AI
A real photo won an AI competition. It was disqualified
You can have 100 users with 12.88 tokens a second each going using a 3090 RTX.
Google has custom LLMs
ChatGPT is watching you. This article is a year old
Vector Databases for AI
What we learned from a year of building LLMs
ChatGPT 4o has cloned users voices for some unknown reason
Math answers to AI by wired
Grok 2 is out on X right now. It requires a paid edition of X, but has very few restrictions. Kamala Harris bearing Donald Trump's baby has already been made.
https://ghostarchive.org/archive/hYvtc
Google DeepMind can solve difficult math problems now
OpenAI has a new SearchGPT that will become a search engine for you
An old google search trick might help with hallucinations
Grow has a new overpowered LLM
Games used to train AI
Uncensor any LLM with Abliteration
Claude 3.5 vs ChatGPT-4o
AI Beauty contest
Generative AI imaging done on a Commodore 64
Raspberry Pi AI edition
New super computer hoped to be used to reach AGI. Could be in computing as well.
AI: Image 3
Imagen 3 is now open
It's on Google Deep mind and is doing fairly well. I haven't tested it.
AI: Flux
Flux has entered the image creation market
Another article on Flux
AI: Stability AI
New things to do with Stability
Stable Diffusion 3 released
The company went through a rough spot in May
SAI: Agreements
Stability AI has a new license
Stability AI has a much nicer user agreement program
AI: Google
Google has a new LLM… again …
Google Gemini will help create. An interview with a promoter of the program
Google shows off what it intends to do with AI and it's other services. I asked it how many links there are in this file, and it said it couldn't do that.
These English PhDs helped create the newest Gemini. They couldn't make tenure with that either.
Ethics in AI chat with Google tech leaders
You can open Gemini in a mini window
Gemini has Imagen 3
AI: ChatGPT-4o
Voice stuff for ChatGPT-4o will be released soon
Open Source LLMs that can beat ChatGPT-4o
It has a face!
How good is it at programming
AI: Agents
Agentic AI will soon take over for the AI we have now.
How Agents work
AI Agents may be the future
How Agents are taking over
Agent engineering
AI: Creations
A user tried to create a videogame character that has visual and logical stimuli as if it was actually in the game. It was Lara Croft. This article is a year old.
Gaussian captured forest with Millennium Falcon placed and animated in Unity. Video is cool, but not archived.
https://ghostarchive.org/archive/mGH71
AI: Lists
News update on July 3rd. Elevent Labs has famous voices now. Runaway has an updated video maker. Other stuff.
https://ghostarchive.org/archive/OUip9
LLMs being used in the real world
Introduction to Language Models
From a year ago. This is a list of words to control MidJourney views to make things more cinematic
Various tutorials to learn AI
Flex vs MidJourney
Lists of lessons on LLM education
AI: Tutorials
Using AI to create 3D printable objects.
Isometric videogame scene. This is a year old, and look at everything that got archived.
Interpreting Weight Distribution
How to train AI without breaking the bank
Simulated Data in Python
How to install Flux to ComfyUI
Creating graphics and coding for a videogame with AI
https://ghostarchive.org/archive/AdlAr
How to test Machine Learning systems
Character consistency in MidJourney a year ago
How to turn your AI into a scalable reality
How to become a deep learning specialist
An exhaustive introduction to BERT or baking into LLMs
Creating a music video
https://ghostarchive.org/archive/Y52jn
Building a virtual world. It's also an interview.
Using a Small Language Model on the Edge Mini Computer
Make an animated movie by making quick still arts, add AI to create the actual image, then another AI to animate.
AI: Sref
MidJourney has a new system where you can source reference an art style you like. This Twitter thread talks about how it has changed things. Nick st Pierre would know.
https://ghostarchive.org/archive/0PwGZ
How srefs have changed MidJourney
A list of srefs
Top 10 srefs from my favorite stef guy. He also has prompts to control the image a bit as well. So there is some really nice stuff.
https://ghostarchive.org/archive/KCmOt
Groovy Sref
Old photo Sref
AI: Videogames
Tiny Glade creators discuss procedural generation in game design
Procedural generation tutorial for unreal engine 5
Procedural Green Fence maker
Actors for videogames are striking and Amazon says AI needs to step in to replace them.
Former GTA Dev says games should use gravity and AI to create animations
Odyssey is a Minecraft AI program
DeepGame is an AI based text game
Procedurally Generated cities
AI videogames and how they show the way we think
Nintendo thinks Gen AI is cool, but they don't need it because of their skills
Doom made by AI in real time without a game engine at 20 Frames Per Second. Google calls it the GameNGen.
Flux is overrated. It is a huge model, at least compared to Stable Diffusion, which does allow it to produce "better" images. That is as long as you like very narrow depth of field and blurry images. It also does not have a negative prompt, which is what allowed Stable Diffusion based models to be so competitive with stuff like midjourney while being able to be run on something like a 3060.
There is also this huge double sided push to get AI off of consumer hardware, and into the cloud. On the video card side, Nvidia isn't making cards with more VRAM, and on the other side the AI companies are making their models just large enough that you cannot run them on consumer gpus and thus have to pay them for tokens. The price difference between a 4090 (24 gb for <$2000) (which CAN actually run flux and LLAMA 13B) and a business class gpu (RTX A6000 48 gb, for $5000-$6000k, or a A100 80 gb for ~$15000 which can run LLAMA 65B) is insane.
I didn't see the second paragraph. I agree completely with you. I want to get more into running on my own rig, but that's a lot of money I don't have.
Don't worry bout it, it was an edit. I just thought i would add some more context to the current ML hardware side.
It is an annoying thing. If I do any nda work, I have to use something like SD or not at all for concept art.
I think SD is useful because of how adaptable is. It's a perfect in between AI