eevBLAB 106 - ChatGPT AI Has Changed EVERYTHING

[EEVblog]

ChatGPT has changed everything in the space of a month.
2023 will be the year of Conversational Search and computer code generation.
And how MidJourney and DALL-E have rendered artists obsolete, or at least turned ANYONE into a great and capable artist.
This video thumbnail was generated in seconds by Dave who has ZERO artistic skills.
Why The Chaser is setting up a futile paywall to stop the AI invastion.
https://chaser.com.au/uncategorized/why-the-chaser-is-going-to-put-up-a-paywall/

[Ed.Kloonk]

Maybe it's time to let the robots "have a turn".

I mean, look at what the hoomans have done to the place in the past few years.

[MK14]

Thanks for making the video, I enjoyed watching it, and learnt a fair bit about it.  Those automatically generated pictures, are just amazing.

Even though, I've known about it (ChatGPT, and some picture generating thing(s)), for quite some time and played around with it, in the current and earlier release(s), some time ago.

I'm both very exited/interested in this ChatGPT (and related stuff), and sometimes/somewhat/partly frightened/scared by it, at the same time.

I wonder how it will affect people's employment situation, as time goes on.  So many peoples jobs, are potentially on the line. But I suppose it is a bit like, when people perhaps thought that wordprocessors, might/would replace secretaries/typists in offices.

In some respects it has.  How many office workers, use an old mechanical or electrical typewriter these days?
But there are still plenty of office workers, it is just the nature of the jobs have changed.

Even the internet and google (searches), haven't stopped people from getting jobs, in general.  So, I would guess the jobs market will still survive.  But may just see some changes.

[madires]

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/).

[kripton2035]

try to ask it to show an electronic schematic of ... whatever.
they are pretty useless (till now !?)

[ataradov]

For programming, I'm not worried at all. I have not seen it do anything actually impressive. I'm not sure what non-programmer understanding of what programmers do on a daily basis, but typing boilerplate code is not it. Plus the code is far more precise than art. For art "something close" is often good enough, which is not the case for programming.

How would you expect new art and art styles to appear if everyone "just generates" stuff? Sure, there will always be a need for art with very precise requirements, but how would new artists get to that point?

I'm also interested to see how the modes would evolve (or degrade). Currently all the models are trained on clean data before AI explosion. Imagine you have to move forward and need to update the model, but now most of the code committed to GitHub and used as training data is AI generated crap? It will further learn and will generate even bigger crap. They would really have to figure out how to filter out that stuff, or things will degrade quickly.

[golden_labels]

Is The Chaser saying their authors can be replace by an IQ 88 entity?

I do not want this comment to be taken as claiming that this technology is not revolutionary. It is! But don’t be blinded by the hype. Do not focus on unbounded extrapolating, as that virtually never works. I expect smortnets will bring huge changes, but I would focus more on the philosophical side of it and how everything around smortnet deployment affects society.

At the current state of machine learning development, smortnets are as capable of replacing artists as calculators are capable of replacing mathematicians. Maybe in the future, but not yet. Being an artist is not about merely creating an image (or other work). It’s about being the taste. Aesthetics, conveying the message, connecting emotionally to other humans — choose whichever you want, smortnets can’t do that yet.

We are not going through anything new. The same old story, just with different characters. For now smornets speed up a mechanical, tedious process that has little to do with artistic expression. My great grandfather was drawing posters for money. He did that, because printing small batches of graphics was prohibitively expensive, and because people had no tools to quickly stich a few images and text in a text editor. Offset printing⁽¹⁾ and DTP made that kind of job obsolete. Equally, photography made commercial portrait painting no longer needed. But neither of this affected actual artistic work. Of course it made it impossible to earn money on doing that boring, skillful job other people were not willing to master. It also nuked the market of people, who lacked any artistic skills (and often any skills), but served customers with very limited requirements. And I suppose smortnets will have a similar effect. But be aware that artists are already jumping to use generated pictures and that still requires their artistic skills: because they still must use their critical thinking to tell which images they want and how they want to affect it. It is already confirmed in legal cases: while smortnet-generated image is not a subject to copyright in US, the mere process of selection by the artist makes the selected picture a copyrighted artwork.

The reason machine learning is unable to replace artists is the same it can’t replace software developers, writing valuable essays or producing documentaries. I do realize smortnets seem magical, but please recognize that: those are nothing more than glorified statistics. Just much less understood and delivering more accurate predictions for arbitrary data. They seem to do much more than that because… imperfection of human brains. Not only there is a strong selection bias in reports/exposure, your brain is “patching bullshit”. It’s no different than people seeing faces on Mars or TA “experts” describing market behavior from random walk. The brain loves patterns and ignores elements that don’t fit. To the point that people were fooled to talk to a markov bot on mailing lists, despite it should be obvious it’s not a human. With smortnets this goes to extreme.

