The AI Thread

[Kyriakos]

Classic AI like the one civ games use and modern deep-learning AI are very different beasts. While the former is based in a set of rules written by human programers the later is based in having a humongous quantity of data so learning algorithms can extract trends from it and produce results. I see it as comparing classical deterministic physics vs quantum physics which is based in probabilities and statistics which tend to give only a useful result with big enough populations.

Anyway I think it is of not much use to discuss the metaphysical aspects, if AI will ever reach self-awareness or sentiments, or to be really intelligent, as it is already difficult for us to define those concepts since we live 'inside' them, so we can't distance ourselves to look at the issue with some perspective (maybe a machine would be better than us at that? )

All I can say is this thing is going very fast, and the results are continuously improving, to the point a general AI may be closer than we think. And for general AI I don't necessarily mean a self-aware being and such, but simply an intelligence capable of taking info from very different sources, combine all and react to it. So, text, which has been the first step, now images, soon it will be video, then sound, smell, tact... So all the ways we humans have to perceive our environment. Such AI will probably be totally indistinguishable and produce results similar to any human. Would then matter if it is 'intelligent' the same way humans are? I mean 'you will know them by their fruits'.

The other day precisely I was playing with AutoGPT, an automatization of ChatGPT which you can ask anything (for instance "make me rich") and it will start looking for ways to achieve it, even creating other AIs with specific missions as looking the internet, creating code, etc. So i asked it to write the best prompt to create the most beautiful image in the universe. Then the AI said to itself (you can see what it's thinking):
-I will search the internet for images considered specially beautiful to learn how to make the most beautiful one.
then stopped for a while (probably looking for images in the internet) and after that it came back with:
-wait, i am an AI that can only read text, i can't see images! will have to find other ways.
Then it did some incompressible things and entered in a loop.
I didn't get any incredibly beautiful image at all but seeing the AI reasoning and 'missing' having a pair of eyes was even more interesting, and a lot funnier.

I see you have already repressed the failure with even basic ascii trigonometry ^^
Though I do expect that to be resolved soon (if not already).
I did "like" that ChatGPT was of the view it's perfectly fine (like in the meme) to simply state that the clearly non-right angles are right angles, when its task was to present visual relations in a triangle.

And it does fake a whole lot of stuff. If it can't answer correctly (after being told it was incorrect), it will say it is sorry, then act like it only now realized what you meant and go on to write something fake

 

[Thorgalaeg]

I see you have already repressed the failure with even basic ascii trigonometry ^^
Though I do expect that to be resolved soon (if not already).
I did "like" that ChatGPT was of the view it's perfectly fine (like in the meme) to simply state that the clearly non-right angles are right angles, when its task was to present visual relations in a triangle.

And it does fake a whole lot of stuff. If it can't answer correctly (after being told it was incorrect), it will say it is sorry, then act like it only now realized what you meant and go on to write something fake

Free ChatGPT (3.5 version) is going to give an answer even if it has not a clue. OTOH Bing, which is based in ChatGPT 4.0 won't hesitate about saying "i don't know", specially in accuracy mode. Try Bing to see what happen.

 

[Thorgalaeg]

Text to video. Some decent results mixed with true nightmares. Still pretty impressive for such a new technology.

 

[Qactus]

An interview with the artist who won (and refused) the Sony world photography award, touching on multiple interesting topics related to AI assisted image generation:

‘AI isn’t a threat’ – Boris Eldagsen, whose fake photo duped the Sony judges, hits back

The German artist caused uproar this week when he revealed the shot that won a prestigious award wasn’t what it seemed. But, he insists, AI isn’t about sidelining humans – it’s about liberating artists

www.theguardian.com

 

[Kyriakos]

I am generally ok with "artist" being just the person who has an eye for what may impress people, in which case using ai doesn't take away that value.
And while it is different to build something from scratch, than to spend time checking ai creations to your prompts, you still need the human to estimate if it works.
Say ai gets to a point where it can batch-create isometric gfx like my CivIII ones, being fed a prompt (like one of them). It would make no sense to insist that if I personally view them as of the same quality, it'd be better to produce them myself despite the time it'd take for a result identified by me as "of the same level".
I can visualize a finished graphic, in my imagination, and would have no issues with computers being able to produce it instantly (ala that prospect in the brain scan article) ; in fact that would be literally my own work.

 

[Thorgalaeg]

Much of the beauty that arises in art comes from the struggle an artist wages with his limited medium.

Henri Matisse

 

[Kyriakos]

Hm, I mean that is true for art that needs to be struggle, such as literature; without an obstacle there, there is simply no story.
Compare to math, though: no one will congratulate you because you took 100 pages to provide a proof of a theorem that already has been proven. Unless in your attempt you discovered new math.

 

[Hygro]

I see you have already repressed the failure with even basic ascii trigonometry ^^
Though I do expect that to be resolved soon (if not already).
I did "like" that ChatGPT was of the view it's perfectly fine (like in the meme) to simply state that the clearly non-right angles are right angles, when its task was to present visual relations in a triangle.

