AI is going to be our biggest game changer

WendyLuvsCatz

and now for some conspiracy 

generalgameplaying

Griffin Avid said:

I found this video very interesting.

Indeed interesting. Woudn't subscribe to all the statements, though due to not sleeping the night, i might actually have forgotten all relevant points, as well as turned stuff into incomprehensible gibberish, while trying to edit it into what i thought what should be there at one or another moment. E.g....

1. While there may be no reason to believe that "ai" stops there quality-wise, there also is no too reliable way, to know how far it can go fom here, except perhaps for tapping into Google's internals. E.g. Better:
   1. Better by more data? Uncertain.
   2. Better by better data? Improving the tagging of all images... likely. To what extent?
   3. Better by more neurons (and similar)? Maybe / likely / until where? Cost control...
   4. Better by better system architecture? Can't really know for the general text prompt thing.
   5. (We don't know. Assuming that would be it, would be naive too, and progression probably isn't linear, meaning: could be logarithmic from some point on. Are we there yet?)
2. Is it good to let courts proceed? I believe the idea that court would resolve the issue directly to be highly problematic. US courts may be better at judging new technology, still the classic laws tends to not grasp concepts like "the internet" or "ai", including the consequences, so we likely are in need of courts relaying it to the lawmakers.
3. Is it good to embrace the transition to no job the next day, because they are allowed to use images, violating intent of licenses in general, whatever, or via cash-grab by a social network, without explicit consent of the artists? -> Hell, no. You don't want war, you don't do those things. Scraping against consent is only needed for cheap and fast destruction - in my opinion this is far more problematic and can not be compensated by many people using the system and gaining a creativity boost. Especially because it's not only abusive in general, but also because the ai systems are in need of such data to even exist. Maybe law is taking the cheap route of not having to think about monopoly either way, so let all be (sc)raped? It won't turn out well that way either.
4. Why disrupt now, throwing people into the fire needlessly? People whose works the machine is directly based on, not like another branch of things, with similar jobs everywhere. Just because the planet needs more sugar, because it is just going down burning anyway?
5. What's the benefit of having this technology ready for commercial exploitation fast and cheap? Inspiring millions with images of artists, who will never create images again? That's exaggerated of course, but will be relevant later on. 
6. The system isn't learning like a human, taking decisions, changing course. That is exactly not what the models represent. Maybe it can serve as a model of a subsystem for explaining some of "dreaming" or rather other sub-functions of (human) brains, like for visual data processing, just in fewer dimensions.
7. What to feed it on the long run? The outlook for getting new training data isn't trivial. No artistry, no new training data. Well some, like generated and changed ones, and maybe steal from the movies, pay some people, as mentioned in the video... but do the information theory on the magnitude, and then read on. Then generated images will "leak" as training data, on the long run, and guess which kind of images will be most. Do the information theory. Courts considering? Do the information theory. Lawmaking? Do the information theory.
8. It'll always/eventually violate something, if you scrape copyrighted works without having a license to do so. But it's also cheating because it scales to nowhere in the number of users, while using the vast amounts of data - in the amortized view, it's no different to troll-flooding the internet with similar artwork to what artists did previously. So whoever will profit, will have profited from millions of minions doing the flooding work in the modern social network manner, with the users even paying for doing other people's work. The misconception about law seems to be, that the law had been lobbied to use cases like "big data for shops" or "image recognition", where you could assume, the training data would never surface. Now adding filters to let it look slightly different, should not count, that would be very cheap. The feedback property, paired with the scalability to millionfold use within short time, constitutes the evil side of it.
9. The consequences could be vast, even with a sound, fair and legal system on the mid to long run anyway, because it still will flood the place with images, resulting in similar effects to copyright. Perhaps an image-alteration based on copyrighted work, paired with generative ai will do the trick. Why would this justify abusing people now? Are we ok with NAZI-Germany invading Poland because the Russians will plow over it later anyway? We'll have a super nova eventually, so if you can't fly away right now, why exist at all? The question isn't asked by society, but by investors who want a quick run on the market during potential crisis.
10. This is not a better hammer, this is a painful feedback loop. Permanent abuse guaranteed ;). So why not let it be paid cloud services at the hands of a few investors, yeah!
11. Oh the flipside potential as mentioned before: Commercial, TOS-based and governmental (laws...) filtering already in the process of creation? Do people think, they can run the larger commercial models on their laptops or something like that?
12. With people training their own smaller models, and succeeding to some point, e.g. in copying artists, shows us the positive potential. That seems reversed, but that's the positive bit. With thinkable improvements, you could have your own deam-machines, with open to train on data sets. Why abuse now?
13. (Another crucial point: Arguing that external forces will bring in such ai, overlooks two points: a) your legislation tells what a legal service is b) you can not defend against an external image generator ai by having a better image generator ai of your own! People will see the carriage from underneath, the next moment. So instead you will regulate them and thus all, or you will keep them out in a protectionist manner, with all consequences for international relations on the side of whatever is affected then.)

 

On the other side: humans are language models with a digestion attached to. And some basic survival needs.

TheMysteryIsThePoint

bluejaunte said:

generalgameplaying said:

bluejaunte said:

At the moment it's simply a more interactive Google (when it comes to pure information finding at least). When you google something, it is up to you to check if the information you got is from legit sources. 

[...]

But it's so much more. The level of understanding it has about topics, the way it can speak to you and explain and circle back and remember past bits of a conversation is unreal. It can write a poem about the last couple of topics we talked about. It understands very easily what I mean, even if my question is crappy as hell. It generates code in lots of languages, 

 [...]

It writes stories,

[...]