In the nutshell, the best they can currently do is performing very complicated mapping between two pieces of information. Which permits them to replicate a single, separated feature of a “thinking” system. But for now they miss the ability to emulate such a system. In particular, and that is of uttermost importance, they lack ability to be critical of their own outputs. I am pretty sure they will acquire that ability in the future, but in 2022 it’s not there. And that is the primary limiting factor for now. A smortnet may write you a simple program, but in many cases it will will be “not exactly right” and the algorithm has absolutely no way of telling that.

Finally, what smortnet operators do not tell you: the energy usage of most advanced algorithms is insane compared what a human uses for the same task. That’s why — aside from having control — ML based mobile apps are usually offered as service running on a remote server and why self-driving cars are employing solutions that are way behind what is supposed to be the current state of the art.

⁽¹⁾ Early offset printing was invented during his childhood, but it didn’t reach widespread commercial use in Poland during his lifetime.

[artag]

What would be more useful than a code generator is a spec-parser. Code that's wrtitten to a vaguely expressed spec will be no better than if you give that spec to a bunch of programmers.  The first stage therefore in creating a code writing bot is to have it identify all the ambiguous elements of the spec and discuss them with you.

At that point, what you've got is the code, written in a high level language. So obviously, it can be compiled to something that can be executed. But understanding the requirements and expressing them in an unambiguous way is the hard part and that's what programmers do now.

[ataradov]

Then "wring the spec" just becomes new programming. And I doubt that for a complex system the spec would be much shorter than the code.

And another thing to consider here is changes later in in the lifetime. You would have to be mad to completely regenerate the code with minor changes to the original spec. So, there must be some way for the AI to know the current state of the code base, which inevitably would include manual changes, especially to support stuff that was invented after AI was trained.

And then you also need to consider how you would approach revision control. Are you just going to give OpenAI control over your whole future? Those models may be very resource-intensive to run, and they are prohibitively expensive to train.

[artag]

Then "wring the spec" just becomes new programming. And I doubt that for a complex system the spec would be much shorter than the code.

Exactly. Programming has always been this : to write a spec for the operations required in a language the hardware can execute. Advanced compilers can accept a higher level language but it still needs to be unambiguous.

[IanB]

Exactly. Programming has always been this : to write a spec for the operations required in a language the hardware can execute. Advanced compilers can accept a higher level language but it still needs to be unambiguous.

The problem is that any spec you write will always be wrong. Because the spec defines what the software will do, and what the software should do is always a matter of opinion. One person's bug is another person's designed behavior.

The art of computer programming is to discover a spec that satisfies more people than not, more often than not. The first specification you come up with is guaranteed to be different from the one implemented in the end.

This is not likely to be a problem that machines can solve any time soon.

[bob808]

@EEVblog And we're not at video content generation yet. That's going to be a whole other thing in its own right.

[EPTech]

Sarah Conner! This is your cue!

[Bud]

Call me when that gizmo will be able to route PCBs for me.

[bob808]

I don't think we're far off from asking Alexa to design&order simpler projects. 
This might be controversial but how hard can it be to reach a point where the AI would be able to arrange the parts on the PCB and route everything, having been trained on a gazillion designs?

[golden_labels]

how hard can it be to reach a point where the AI would be able to arrange the parts on the PCB and route everything, having been trained on a gazillion designs?

Have you noticed, how the entire hype revolves around aiming at the low hanging fruit of fooling human brains and not solving any actual, computationally hard problems? Guess why!

There are some smortnets that can solve strategic games, aid protein folding, and spot optimization opportunities in data fed to classic algorithms. But that’s it. I believe that machine learning solutions will be able, perhaps even by 2030, to greatly outperform existing routing algorithms. But that’s an observation not related to the recent hype. ChatGPT, Stable DIffusion, even AlphaZero — they are absolutely incapable of providing outputs meeting any strict constraints, if the result is to be evaluated precisely.

The closest they get is being able to produce something resembling floor plans. What reporters forgot to mention: there is no real limits set for the layout, you must filter out manually the most absurd responses, get used to 30cm wide doors here and there, entering your home through the restroom, and having a wall in front of your living room window. But you may appreciate an array of chimneys in your garage — perhaps an art piece. Garage has a gate leading to your kitchen.

