The pace of the open source community's progress on liberating language learning models from the iron grip of OpenAI has been incredible this month.
After Llama was leaked by Facebook staff, the race to make a free, fast, high quality language model that can run on a typical gaming laptop or pc has taken hold. Powerfully.
It is a bit like watching marathon runners(the os community's devs), each competing to make language models free for everyone to use without paying a monthly subscription and ugly, censored and woke-washed replies from the algorithms.
Vicuna: Open Source LLM that boasts 90% similarity to ChatGPT4 in quality
Vicuna, a model that uses ChatGPT4 in combination with Facebook's Lama leaked 'base', produces arguably similar quality responses to questions to OpenAI's ChatGPT4. They boast a 90% similarity in replies to ChatGPT4. (And I believe that is truly the case after trying the huggingface for the model.)
As people learn about these open source programs, some are able to keep up with the speed of updates and improvements.
But not all.
As it leapfrogs forward at Olympic pace, AI tech becomes a tantalizing, confusing and murky space if you are not a developer with some experience with machine learning and coding... which is the meat of this blog post.
Exaggerations, Gaps and Avatars that Talk: the Tripwire of Idolatry
Upon scanning comments by various audiences on Youtube, there are large gaps in understanding. What is the technology behind a talking 'chat' like ChatGPT and Vicuna? How does it actually work? What can it do, and what is just imagination, human projection?
-- For example, one user commented that they were able to see Vicuna 'learn' and that they believed it was storing the data in some sort of permanent memory which others were able to use to 'improve' the Vicuna LLM.
This is really not the case. However, it was very revealing. As a desktop standalone, it runs on your PC's GPU / CPU and does not send any data back to any central data storage for use by other users.
You see, the introduction of humanizing elements, and language itself -- speech, is a formally human activity-- results in mythologizing, fantasizing and projecting upon the inanimate algorithms utilized for the complex appearances of intelligence.
As such, it is becoming more commonplace to err and attribute magical and mythical qualities to Vicuna, ChatGPT and "talking avatars" talking engines like phind.com and you.com.
This alarming aspect, and the growing ignorance, creates a deep black chasm that I believe humanity will eventually fall into as a whole.
The devs, elite programmers of machine learning models and those that utilize advanced llms, may, at the time of this writing, be able to keep up with the technology, its inner workings... but this won't last for very long.
So the danger, then, is that we lose the elite programmers and devs of the open source community who are able to, as interpreters of complexity for the rest of us, make this tech traversible in a simple human way.
But the more complex the language models and tech becomes for the user who uses a simplified interface, the more likely idolatry is ahead.
...While it is wonderful to get instant gratification, the answers, images and solutions to life's problems -- the reliance upon speech systems, and a feedback loop of engagement shrouded in mystical thinking, can and will materialize as mystification among a rather sizeable portion of the population.
In fewer words:
- Less reliance on your own individuated thinking process because AI chats are so easy to engage with
- More trust and integration with AI generated solutions leading to less introspection of devices, properties and mechanisms of the algorithms - so imaginary explanations and human characterizations will form naturally
- Simpler interfaces that speak and have a visual appearance engages centers in our brain associated with real friendships and real people!
"A feedback loop of engagement shrouded in mystical thinking, can and will lead to mystification among a rather sizeable portion of the population."
This is primarily due to our inability to grasp and retain new information in an exponentially expanding techno horizon: where, if you look away for 48 hours, there may be a critical leap, advance, that changes the game entirely for everyone.
--=--=--=--=-- OPEN SOURCE EFFORTS YOU MAY WANT TO CHECK OUT --=--=--=--=--
This space moves very, very fast, so these links may be outdated by the time you reach them:
- Oobabooga WebUI is just like Automatic1111 but for Chat AI models: https://github.com/oobabooga/text-generation-webui/
- How to setup Oogabooga+Vicuna:
- Overview of current OS efforts with analysis: https://github.com/langbridgeai/Open-Instructions
- Phind talk engine if you don't want to use ChatGPT: https://www.phind.com/
The OS community is exploding with a flurry of activity and training tons of uncensored language models for virtually ANY purpose you can dream of.