eevBLAB 106 - ChatGPT AI Has Changed EVERYTHING

[bob808]

General AI is a whole other subject. With MidJourney you have the final say on what is the final art piece. I see it as a tool and nothing more. 
I think there's two aspects to what we (used to) call art. There's the execution and there's the idea itself. Clearly execution augments the idea. But that is not to say that many more people don't have that or even greater ... "ideas"? or powerful internal imagery that they have no way of expressing via classical means, be it drawing, sculpting, music. I see these new tools as interfaces for expressing certain ideas, some of which we celebrate as art. Makes it easy for more people to express themselves, kinda like Arduino did for some artists. Or python for scientists. 
So far great art has been heavily gated by the ones that both had interesting things to express and also possessed the means to do it greatly, at the same time.

[Unixon]

General AI is a whole other subject.

If it doesn't fit the definition of strong general AI, it is not an intelligence at all, just an overcomplicated tensor function.

The problem is most of inadvertent users of weak AI systems are not mathematicians or AI specialists and they take it for appearance of a strong AI behind the curtains.
What's even worse, people in governments and evil corporate management began to think it is already good enough and they're already putting these artificial idiot kind of chat bots everywhere.

[adx]

What got me hooked on all this stuff was this thread:

https://www.eevblog.com/forum/chat/the-long-road-to-singularity/

I had heard the hype and assumed it was just hype, I had no idea how far it had come. I used to play with neural nets and took me a while a few years back to even recall the name of the company that made the software I bought (I think!): California Scientific Software. Brainmaker I see, DOS, late 80s IIRC. Being not unaware of advances in computing since then, I've wondered what came of that sort of stuff, given the power of even a 100 neuron model. AI this AI that, but nothing which really seemed to live up to the hype with real functional demonstration.

Until now.

[PA0PBZ]

Just launching the idea... 

[MK14]

I asked (I can't remember, exactly, but very approximately this) it to design a new EEVblog forums, Logo.  With bright, colourful flashing LEDs, and transistors.  This is what it came up with:

[PA0PBZ]

This is interesting, it will tell you the used source(s). The question used to demonstrate it is silly of course, but it also does it for pieces of code etc.

[SiliconWizard]

Just another round of buzzword bingo. For example, AI generated code: Code-generating AI can introduce security vulnerabilities, study finds (https://techcrunch.com/2022/12/28/code-generating-ai-can-introduce-security-vulnerabilities-study-finds/).

Yes, videos about ChatGPT are currently everywhere on YT (and other platforms), it's a frenzy that makes tons of views.
This is only benefiting people making views about it and OpenAI.

And as it has often been mentioned here and elsewhere, AI is a misnomer.

[EEVblog]

I asked (I can't remember, exactly, but very approximately this) it to design a new EEVblog forums, Logo.  With bright, colourful flashing LEDs, and transistors.  This is what it came up with:

I also tried doing company logos and apparently it's not at all trained on doing that. I couldn't even force it to put the EEVblog text. Maybe I was doing it wrong 

[IanB]

Maybe I was doing it wrong 

Or maybe it lacks intelligence? 

[MK14]

I asked (I can't remember, exactly, but very approximately this) it to design a new EEVblog forums, Logo.  With bright, colourful flashing LEDs, and transistors.  This is what it came up with:

I also tried doing company logos and apparently it's not at all trained on doing that. I couldn't even force it to put the EEVblog text. Maybe I was doing it wrong 

I think if I/we could get proper access to it and/or supply our own extra images, it could do better.

They seem to be limiting the free accounts to 50 'tokens' (or whatever they call them), then 15 extra tokens (but balance won't go above 15 per month, if I understand their details).

Which doesn't give much of a chance to experiment and learn with it.  As it messes up the drawing process, most times, so you would use up many of those tokens, just to make one semi-decent picture.

It seems 1 picture = 1 token, each clicked retry, costs a token.

Tokens are something like $10 for 115.  So rather pricey, unless doing it for serious business reasons.

If it was continuously free (as a trial, perhaps with 5 free tokens per hour), I suspect that with a huge number of attempts, which would mean gaining experience on how to use it well.  Semi-reasonable results could be obtained.

