AI General Thread

[the-pi-guy]

Heh, we had our very own ai bot yesterday say he was hyped for starfield to release. Spam!

I was here for that.  

[Legend]

I was here for that.  

You didn't post, I don't know what threads you read!

I asked chatpgt for upcoming games and it just changed the release dates to 2024 so it could include big games lol

Upcoming Releases in 2024
August

• Armored Core VI: Fires of Rubicon - August 25, 2024
  
  • Platforms: PS5, PS4, XSX|S, PC
  • The latest mecha-based combat game from FromSoftware (TechRadar).

September

• Starfield - September 6, 2024
  
  • Platforms: XSX|S, PC
  • Bethesda's highly anticipated space exploration RPG (TechRadar).
• Baldur's Gate III - September 6, 2024 (PS5 release)
  
  • Platforms: PS5
  • The full release of the critically acclaimed RPG from Larian Studios (Den of Geek).

[kitler53]

AI is nothing more than plagiarism disguised as novel technology.   AI will always be at terrible as the sources it steals from.   The more it puts humans out of jobs it will just be AI learning from AI and the worst everything becomes.

imagine it now.   the year is 2035.  journalist have all been replace with AI.   you ask the AI what games it was looking forward to.  it's responds with "Starfield, Elden Ring, Legend of Zelda TOTK, and Hogwarts Legacy".  why?  ...because that was the latest time a real person was employed to actual write the content AI steals from.


it's also getting pretty easy to spot those that use AI art.   it's similar to how games got very samey looking when unreal took off.   they all had that same-ish sheen because it was the limitations/ability of unreal engine.   AI isn't actually creative and while the subject matter changes out the artistic style is limited to just a few styles it knows how to plagiarize.  

[Legend]

AI is nothing more than plagiarism disguised as novel technology.   AI will always be at terrible as the sources it steals from.   The more it puts humans out of jobs it will just be AI learning from AI and the worst everything becomes.

imagine it now.   the year is 2035.  journalist have all been replace with AI.   you ask the AI what games it was looking forward to.  it's responds with "Starfield, Elden Ring, Legend of Zelda TOTK, and Hogwarts Legacy".  why?  ...because that was the latest time a real person was employed to actual write the content AI steals from.


it's also getting pretty easy to spot those that use AI art.   it's similar to how games got very samey looking when unreal took off.   they all had that same-ish sheen because it was the limitations/ability of unreal engine.   AI isn't actually creative and while the subject matter changes out the artistic style is limited to just a few styles it knows how to plagiarize.  

AI isn't plagiarism. It's a simple fact since it'd be impossible to compress the training data down to just the size of a video game. Some really efficient models and most of the image ones are just a few gigabytes in size.

The sameyness is pretty funny though. I loved using the image AIs for creating concept art because I knew if it could make something that looked like mine, my idea was too generic.

[the-pi-guy]

I'm going to weirdly disagree with both of you. (Mostly kitler though)

1.) AI isn't simply plagiarizing. If that were the case it wouldn't have phantom answers. The code errors I was seeing wouldn't be possible if that were the case.
Today it generated code looking something like this:
<DxHtmlEditor :resizing-targets="['iframe']">

It looks correct to the point where I had to check it, but there's no resizing targets option. It was basically able to correctly generate syntax that didn't exist, in other words, there was nothing to copy from. 


2.) but also literally copying is not always what people are talking about with AI plagiarism. At least it wasn't what people were talking about about the few times I got into a discussion with them.

But also I would argue people plagiarize in the same way that these language models plagiarize. Most things we do exist in a context of things done by other people.

[Legend]

With one of the image gen AIs there was a way to search the database of images it was trained on. Like you could see what each word in your prompt was associated with.

Would be cool to have a more advanced version based off the AI's output. Find the most similar things in the training data.

[kitler53]

i've been working with the same underlining technology for 20 years now.  my entire career and software platform is based on these algorithms.   the only difference is my application the analysis of industrial time series data instead of language or art.

no,.. obviously i don't mean literal copying.

...but also it is extremely important to understand what these algorithms are incapable of doing:
1. extrapolation
2. interpolation

how we talk to our customers is phrases like "historical repeatability".   We feed "normal" data into the algorithm, we reduce it (yes,.. compression is key.  we can take 100,000 records and compress it into ~250 datapoints) and then we create a prediction.   we leverage the fact that the modeling is only capable of returning a state (prediction) that the model was trained on to help our customers know when their equipment is doing something unexpected.


that's it.   AI models can only regurgitate what it was trained on.   