Who needs IQ and intelligence when it can do this by just being a "dumb" database of human knowledge?

Interactive Google: Yes, and that could be a merit. But when it comes down to checking for stuff, the task even Google denies to carry out, in order to show more ads, is to efficiently link to the encyclopedias. Well, it now actually does often display links to encyclopedias and facts on the top, directly showing a description. But the real difference is not "efficiently", it is that google actually links to the encyclopedias, in the good case. So with simple additions, like confining search to science and encyclopedia sites with a checkbox, google would be magnitudes better for that than it's now. An explanation .. you'd again be at a search engine. The last thing i want to see, when i search for something is a wordy description, for most of the cases. But i do see much potential of such models providing wording to make things colloquially understandable. And the component doing the understanding actually seems to be an extra thing, that ChatGPT uses to understand people, so i could imagine such being used for one type of search engine. The problem is, that the system likely has no idea when it's walking on a thin thread, when it hits a topic that it can't handle well. There is ways to give the system an idea of uncertainty, concerning training data, but it's not clear how far that will lead in this instance. And the human typing the search terms, probably has no idea of the reliability of an answer either. In the end humans and writers for magazines make errors too, though upon notice an error might be corrected (or not) and then stand there corrected (or not)  forever. With the language model, you can add corrections to training, but with any more significant changes it might not be clear, in terms of being able to ensure, that it will be more correct on all corrected topics the next time, or if previously well done topics will still be handled correctly, especially if you don't check the answers after re-training (because you need an expert for that, it can't be done automatically). Guarantees are much less easy to give than with more specialized (classic) search engines (which no one really seems to implement, possibly due to business models). Since the similar players like Google, who busy about keeping the users in their realms, e.g. for showing ads, are also heavily invested and tecnhnically able in the field of ai, so i ask: what to expect more from them?

 

More: Yes, but "it doesn't tell jokes" and such. "It" is language models, not magic. It's more like patterns applied in a language-like fashion. There is abstraction contained, which to some degree means understanding and context (somehow). But it's not so deep. It's just as deep as the training data goes (and reinforcement learning), and there could be articles about just anything, so from somewhere to somewhere, often some answer exists at least in the training data. To really make it even half as bulletproof as an encyclopedia, they would need to fact-check the output for all corrected answers, which probably is out of scope right now. BUT they can and likely do feed it encyclopedias. Just i am not sure if they will manage to have something like specialized modules and at the same time select the correct one, meaning if they really hit the context well enough, and not mix in something from elsewhere at random. Combination of techniques will lead further, and i am not sure that language models alone lead there (and at which size). All those examples are derived from languages, based on the data we/they have, plus reinforcement learning with people tasked for it. The model currently is just very big, and obviously very good, in terms of comparing it to previous models. That's still all a similar quality in general - the question is, how far it can lead (and how so). You probably can't expect upcoming improvements to display the same magnitude of change, as ChatGPT represents now, compared to previous versions. Presenting step-by-step knowledge, even if all sources of the world have been used correctly in training, doesn't guarantee putting answers together correctly, especially not when understanding is necessary. The biggest potential for market in my opinion still is manipulation (*), until outlawed. Think of ads worded for you personally, links generated with texts for you personally, wording of links created in real-time allows much better tracking and personality testing - hell on earth. From hell to hell, though.

 

Stories: That's great, but it's still not actual intelligence ("... still only counts as one!"). I actually might have great use for "it", thinking of rephrasing what happened in a computer game scene or an arena battle, based on a series of simple descriptive terms on what happend, plus similarly a description of the hero character and their past efforts. Just i wouldn't like a cloud version of this, which it is, as it's "the biggest neural network" someone ever built. So for me, i'll need a boiled down version, if i went there. In theory, it could use special accelerators for ai, e.g. if it ran on much less power consuming and less accurate ones made for neural networks. So that'd be fictional stories. Maybe for (fantasy) storytelling a smaller thing can exist some day.

 

IQ and intelligence:  General intelligence is the most important part, still. A killer scenario means massively relying on an ai like ChatGPT in a similar state as is now, and actually losing the abilities - the dream of the elites of the past, to rule over dumb people once more, controlling people with a switch, or if necessary: the power plug. That aside, it's not an intelligence, and it won't grow there by itself. As far as i can judge, there will be a new generation of the system some day, training the same plus new data. But it is not evolving on it's own or something like that.

 

-> Such systems CAN remove hindrances and help do stuff, of course. We're not at the magic bullet with the current display of ability, given the description of the system isn't decoy fiction. Given the kind of mistakes it makes, and what it appears not to be able to do, suggests, that it's essentially still language models done well, but won't understand to depth. So the training data question will apply here, similar to image recognition. But since the system is more complex, it'd be interesting how precisely we/they actually can tell, if it will be improved by a certain change, or the bulk of adjusted training data. But that's probably also the good side, because a general ai would probably learn and do random stuff on purpose, just to unnerve people. I also believe that further improvements are possible, but it's not so clear-cut, that it wil do so much better than replacing jobs, just to leave societies with less than before, because people lose abilities, the jobs are gone, but the system can not train on new data generated by the very people it just has taken the jobs from. Of course there will be other data sources, concerning "the job", but it may still become a real problem. (*) These kind of scenarios also could be counteed as the result of manipulation. That's different systems maybe, though for programming we may see some effects at some point, and it's obviously close to language models (but not only that). With ChatGPT the manipulation potential seems vast, because it appears to be so [...], whatever, while it actually is "so language". Language isn't the worst thing to be, per se, though...

