AI on Cloudron

[LoudLemur]

Prompt Engineering World Championships:
https://app.openpipe.ai/world-champs/signup

RoboCup 2023 - AI/Robot football:
https://www.robocup.org/
https://vid.puffyan.us/watch?v=vwIuQKKg-sY&quality=dash

[LoudLemur]

Nvidia reveal new chip:
https://nvidianews.nvidia.com/news/gh200-grace-hopper-superchip-with-hbm3e-memory

[LoudLemur]

Useful collection of AI prompts:
https://huggingface.co/datasets/fka/awesome-chatgpt-prompts

[LoudLemur]

Has anybody had success finetuning a language model so that it becomes an expert on local data? What free software tools were used and was it an enjoyable, fruitful process?

[micmc]

@LoudLemur said in AI on Cloudron:

Has anybody had success finetuning a language model so that it becomes an expert on local data? What free software tools were used and was it an enjoyable, fruitful process?

For my part, I'd been trying to play a little but not much yet, I'm actually busy working on developing and launching a SaaS Web UI through API connections to different LLM providers.
The service makes it very easy to anyone to create content (and well, much more like consulting with powerful coach bots in several fields) with a few clicks, and answering a few questions.
Local data labs is next...

[LoudLemur]

@micmc https://www.deeplearning.ai/short-courses/finetuning-large-language-models

[LoudLemur]

Fine-Tuned CodeLlama-34B now beats ChatGPT4 on HumanEval.
https://www.phind.com/blog/code-llama-beats-gpt4

https://huggingface.co/Phind

https://huggingface.co/sbeall/Phind-CodeLlama-34B-v1-q5_K_M-GGUF/tree/main

You can run it on Ollama today:
https://ollama.ai/library/phind-codellama/tags

OpenRouter allows you cheap access to Free Software Language Model APIs:
https://openrouter.ai/docs#models

[LoudLemur]

Falcon-180b (that is 180 billion parameter) Free Software model has now been released.
This is the chat version:

https://huggingface.co/tiiuae/falcon-180B-chat

The open source dataset was created in the UAE through a process of web crawling and stringent filtering out of "adult" sites based on their URL. It is multi-modal friendly with image tagging.

The main site is being absolutely hammered at the moment. Archive:
https://archive.ph/trCCZ

This is by far the largest Free Software model available at the moment, and it is outperforming Llama 2.

500GB download, 2.8TB storage after unpacking, 400 GB of memory will be needed to swiftly run inference o Falcon-180B.

"I think we are going to need a bigger boat."

Demo: https://huggingface.co/spaces/tiiuae/falcon-180b-demo

[LoudLemur]

Petals.dev

Run large language models, including Falcon-180b, bittorrent style:
https://petals.dev/

[micmc]

@LoudLemur said in AI on Cloudron:

500GB download, 2.8TB storage after unpacking, 400 GB of memory will be needed to swiftly run inference o Falcon-180B.

I'm going to install 2 right away!

[LoudLemur]

OpenAI ChatGPT Enterprise:
https://openai.com/blog/introducing-chatgpt-enterprise

We’re launching ChatGPT Enterprise, which offers enterprise-grade security and privacy, unlimited higher-speed GPT-4 access, longer context windows for processing longer inputs, advanced data analysis capabilities, customization options, and much more.

[LoudLemur]

Talk to President Obama: https://gptcall.net/

[LoudLemur]

Chat with AI Characters Offline.
Runs locally. Zero-configuration.
https://faraday.dev/

This library aims to help make gguf downloading/management easier:
https://github.com/ahoylabs/gguf.js

[nebulon]

@LoudLemur said in AI on Cloudron:

Talk to President Obama: https://gptcall.net/

This is indeed quite fun!

[micmc]

@nebulon said in AI on Cloudron:

@LoudLemur said in AI on Cloudron:

Talk to President Obama: https://gptcall.net/

This is indeed quite fun!

Yep, and in reality it's relatively easy, and fun too of course, to build such web UI. At the base, it's simply built through prompt engineering. Meaning a fair knowledge of how to talk to the machine, to make it spit according to your will loll
Which means such cool site could also be built by almost anyone with a good idea, using ChatGPT, or almost any good GPT LLMs, and incorporating the right "prompts" in a database like Notion (or even Directus on Cloudron? I'm testing it...), that has capabilities to publish public websites based on their databases.
As anyone ever tried, for example, to just tell ChatGPT to act as Obama or any well known personality, just before starting a conversation with it?

[marcusquinn]

Seems to be an above average chance that all our digital content creation is going to become an AI Avatar that future generations will be able to chat with. Wild times!

[LoudLemur]

RTX3090 and Llama 70b @ 20 tokens/second:
https://old.reddit.com/r/Oobabooga/comments/16gxqzt/exllamav2_20_tokenss_for_llama270bchat_on_a_rtx/

Haven't yet verified

https://github.com/turboderp/exllamav2

[LoudLemur]

@marcusquinn Somebody just created a Lora which they trained on all of the art they created at school, pencil drawings. The AI can now make art as though it were their younger self.

[marcusquinn]

@LoudLemur We're being uploaded to the Matrix! Maybe we'll become the aliens popping in on other planets

[micmc]

@marcusquinn said in AI on Cloudron:

@LoudLemur We're being uploaded to the Matrix! Maybe we'll become the aliens popping in on other planets

That's why we 'start' lurking in the IPFS direction