apps and further dividing the space
If by algorithms you mean things like GPT, all data on the fediverse is effectively public and arguably even easier to be collected than the likes of reddit, and is almost definitely going to be used to train models whenever or not the fediverse federates with threads.
There's still significance in defederating though, specially when it comes to preventing "Embrace, extend, and extinguish"
That is already possible on kbin (browse to kbin.social/d/instance then block there, e.g. https://kbin.social/d/vlemmy.net ), I don't think that it is supported on Lemmy though.
(Mentioning since this discussion is on kbin, though it seems like you are using lemmy. rip)
Out of all things to hate Reddit for, giving data to AI isn't something fediverse users can really criticize it for, though making money from it perhaps.
Remember: All data in federated platforms is available for free and likely already being compiled into datasets. Don't be surprised if this post and its comments end up in GPT5 or 6 training data.