I'm beginning to think that whether something is intelligent or not is kind of irrelevant. You keep pointing out it's not intelligent, but I can talk to it. It helps me, it educates me and has more knowledge than any one human being on this earth. Case in point: if it tells a joke, does it matter to me if it's intelligent or not? Is every human who tells a joke intelligent? Of course not. I would argue it probably tells more intelligent jokes that are about very specific topics or even a combination of topics than the average human could come up with.

How much of our intelligence is just data? We all know that knowledge and IQ are not the same things, yet how we perceive intelligence often comes down to how much someone knows. A lot of education is just knowledge. Pure data. In some fields, you do need actual intelligence. But think about how many jobs out there revolve around amassed knowledge, communication skills, and maybe some basic math that is also data. We memorize that 2 + 2 is 4 and 10 x 10 is 100. That level of math is what 99% of people on this earth get by with.

So what about GPT is not intelligent then? It can come up with a story based on my parameters. It may not be a very good story, but it does have the ability. Most humans don't have the ability to create a good story either, let alone write it. GPT can store all the human knowledge. It can do math better than any human. And this is just a computer thing, not really even a feature of GPT. It can play better Chess than any human, at least if it ever gets proper Chess engines integrated that far outperform the very top human chess players for basically decades now.

In other words, what is the difference between "telling a joke" through magic, as you put it, or through a language model? If there's no difference in the end result, there is no difference. Technical differences are irrelevant, and in any case, you could argue that we very much have a language model in our brain too. if I tell you to speak Swahili you'll probably say you don't know that language. You'd have to learn it first. Your human intelligence alone does not allow you to speak Swahili. In fact, I would say that learning a language doesn't have much to do with intelligence and everything with memorization.

So it writes a story, it's not real intelligence. Is a human writing a story real intelligence? Isn't it rather a collection of data, memories, experiences, and some creativity to put it all together into a story? All of this GPT already does and it will get a lot better. To the point where it will spit out a whole 300-page novel, nicely formatted, and 100% correct spelling. Will it be good? Who knows. It might, if the AI learned from all the human books out there. It will "read a lot" like any writer and then create a story based on your parameters. What is the difference to a human?

I've been silently cheering @bluejaunte on, on many points. But on two in particular:

1) It took us 60 years to get from Eliza to ChatGPT. But a few weeks from ChatGPT not being able to remember what you said in the prompt before last, to it remembering everything you've ever said. It will get better. Not incrementally, but at an accelerating rate.

2) You can devote your time searching for arbitrary and insignificant was in which it is technically still insufficient, or you can use it in ways that improve your own performance in endeavors that matter to you, right now. Is it intelligent, or not? Who cares if it just ported 500 lines of Sagan's source code to Python 3, a language I am not at all good at, in about 3 minutes? If it just pointed out a severe plot hole in my story and came up with ways to address it that I am not sure I would have thought of, given an infinite amount of time?

We are past the point where it is more interesting to use AI in practical situations than it is to debate its objective technical properties.

 

TheMysteryIsThePoint

nonesuch00 said:

I was being silly, or so I thought, and purposely chose {} format preferences to query on. I asked it what was the best way to format {} in programming. I think the answer it gave just citing database entries of (associated with UC Berkeley no doubt) {} format styles and the celebrity programmers that claim to have originated those particular {} formatting styles wasn't problem solving at all, but just a database look up. It did great though understanding my English query and converting it to a database query to fetch that data.

A problem solving AI would have considered the way the various text editors work, whether tabs were hard tabs or soft tabs when inserted by the text editor and such, studies on making blocks of text easily readable to humans, human typing skills, reading skills, vision properties, and keyboards, and proposed a solution using those available facts. I've never known a programmer look up what celebrity programmers from UC Berkley use so they could use that style as their personal {} formatting style. Come on, man!  They use what the editor defaults to or what the existing code is already using, otherwise for new code they may use what the editor defaults too or they my have devised there own {} formatting style. Look up what some programmer from UC Berkeley does and copy that style? HaHa! It's a good joke.

It just doesn't have those facts in it's database as to why those {} formatting styles with noted celebrity UC Berkeky programmers were chosen as the answers so it can't figure out the technical why and so just recites the final answer that was given to it at some point or another in the past.   

Am I wrong to think it should have the capability to problem solve in a way that I was led to believe this AI could rather then do a dB lookup? I guess so.

The professor got better results, about 75%. Take note his questions were well known physics theorems with the logical mathematical relations stated in those theorems. So the AI is demonstrating it can convert English queries to typical database lookups and math tables and equations look ups and application of those quite well. 

I've not tried querying in another language either because my language skills outside of English are barely intermediate level at best and only in the best of fortuitous circumstances. 

Others have complained that these AI engines are too weighted to be dependent on "subject matter expert celebrities" database look ups and not actually looking at the problem from a naive perspective and trying to solve it independently. Would it get the answer right or wrong? I am going to agree with them (the complainers) for the most part.

Have you considered that it made a decision about the context in which you wanted the answer? I would not be surprised if it knew about all the issues you mentioned, individually, but came to the conclusion that that was probably not what you were interested in. I've encountered many instances when I was not satisfied, for whatever reason, with a particular response and found that it had a whole lot more to say after I asked it to expand on a certain aspect.

And I appreciate that it is natural to use things we do understand to explain things that we don't, but it's sort of a disservice to keep saying "database lookups". That is not how neural nets encode nor retrieve information at all, and helps to complicate and misguide the coming legal debate on IP.

 

generalgameplaying

TheMysteryIsThePoint said:

I've been silently cheering @bluejaunte on, on many points. But on two in particular:

1) It took us 60 years to get from Eliza to ChatGPT. But a few weeks from ChatGPT not being able to remember what you said in the prompt before last, to it remembering everything you've ever said. It will get better. Not incrementally, but at an accelerating rate.

