That page seemed outdated, but: From further down that page:
The recommended strategy is to share the keys automatically only to verified devices of the same user
This is the same situation where the key backup is accessible - which is not described on that page, but it's a key store of all the megolm keys. This is what is now generally used instead of that as it doesn't require devices to be online and allows recovering keys if all devices are lost.
I host mail via Stalwart, which makes it pretty damn easy - it handles most everything, just giving you a big block of DNS records to upload with all the DKIM SPF MTA-STS nonsense. However, spam filtering from big providers is still occasionally an issue. I still occasionally get reports of mail making it into Gmail's spam inbox, for example.
A few days late, but I have a pretty similar usecase to you on https://forgejo.ellis.link/. My solution is go-away, https://git.gammaspectra.live/git/go-away, which just sits as a reverse proxy in between traefik and Forgejo. I haven't enabled fancy stuff like TLS fingerprinting. It's been effective enough at killing the bots downloading archives and DDoSing the server from residential IPs. My config is based on the example Forgejo config, but with a few tweaks. Too long to post here, though, so message me if you need access
You might want to check out https://matrixrooms.info/, which is good for a search around. Some project communities also have offtopic rooms that are good to chat in. Even very small rooms can be very active.
"can't guarantee the authenticity of this message" just means it was restored from backup. In the same vein, if you can decrypt a message in any client, it should upload the keys to the message backup so it can be decrypted on other clients, even ones that haven't logged in.
Continuwuity developer here - Conduit is reviving itself, and you can no longer move from Conduit to tuwunel or Continuwuity. You haven't been able to for as long as either project has existed. You might be confusing conduwuit with Conduit.
Basically: record every song you listen to, and when you listen to it (plus some more metadata), and then add it to a giant public dataset. Open source software then uses that to make music recommendations based on your and other people's listening, and to give you interesting stats about your listening.
That kind of captcha is trivial to bypass via frequency analysis. Text that looks like language, as opposed to random noise, is very statistically recognisable.
Finetuning, self-hosting and whatever decentalised network they've got going on there aren't going to change the core of the technology. Oh, and it's a tiny local model (about 1/100th the size of cloud models), too, it's going to perform poorly compared to SOTA models anyway.
These will still fall prey to the reason that LLM summaries are bad: LLMs pick up the average, what is common, rather than what stands out and is genuinely important or new. Your writing will end up averaged out and the key things will be missed, only what is repeated again and again.
In my experience you can use a LLM to point out typos or grammar errors, but not to actually edit or rephrase your work. And at that point it's effectively just a slow and expensive, but better, spelling/grammar checker.
That page seemed outdated, but: From further down that page:
This is the same situation where the key backup is accessible - which is not described on that page, but it's a key store of all the megolm keys. This is what is now generally used instead of that as it doesn't require devices to be online and allows recovering keys if all devices are lost.