[Jacon]

Another technological fever  , which raises interesting law problems:

https://www.computerworld.com/article/3684408/the-one-real-problem-with-synthetic-media.html

[Unixon]

Another technological fever  , which raises interesting law problems:
https://www.computerworld.com/article/3684408/the-one-real-problem-with-synthetic-media.html

In short: F the IP, switch to crowdfunded PD.

[Unixon]

Why is everyone so excited about these developments? The pictures are great I must admit, but these are at the end of the day are just manually tuned approximations / carefully selected learning datasets.

[james_s]

The AI generated pictures are fun, but ultimately they all have a similar look, like AI generated pictures. I suspect there's a bit of a fad there and people will eventually get bored of it.

[Bud]

When AI will start make everything and humans will stop producing content because everyone will be using AI, then what AI is going to be trained on if there will be no human produced content? 

[IanB]

When AI will start make everything and humans will stop producing content because everyone will be using AI, then what AI is going to be trained on if there will be no human produced content? 

This is easier to answer if instead of calling it AI, you call it machine learning (ML). Machines have been shown to be able to teach themselves (for example, the Go program that beats the best human players).

The problem now becomes one of goal setting rather than training, and the challenge will be to provide appropriately defined goals. If the goal is quantifiable, such as winning a game, or maximizing the number of successful outcomes, that is one thing. If the goal is artistic (produce good pictures), that is harder, especially since humans do not agree among themselves what is good art.

In the latter case the quantifiable goal will be to produce art that people are willing to pay money for, and that will be interesting for the art world, since art will have to be judged on its merits and not on the reputation of the creator. Or maybe certain programs will get their own reputation and have name recognition like human artists?

[golden_labels]

Another technological fever  , which raises interesting law problems:

I do not know, why is the article so alarmist, talking about the topic… which now has the least unknowns.

Berne convention is very clear in this case, it has been tested multiple times in court, and there is at least two US cases confirming the situation for generated images. They are by themselves not eligible for copyright, because they are not made by a human or a group of humans.

There has been uncertainties regarding artistic input from the human controlling the smortnet. This has been theoretically on the side, that artistic choices make the work a subject to copyright. The active human choices guiding the algorithm are what makes it intellectual property. It’s not a new subject and has been beaten to death for a century: with passport photos, medical imaging, photographic reproductions of old paintings; more recently CCTV, dashcam recordings, and automated satellite photos. It is already tested in US with smortnet-generated images, with the expected outcome.

Most photos do not have any recognizable style. For the remaining minority of art, the style itself is not a subject to copyright. Even the most extreme IP supporters among artists did not call for that. It would prevent them themselves from publishing works, because a single author would sit on the entire genre. With art being mixing and recycling, it’s also impossible to tell, who would that be. You can’t tell who really created the first metal or jazz piece.

The problem is with slurping the learning dataset from unauthorized sources, hence not getting legal clearance for producing derivatives. That’s currently the major issue with open source content being taken from GitHub to teach Copilot. And there is another minor point: accidentally generating something closely resembling an existing work. See the attached picture, which I made with IIRC Stable Diffusion in September. The clothing is different, the face proportions are a bit off, the contrast differs; but otherwise it’s a copy of this photo down to single wrinkles on her face. Most importantly, smortnet copied the crucial and unique element of the original: her facial expression and the look in her eyes. But that is not different from the current situation, where you are likely to trespass someone’s IP by accident.

[bob808]

these are at the end of the day are just manually tuned approximations / carefully selected learning datasets.

 
Isn't that what Art school is as well?
What is the difference between an artist drawing something by hand, an artist using a tablet or anyone using MidJourney to express some idea in a visual form. Does the technique matter over the result? 
Sure you could argue that using MidJourney there's not much effort, and you maybe settle for something that comes out randomly great. But also some artists operate the same way. They start the creative process and they're not sure what they'll finally come up with.

[Unixon]

Isn't that what Art school is as well?
What is the difference between an artist drawing something by hand, an artist using a tablet

The major difference I can point to immediately is that an artist knows in some shape or form what it wants to achieve, while a neural network just performs a numerical conversion of some meaningless data.
An artist can explain why he/she/it wants to pursue a certain path, probably without being too enthusiastic about giving such explanation because he/she/it "just feels so".

or anyone using MidJourney to express some idea in a visual form. Does the technique matter over the result? 

In case of an artist using the MJ for an instrument the selection of final image happens on the part of the artist, and that's the point. MJ has no effing idea what it has to produce, the only thing it has is just a huge number of weights in the network. Of course it could be argued then that a real biological neural network does nearly the same and I would even agree to this argument but only if a number of strong additional conditions are met (self-reflection, common cognitive basis, etc...).