How much steam does the current AI trend have?

[Some_porcupine]

anim from poses

 

[bnuungus]

My guess is that the only industry that may be mostly displaced from current ai tech is animation, though probably not for a few years and definitely not from any current ai image generation program. You would basically need something trained on anatomy and programmed to take and impose characters based on detailed concept art onto basic planning drawings to create high detail animation frames. It's something I think would have to be specifically designed for the task and would displace the small armies of animators currently needed to create animation down to just a few.

Corridor Crew has a video where they make an AI generated anime short. It's interesting because it seems like the amount of work required to create this is the same as if you animated it by hand, but the work is just different. Instead of taking care to draw the images you are taking care to make sure the AI works right. And surprisingly they have to do a lot of VFX work themselves. I'm not sure if the technology will cut down on the work required but I think there will always be an aspect of human interaction involved.

 

[Sketch Relics]

Corridor Crew has a video where they make an AI generated anime short. It's interesting because it seems like the amount of work required to create this is the same as if you animated it by hand, but the work is just different. Instead of taking care to draw the images you are taking care to make sure the AI works right. And surprisingly they have to do a lot of VFX work themselves. I'm not sure if the technology will cut down on the work required but I think there will always be an aspect of human interaction involved.

That's rather impressive, as the tool becomes more fleshed out it should become easier to use and produce better results like how flash animation or 3D model animation has progressed over the years. Though the way it's being used here is a more advanced version of rotoscoping rather than what I was talking about. Kinda similar in principal though.

Oh that actually reminds me, we can also use this to recreate 80's~2000's lighting effects in animation that can't be done anymore due to the move to digital space. Stuff like this.

View: https://www.youtube.com/watch?v=6sFr8xTJmkA

Computers apparently suck at modeling stuff like the lighting used for the sun at 00:25/00:45 and that neon laser effect at 00:10/ the transformation sequence/the laser blades/ panels in space that popped up alot in the 80's and 90's. Due to it being done actually by lighting tricks through the animation cells.

 

[Eden]

I watched this TED Talk about Khan Academy integrating GPT-4 this week:



Later in the week, my school got an overpaid consultant come to talk to us about proooooompt engineering with it's own unique acronym and shid because ofc:

'P.R.E.P - Prompt, Role, Explicit, Parameters' (incredibly cringe article, beware)

But hey, at least I got a blog post out of it:
HERE BE AI
¯\_(ツ)_/¯

 

[Some_porcupine]

to me, it is just browser. like, it is that people are comfortable and all that.
it is just tool.
just searchbar.
someone has to make content for it to scoop answers from.
until it cant think on its own, which could give it empathy, only our flaws are our demise, not tool.
only its misuse.
it cant use itself...

 

[redacted345]

Google "We Have No Moat, And Neither Does OpenAI"

Leaked Internal Google Document Claims Open Source AI Will Outcompete Google and OpenAI

www.semianalysis.com

apparently leaked from google memo saying open source will get ahold of tech innovation on AI. The timeline at the end reads like the intro of a scifi dystopian movie.

 

[qwerty]

I'm not sure if the technology will cut down on the work required but I think there will always be an aspect of human interaction involved.

Just wait until they use AI to create prompts for AI...

 

[W1NTER]

I think AI related technology is still in its infancy when it comes to widespread public usage but will slowly get more and more popular and ingrained within modern day media and online interaction

 

[alCannium27]

Just wait until they use AI to create prompts for AI...

Already done

bigcode/ta-prompt · Datasets at Hugging Face

We're on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

Magic Prompt - Stable Diffusion Prompt Generator

Generate and improve your Stable Diffusion prompt magically! Simply enter a few concepts and let it improve your prompt. You can then diffuse the prompt.

stablediffusionweb.com

And these are just the ones i can find with 5 seconds googling for a link and quick summary. Of those that's open sourced, text-to-text models are very common, and there are now prompt generators for many scenarios, such as program code generation and image generation like the above. I recall at least one ChatGPT based prompt generator for Midjourney, and possibly for Stable Diffusion -- certainly at least one openGPT prompt generator. Other niche applications are sure to exist, just off the radar at the momemt.
Apart from their current premature state and high cost, what they lack is the ability to understand orders and cardinal numbers, though the largest LLMs perform better in that regard than Diffusion models -- t's after all, very recently that Midjourney learnt humans have five fingers. Another weaknessthey have no initiative - but is that a bad thing? Do we need a screwdriver to have wants to be useful?
There have been very promising deveopments in low-cost ML training methods, such as LORA and hypernets, that can be trained with a significantly smaller dataset and computing power on one GPU, and though they can't function by themselves, they can perform just as well as big models when they are used to augument existing large models, which is what the Google leak soc was talking about ofc -- this has the potential to decentralise ML industry, after all； will it? i guess only time would tell.
Also promising is the fast development in local deployment of AI models. My first post here a little while back had me pondering how long before we can finally have AI models run on one's personal computer (the example I used was a personal AI voice assistant that doesn't require constant wifi connection to finction), and a week ago I managed to do just that on a RTX2080. These are lightning fast developments, and yesterday someone just figured out how to generate coherent videos with "unlimited" length and put their model out into the wild. May have to try that one out myself.

