Does anyone know of any attempts out there of building more ethical AI models for things like text LLM, image or audio processing?
To clarify, I am talking about an AI model which is trained on a dataset of public domain, creative commons, licensed, or more intentionally fair use sampling methods.
Personally, I believe that the way the modern AI industry has built their giant models is incredibly exploitative, unethical, toxic to society and culture, and arguably not even consistent with the letter or spirit of the existing rules by which we have all been asked to play.
As a tool, machine learning may not be fundamentally bad, but who, if anyone, is out there trying to do it in better, more legitimate way?
Fair. I guess I’m curious if any people who exist outside of the profit-driven investment hype “AI” bubble are trying to see what can be done with more legitimate sources of data.
It’s possible that nobody is even trying it…
Hey so my research was on AI. Your specifically talking about LLMs which is a very very small subset of AI. There’s a metric TON of public LLMs that are specialty made. The more recent one was a person who made a tiny one just for one off bash scripts based on his own massive 20+ years of dot files. Hugging face has some but others are just all over the internet. The training sets ate usually public sources. Creative commons or data specifically for that purpose.
Most AI work is happening at university and open source levels.
I don’t really mind if it’s LLMs or other “AI” models for things like image or audio processing. My question is really about whether there are any noteworthy models that have been trained off an more ethical dataset than the big ones that have been trained on massive troves of scraped data.
That’s interesting.
Does “public sources” mean public domain, or just like visible on the internet?
In research (as in papers) you have to use curated results or peer reviewed datasets. So its a grab bag. But some are ONLY available in acadamia. Its a strange world. The stuff AIs are doing are novel but there is many more nitch stuff for very specific reasons.
For example my research was on using a variant on the Probabilistic Latent Semantic Analysis for specific applications (cancers/medical applications). It was trained on datasets and proven to work. It was 90percent on other peoples work and was more a proof of concept. And there was work for the multi-model version at the time.
Datasets were all over, some public domain, some very specific medical data with permission.