From using up my 50 attempts (tokens), it seems it is a fun toy, and learning exercise on what AI can do, now.  But is not really useful for most tasks (pictures), at the moment.

But I'm sure tools like that, will rapidly improve.  Eventually becoming useful.

They perhaps are good enough to generate book or report, illustrations.  Without needing to pay a professional illustrator, to do it for you.  Like you said, it can be useful for thumbnails, or to illustrate a point in youtube videos.

[adx]

If you have a GPU with enough RAM (4GB), you can download Stable Diffusion and run the art models locally, including giving it an image prompt, or if keen, train it on new art (with more RAM and enough time).

I tried finding an easy way to run it on a NAS with weak GPU but gobs of RAM it can probably access. But could only find CPU versions which would be nasty slow on it (1 hour I think) so I put that idea on hold.

It can run on an iPhone.

Based on my earlier calcs the model size of something like that or DALL-E 2 is only around the size of a bee's brain including the language comprehension side. That means in theory a bee could understand all those words and dream those same images in its little head (if it didn't have to control arms and legs or whatever they have).

[MK14]

If you have a GPU with enough RAM (4GB), you can download Stable Diffusion and run the art models locally, including giving it an image prompt, or if keen, train it on new art (with more RAM and enough time).

I tried finding an easy way to run it on a NAS with weak GPU but gobs of RAM it can probably access. But could only find CPU versions which would be nasty slow on it (1 hour I think) so I put that idea on hold.

It can run on an iPhone.

Based on my earlier calcs the model size of something like that or DALL-E 2 is only around the size of a bee's brain including the language comprehension side. That means in theory a bee could understand all those words and dream those same images in its little head (if it didn't have to control arms and legs or whatever they have).

Thanks, that is VERY interesting.  If it makes any kind of sense, I have significant aversions to paying them money, as I'd rather do those things on my own equipment.  Also, I don't like cloud services, partly because the monthly subscriptions, can be huge, when properly perceived/calculated over a realistic time period.  E.g. $20/£20 per month = Over 10 years = 12 (months) x 10 (years) = x120 = $2,400 / £2,400 = A large amount of money.

Also, cloud services can suddenly change (interfaces and/or functionality and/or ownership), and dramatically increase in cost, over time.  Additionally they lack privacy.

One of the worst features, is that they can suddenly disappear over-night, or at too short notice.  Along with all your data/projects/stuff.  Which can be in proprietary formats, and not supported by anything else.  Which can be a real pain.

The requirement you specified, don't sound too bad.  A modern GPU (I think it is best, to carefully check the required specification of it, especially as regards video memory size, as even the biggest domestic use ones, might be too small, for some of the AI stuff, you want to use on it, e.g. training certain stuff) and computer to go with it, and patience to wait for it to process the results.

The Wiki on the software system you mentioned.  Here:   https://en.wikipedia.org/wiki/Stable_Diffusion

Seems to say:

it is difficult for individuals to run models in consumer electronics. For example, the training process for waifu-diffusion requires a minimum 30 GB of VRAM,[32] which exceeds the usual resource provided in consumer GPUs, such as Nvidia’s GeForce 30 series having around 12 GB.

So care is maybe advisable, before buying stuff to do things like this on it.  Depending on your requirements and intentions, which will dictate how big and powerful a system(s), you may need.

[PA0PBZ]

Download from here: https://github.com/cmdr2/stable-diffusion-ui#installation

I have it running locally, installation was easy. Once running it gives you a local webpage (localhost:9000) where you can put in your text prompt and hundreds of settings to play with.

[bob808]

I suppose there's other algorithms that get trained on what we do with ChatGPT and MidJourney and the likes.  There's interesting data to be extracted from that as well.

[Fungus]

There's all sorts of things the image generator can't do that a human artist can:

eg. Try making a series of images for a book using the same character but doing different things.

PS: Would you trust an AI to generate a complex PCB?

[MK14]

There's all sorts of things the image generator can't do that a human artist can:

eg. Try making a series of images for a book using the same character but doing different things.

PS: Would you trust an AI to generate a complex PCB?

I don't think the AI image generator, is even being artistic at all.  It is relying on the user of the AI package, to choose an image, because it looks the way they want.  I.e. all the artistic taste, comes from the human user of it, NOT the AI system.