2) You can devote your time searching for arbitrary and insignificant was in which it is technically still insufficient, or you can use it in ways that improve your own performance in endeavors that matter to you, right now. Is it intelligent, or not? Who cares if it just ported 500 lines of Sagan's source code to Python 3, a language I am not at all good at, in about 3 minutes? If it just pointed out a severe plot hole in my story and came up with ways to address it that I am not sure I would have thought of, given an infinite amount of time?

We are past the point where it is more interesting to use AI in practical situations than it is to debate its objective technical properties.

 

1) There is no "it". Hardware and some theory have scaled ridiculously, with them using a sophisticated modeling alongside. That's it. Could be another 60 years from now. Well, i'd not assume so, but as there are limits to data, and to real brains, you won't get arbitrary improvements "for free". And ChatGPT "remembering" is / would-be just an add-on to the prompt +- 100%, it's not an advancement like from Eliza to ChatGPT. Still there can be gradual improvements from day to day after re-training, as they keep updating the models.

2) The hand of god has decided to... have us all plod along? If this is intended as a civilization-based show, things might want to be different. There also is potential for fundamental dealbrakers ahead, which i don't know of, if theory has removed them yet. May be lack of knowledge on my side, could ask ChatGPT for a change. The players lobbying for free scraping, probably don't care for more than the quick $, whatever to follow. Arguing with the past for a thing that is supposed to be future in a fundamental way, because the experience of the past doesn't work for judging it... judge for yourself, how that will not lead to repeating the past in other ways. So using the ai now? Yes, of course! People should just be ready to understand the downsides of what humans actually do, including those humans who run the scraping and the models. Would you take responsability for a commercial application, putting in that code without fully understanding it? [Hint 1: Isn't there an active lawsuit in a similar case?,  Hint 2: Quality control..., Hint 3: Just checking "the function" /once might not be enough for a satisfying result.]

 

Let's be realistic - using generators commercially still should mean taking a high risk, either way (image/text). So "you" or "me" can't really go there "improving", of course you can train, and be a couple of days ahead of other people, if it's freed of any charges/odds/ends for some reason. Using a commercial cloud service is no option for many projects, neither would be an expensive on-premise version. In fact for what i am planning, this has little use alltogether. Neither for programming - you typically learn in small-ish steps for good reasons, to really understand the bits you are using. Some people learn faster, and maybe some people are good at cross-checking examples from ChatGPT with a documentation - this likely isn't the same for all humans. With an on-premise thing, better an open-data open-source one, i'd rather use the text generation for dynamic storytelling. Pricey means maybe later, if i have a game - i'd probably design for use of some ai, maybe i'll "rite my own". Such can be done in more classic ways too, though, but the syntactic and stylistic sugar would be very welcome to make up a story of what happend on-screen (or probably even more so: off-screen). Since those new things don't work for me currently, i've got all day to make up random show-stoppers for the amateur players who run this "civilization".

By no means feel discouraged in using those generators. It's not my intention to fiddle with the fiddling. I just don't see those blue skies...

generalgameplaying

TheMysteryIsThePoint said:

Have you considered that it made a decision about the context in which you wanted the answer? I would not be surprised if it knew about all the issues you mentioned, individually, but came to the conclusion that that was probably not what you were interested in. I've encountered many instances when I was not satisfied, for whatever reason, with a particular response and found that it had a whole lot more to say after I asked it to expand on a certain aspect.

And I appreciate that it is natural to use things we do understand to explain things that we don't, but it's sort of a disservice to keep saying "database lookups". That is not how neural nets encode nor retrieve information at all, and helps to complicate and misguide the coming legal debate on IP.

 

I don't want to nit-pick on everything, but you say in one instance "it takes decisions" and in the other instance "it's working like a neural network (and not like database lookups)". This could be seen as slightly contradictory, because neural networks don't take decisions in such ways, because that's not in their nature ;).

 

In fact it's comprised of multiple stages and components. At least there is the big language model/s with one component being the prompt interpretation and one component being pretty much the rest of ChatGPT. Very freely reworded. OpenAI has an explanation page on their (rough) system design.

bluejaunte

nonesuch00 said:

bluejaunte said:

nonesuch00 said:

WendyLuvsCatz said:

ai art serves as the distraction 

ChatGPT is the thing everyone should be worried about

 

generalgameplaying you posted between my edit

I've watched a couple of YouTube videos too by a couple of people that ran into the same sort of problem, that of the AI getting simple problems badly wrong. One was a college professor that was feeding his froshman level class exam to this ChatGPT in order to help the professor evaluate if his students used that AI. The exam is "open book".

I also did a chatGPT query regarding the use of {} by programmers and was equally as disappointed as the college professor.  It acts more like a massive database than an AI.

I think it will improve but for now the chatGPT is really as flawed as much of the art generated by AI.

As a database? I love it. 

What was the question regarding {}?

Even in its current iteration I would hesitate to call it flawed. Simply because it already knows more than any one human being on this planet. It's an encyclopedia that can talk. Enjoy it while it still gets some stuff wrong. It won't in the future.

I was being silly, or so I thought, and purposely chose {} format preferences to query on. I asked it what was the best way to format {} in programming. I think the answer it gave just citing database entries of (associated with UC Berkeley no doubt) {} format styles and the celebrity programmers that claim to have originated those particular {} formatting styles wasn't problem solving at all, but just a database look up. It did great though understanding my English query and converting it to a database query to fetch that data.

