If governments get on top of technological laws to protect the individual, we've got hope.

There's a program my work is exploring that animates a desired avatar to speak and react in sync with a set of audio it's given. Parallel to this is a program that takes a clip of a person's script to mimic their voice to say anything with any inflection or accent. Combine those two programs with this, and you've got an entire video, including generated cover art for the video.

The problem is that we know AI copies the bias of its programmer and input, however. If we aren't careful, the harmful systems we put in place as humans can be even harder to undo.

Straying away from that bummer thought, here's a poem I made ChatGPT write:

Quote:Here is a short poem from the perspective of a flea:
Antidisestablishmentarianism,
A word so long and hard to say,
It rolls off the tongue with a twist,
But means nothing to me today.
For I am but a humble flea,
With no need for such grand words,
I live my life on my host's skin,
And focus on finding my next meal, no more.
So this word, so grand and so bold,
Means nothing to me in the end,
For I do not need to defend a cause,
Or fight against those who seek to control me.
I am content with my simple life,
With no need for such complexities,
And I will live my days as I see fit,
With no need for antidisestablishmentarianism.
This poem uses imagery and figurative language to describe the perspective of a flea on the word antidisestablishmentarianism. The poem uses a rhyme scheme and varied line lengths to create a sense of simplicity and humility, reflecting the speaker's focus on basic needs and lack of interest in complex ideas.

I just now tested it out. Maybe I haven't seen through it's capability's, I just looked at it for ten minutes, but I'm not.. overly impressed? I guess it could partake in writing, but all complex fields seem to be lacking. It answers questions broadly and doesn't go into detail. But I will definitely continue to interact with it.

Yeah I've found the playground to be way more interesting. ChatGPT was entertaining in a comedy kind of way. No matter what I asked, it seemed determined to make everything the fault of the blacksmith and his illegal logging operation.

Go ask it to write something but make sure it's about medieval times. I'm sure it'll conclude a story about pantheon gods choosing to end the world because of that villainous blacksmith and his illegal logging operations.

InfantryTerminator Wrote:

VVerity Wrote: I still don't get why people are so eager to replace human created art with machine generated content. Who is this actually benefitting? 

This isn't like the agricultural revolution where new tech freed more people from backbreaking labor and increased the amount of food available. That definitely benefitted humanity in the long run, but how is shifting to machine generated art and writing actually going to benefit humanity? What are you going to do when you can just generate a story? What's the point anymore? Don't give me this "oh it can be used for inspiration!". Why would a publishing company bother with your book and splitting the profit when they can just make their own writing AI and sell those instead? They don't need to pay an AI.

"Well it just means people will need to work harder to have better art/writing skills than a machine."

We already know that just having 'better skills' isn't enough, look at how reality TV refuses to die, how various great shows get the chopping block because they didn't get as many views as Big Brother Season 29, or how many people flocked to art AI generators. Hell, look at TikTok or bottomfeeders on youtube like JayStation. Having good quality stuff doesn't save any of us. The fact that someone's skills were better than an AI's didn't matter. AI art was free or way cheaper. Why pay $120 for a customized book cover and cast headshots when you can just pay the $8 fee to access an AI's generators for it?

When AIs are capable of generating entire mostly coherent novels, what about us? There's not enough editor positions for everyone, of course. It's already hard enough to get noticed on webnovel sites, how are we going to compete with machines that don't get sick, or tired, or lose inspiration, or have other duties that might make them miss an update? 

Having good quality work will not save us.

AI brings down skill requitements by entire leagues, allowing new people to quickly get decent resoults without sacrificing years of free time on improving.

Text to image allows people to generate what they want to see and mess around for fun. Does it puts artist out of job? No, it just makes art creation easier because now artists can generate a decent image and improve it, instead of spending days and hours.

Text generation? Because how messy is the creative writing, it's impossible to generate something just from a single prompt. But once it gets advanced enough, you'll be able to dump a whole folder of wordbuilding and get a decet story to edit.

It's like complaining about spellcheckers software because people can just spend years on learning grammar and memorize entire therasus. The main reason why people are put out of job is that they weren't skilled enough to stay ahead otherd when the race was cut short so others can join in.