Hence the need for it to give, four sample images each time, in the hope that at least one of them, might hit the mark.

I'm not even convinced, a system which only uses relatively short text descriptions, is going to get the right/desired image.  Especially with very tricky things.  Whereas a real artist, would probably (I guess) spend time with the customer, and do very quick rough sketches, get feedback on those, to narrow down what the customer really wants.

Maybe it is a bit artistic, especially with the oil paintings.  But I've seen too little and know too little about it, to know, either way.  I remember the old/previous one (or something), could make these sickening (at times), pictures of cats, which looked rather horrifying.  There still seems to be a bit of that left, in the system.
Because although humans, look ok, at a quick glance, a more careful look, especially around the eyes area of the face.  Doesn't look right, and even looks a bit spooky, in a bad way.

E.g. Posted earlier in this thread.  It looks a bit odd/funny, if you look at it carefully.  Especially the face, and the eyes.  It doesn't look right or good, to me, at least.

 

[Fungus]

E.g. Posted earlier in this thread.  It looks a bit odd/funny, if you look at it carefully.  Especially the face, and the eyes.  It doesn't look right or good, to me, at least.

I'm sure there will be a lot of "uncanny valley" artifacts in there.

FWIW I Just paid a student to do some artwork for me. I gave Dall-E a try at making something suitable and couldn't get anywhere close to what I wanted. Dall-E doesn't seem to know what an Arduino is, and it like putting pictures of children into "Arduino" images even when I ask for "space monster". The student got it pretty much right on the first try.

My take: The demos are pretty but there's a looong way to go yet.

PS: I've been reading about computer program generators taking all the programming jobs since the 1980s (check articles in the "References" section for some good chuckles).



(from: https://xkcd.com/568/)

PPS: Spooky coincidence? http://smbc-comics.com/comic/art-5

<exaggeration>The biggest danger from AI right now is students thinking it can do their homework for them.</exaggeration>

[Ed.Kloonk]

I asked (I can't remember, exactly, but very approximately this) it to design a new EEVblog forums, Logo.  With bright, colourful flashing LEDs, and transistors.  This is what it came up with:

I also tried doing company logos and apparently it's not at all trained on doing that. I couldn't even force it to put the EEVblog text. Maybe I was doing it wrong 

My spidey-sense wonders if that result might be intentional. Company logos are a minefield. Trademark infringement is quite different from copyright infringement. You'd be sourcing your logo from other brand's logos (and trademarks). Woah, boy...

[Brumby]

Company logos are a minefield. Trademark infringement is quite different from copyright infringement.

It certainly is.

You'd be sourcing your logo from other brand's logos (and trademarks). Woah, boy...

I get an image of diving into a wood chipper......

[Fungus]

My spidey-sense wonders if that result might be intentional. Company logos are a minefield. Trademark infringement is quite different from copyright infringement. You'd be sourcing your logo from other brand's logos (and trademarks). Woah, boy...

That's not the AI's problem... 

[Nicholas1]

This AI reminds me of a program called "GitHub Copilot". It works on multiple programming softwares such as Visual Studio Code and PyCharm and to a very accurate degree if one describes the problem in a detailed but concise manner. I'm an Electrical/Electronic Engineering student, and my friend tested it on our embedded systems exam after we completed it. He noted that it worked quite well and wrote the code rather quickly. 0s and 1s may be our new world...

[bob808]

MidJourney is a show-off tool atm. It isn't optimized for certain workflows. But I am pretty sure many more tools are coming which will be optimized for different industries. Like logo generation or website graphical kit generation. Or get trained on being able to use the same character of interest but in different contexts in each new photo.
I think I remember reading something about there being no IP for whatever the AI is generating as it's not human generated. Pretty sure there's a lot of discussions on this topic atm, from a legal point of view. They'll sort it out somehow.

[thm_w]

My spidey-sense wonders if that result might be intentional. Company logos are a minefield. Trademark infringement is quite different from copyright infringement. You'd be sourcing your logo from other brand's logos (and trademarks). Woah, boy...

That's not the AI's problem... 

The point is, they may have removed any obviously copyrighted logos from the learning material, to make any future lawsuits against themselves, less cut and dry.

wizardtim had a good relevant post on youtube.