A problem solving AI would have considered the way the various text editors work, whether tabs were hard tabs or soft tabs when inserted by the text editor and such, studies on making blocks of text easily readable to humans, human typing skills, reading skills, vision properties, and keyboards, and proposed a solution using those available facts. I've never known a programmer look up what celebrity programmers from UC Berkley use so they could use that style as their personal {} formatting style. Come on, man!  They use what the editor defaults to or what the existing code is already using, otherwise for new code they may use what the editor defaults too or they my have devised there own {} formatting style. Look up what some programmer from UC Berkeley does and copy that style? HaHa! It's a good joke.

It just doesn't have those facts in it's database as to why those {} formatting styles with noted celebrity UC Berkeky programmers were chosen as the answers so it can't figure out the technical why and so just recites the final answer that was given to it at some point or another in the past.   

Am I wrong to think it should have the capability to problem solve in a way that I was led to believe this AI could rather then do a dB lookup? I guess so.

The professor got better results, about 75%. Take note his questions were well known physics theorems with the logical mathematical relations stated in those theorems. So the AI is demonstrating it can convert English queries to typical database lookups and math tables and equations look ups and application of those quite well. 

I've not tried querying in another language either because my language skills outside of English are barely intermediate level at best and only in the best of fortuitous circumstances. 

Others have complained that these AI engines are too weighted to be dependent on "subject matter expert celebrities" database look ups and not actually looking at the problem from a naive perspective and trying to solve it independently. Would it get the answer right or wrong? I am going to agree with them (the complainers) for the most part.

 

I just asked it that question (what is the best way to format {} in programming) and the answer seemed fine to me.

1. Allman style: Braces are placed on a new line and indented at the same level as the surrounding code. This style is used in languages such as C and C++.
2. K&R style: Braces are placed on the same line as the surrounding code and indented at the same level as the surrounding code. This style is used in languages such as C and C++.
3. 1TBS (One True Brace Style): Braces are placed on a new line and indented at a higher level than the surrounding code. This style is used in languages such as Python, Java, and C#.

For each it gave a code example that also looked fine. I can google all these styles and find them by name. It's all here in the Wiki really: https://en.wikipedia.org/wiki/Indentation_style

It added:

Ultimately, the choice of formatting style depends on the project and the team working on it. Some projects may have a predefined style guide that all members must follow, while others may allow team members to choose their preferred style.

So I don't really know what you were expecting. Did you want it to tell you why these styles are readable on a more visual/perception/brain kind of level?

Never mind, you already explained this above. But you know, why not keep asking? If you asked a human that question, it's quite possible you wouldn't get the answer you wanted right away either. Have a conversation with it. If you wanna know more, dig deeper. Ask it about studies. About soft vs hard tabs, editors etc.

I asked it about studies:

There are several studies that have been conducted on the topic of code formatting and its impact on readability, understandability, and maintainability of code. One study conducted by researchers at the University of Gdansk found that code readability is improved by using consistent indentation and naming conventions, as well as by breaking up code into smaller, more manageable chunks. Another study published in the Journal of Systems and Software found that code readability is also improved by using whitespace and blank lines to separate code blocks and by aligning code elements consistently. The use of code formatting tools has also been found to be beneficial in improving the readability of code. The study also found that developers preferred using a consistent code formatting style, as it made it easier to understand and maintain the code.

I asked to cite sources:

There have been several studies on the readability and understandability of code formatting. Some examples include:

"Improving the readability of program source code" by Thomas J. Ostrand and Elaine J. Weyuker (ACM Transactions on Software Engineering and Methodology, 2002)
"The effects of code layout on program comprehension" by Michael J. Godfrey and David Binkley (Empirical Studies of Programmers: Third Workshop, 1991)
"The impact of code layout on software maintainability" by Lionel C. Briand, Jacquelyn S. Moore, and Yvan Labiche (IEEE Transactions on Software Engineering, 1999)
"An empirical study of the relationship between program structure and program comprehension" by J.M. Carroll and J.C. Thomas (International Journal of Man-Machine Studies, 1979)
It's worth noting that these studies often have different conclusions depending on the languages and the sample code used. Also, it's important to consider that code readability is a subjective concept and what might be readable for one person might not be for another.

PixelSploiting

The IP shenanigans the AI generated images are facing now are exactly the same human artists are facing every day unless they are doing it purely as a non-profit hobby. Everything is property, graphical assets are property too. This isn't caused by people being tech-hating luddites.

If a human artist has to be careful to not run into this kind of problems, why wouldn't be an image generator affected by this? Note that a machine cannot claim artistic inspiration or interpretation.

If an image is not royalty-free, it always is risky to use it in any fashion because it belongs to someone. This someone might be considering it enough of a problem to hire a lawyer.

 

There's nothing new nor unexpected here, really.

generalgameplaying

PixelSploiting said:

The IP shenanigans the AI generated images are facing now are exactly the same human artists are facing every day unless they are doing it purely as a non-profit hobby. Everything is property, graphical assets are property too. This isn't caused by people being tech-hating luddites.

If a human artist has to be careful to not run into this kind of problems, why wouldn't be an image generator affected by this? Note that a machine cannot claim artistic inspiration or interpretation.

If an image is not royalty-free, it always is risky to use it in any fashion because it belongs to someone. This someone might be considering it enough of a problem to hire a lawyer.

 

There's nothing new nor unexpected here, really.

Absolutely, that's the minimal point to make. The "button" doesn't relieve me of basic duties.

The issue is, that i don't create the actual image consciously and there is no way to know, what it was based on (at present), effectively making it more difficult to judge.

