

Sure, if you want… ¯\_(ツ)_/¯


Sure, if you want… ¯\_(ツ)_/¯


At first, I wrote “Element/Matrix” and decided to not be too pedantic… But if you want to be complete: the messaging protocol is, of course, Matrix. You could say there is actually no such thing as a Matrix server either, because it’s a protocol. The server must probably be Synapse-based, I guess. But there is an “Element-based server” in the sense that the web interface of Tchap (and phone apps) are very clearly forked from Element, which is what I meant.
Visio is based on LiveKit, which Element Call is also based on (as far as I understand). It lives outside of Tchap. The DINUM never mentioned it was based on Element Call. Do you have additional information? (Not that the difference matters much I think)


I guess. There must be a reason for them turning to something else than Jitsi given they already had that running (and the French service responsible for that, DINUM, is surprisingly extremely open-source friendly for a state service), but I don’t know which one.


Tchap is the Element-based server. Visio is a different thing. Although you can trigger Visio from Tchap (like you can do it with Jitsi from Element).


Actually, there’s also an official Jitsi server (called Webconf). And there’s another Big Blue Button one. I think Visio has features not available in Jitsi though, like AI-produced transcript of the call.
The ingredient you might be missing for how common the CLT is applicable is the following: in most complex systems (e.g. biological systems), any variable you measure is likely influenced by a lot of other hidden variables. Because there are so many variables at play, each effect is likely to be small, and the way their effects are compounded is likely somewhat additive (this one comes from things like series expansion). Hence summing up effects between variables must be relatively common and account for the bulk of the variation in a response variable.
A last bit is this: most statistical methods like the linear model are relatively robust to deviation from the normal distribution. So, you don’t need exactly a normal distribution, you just need close enough. It turns out the CLT often produces “close enough” quite quickly (i.e. with a few variables added together).