I wouldn't say 2022 is the year of mainstream AI, it's the year of mainstream justification for large scale data theft to train AI. As you say garbage in, garbage out, but the prompt is a pittance compared to the training data which is a very large component of how good any AI model is, however this data is very difficult and expensive to obtain.

A couple years ago I worked on an AI model to detect and classify certain sounds, we had a DNN model architecture that was fine tuned for the task but the thing we really needed was lots of training data, we had a couple hundred training examples and were able to train a decently competent model with that but it still wasn't very good. However in the process of developing it I obtained A LOT more data, both scraped from the internet as well as taken from other researchers, I trained the model on this data as it was MUCH more accurate, but I deleted that trained model and in the end we used only the model trained on the data we paid for, we paid about $100k for people to go out into the world and record the sounds for our data set over several years. If we just used the several thousand we obtained for free it would have been a massive game changer but it would have been theft, but those large AI companies that have recently come to popular success (like Midjourney) have stolen literally hundreds of millions of training examples to train their AI model and they don't even try to hide it, they brag about it. When they were stealing data to make free open AI models that just made low quality funny memes it was all fun and games, but now they're closed source and asking for money to do the work of those they stolen works from. In particular with AI art, there are now tutorials on how to remove watermarks and artist signatures which have leaked in from the stolen training images so that you can 'make a business charging for your images as an AI artist', this is truly disgusting to me. 2023 will be the year of AI training data litigation, private art feeds, tag vandalism, intentional training data contamination, increased DMCA on image boards, and the end of many creators posting online at all. There are now websites that index the known datasets of stolen works those AI companies use, a lot of works I've posted have been stolen (and your too Dave, EEVBlog forum posts, video thumbnails and frames and flickr images), I can see why so many are privating their works and no longer posting because all they would be doing is supplying free training data for those AI companies and immoral 'AI artists' looking to clone an artist's work and style.

But seriously, why the hell did I kept with my morals back then and delete that trained model? I should have just stolen all that data and we would have had the best model in our entire field, we would have exceeded our performance targets and gotten a lot more funding, we probably could have commercialized it or sold the IP with performance that good! Why the hell did I do the right thing?!

Although I don't know if I'm in entire agreement, as, what is the different between AI looking at a sample of copyright art and a human artist looking at those same samples with their eyes, and then using that for inspiration. Not sure.

[golden_labels]

E.g. Posted earlier in this thread.  It looks a bit odd/funny, if you look at it carefully.  Especially the face, and the eyes.  It doesn't look right or good, to me, at least.

But that is not an inherent limitation of the technology itself. I was also focused on finding a particular output and not caring about quality. The point I want to make is that it clearly copied something from the training set. It might’ve mangled the picture a bit, but all the important features remain intact.

[MK14]

E.g. Posted earlier in this thread.  It looks a bit odd/funny, if you look at it carefully.  Especially the face, and the eyes.  It doesn't look right or good, to me, at least.

But that is not an inherent limitation of the technology itself. I was also focused on finding a particular output and not caring about quality. The point I want to make is that it clearly copied something from the training set. It might’ve mangled the picture a bit, but all the important features remain intact.

I wasn't trying to criticise your earlier post.  I'd just stolen borrowed your picture, to illustrate, that at the moment, the ChatGPT's, DALL·E 2 picture generator, via a word description.  Is (N.B. opinions can easily vary, on its suitability) not really ready for more serious uses, of such a picture.
But it is perhaps good enough for messing around, fun, draft/early copies of stuff.

Because pictures are so important, valuable and used all over the place (as others have also stated, perhaps elsewhere on this forum).  It is likely that these text to picture offerings, are going to rapidly improve and evolve to the point of being increasingly useful.

The pictures themselves, are actually sometimes, quite good.  But the text to picture mechanism, seems to be fiddly, imprecise and not really listening to the instructions, like a professional human artist, would.

I suppose putting it in different words.  It is a bit like encountering what is going to be a really good photo processing application, such as Photoshop.  But it is still in a very early alpha software state, and has a huge number of bugs, failings and missing features.
But it does look really promising, when it is much closer to the finished article.

Some people may have mentioned, that there are better ones (AI text to pictures), but I don't think I've tried them, or not recently.