Oskar

joined 2 months ago
[–] Oskar@piefed.social 5 points 9 hours ago

Pixelfed has been around for a couple of years. I created an account in the summer of 2022. But it didn't become popular until there was an app. A web app that does the same isn't enough. If it isn't in App Store/Play it doesn't exist for most people (which is something all Fediverse devs must think about).

Anyway, plenty of information about the development pace was available for anyone who wants to make an informed decision before they send their money somewhere. Apparently >2000 people thought it was worth it.

[–] Oskar@piefed.social 12 points 2 days ago (2 children)

A decentralised Bluesky won't work. We have all heard the arguments that it's totally impossible to decide which Fediverse instance to join. Surely that problem will doom Bluesky's decentralisation efforts.

/s

[–] Oskar@piefed.social 4 points 4 days ago

Interesting, I had missed that there are "non-official" models that can be used with Ollama just like the official ones. e.g. https://ollama.com/huihui_ai/deephermes3-abliterated

And it gave a good explanation to my "lithmus test" code snippet

[–] Oskar@piefed.social 3 points 4 days ago

Ollama's Python API works well and there's a lot of examples. However, I've not gotten the Ollama REST API to work, the response doesn't unpack into json for me.

Just design the rest of your system so that it doesn't have to know anything about the implementation (only prompts and responses) and you should be able to easily replace the LLM part later.

[–] Oskar@piefed.social 4 points 1 week ago* (last edited 1 week ago) (1 children)

If Bluesky becomes federated with multiple instances, it will be just as impossible to enter as ActivityPub-based services apparently are since instance-selection is a blocker.

RIght?

;)

[–] Oskar@piefed.social 1 points 2 weeks ago (1 children)

Interesting, lots of "bang for the buck". I'll check it out

[–] Oskar@piefed.social 1 points 2 weeks ago (1 children)

of course, I haven't looked at models >9B for now. So I have to decide if I want to run larger models quickly or even larger models quickly-but-not-as-quick-as-on-a- Mac-Studio.

Or I could just spend the money on API credits :D

 

Thinking about a new Mac, my MPB M1 2020 16 GB can only handle about 8B models and is slow.

Since I looked it up I might as well shared the LLM-related specs:
Memory bandwidth
M4 Pro (Mac Mini): 273GB/s M4 Max (Mac Studio): 410 GB/s

Cores cpu / gpu
M4 pro 14 / 20
M4 Max 16 / 40

Cores & memory bandwidth is of course important, but with the Mini I could have 64 GB ram instead of 36 (within my budget that is fixed for tax reasons).

Feels like the Mini with more memory would be better. What do you think?

[–] Oskar@piefed.social 1 points 3 weeks ago

For lots of people it'll be less confusing. They open their app store and search for this Pixelfed they've heard of and find an official app. That's what they expect.

[–] Oskar@piefed.social 1 points 3 weeks ago

I finally found a Python-specific model. Searching is tricky when almost everything around local llms involve Python. https://huggingface.co/codellama/CodeLlama-13b-Python-hf

 

There are lots of general-purpose models to use locally, and also coding-specific models.

But are there models specialized in one programming language? My thought was that a model that only needs to handle one language (e.g. Python) could be faster, or be better for a given size.

E.g If I need to code in Rust and is limited to an 8B model to run locally, I was hoping to get better results with a model that is narrower. I don't need it to be able to help with Java.

This approach would of course require switching models, but that's no problem for me.

[–] Oskar@piefed.social 5 points 4 weeks ago (1 children)

I don't have any recommendations, but I'm sharing information about train travel through a Mastodon account in Swedish https://mastodonsweden.se/@interrailinfosvenska I wish you good luck creating a new community in the Fediverse! And I have posted a challenge for a similar thing in Sweden.

[–] Oskar@piefed.social 6 points 1 month ago (1 children)
view more: next ›