Yes

The training by using AI for training is distilling the used model. It may be their own or another. If the workers were not instructed to use AI, I suppose they choose their own favorite AI service.

Distilling your own models is not uncommon (usually a bigger one into a smaller one), but there are methods for it without (paid) human in the loop and they obviously did not pay them to do distilling but to add a human opinion to the mix.

If you want so, the contractors were paid to contribute their brain. Many clickworker tasks are about human alignment: Which answer is better? Is the image crap or just unusual? All the questions that current automatic systems cannot answer. You can rate mathematically aesthetics like symmetries, but many human opinions are not that straightforward to formalize. You can formalize some of them, by training a model that answers the same as the human - and for that you need the human with an honest human opinion instead of a copied AI opinion.

You can also have a look at all the "AI as a judge" benchmarks. Which model produces better prose? We let them create 100 stories and let a stronger model rate the outputs. Yeah ... but even the strongest models are no good judges what is "good prose". If you align using AI, you get something aligned with AI. You most the time want something aligned with humans.

It always was. But that never was a limitation. The neural network knows a lot of stuff, and the way to extract it is to let it continue a text, just like you write down your knowledge word by word. Nothing's wrong with that. Guessing is not the best word, though. Let's say a word selecting machine.

> I'm curious which use case you think they're limited in.
Writing prose. And the persons who hate AI may think they criticize AI the hardest, but the people who actually like using it are waay harsher when they talk about how bad AI sounds when writing creative texts. First it lacks diversity (in topics, names, style), and second it fixates on certain phrases like "it is not X. It is Y". It is not completely clear, if this is a way harder problem than coding, or if it just does may less money than coding. Sometime a company needs to really try to train a creative writing model for us to know.

Basically, because you don't want output that's pleasant to AIs, but output that's pleasant to humans. If you don't align it with human preferences, humans will not like it. There are some proxies for human preferences (aesthetic scoring, etc.) but alignment best works with actual human feedback.

Funny thing: Too much human feedback isn't good either. Remember earlier chatgpt glazing people too much? That's the result of a feedback loop in which humans rate responses and like these that glaze them. People do not always know what is best for them, so one might also want to curate curation - did the curator make a unbiased decision or will their opinion lead to an echo chamber?

That's just what these people are paid to do. If you read the model collapse paper you know that the point is not that something suddenly collapses, but that the training data quality is the ceiling. The quality (e.g. image, or writing style) of automatically generated content (i.e. no human steering, not manual prompts, no cherry picking, no rating) cannot be better than the training data, but may be worse. If you train on a mix of current ceiling plus worse input, your model can not get better but may become worse.

The solution is not necessarily human data, but for example human curation. Click the better of these two images. You regenerated the chatgpt response three times and then wrote the next question under the second generation. Input a description of the image. Write a prompt and select the best image of 20.

You need to treat the problem statistically (therefore mode collapse). Your AI generated in-distribution data and to widen the distribution you need to add out-of-distribution data. The AI can generate out-of-distribution data, but it needs to be nudged to do that. You know the in-distribution data is good (but lacks variety), but you don't have an idea about the out-of-distribution data, so someone needs to sort it.

And someone could argue you're distilling their model. Or are only the chinese doing that?

The thing is, anything that can be automated is automated cutting the middle man. If they are paid, they are paid because they are adding some value AI cannot add. If they instead use AI, they don't do their job.

Stop citing 404 media. They are either not understanding it or do not want to understand it (they seem to hate AI).

If you need to train an AI for some task, because current AI cannot do the task as good as you hope your future AI will, you cannot train on the output of the current AI. If you could, they would automate it and wouldn't need humans at all.

If the current AI is limited, there is no simple way to use it to create a AI that's better. Again the argument is simple to see: If you could use it to create decisions for a better AI, you could also use these decisions directly without training another AI.

So humans are paid to train AI, because they can do tasks current AI can't. If they cheat, they don't do their job. And if you don't do your job, you get fired.

The EU *would* care, but it will take five years for any action of them to take effect. Google will pay a fine that's no problem for them and in these five years they already got all ex-F-Droid users to says "Fuck, using F-Droid is too complicated now, let's go back to play".

If you get F-Droid from Play, you cannot upgrade it using F-Droid later (different signature keys). So this means you rely on Google for upgrading your non-google store.

I'd say it is more customizable than ever.

You have to think of AI as a (mathematical) function. Same input, same output. You have some random seed, but that picks from a distribution that is still determined from your input. If your input is too simple, you sample near a point where everyone samples, so the variation between the works of different people is small (which they don't know without comparing their works to others). And for the current generation of ChatGPT, the point where you land with unspecific prompts generates images I do not like, but that may be a personal taste and others wanted this (not unlikely as post-training involves human feedback).

I think you see the "curating and distributing" thing good with scientific journals. They charge money while the authors and editors work for free. Why does anyone participate in this predatory model? Back in the days people got their new scientific articles by subscribing to a real paper journal and the journal did real work for science communication.

Today everyone can upload a PDF and a good community could curate what's crap and what's insightful. Scientific journals are only not dying out of tradition. If it's published in a high-ranked journal it must be good, so we must publish in the journal ... while the publisher doesn't do much work anymore. Printing is automated, hosting is near free compared to the prices they charge for articles, but the science community still believes in their reputation.

The overall problem you see at many places is filtering. When people talk about all the "slop" on their art site, the problem is not someone posting low-quality works or needing storage for it, but that their feed isn't filtered to include only high-quality works (or works they like for other reasons). The same applies for books, scientific articles, and many other works.

Having one authority, like with scientific journals, is one option for filtering. Having publishing being costly (like with paper books) is another one. Both are only proxies for quality and not always good proxies. Maybe AI curating your feed could be a long-term solution? I am not sure if the systems are there yet and not sure how much work it is to clearly formulate your criteria (on the other hand, just as you believe in publishers you could believe in criteria-presets curated by others), but at least these things can process large volumes of works to find the gems, given a good enough definition of what you consider a gem.

Yes and no. If the AI has a good voice for the story (let's assume common problems like slop phrases, uniform style, etc. are solved a bit more) than there is no difference between a human author's voice and an AI author's voice. Either you like it or not. And not liking it doesn't mean it is bad, but that it isn't your taste. I have authors I like and authors I dislike and know people who have different opinions on them.

In the end there are more details one can discuss, but the blanket "I do not read it when it is not human written" is something I can see when it comes to internet arguments ("I let ChatGPT refute your comment, read this wall of text") but not when it comes to creative works I read, because I like them.

For less popular works this already happened with a lot of people being angry at not noticing before. Also currently a lot of indie games are review-bombed for mentioning that they used AI somewhere in the process, no matter how minor the use was. I think that will go away, in particular with increasing quality. You can already read now "I hate AI slop, but this slop is sooo cute!" when someone posts a great AI video clip. Yeah, the people think they hate it, but when they see some high quality work, they suddenly like it.

I also agree with people hating ChatGPT graphics, at least the ones that you get without specifying details. ChatGPT always had distinct style and ugly style when you don't direct it away from that. Last year it was the "piss filter" this year it is just overly detailed. I am not sure how you can direct it well as I don't want to support commercial generators, but from what I heard you *can* ask it to use another style, but you have to invest some thought into it.