Something like this would be interesting because it would make the other models seem even less ethical.

It would no longer have the problem of being based on stolen works, but it would still have all of its other problems, such as frequent hallucinations and logical errors, higher energy costs than software built for purpose, and the really big one, evident cognitive decline in frequent users.

It would still make junior developers stupid.

Would eliminating one item from a long list of problems delete all the other problems from the list?

yeah, it would be a large use of resources for not much gain

I think what the current LLM clusterfuck is showing is that our little world or copy left and copy right we duct taped together since the printing press is all a bit silly.

Barring environmental concerns, no, I don't think so.

I keep thinking that we would be in a much better place if datasets were curated and models trained by respectable institutions like universities and foundations, I would "hate AI" a lot less and just use the LLMs that themselves have done nothing wrong.

I think the "copyright" shit is the biggest non-issue with LLMs. Abolish intellectual property. If you put a piece of software out there, you shouldn't be able to stop anyone else from using it for the public good.

The issues I take with what's being called "AI" right now is the volume of slop people have to drudge through. It particularly makes it hard to look for new software as there's so many vibe-coded projects getting published then abandoned. It's the aggressive and obnoxious marketing campaign that insists I pay for your shitty SaaS LLM, that I don't get to keep or run on my own machine, that I'm supposed to use for all sorts of tasks where I don't want a non-deterministic probabilistic machine (people use LLMs to do number-crunching?? you're using a machine designed to do math, interfacing with the one part of it that can't do math!). And it's the ridiculous computing power people are using to do activities that do not require anywhere near that amount of power if they just do things the normal way. I could not care less about "copyright". Nobody should be able to own a codebase anyway.

I think AI's a pretty good candidate for the Great Filter. There's a quote I think about often.

Each and every individual in such a species must eventually become capable of destroying their entire species at any time. Yet they must learn to control themselves to the degree that they can survive even such deadly insight. And frankly, they’re the only ones we really want to see leaving their solar systems. - Conversation with God by "Harry Stottle"

Would it 'hallucinate' any less often?

Given that FOSS licenses are premised on copyright, yes, the same ails would still exist: 1) AI washing of licenses (including transforming one license into another), and 2) the vagueness of whether LLM outputs can be copyright, which threatens the validity of a FOSS license upon that output.

The first ail can be seen even without LLMs: the BSD variants have gone through great pains to remove GPL-licensed code from their base repositories. This basically involves reimplementing utilities and functionality from scratch, using only the ideas that are in common with the equivalent GPL code, but never copying that code directly. This is properly considered a reimplementation, which can then be licensed permissively (eg MIT license).

If an LLM were to train on GPL code but the output were licensed with MIT, then that could be a GPL violation because GPL mandates that remixes continue to keep the GPL license.

Maybe you could avoid this fate by limiting the LLM to only train on permissively licensed code. So that it would be permissive licenses going in, and permissive licenses coming out. No GPL problems here. But that brings us to ail #2.

Some jurisdictions have rules against granting copyright for computer-generated works, in the same vein as works generated by non-humans (eg a macaque). If this LLM fell into this situation, then the output is not copyrightable. And if there is no copyright, a license like MIT or GPL simply cannot apply, because its terms couldn't be enforced.

Well, to be clear, the copyright parts of those licenses would be unenforceable. Some parts of the license may still be enforced under a contracts claim. But in any case, the things we refer to as "FOSS licenses" cannot attach to uncopyrightable works (with the possible exception of the CC0 license, which is essentially the absence of any license whatsoever).

It would still be problematic, yes. You can’t copyright the output of an LLM, so you can’t actually enforce any license on it.