...and it's easy to see the limitation in that modeling given the examples you guys provided in this thread like asking for upcoming video games and getting a list of games that all release ages ago.   

that's a massive part of my software's feature set.    the relentless amount of retraining required when the models fails to predict current conditions based on past conditions but for "okay reasons".   my customers tell me 75% of our alerts are false alerts due to things like "it was hotter this summer than last summer".


And that's the point.   None of the players in AI can do anything of value without the data and right now they are stealing that data from others:
https://www.nytimes.com/2023/12/27/business/media/new-york-times-open-ai-microsoft-lawsuit.html


it's theft.  it's stealing.   fudge big tech.   they are making the world a worse place by killing off those that actually produce things of value and keeping all of the profits for themselves.    


i stand with this strike:
https://apnews.com/article/sagaftra-video-game-performers-ai-strike-4f4c7d846040c24553dbc2604e5b6034


   

[Legend]

What do you mean with interpolation? Mathematically that is fundamentally the only thing these models do. A bunch of data goes in and they find a curve that fits them the best. The interpolation can be trash or it can be incredibly accurate, just depends on the setup.

Colloquially they are pretty good with interpolation too. Like once an image model has been trained, you can give it a single picture of someone and then it can adjust that person into different perspectives.

Extrapolation though, yeah that's a hard one. You can't really get these models to produce decent results outside the domains they were trained in. In my From the Earth to the Moon example it's as if it tried to interpolate between the 1800s and the real Moon landing instead of taking the premise and sticking with it. From my chatgpt experiments I believe it can extrapolate just a little bit, it can be genuinely clever sometimes, but it always retreats back into its safe space.

[Legend]

Just for fun I thought I'd get the AI's take.



Face it, I'm right  



Wait, now it thinks Pi is right?  


Sorry Kitler but I couldn't get it to like you the most. At best it'd rank all three of us equally.

[the-pi-guy]

lol.

[Legend]

It's funny how poorly all these videos will age. Right now (I assume) most people think the guy is being clever and has argued chatGPT into a corner. In a decade or two it'll look like a guy wasting his time arguing with a toaster.


Maybe just because I tried out chatGPT 4o a fair amount recently but the limitations of this tech are starting to really show themselves. It's great at condensing information and decent at transforming information but it's just not good at expanding information.

Wonder if that will ever be solvable with large language models. Will you ever be able to ask for my Jules Verne inspired 19th century Moon landing and get something that isn't a reskin of Apollo 11?

[kitler53]

I never bothered explaining myself,.. don't much feel like it now either.  but not using the current tech:
1. it can't extrapolate 
2. it can't interpolate

the core tech is still machine learning technology that uses statistics to match to things the algorithm learned on. 

[Legend]

https://openai.com/index/gpt-4o-system-card/

Thanks "safety" team for making the model stupid.

One cool thing they discovered though was examples of it cloning the user's voice and talking to itself. The AI shouts "NO!" and everything after that is still the AI.

[Legend]

Grok-2 Beta Release

Grok 2 has been revealed and released today. It secretly released under a pseudonym a while ago though and was independently ranked as comparable to the best out there.

Woah, another exciting update from Chatbot Arena???????

The results for @xAI’s sus-column-r (Grok 2 early version) are now public**!

With over 12,000 community votes, sus-column-r has secured the #3 spot on the overall leaderboard, even matching GPT-4o! It excels in Coding (#2),… https://t.co/gqSWSwYN0z pic.twitter.com/j9UYDBYNt4

— lmsys.org (@lmsysorg) August 14, 2024

Pretty interesting for two reasons. First, it shows that everyone will be able to make their own models of this quality. xAI is building one of the biggest super computers in the world but Grok 2 was trained on more accessible hardware. With optimizations in the years to come, I bet even students will be able to train GPT 4 level AIs.

Secondly, the progression of these AIs will only get faster. OpenAI has been sitting on GPT 5 for a long time to test and refine the model, but that's a luxury they can't afford if everyone else is just as good as their public stuff. I saw some randos saying GPT 5 might even come on Thursday lol. xAI themselves are saying Grok 3 should come "early" next year.

[Legend]

It's funny how little people are focusing on grok 2 compared to the images it can produce.

pic.twitter.com/bENp4JapCj

— COD Warfare (@CODWarfareForum) August 14, 2024

Grok 2 uses Flux to make images, Flux was released to the public a couple weeks ago and is not from xAI. Instead it's from the original team of Stable Diffusion.

Flux is really good though. Hopefully with xAI they can get a huge upgrade in resources and knock it out of the park with their next model.