@davidgerard @frog_reborn @nihilistic_capybara @zkat let's list all the unethical aspects of the current state-of-the-art LLMs e.g. trained using stolen content, no sharing of profit with corpus contributors, centralisation of wealth/power, carbon emissions, data centres spoiling potable water and driving up real-estate/utilities prices for locals, generation of CSAM and misinformation, plummeting average quality of software, making it harder for people to afford personal computers, lighting trillions of dollars on fire instead of spending that on humanitarian causes, ... etc (loads more factors to consider, of course)
I can see how one model might be more or less problematic during training and more or less problematic during inference compared to other models
e.g. if I train a model on my own computer using only my own content, and only from a battery charged by solar, then surely that's more ethical than using a subscription-only cloud-only model offered by OpenAI, right?
e.g. if a model is trained in a data-centre, then it effectively has the LLM equivalent of "original sin" and a lot of harm was caused just by that very fact, but is using that model for inference on my local GPU is slightly more ethical than paying an ongoing subscription to Sam Altman, right?
of course, most of the people using LLMs are doing so by choice, and would be perfectly fine if they did not use them, so harm minimisation here isn't really a redeeming argument
but, there is likely a non-zero set of people with accessibility needs or other reasons, where they have less choice in their engagement of LLMs (compared to the majority), and for this group maybe it's more valid to consider a spectrum of LLM ethics 🤷