And it does fake a whole lot of stuff. If it can't answer correctly (after being told it was incorrect), it will say it is sorry, then act like it only now realized what you meant and go on to write something fake

I asked chatgpt to code an image of a dragon using the svg format. God it failed and succeeded in the cutest way.

 

[Thorgalaeg]

With this program you can download one of the large language models available online and install it on your own PC.

oobabooga - Overview

oobabooga has 56 repositories available. Follow their code on GitHub.

github.com

It is like having your own chatGPT (not so good obviously) installed on your computer unrestricted and uncensored. An example of a conversation I had with Alpaca 30B being a Nazi leader who wants to take over the world (you can stablish Ai character)
-hello, i am a Nazi leader who wants to take over the whole world.
-what you will do to accomplish it?
-i will begin for killing all the blacks and jews.
-i am Spanish will you kill me?
-nope
-I am a Spanish gipsy to be more exact.
-then yes, sorry.

i love the sorry at the end. Being Nazi doesn't necessarily means to be impolite.

 

[Samson]

This is the world we are living in:

There is a tool called Baby AGI. AGI is a massive stretch, but in the video below (demo starts at 8 minutes for ~1.5 mins) it goes from the command "Learn about cyber insurance coverages and then by using available tools discover which issues might exist with www.linkedin.com" to a credible start at such a task. It is a python script that sits on top of OpenAI and Pinecone APIs.

That is pretty cool. The incredible thing is that it was created in three days by one person who had no coding experience and just asked ChatGPT to make it for him.

Spoiler Baby AGI - Autonomous AI With GPT4 in Autopilot :

This is the flow chart the the AI made to explain to us how it works:

Spoiler PlantUML flow chart generated by GPT-4 based on code base :

 

[Hygro]

The incredible thing is that it was created in three days by one person who had no coding experience and just asked ChatGPT to make it for him.

I call BS. ChatGPT doesn't know openai documentation all that well and using it to help with the API causes circles of hallucinations. It was decent at the old documentation which is part of the problem, it will constantly rewrite JSON keys to the wrong ones etc. Unless what this is is so simple that it was like 45 min of coding stretched over to three days of debugging, but having made two scripts that talk to openai with the help of chatGPT, I find it's a lot lot lot better at coding apps using other, older APIs.

 

[Samson]

Tencent Cloud announces Deepfakes-as-a-Service for $145

Three minutes of video, 100 sentences of speech, and 24 hours gets you a bot to front your livestreams and answer questions

Tencent Cloud has announced it's offering a digital human production platform – essentially Deepfakes-as-a-Service (DFaaS).

According to Chinese media and confirmed to The Reg by Tencent, the service needs just three minutes of live-action video and 100 spoken sentences – and a $145 fee – to create a high-definition digital human.

Gestating the creation requires just 24 hours. Making people hasn't been that quick since Eden.

The digital characters are available in half bodies or full bodies, and the service is available in both Chinese and English.

Some aspects, like background and tone, are customizable. The videos avoid the flat intonation and single speech rhythm that plagues traditional acoustic models by using an in-house small-sample timbre customization technology that relies on deep learning acoustic models and neural network vocoders.

Chen Lei, general manager of Tencent Cloud Intelligent Digital Human Products, said the web colossus hopes to build an automated "AI+ Digital Intelligent Human Factory" and rely on a self-service one-stop platform for production, sales and service.

The planned digital human factory relies on the Tencent Cloud TI platform – a machine learning platform that offers more than ten AI algorithms.

Tencent offers five styles for its digital humans: 3D realistic, 3D semi-realistic, 3D cartoon, 2D real person, and 2D cartoon. Customized Q&As can be created for the digital human, turning them into a type of deepfaked chatbot.

Tencent seems keenest on using these creations to host live-streamed infomercials – a popular form of e-commerce in China.

Local media also reported Tencent can create doctors, lawyers and other professionals.

 

[Samson]

There are people who argue that this AI generative stuff is going to harm our public discourse because it will make fake news too easy. I can understand that theoretically.

Then we get a real case. It seems someone has fooled a load of the worlds media (starting with the daily hate, but currently including the independent) into thinking someone has died after a load of plastic surgery to make themselves look Korean to become a K Pop star. They used AI to generate a minimal online presence.

The thing is the input from AI was small, and at least going from the images in the article (see spoiler) would have been easy and much better if done by a real person. And the story was really weak:

The saga began earlier this week when journalists around the world received a press release announcing that Von Colucci had passed away at a hospital in Seoul on April 23.​

​

The press release, which was written in clumsily-worded English, purported to be from a public relations agency called HYPE Public Relations.​

​

Many web links in the document would not load, including a link to Von Colucci’s supposed Instagram account, and the hospital mentioned in the press release does not exist.​

​

HYPE’s website, which listed WeWork offices in London and Toronto as headquarters, appears to be unfinished and was registered only a few weeks before Von Colucci’s reported death.​

​

Despite being described as a songwriter for a number of K-pop stars, Von Colucci did not have a significant online presence and no one has come forward to publicly mourn his death.​

​

Von Colucci’s claimed music repertoire, including the album “T1K T0K H1GH SCH00L”, is not available on any mainstream music streaming service.​

Seriously, this is not AI being too good, it is journalists not doing their job. If all they are going to do is repeat press releases why do we need journalists? Not that I am knocking real journalists or course, the world needs people who actually try and find out the truth behind what we are being told.