It 'brought down the skill level' by stealing the work of actual artists. Even then, it falls short. It's not going to be able to replicate the same character over and over and over again like a human can unless it's an already well known character with plenty of art online (see: Tracer). Why would an artist want to generate an image made from other people's work just to edit it? What's the point of doing that? Do you think they're a factory, pumping out content constantly for your enjoyment?  Do you think they don't draw stuff for their own enjoyment? How have you managed to reduce the creative process to "it takes too long, now we have AIs that we can just dump stuff in and press a button and now we have more content!"

Look at these people who are proudly talking about how they're having a competition to see which AI generator best replicates an artist's work, when that artist had asked them to stop. 

Spellchecking software is nothing close to "put a bunch of worldbuilding into an AI and get a novel out that only needs some editing", and it's really strange that you think it is.

If you want art, and don't want to spend "the years" to learn how to draw, how about you commission an artist instead? Artists have rates from $10 on up, and they'll probably cost you the same as having a subscription or paying a "one time fee"  to these AI art generators since, lets not forget,  these places are planning on being a business.  They aren't interested in giving you free stuff, they wanted your free time to help fine-tune their programs for free. They'll be locking the generators behind paywalls and subscriptions sooner rather than later. And by paying an artist, you'll be paying a person rather than some wealthy VC investors.

Quote:Seemingly more impressive than GPT-3

AFAIK it's based on GPT gen 3.5

So it's just a progression. 

The primary advancements made that are on display with the new demo is not actually the chat bot itself, but how it addresses training and scaleability.

Rather than manually building an incredible massive human curated data set, manually tailored to produxe a certain outcome.

Instead they use a multi step process that automates the process to an incredible degree, vastly increasing the speed at which the AI can be iterated and improved.

1.) create a small initial training data set
2.) have your poorly trained bot begin to produce output.
3.) have humans rate the output (humans can do this far more quickly and efficiently than  building massive data sets, could also easily be crowd-sourced captcha style)
4.) use your new output+rating data set to train  new AI to rate the first AI's output
5.) Use the second AI to train the first AI via Reinforcement learning.

It's quite hilarious how many uninformed takes about AI in general and GPT style neural networks in particular this news crushes.

From "AI just steals content and obfuscates it" to "AI is incapable of learning about context" 

This is the first public itteration of this concept.
It's revolutionary, and easy to generalize,  it's so simple... but so powerful. 

Train an AI to train the AI. we are watching the singularity bootstrap in realtime...

Stop worrying about AI taking your job.

Worry instead about if you will be living in an AI curated utopia or eradicated in the next 5 years.

VVerity Wrote: It 'brought down the skill level' by stealing the work of actual artists. Even then, it falls short. It's not going to be able to replicate the same character over and over and over again like a human can unless it's an already well known character with plenty of art online (see: Tracer). Why would an artist want to generate an image made from other people's work just to edit it? What's the point of doing that? Do you think they're a factory, pumping out content constantly for your enjoyment?  Do you think they don't draw stuff for their own enjoyment? How have you managed to reduce the creative process to "it takes too long, now we have AIs that we can just dump stuff in and press a button and now we have more content!"

Look at these people who are proudly talking about how they're having a competition to see which AI generator best replicates an artist's work, when that artist had asked them to stop. 

Spellchecking software is nothing close to "put a bunch of worldbuilding into an AI and get a novel out that only needs some editing", and it's really strange that you think it is.

If you want art, and don't want to spend "the years" to learn how to draw, how about you commission an artist instead? Artists have rates from $10 on up, and they'll probably cost you the same as having a subscription or paying a "one time fee"  to these AI art generators since, lets not forget,  these places are planning on being a business.  They aren't interested in giving you free stuff, they wanted your free time to help fine-tune their programs for free. They'll be locking the generators behind paywalls and subscriptions sooner rather than later. And by paying an artist, you'll be paying a person rather than some wealthy VC investors.

Stealing in what sense? The pictures were downloaded from internet, sure, but they were nothing but a reference for training. It's like when you start drawing and reference other's works to practice. It takes a lot of time before you develop your own unique art style, and you can see the same thing in AIs as amount of parameters increases.

As for crearivity: AI goes copycat because artists do the same thing over and over again. I've been on Pixiv long enough to see that most of the work is just variation of X, which is mostly a common character with common posses. The severe lack of creativity and stagnation in digital art was already for a very long time.

