Mauve 👁💜 @mauve@mastodon.mauve.moe

This audio book trained TTS gets some sort of brain spasm when it encounters characters or words it can't easily pronounce. Feels like they should have trained it on IPA instead of regular token streams.

Here's what happened when I tried to get it to say a few tongue twisters ripped from wikipedia. It sounds drunk or like it's language center is being put through a blender.

https://huggingface.co/spaces/parler-tts/parler_tts_mini

Brushing up on #Svelte to do code reviews for a project and I gotta say it's pretty magical. :P

https://svelte.dev/docs/svelte-components#script-3-$-marks-a-statement-as-reactive

Watching my partner play Minecraft Legends. NGL it's kinda confusing. and this is coming after playing Scorn whicb seems to have zero guidelines at the beginning. #gaming

After struggling for a week to get ANTLR4 to parse/transform/serialize this obscure programming language, I've ditched it and got everything working in about two days of work.

It's not perfect since I'm skipping tokenizing some of the language constructs, but it's easy to iterate on and has just enough for the analysis/transform stuff I need to do.

I sometimes forget that my experience with AI is totally different from either the corporate or consumer experiences. Folks are talking about gpt 4 or claude or whatever while I'm focusing on tinydolphin and hermes 2 pro.

Shit I might get a beefier version of my GPD Win 4 so I can fit larger models into video ram. 😭

@Filene The solution is to expose a query layer to pull data for the specific view you're displaying and to sync any extra data on an as-needed or background basis. This is why I think p2p apps should be building on indexed databases as a building block and performing queries on remote datasets instead of "catching up" with all the remote data before showing it to the user.

Long read: https://blog.mauve.moe/posts/peer-to-peer-databases

@Filene So the main issue is with applications that pull oplogs from remote systems in order to calculate local state in a local index. In the case of matrix your client pulls all your new messages at boot and queries the state of all your rooms (and I think users in those rooms?). This is fine when your client is always online and is just listening for updates. It's also fine when you're in a small number of rooms. I am in hundreds of rooms with thousands of new messages.

@narF Yeah, but sadly it's all still very experimental. Maybe a few years from now if the effort doesn't get abandoned it will actually be practical.

Whoaaaa. Eclipse

My matrix server is like a hungry beast. I feed it CPU and RAM and it just keeps asking for more. Then again it is handling thousands of chat rooms and accounts at this point and needs to calculate a huge diff to send to my client every time I load. :P

Heads up: Matrix like all append only log systems sucks when you're not always online and loading data.

What are folks using for subscribing to #email #MailingList software? We basically want something to manage email signups which we can use with an API to send out emails to everyone. Ideally they should be able to unsub if they want. We'd want to be able to send either html or markdown via api.

Bonus points if the tool can subscript to an ActivityPub actor. :P