Truly open source ai model

Does anyone know about a truely opensource ai model and trained with open data set and usable good quality.

I find llama2 meta
Mixtral 8x b
And openchat openchat.team/

The Allen Institute published OLoMo:

2 Likes

Llama2 by meta is also fully opensource with opensource licence what is the difference between this two.
And openchat is also fully open-source
What is the difference between them also.

:thinking: Did you read through it?

1 Like

Can you elaborate

do you guys think in the short time these open source models will be integrated to linux desktops like how windows is doing ?

Although Meta refers to it as “open source”, it is not under an open source license. See https://opensource.org/blog/metas-llama-2-license-is-not-open-source

3 Likes

Different tools already integrate them, I saw this for vim, for example:

I guess by “Linux desktops” we mean the different desktop environments—Gnome/KDE etc. I haven’t looked to see if they are looking at integrating these.

For any one that has used these on Windows—what new functionality does it really give?

Also as far as I remember Meta had its model leaked, and opensourced it afterwards to turn around the scandal. Pretty smart, but I dont have sources yet.

There are certainly some very interesting things to navigate here! OLMo is impressive in that they (as far as I can see) meet everything in the Open Source Initiative’s current draft definition of Open Source AI — the model and weights and parameters, and training code and details, and the training set.

Mistral, by contrast, has an openly-licensed model and weights, but it gets murky past there.

Our current stance on this is that these models constitute content rather than code (see Fedora Legal Docs: Licenses Allowed for Content). (And, you’ll find llama on the Not-Allowed LIcense List.

We may develop other guidelines about we want for Fedora policies above and beyond this, though. I think we want to strongly encourage truly Open Source AI, but the “content” policy makes sense to me.

1 Like

Why not for those have the hardware they can toggle that on or off to install a small model which is around 2b like 2gb and use it locally.

If Fedora, Ubuntu, Red Hat, SUSE don’t integrate AI features mainly llm chat like stuff into their desktops within the next two years, Linux desktops could fall behind the competition.

As you are seeing the growth of desktop to keep that constant this is important.
And i think @mattdm @ankursinha will also agree with it.

With data privacy as it runs locally.unlike copilot which collect data you use it or not even in server os.

1 Like

I don’t quite see how distributions can “integrate AI features”—they don’t develop the software. As in, what part of the distribution should be AI enabled? Different apps can be, even desktop environments, but Fedora doesn’t develop any of that—upstream developers do.

I haven’t yet seen a use case where “AI” has helped on a desktop computer—what does it do there, even on Windows? It’s cool yeh, but what functions does it carry out?

No that assumes everyone wants AI integration.

I think:

Good GUI > AI assisted GUI > AI assisted TUI > TUI

(GUI means graphical user interface, the desktops. TUI means the terminal)

I rely on ChatGPT a lot to do complex stuff or just to get the needed syntax for CLI tools.

Problematic here is

  • the data is random and often wrong
  • the model is huge and not single purpose (waste of money)
  • the data is outdated

Having a LLM trained on Fedora stuff would be cool, but good documentation or an easy “poweruser friendly” GUI (that doesnt require users to go to the Terminal) are way more important.

And because of that I am a huge supporter of “Fedora Workstation KDE”, as KDE Plasma is exactly that desktop.

I don’t really see a lot of use for large-language model AI in a desktop environment, although I’m interested to see if a carefully-tailored tool could help with infrequent complex tasks.(@boredsquirrel, do you have some specific examples where you found chatGPT to do a good job?)

I do think that there’s a lot of value in on-device voice recognition and speech-to-text, and in possibly finally getting decent open source text-to-speech. These are valuable as assistive tools for accessibility, but they’re also helpful for a large number of people who don’t require them in that way — just like wheelchair access rules make it much easier and safer for everyone to get around.

I won’t be terribly surprised if most children in the next decade are more comfortable interacting with a computer via voice than by keyboard (whether physical or on-screen). How do we make Linux and open source available to them?

Another idea: something could identify untranslated strings in software we distribute, and (optionally, of course) provide a possible translation and ask users if that translation seems right. This input could then be collected and used by our translation team to get better coverage in more languages (and perhaps draw more users into becoming part of that effort in a more active way).

1 Like

And, there’s another obvious place, not on the desktop, but in Fedora ioT — I’d love to have an out-of-the-box voice assistant that runs everything locally under my control and which can interface with Home Assistant.

1 Like

i use daily AI Google Gemini 1.5 advanced now on desktop usage and phone for daily normal tasks to get basic stuff faster and i use Coding AI on intellij IDEA for help debug and make life much easier and i like how my setup flows already and cant wait to get more advanced integrations for Gemini that is launching now Chrome integration and later desktop integrations so the future is defo in AI on daily use i have tried copilot pro and github and now using other ones and difference is night and day compared to copilot and Gemini/jetbrains AI and the use cases are coming more and more automatic for now examples to create emails and email templates, searching and automating stuff integrating everything to Smart home HUB and i actully have just random chats and jokes to AI like it has become part of the workflow and daily life that actually sometimes scares me that i tell jokes and talk to AI

The data privacy of running locally is a very important difference from anything that requires and sends data to online services.

I’m eagerly anticipating the release of the fully open-source PLLuM, slated for late 2024. In the meantime, I highly recommend LM Studio. It boasts an extensive array of downloadable models that can be run locally, all within a user-friendly interface.

(However, it’s worth noting that there’s a known issue with LM Studio version 0.2.20 beta on Fedora. Fortunately, the earlier iteration, version 0.2.19, functions flawlessly (for me). Additionally, you might find it necessary to deactivate GPU Acceleration in the Hardware Settings.)

This does not require any data send to any server everything is locally.