 

[LostintheCycle]

I watched this TED Talk about Khan Academy integrating GPT-4 this week:

Wow, and I was just about to hop onto Khan Academy after a glowing review from a friend. I don't trust AI to teach me calculus, I'm gonna stick with books. GPT is terrible for anything technical, no matter what undergraduate using it to solve basic loop questions tells me.
I checked out your blog post too. Can't agree more on the point of education. People have been told that somehow, more advanced technology makes for more effective systems if applied willynilly. They are so caught up in this fantasy that they've forgotten what the problems are in the first place, if they think being able to talk to a river will solve it. What we really need is to investigate more fundamentally alternative systems, like Finnish education, or Waldorf schools.

 

[redacted345]

, if they think being able to talk to a river will solve it.

"Hi Matthew, I'm the representation of the Rape of Nanjing. I'm kinda sad but you can ask me anything...my favorite color? Uh that's a hard one! let's see..."

 

[Eden]

"Hi Matthew, I'm the representation of the Rape of Nanjing. I'm kinda sad but you can ask me anything...my favorite color? Uh that's a hard one! let's see..."

Fuggin kek

Edit: Lol, sorry I'll come back and make this post more substantial.

Edit2: Man, was reading through the Character.ai thread and it was pretty funny too. U fuggers all a barrel of monkeys.

 

[Andy Kaufman]

View: https://www.youtube.com/watch?v=xoVJKj8lcNQ

 

[Dr. MacGutsy]

Holy guacamole man.

https://www.whitehouse.gov/ostp/ai-bill-of-rights/

 

[pickleman]

I'm kind of baffled that this hasn't gotten any responses yet, is Agora dying? To answer you're question though, I wouldn't say we're near a singularity yet, and I'm still skeptical of whether we ever will reach a true singularity where Tech and AI truly manages to replace the utility of human beings. But I totally agree that we are witnessing something new, and paradigm shifting (please excuse the cliched terminology).

It feels that it was just a year or two ago (because it was) that people still viewed AI as something off to the distant horizon of the future. But AI's demonstrated it's abilities and there is no denying that its far more impressive than many imaged AI would be at this point from the perspective of even just two years ago. The fact that Microsoft is planning to implement AI as part of Bing's search feature (which means that Google might/probably will follow along) means that the net-browsing experience as a whole will change. To be fair, search engines have been increasingly disappointing and these days just point out how hollow the net really is, hence the Dead Internet Theory, which I've started to see even mainstream outlets discuss.

I've been somewhat reluctant, but I've been working recently on a new documentary about this very topic. In many ways it will be a spiritual sequel to my Lain video, as I want to expand on a point I concluded about AI that I see few people discussing. What I find most unsettling about AI isn't Artificial intelligence itself, since their are plenty of types of artificial intelligences that I think are perfectly harmless (NPCs and enemies in video games for instance). It's AI connected with the internet specifically which is a recipe for some pretty spooking horrors. Because the internet itself, especially these days when nearly everybody is connected to it (as Lain emphasized), is really sort of a repository of the collective data output of humanity. What these models of AI are doing, is feeding themselves on all that data, and developing outputs off of it. What this type of AI really is, is a synthetic collective unconscious.

All the outputs of AI art for instance, are in essence the dream images processed by a robot. It's a robot drawing from the well of collective humanity and spitting out results from that based on inputs. Just as the unconscious houses an ocean of data that we aren't conscious of, but strives to communicate with us in our dreams and via certain meditative practices. The unconscious of humanity (as expressed on the web) is being organized by these AI. That's what really strikes fear into me. Since what we're really dealing with aren't robots so much, as robots with access to all the conscious and unconscious expressions of humanity itself that's available online. Inverted cyborgs. Not humans that have mechanical add-ons, but machines with human consciousness as its add-on.

This is a very good assesment

But what about training data fragmentation? Many people are increasingly becoming more private with their data including companies.

Also there is the considerations of silo'ed data. For example, at work we use microsoft products which are obvious sources of training data for AI. But this data is only available to microsoft AI. Likewise consider Google/Apple. They have access to data on your phone, location, texts, etc etc.

Sometimes I think that capitalism is actually mirrage, a means for the deep state to cercumvent laws made to control the government. Since corporations have much more freedom then government and do not have democratic pressures on their leadership (while typically providing neccessary utilities).

The only way to get a proper data set is to merge the data collected from all these companies, and I think the snowden revelations with PRISM and its members (microsoft\google\amazon\apple\etc) you can have a data set which can train a capable AI. Perhaps this is why they say "diversity is our strength" its because of the diversity of training data from many different humans.