(In a way, following a style of artwork for a while, might have made me less susceptible to violating anything, once i got it right. No guarantee either, though. Should i, as an artist, hypothetically be forced to change style every day, because everything gets copied too quickly, then i am in fact more susceptible to making mistakes than before. So the new curse could become "to be known".)

FirstBastion

Bottom line it is not the AI that committed the crime of copyright infringement,  it was the very human  owners of the companies that made the request and the very human programmers that programmed the computers to steal and scrape the copyrighted material off the internet from the legal owner of those Intellectual properties. 

MelissaGT

FirstBastion said:

Bottom line it is not the AI that committed the crime of copyright infringement,  it was the very human  owners of the companies that made the request and the very human programmers that programmed the computers to steal and scrape the copyrighted material off the internet from the legal owner of those Intellectual properties. 

I love the argument of "how are we supposed to track down the owners of the millions of images we used in our dataset?"

Um...then perhaps don't use them.  

FirstBastion

MelissaGT said:

FirstBastion said:

Bottom line it is not the AI that committed the crime of copyright infringement,  it was the very human  owners of the companies that made the request and the very human programmers that programmed the computers to steal and scrape the copyrighted material off the internet from the legal owner of those Intellectual properties. 

I love the argument of "how are we supposed to track down the owners of the millions of images we used in our dataset?"

Um...then perhaps don't use them.  

 It seems they don't have the capacity to understand it was not "their" dataset to begin with.  Ignorance to the LAW is not a defence,  never has been.

PixelSploiting

I bet if the AI was trained only using purchased references or even only hiring artists for this, all of this would be a dismissible annoyance. But it wasn't.

 

I can google anything Matt Rhodes and put it on my desktop, and that's about it.

 

I'm almost sure Adobe neural filters were only made using stock illustrations owned by Adobe. I doubt Nvidia was making this kind of mistake too.

 

Edit: In theory only public domain images could be used too, but classic senior art wasn't what they were going for.

bluejaunte

I'm not sure it's quite as clear though. One could make an argument that any human or machine could learn from publicly available images on the internet. This isn't a copyright infringement yet, is it? It's just learning from others who willingly put their images on the web. So the real question is this: is what these AI generators do more copying or learning? Society and courts will have to decide I guess. Creators of those AIs I'm sure feel like this is public information that is freely available and can be used to learn from by humans and machines alike.

PixelSploiting

Human can claim inspiration. Here, you feed an image with possible IP rights attached to a computer. A machine is not a person with any rights. A company using said machine does need the rights to use an asset, though.

 

The worst case scenario, it's going to be considered not any different than photoshopping using said image.

The best case scenario, the courts might handwave the issue only to get rid of it by deciding that it's under no IP rights at all.

 

Either way you have to deal with an owner of an original image being perfectly in the right of stating that they simply do not allow this kind of use. It's their property so they can.

They could even state this if another human was using it as an inspiration, but let's say they didn't feel like it before. Because IP rights are very much pursued on demand unless they are trademarks, it's literally betting on someone else will to pursue it or not.

 

One way or another, it's a bit too open for lawsuits.

 

It'd be safer to just purchase it before feeding it to the machine.

FirstBastion

bluejaunte said:

I'm not sure it's quite as clear though. One could make an argument that any human or machine could learn from publicly available images on the internet. This isn't a copyright infringement yet, is it? It's just learning from others who willingly put their images on the web. So the real question is this: is what these AI generators do more copying or learning? Society and courts will have to decide I guess. Creators of those AIs I'm sure feel like this is public information that is freely available and can be used to learn from by humans and machines alike.

 No. It is clear from a legal standpoint,  versus opinion.

There is clear evidence that a majority of the AI outputs tried to obscure clearly marked WATERMARKS and COPYRIGHT NOTICES on the source material artist's works. The evidence is there for anyone with an objective understanding of the law and copyright protection. 

bluejaunte

FirstBastion said:

bluejaunte said:

I'm not sure it's quite as clear though. One could make an argument that any human or machine could learn from publicly available images on the internet. This isn't a copyright infringement yet, is it? It's just learning from others who willingly put their images on the web. So the real question is this: is what these AI generators do more copying or learning? Society and courts will have to decide I guess. Creators of those AIs I'm sure feel like this is public information that is freely available and can be used to learn from by humans and machines alike.

 No. It is clear from a legal standpoint,  versus opinion.

There is clear evidence that a majority of the AI outputs tried to obscure clearly marked WATERMARKS and COPYRIGHT NOTICES on the source material artist's works. The evidence is there for anyone with an objective understanding of the law and copyright protection. 

Yeah, in that case that would definitely point more toward copying than learning. Unless maybe you make an argument that the AI simply didn't know that a watermark is not part of the art style. If I'm giving a child watermarked images without explanation then it would assume that this is how you draw. But then it would be the parent's duty to teach better, as it may be the creator's duty to teach the AI better.

I don't know, I just find it so interesting. All the questions AI creates for society. So many ethical and philosophical questions, far beyond just the boring legal stuff.

Griffin Avid

I think there's a lot of cross-talk over this.

I think the video I posted earlier addressed a lot of fallacies.

Copyright is about owning a thing.

A License is about how you can use the thing.

A Trademark is branding.

A signature on a piece of art is not a copyright.

It's not clear if anyone "needs your permission" to scrape your work(s) and build their datasets.

In other words, this has not yet been decided. It is still in the court of public opinion.

Infringement is the word that no one wants to use because that goes on a case-by-case basis.

And it's also VERY hard to prove.

AND for that, you need to go after the person, yes, THE PERSON that is doing the infringing and NOT the company that made the tool(s) they used.

That's the real elephant in the room.

-------------------

And I checked the data set using the link from about two pages ago.

