AI on Cloudron

[marcusquinn]

@timconsidine Tried poe.com ?

[timconsidine]

@marcusquinn said in AI on Cloudron:

Tried poe.com ?

Err, don't think so - will check it out - thank you !

[LoudLemur]

Pi AI from Inflection AI
https://pi.ai/talk

And Sam Altman is back at OpenAI, where they apparently have had a big breakthrough on AI automated math solving in a hitherto secret project Q* (q star).

[marcusquinn]

@LoudLemur If it could remember a conversation for more than 10 minutes, I'd be more impressed.

[LoudLemur]

Stable Diffusion XL Turbo - images fast as you can type
https://arstechnica.com/information-technology/2023/11/stable-diffusion-turbo-xl-accelerates-image-synthesis-with-one-step-generation/

Try it here:
https://clipdrop.co/stable-diffusion-turbo

[marcusquinn]

@LoudLemur Looks great, but couldn't see how to try out without paying.

[LoudLemur]

@marcusquinn said in AI on Cloudron:

@LoudLemur Looks great, but couldn't see how to try out without paying.

That link allows you to type about one sentence as a prompt without payment. As you do so, you can watch the images appear and change.

People like it because it can create so quickly, like in frames per second and it is usable to make video.

[marcusquinn]

@LoudLemur Odd, didn't work for me, just wanted Pro subscription.

[robi]

Have fun playing with this one:

https://github.com/OthersideAI/self-operating-computer

[marcusquinn]

Amazing: https://x.com/javilopen/status/1730986775135305851

[LoudLemur]

@marcusquinn Wow!

[LoudLemur]

This ASR (Automatic Speech Recognition) software is impressive. It can work in real-time and perform transcriptions. I don't know whether it would be suitable for hosting on Cloudron.

https://huggingface.co/openai/whisper-large-v3

[robi]

@LoudLemur not really as it needs GPUs to be fast enough.

There is an even faster version with better quality output: https://huggingface.co/spaces/sanchit-gandhi/whisper-jax

..and you can run them for free on HF.

[micmc]

@LoudLemur

Demo: http://imagebind-llm.opengvlab.com/

Interesting vision model, but still a bit confused
I might try self-hosted version, have you?

[micmc]

@LoudLemur said in AI on Cloudron:

plugin.surf
Thousands of free as in beer ChatGPTs
https://plugin.surf/

Have chosen to not sleep anymore?

Nice find! Thanks for sharing.

[robi]

Someone sent me a message with a streamlined image generation interface for Krita https://github.com/Acly/krita-ai-diffusion

[jagan]

Check out Mistral's new version, Mixtral 8x7B, whose performance has now reached or exceeded OpenAI's GPT 3.5.

https://mistral.ai/news/mixtral-of-experts/

[LoudLemur]

@jagan LMstudio now supports this mixtral model.

It also supports vision models!

#### New
- **Vision models** support! 🎉 (BakLLaVA, LLaVA, Obsidian 3B)
  - Vision supported via chat UI and server API (same format as OpenAI API)
- **Mixtral 8x7B** support! 🎉  (Experimental)
- **Rich Error Messages** for model loading errors (includes system diagnostics info)
- **Model Inspector**: peek into the model file metadata to see trained n_ctx, and more details.
- [Local Server] New client code examples for chat + vision (incl. updated OpenAI SDK example)
- Auto detection for \*.preset.json in Hugging Face repositories
- Support for StableLM Zephyr, Deepseek Coder, Persimmon, Qwen (thanks to llama.cpp improvements)

#### Bug Fixes
- `prompt_tokens` are now correctly counted in API responses
- Fix for multi-part model file downloads```


Andrew Torba, CEO of https://gab.com, says that 2024 will have a totally based AI model for us all to use.

[LoudLemur]

Chat with a (self-hosted) AI that took 40 seconds to learn a Github repo and find out how it works. This might be useful for people who want to package software for Cloudron:

https://app.getonboardai.com/

[LoudLemur]

Enable search (via DuckDuckGo ) on your local llama:
https://github.com/mamei16/LLM_Web_search

Actually, this has some support now for using SearXNG as the search engine, so it would work with a Cloudron package.

https://github.com/mamei16/LLM_Web_search/commit/68a85f0eda681a5e890c0de36fcf5f14f88abb46