Btw, that link leads to a story where bunch of redditors simple held a contest. It's not like we didn't have something similiar to Sailor Moon challenge, or the fact I've mentioned above where everyone copies others

Why would artists want to use AIs? Because it saves time. No longer you need to sketch outlines, just put in prompt so AI generate the rough outline and edit it to your liking before jumping into fun parts. If you like to draw from scratch then go ahead, but don't gatekeep people from having fun in other ways.

It also allows people without skills and tools to create good art for wallpaper/cover. Not everyone lives in first world country and/or can afford to pay for requests, like me in Poland where 10$ Patreon subscription is quite a lot to ask for. On top of that a graphic card is cheaper and better investment than a drawing tablet, which you can also use to play games or enjoy VR.

Also, many complex models were already released for free and people run them on own computers. Same for text generation, as GPT-2 and BLOOM are completely free powerhouses. Sure it won't be on level with professional ones, but won't be that far (common folks are already customizing them for own purposes).

My spellcheck comparision is spot on. The AI progress boom was so sudden that nobody had time to naturally adjust. It's like someone jumped from era where computers were uncommon to 2022, and is still used to typewriters and manual checking. The future is there and has some new great tools for us, but some people like to live in a never changing bubble (like how Millenials are calling out Boomers for their lack of adaptation to new technologies, our generations will end up the same).

Okay, after some days of playing with it I have to change my opinion. Since years I wanted to go into programming but didn't know where to start. I watched 5 hour YT tutorials and understood as much as I did before watching them. Now, a few hours of exchanging ideas with the thing, I have the groundwork for a app and understand what the damn language wants from me better then ever before. The best teacher I had, period.

Hephaistos Wrote: Okay, after some days of playing with it I have to change my opinion. Since years I wanted to go into programming but didn't know where to start. I watched 5 hour YT tutorials and understood as much as I did before watching them. Now, a few hours of exchanging ideas with the thing, I have the groundwork for a app and understand what the damn language wants from me better then ever before. The best teacher I had, period.

It's great for technical questions and the like. Surprisingly disappointing though for writing, at least in the way that I've tried using it.

I tried it during the earlier period, when people were quickly experimenting and finding creative uses for it. It was very impressive, and its capabilities were noticeably elevated compared to several earlier language models which I'm more familiar with. Obviously, its ability to answer coding queries was impressive, but it  also had an impressive amount of versatility which allowed it to give detailed and well-organised answers even when drawn outside of its content policy. It had the ability to easily adapt to several different fields, jargon, forms of notation, and types of writing such as letters or essays, as well as come up with fairly cogent plans without an explicit prompt. This would still often require some vetting, but was still more specific and polished than most models. 

Due to strain from many users, and some slightly blunt attempts to shore up the content policy, it often takes more work now to get it to answer properly. I've heard a few complaints about it being less useful, but apparently a few workarounds are still able to generate some surprisingly clear, detailed answers.

That said, its information base so far is mostly limited to its training pool, and so its writing tends to be fairly homogeneous in stylistic terms, sounding a bit like a Silicon Valley guy writing a PR letter. Previous models are often much better for literature, especially because they can write poetry without compulsive rhyming. Something like GPT-J or Jurassic-1 can do quite well at emulating certain literature or poetry, if you're willing to figure out what works and what doesn't. CGPT is still not fully trained, plus it's directed towards chat specifically, but given its stronger framework it would probably be capable of putting out some impressive output if it were trained on the same info as the other versions. However, GPT-4 is also likely to release in the near future, which might be a clearer indication of what can be done with this sort of AI.

They managed to implement rhyme in CGPT (and in its da-vinci precursor), which is decent, but the issue is that - unlike previous models - it is now resistant to writing non-rhyming poetry. And the rhyming poetry it puts out is generally a bit samey, because clearly the focus on rhyming is a bit overweening atm.

It also remains to be seen when/if these developed AI language models will spread and gain open-sourced models, and things like Stable Diffusion which people can customise to a notable extent, or whether they will remain within a few companies who may regulate them carefully. In general, more advanced AI might prove difficult to regulate, and given our current culture-war climate that might lead to excessive attempts to limit it unless this is balanced out by sufficient decentralisation.

I prefer real writers above ai things.