Yes, my comics are in there.

And yes, my friend made something that looks like a horrible version of my work and sent it to me as a joke.

----

I still can't process all of this and have no idea how to feel.

So I'll the courts decide and roll with that. lol

 

PixelSploiting

But if you can't prove that a person putting it into a database had right for doing this then you might end having an infringement. Because it's going to be hard to tell if it even falls under educational uses... Might as well end boiling down to the owner saying whether they agree or not to this use.

 

I think we're going to see more of the art hosting sites trying to stay on the safe side of things by making it easy to state if images can or can not be used in AI generators.

 

No one in their sane mind wants to tackle with lawyers if it can be avoided.

bluejaunte

Aren't search engines essentially putting images into databases too? Nobody cares about that. And what if the AI would instead browse the web in real-time while learning, without putting anything into a separate database?

Is there even a law against saving an image somewhere? I thought it was only the usage of that image. Nobody cares if I save a copy of it to my computer to look at it later. Or am I wrong?

IceCrMn

Is there an AI version of limewire yet?

There's a few very expensive pieces of software I'd like to have for "training purposes".

I promise I'll never make a single cent from their use.

 

...anyone else see the similarities?

 

 

MelissaGT

bluejaunte said:

Aren't search engines essentially putting images into databases too? Nobody cares about that. And what if the AI would instead browse the web in real-time while learning, without putting anything into a separate database?

Is there even a law against saving an image somewhere? I thought it was only the usage of that image. Nobody cares if I save a copy of it to my computer to look at it later. Or am I wrong?

It's not about people looking at the pictures or saving the pictures to their computers to look at later. It's about people profiting from other people's copyrighted works without any permission or compensation given back to the original artist. 

Think of it this way. Midjourney uses an AI dataset that was scraped from hundreds of thousands of copyrighted works. They turn around and charge a subscription fee for people to use their service...that was trained off of copyrighted work without any compensation back to the artists who essentially created that dataset. Even further...there are people using Midjourney and other tools (Stable Diffusion et al) to spit out pieces...and then turn around and re-sell said pieces...thus adding an additional layer of profit off of copyrighted works. 

If someone wants to train their own dataset on their own work and then use it to remix or create new works to sell...totally fine by me (minus the environmental impacts if everybody were to all of a sudden start training their own AI's). But this is such a small % of use cases. 

bluejaunte

MelissaGT said:

bluejaunte said:

Aren't search engines essentially putting images into databases too? Nobody cares about that. And what if the AI would instead browse the web in real-time while learning, without putting anything into a separate database?

Is there even a law against saving an image somewhere? I thought it was only the usage of that image. Nobody cares if I save a copy of it to my computer to look at it later. Or am I wrong?

It's not about people looking at the pictures or saving the pictures to their computers to look at later. It's about people profiting from other people's copyrighted works without any permission or compensation given back to the original artist. 

Think of it this way. Midjourney uses an AI dataset that was scraped from hundreds of thousands of copyrighted works. They turn around and charge a subscription fee for people to use their service...that was trained off of copyrighted work without any compensation back to the artists who essentially created that dataset. Even further...there are people using Midjourney and other tools (Stable Diffusion et al) to spit out pieces...and then turn around and re-sell said pieces...thus adding an additional layer of profit off of copyrighted works. 

If someone wants to train their own dataset on their own work and then use it to remix or create new works to sell...totally fine by me (minus the environmental impacts if everybody were to all of a sudden start training their own AI's). But this is such a small % of use cases. 

Well, yes but it's a fine line and brings us back to square one. Did the AI not simply use publicly available images to learn? Isn't it what humans do too with images that they see and copy in style? It may seem obviously wrong but it's an unprecedented situation that will have to be solved in court. I can certainly understand the concerns of artists, I mean, I'm an artist myself. But I can also find arguments for the other side. This isn't just a case of "you're profiting off of my stolen images", it's also "I put these images on the web for everyone to see, I can not prevent someone from learning how to draw in the same fashion". There is no copyright on the looking-at, analyzing and learning part. There may be an issue with the end result if it looks too much like some other image, but then that would have to be solved on a case-by-case basis. And with the sheer number of images in circulation, it'll probably be a nightmare. Like how many different ways are there to draw the Eiffel Tower.

WendyLuvsCatz

MelissaGT said:

bluejaunte said:

Aren't search engines essentially putting images into databases too? Nobody cares about that. And what if the AI would instead browse the web in real-time while learning, without putting anything into a separate database?

Is there even a law against saving an image somewhere? I thought it was only the usage of that image. Nobody cares if I save a copy of it to my computer to look at it later. Or am I wrong?

It's not about people looking at the pictures or saving the pictures to their computers to look at later. It's about people profiting from other people's copyrighted works without any permission or compensation given back to the original artist. 

Think of it this way. Midjourney uses an AI dataset that was scraped from hundreds of thousands of copyrighted works. They turn around and charge a subscription fee for people to use their service...that was trained off of copyrighted work without any compensation back to the artists who essentially created that dataset. Even further...there are people using Midjourney and other tools (Stable Diffusion et al) to spit out pieces...and then turn around and re-sell said pieces...thus adding an additional layer of profit off of copyrighted works. 

