FaceDeer

Yeah, the whole concept of "national" TLDs is proving to be a rather poor one in practice. Very few of them actually make sense in the way they're used.

Heh. I fell off of contributing in recent years, but there was a time back in the day when my edit count was in the top hundred or so. Your impression is completely wrong.

Anyway, this discussion here isn't going to affect what the people on Wikipedia are doing, so it doesn't really matter. I linked to the project page above and it's quite clear that even this "AI Cleanup" project is not in any way fundamentally opposed to using AI, they're just focused on ensuring that editors using it are adhering to Wikipedia's guidelines. If you think AI can't do that then clearly your concept of how AI is useful is too limited.

You're probably assuming that someone would just go to an LLM and say "write a Wikipedia article about subject X"? That wouldn't work well, but that's very far from the only way to use LLMs for Wikipedia work.

For starters, it doesn't have to actually write content at all. You could paste an existing article into an LLM and ask it "What facts in this article lack references to back them up? Are there any weasel-worded statements, or statements that don't appear to follow a neutral point of view?" And get lists of things that require attention.

Or you could paste a poorly-worded article in and tell it to rewrite it with all the same information but better phrasing or structure. You could put a bunch of research materials you've gathered into the LLM's context and tell it to write a summary in the style of a Wikipedia article, with references to the sources for each fact mentioned. Obviously you'd check the LLM's work afterward and probably do some manual editing, but this would be a great time and effort saver to get a first draft written. You could take an existing article and tell the LLM that some particular fact had changed or been discovered to be incorrect and ask it to rewrite the relevant parts to account for that.

Wikipedia is in many, many languages. You could have a multilingual LLM automatically compare the contents of different language versions of a Wikipedia article and ask it to spot differences in content or tone. You could have an LLM translate an article from one language to another as a starting point for creating an article in that new language.

You could have the LLM check the references of an existing article - look up each referenced work on the web and see whether it genuinely says what the article that's using it as a reference says. It could flag all manner of subtle problems that way. Perhaps the reference sounds biased, or whoever used it as a reference misinterpreted it, or the link was simply incorrect and points to unrelated material. Being able to have an AI do a first-pass check of all that in a completely automated way would save huge amounts of time.

This is all just brainstorming off the top of my head, so I'm sure there's plenty of other good uses that aren't coming to mind.

From the project page:

The purpose of this project is not to restrict or ban the use of AI in articles, but to verify that its output is acceptable and constructive, and to fix or remove it otherwise.

There's nothing fundamentally wrong with LLMs. Users just need to know their capabilities and limitations and use them correctly. Just like any other tool.

They're not talking about the same thing.

Last week, researchers at the Allen Institute for Artificial Intelligence (Ai2) released a new family of open-source multimodal models competitive with state-of-the-art models like OpenAI’s GPT-4o—but an order of magnitude smaller.

That's in reference to the size of the model itself.

They then compiled a more focused, higher quality dataset of around 700,000 images and 1.3 million captions to train new models with visual capabilities. That may sound like a lot, but it’s on the order of 1,000 times less data than what’s used in proprietary multimodal models.

That's in reference to the size of the training data that was used to train the model.

Minimizing both of those things is useful, but for different reasons. Smaller training sets make the model cheaper to train, and a smaller model makes the model cheaper to run.

If I could get glasses that told me "that guy enthusiastically greeting you by name right now is Marty, you last met him in university in such-and-such class eight years ago" I would pay any amount of money for that.

"Doxing people" and "recognizing people" have a pretty blurry border.

And yet it's accomplishing those tasks. I guess that means "understanding" wasn't necessary for them after all.

Alright, since you find this such an important issue, consider the first bullet point cropped off of my humorous list of milestones.

Doesn't change the underlying point.

Phone operators weren't call center staff, they were literally routers in human form. Secretaries were your email program, calendar, and your folders full of word documents.

• Computers might be good at numbers and typesetting, but we'll always need human secretaries and phone operators to keep things running.
• They might be able to beat a novice, but no computer will ever beat a human grandmaster at chess.
• Okay, then they can't beat humans at Go or poker.
• Any non-trivial task requiring creativity and understanding is beyond these tools. ← you are here
• AI-run corporations will never be able to outcompete ones with ones with human boards and CEOs.
• An AI scriptwriter could never win an Oscar.
• I'm voting for the human candidate for president, I don't think the AI one is up to the task.