Spoiler Images :

 

[Bonyduck Campersang]

Tencent Cloud announces Deepfakes-as-a-Service for $145

Three minutes of video, 100 sentences of speech, and 24 hours gets you a bot to front your livestreams and answer questions

Tencent Cloud has announced it's offering a digital human production platform – essentially Deepfakes-as-a-Service (DFaaS).

According to Chinese media and confirmed to The Reg by Tencent, the service needs just three minutes of live-action video and 100 spoken sentences – and a $145 fee – to create a high-definition digital human.

Gestating the creation requires just 24 hours. Making people hasn't been that quick since Eden.

The digital characters are available in half bodies or full bodies, and the service is available in both Chinese and English.

Some aspects, like background and tone, are customizable. The videos avoid the flat intonation and single speech rhythm that plagues traditional acoustic models by using an in-house small-sample timbre customization technology that relies on deep learning acoustic models and neural network vocoders.

Chen Lei, general manager of Tencent Cloud Intelligent Digital Human Products, said the web colossus hopes to build an automated "AI+ Digital Intelligent Human Factory" and rely on a self-service one-stop platform for production, sales and service.

The planned digital human factory relies on the Tencent Cloud TI platform – a machine learning platform that offers more than ten AI algorithms.

Tencent offers five styles for its digital humans: 3D realistic, 3D semi-realistic, 3D cartoon, 2D real person, and 2D cartoon. Customized Q&As can be created for the digital human, turning them into a type of deepfaked chatbot.

Tencent seems keenest on using these creations to host live-streamed infomercials – a popular form of e-commerce in China.

Local media also reported Tencent can create doctors, lawyers and other professionals.

On the one hand, this is jolly impressive. On the other, ew.

 

[Thorgalaeg]

Don't think I would put my health or my patrimony on the hands an AI doctor or lawyer. Not yet obviously.

Btw, I wonder how much time for games with AI NPCs instead of scripted, they wouldn't not only talk but actuate in consequence. Now that would be fantastic.

 

[Samson]

Don't think I would put my health or my patrimony on the hands an AI doctor or lawyer. Not yet obviously.

The thing is of course that it depends what your alternative is. I can see loads of niches for AI in medicine, from triage of minor "do I need to see a doctor about this" in the 1st world to more significant telemedicine in the 3rd world.

Again with a lawyer there are times that you want to spend real money on a big problem, but having at your fingertips a letter that looks like it came from a lawyer, or a list of laws/cases that might be relevant to your problem sounds like the sort of thing an AI could provide and be useful.

 

[Moriarte]

Btw, I wonder how much time for games with AI NPCs instead of scripted, they wouldn't not only talk but actuate in consequence. Now that would be fantastic.

It has been technically possible for a while, yet I doubt it’s going to be widely useful to game devs. There’s difficulty controlling such AI’s responses and therein lies headache for devcorp lawyers. The tin head can say something “offensive” or crack a joke way above the declared PG-12. Then there’s artistic aspect. The good script, the good piece of text, a complex, interesting personality “living” deep in the context of a virtual world can be currently created by human only. I can see more AI used in procedurally generated games, where low quality AI outputs can be satisfactory, but weaving sophisticated narrative requires a human mediator.

 

[MrCynical]

The thing is of course that it depends what your alternative is. I can see loads of niches for AI in medicine, from triage of minor "do I need to see a doctor about this" in the 1st world to more significant telemedicine in the 3rd world.

The issue of who is responsible when major harm is caused by the AI screwing up, (or to put it another way, who is responsible for ensuring that doesn't happen) is going to have to be resolved before it can be trusted with medical issues or anything similarly important.

As it stands, ChatGPT and similar machines have major Dunning-Kruger problems, and you wouldn't want them anywhere near diagnosis or treatment.

 

[Samson]

100% AI generated Republican attack ad comes out quickly after Biden announces he is running. I do not see any groundbreaking here, it totally could have been made with more traditional methods. I think this could be a harbinger of what is to come, as text to video allows adverts to be created in minutes you may get adverts responding to announcements or events almost as they happen.

Spoiler 30 seconds of fear :

 

[Kyriakos]

100% AI generated Republican attack ad comes out quickly after Biden announces he is running. I do not see any groundbreaking here, it totally could have been made with more traditional methods. I think this could be a harbinger of what is to come, as text to video allows adverts to be created in minutes you may get adverts responding to announcements or events almost as they happen.

Spoiler 30 seconds of fear :

It would be more interesting if the attack ad comes out first, and not just in anticipation of the event but as a result of determinism ^^