If someone wants to train their own dataset on their own work and then use it to remix or create new works to sell...totally fine by me (minus the environmental impacts if everybody were to all of a sudden start training their own AI's). But this is such a small % of use cases. 

well I mostly use my own renders and photographs, videos as inputs for im2img,

the ai I use as a filter to apply a style which is never one artist if I use any, if I do it's usually 5 or more

I run Stable Diffusion on my own machine and is much faster and uses less power than a DAZ studio render

I think I can safely say anything I do with ai is very transformative,

I don't refer to it as my art though, I don't consider myself an artist even using DAZ content, I am more a producer of entertainment, 99% of my stuff is YouTube videos 

I shot a video of me sitting on a chair an hour ago

I am now running the separated images through Stable Diffusion as  batch render

Prompt
an ugly old medieval hag with  saggy old skin Shub-Niggurath, highly detailed, jacek yerka gaston bussiere, craig mullins, j. c. leyendecker Ernst Haeckel Mark Ryden James C Christensen

size width 1088 height 1920 resize mode resize and fill
50 sampling steps
CFG scale 16
Denoising strength 0.4

 

MachineClaw

Getty Images is suing the creators of AI art tool Stable Diffusion for scraping its content. search web for article.

 

Artists are sueing  - Stability AI, Midjourney, and DeviantArt are being sued by a trio of artists who allege that the companies' AI art models violate copyright law.  search web for article sand read lawsuit.

 

While AI might be the next big thing, implementation and copyright is a thing.  reminds me of the MP3 craze and Napster. 

bluejaunte

Yup, and the courts will settle it. Then we know. Right now we don't.

MelissaGT

bluejaunte said:

MelissaGT said:

bluejaunte said:

Aren't search engines essentially putting images into databases too? Nobody cares about that. And what if the AI would instead browse the web in real-time while learning, without putting anything into a separate database?

Is there even a law against saving an image somewhere? I thought it was only the usage of that image. Nobody cares if I save a copy of it to my computer to look at it later. Or am I wrong?

It's not about people looking at the pictures or saving the pictures to their computers to look at later. It's about people profiting from other people's copyrighted works without any permission or compensation given back to the original artist. 

Think of it this way. Midjourney uses an AI dataset that was scraped from hundreds of thousands of copyrighted works. They turn around and charge a subscription fee for people to use their service...that was trained off of copyrighted work without any compensation back to the artists who essentially created that dataset. Even further...there are people using Midjourney and other tools (Stable Diffusion et al) to spit out pieces...and then turn around and re-sell said pieces...thus adding an additional layer of profit off of copyrighted works. 

If someone wants to train their own dataset on their own work and then use it to remix or create new works to sell...totally fine by me (minus the environmental impacts if everybody were to all of a sudden start training their own AI's). But this is such a small % of use cases. 

Well, yes but it's a fine line and brings us back to square one. Did the AI not simply use publicly available images to learn? Isn't it what humans do too with images that they see and copy in style? It may seem obviously wrong but it's an unprecedented situation that will have to be solved in court. I can certainly understand the concerns of artists, I mean, I'm an artist myself. But I can also find arguments for the other side. This isn't just a case of "you're profiting off of my stolen images", it's also "I put these images on the web for everyone to see, I can not prevent someone from learning how to draw in the same fashion". There is no copyright on the looking-at, analyzing and learning part. There may be an issue with the end result if it looks too much like some other image, but then that would have to be solved on a case-by-case basis. And with the sheer number of images in circulation, it'll probably be a nightmare. Like how many different ways are there to draw the Eiffel Tower.

But the AI physically saves and and breaks down the images into the format it needs to learn. And then the algorithm spits out those pieces (along with pieces from other consumed parts) based on the prompt. The process is described a couple pages back and is wonderfully put together. It's not like a person who looks at images and then tries to emulate someone's style. Not like me when I first started out looking at Shiba Shake's artwork and saying something to the effect of "wow, I really love their style...that's totally the look I'm going for"...and then practicing and practicing and practicing til I developed my own style. When the AI responds to a promt, it's not like Bob Ross thinking to itself that it's going to put a pretty little tree over here this one time just to see how things go. There's no thought/feeling and no anthropomorphization. It's a machine that is performing calculations based on following a prescribed algorithm. As already explained, the same prompt and seed will always result in the same output regardless of what machine it is run on or who inputs the promt.  

bluejaunte

I completely understand the reaction. Aren't these wonderful philosophical questions though? What even is learning? Is it not learning anymore if the AI can do it a billion times faster than any human could? Is it not learning if all it does is execute algorithms? Are we not also essentially executing our own human algorithms and storing the information in our brains? 

SnowSultan

 And then the algorithm spits out those pieces (along with pieces from other consumed parts) based on the prompt. 

Except for the millionth time, it never "spits" them back out in the exact same configuration as the image that it was trained on. Photoshop tutorials literally tell people to search for a texture on Google, save it to your computer, then paste it into your scene where necessary for a background or to add grunge or roughness to your scene - which is closer to actually stealing art than what the AI is doing. Also, you can "put a pretty little tree over there" wherever you want using inpainting, and the AI will adjust the lighting on that tree to match your scene as well (which it couldn't really do if it was actually copying a tree it got from another image, could it?).

Richard Haseltine

SnowSultan said:

 And then the algorithm spits out those pieces (along with pieces from other consumed parts) based on the prompt. 

Except for the millionth time, it never "spits" them back out in the exact same configuration as the image that it was trained on. Photoshop tutorials literally tell people to search for a texture on Google, save it to your computer, then paste it into your scene where necessary for a background or to add grunge or roughness to your scene - which is closer to actually stealing art than what the AI is doing. Also, you can "put a pretty little tree over there" wherever you want using inpainting, and the AI will adjust the lighting on that tree to match your scene as well (which it couldn't really do if it was actually copying a tree it got from another image, could it?).

and that would not be allowed in an environment where rights management mattered, because it would stand a good chance of opening the door to being sued. Something being suggested in a tutorial does not guarantee that it is legitimate.