That doesn’t take into account the extensively researched plagiarism concerns. It’s not just that it’s low quality slop but that some of us think the GPL doesn’t work if you can just train LLMs on GPL, then have it spit out GPL snippets un-GPL’ed.

Some people literally un-GPL projects via AI in one go. While that’s the egregious version, any LLM use seems to risk having a similar effect in a smaller scope.

This isn’t only a legal question. At least if you think the GPL has societal and moral value.

We need to be pointing to good applications of AI Freel free to do so, but studies are not on your side.

The only somewhat clearly useful use case appear to be code reviews, but then you don’t need to actually allow submitting any LLM rewritten code or text since code reviews can be done using natural language. And if you use server-side LLMs, you’ll probably agree to ToS that they steal your data.

And they seem to be amazing at plagiarism.

You, like a large portion of the ‘fuck AI’ community are angry at LLMs or image/video generation models and their associated capitalist bubble. Yes, LLMs produce poor quality output compared to humans and yes the current marketing and capital explosion is bad for everyone involved that isn’t otherwise independently wealthy.

The reason that these are the AI that you’re aware of is that AI needs a lot of data to train and the only source of a huge amount of data, the Internet, is primarily text, images and video. So the first large transformer-based neural networks were trained on that dataset.

ChatGPT and Sora are toys, they were just the toys that were easiest to make given the data available when transformers were discovered.

If you train neural networks on different kinds of data you get different models. For example, if you train neural networks on protein folding data, you get neural networks that can predict protein folding based on an amino acid sequence. This is a thing that human-created software has not had great success at.

People may be familiar with Folding@Home, a project which attempts to leverage donated computing resources to brute force the problem. These projects have consumed thousands of person-hours of our best scientists and engineers and the results are pretty poor.

However, since we now know how to train neural networks on data, we can train an AI to predict the protein structures and the resulting networks such as AlphaFold (en.wikipedia.org/wiki/AlphaFold) produce results much higher than human engieered software.

In addition to predicting the structure, other scientists have used diffusion models (similar to how consumer AI products generate images) to go the other way. Now a scientist can describe a protein’s properties in a prompt and instead of generating a picture the network outputs the sequence of amino acids that are most likely to fold into a shape with those properties.

Robotics are another field where AI is making an impact unseen to the public. There isn’t an Internet full of bipedal motion or limb-positioning data, so it is much harder to train an AI to operate robotics. There are many projects which are working to create that data and the results are pretty impressive. This is a bipedal robot which has been trained on human motion: www.youtube.com/watch?v=I44_zbEwz_w compare that to pre-AI motion: www.youtube.com/watch?v=LikxFZZO2sk

Weather forecasting is another field where AI is useful. Predicting weather requires identifying patterns in huge amounts of data and AI is uniquely able to deal with that level of complexity.

None of these uses of AI can talk to you, or produce pictures. They cannot understand sentences or write e-mails or generate code. They’re trained on data generated specifically for their purpose, not on public data scrapped from the Internet. Their output allows us to develop medicines faster, automate dangerous jobs and predict weather disasters.

I’m with anyone who’s concerned about the capitalist frenzy over LLMs and image/video generation products. This is clearly another dotcom bubble and the spending frenzy and disruption in the job markets is damaging the economy and hurting workers at a large scale.

I do not lay the blame for this at the feet of neural networks. The blame lies with the human beings making the decision to take a promising technology and to dump trillions of dollars into it without any endgame other than market dominance.

The community should but ‘fuck AI executives’, AI has many uses outside of LLMs and image generation and people are completely missing all of the amazing things that this technology is making possible.

Then the PR can be evaluated, rejected if it’s nonfree or just poor quality

I don’t get the difficulty of rejecting “if it’s nonfree or just poor quality or known LLM code”.

I don’t think it’s a vague criterion at all. And for many projects, if you tell them it’s from a StackOverflow post, unless you can show it’s not a direct copy they will reject it as well. I don’t see the difference. Now whether you think LLMs are worth the trouble to use is a different discussion, but your argument doesn’t convince me. Many bans aren’t easy to enforce, that doesn’t mean it’s a bad idea.

There is also a responsibility and liability question here. If something turns out to be a copyright issue and the contributor skirted a known rule, the moral judgement may look different than if you knew and included it anyway. (I can’t comment on the legal outcomes since I’m not a lawyer.)