AI .... AI everywhere!

Right after I posted my previous post on Hive, I got back in touch with a lot of "blockchain friends" and acquaintences, mostly via Discord DMs. One of them is @ausbitbank and we soon began to discuss the recent developments in various generative AI projects, such as:

• AI systems capable of creating (realistic) images and art from a natural language description, such as Stable Diffusion, and MidJourney, or DALL·E 2
• AI systems capable to interact in a conversational way with follow-up questions and responses, including code reading and writing in various programming languages, such as ChatGPT, and "New Bing"
• AI systems specifically designed for code generation / software development assisting, such as GitHub Copilot and Tabnine.

This list is far from exhaustive as new AI projects are popping up every day in various subfields and to test them all, let alone getting to use them all in-depth, is literally a full time job. It even took me a few minutes to try and describe / categorize these few listed AI projects in a way that does them justice,because they're so new and all-encompassing. A "new paradigm" doesn't even begin to describe them, really.

AI image generators

In his Discord server, @ausbitbank has implemented a bot which is connected to his own GPU-powered computer which in turn is running a version of the software Stable Diffusion. Like DALL·E 2, but open sourced, Stable Diffusion is an image generative AI model that accepts (natural language) user input "prompts" describing what should be visualized in the image.

How this works:

As a user in @ausbitbank's Discord server, you can type in an input prompt, and his bot then sends that input prompt from Discord to his GPU-server where the AI software running on it, Stable Diffusion, generates or "imagines" / "dreams" an image based on its trained model and said input prompt. Once an image is generated (a process right now taking up about 20 seconds or so, depending on the server load etc), the bot pushes the AI generated image back to @ausbitbank's Discord server, for the user to download, re-generate or further modify with additional commands. (This is just a high-level explanation of what is, in fact, extremely sophisticated computer technology.)

After reading how it works by analyzing some prompts other users in the Discord server had used, I entered the following command / prompt myself:

/dream `prompt low poly art flat design minimal barack obama huge sunglass Elton John style looking very serious brown golden shade colors --steps 10`

I think the generated image of former US President Barack Obama looks absolutely amazing and is very impressive, but that it would be very hard but still doable for a skilled designer to manually draw the same image.

However, observe the following prompt I gave it and the quality of the generated image:

/dream `prompt ("Donald Trump","Vladimir Putin").blend(0.8,1)`

The longer I look at this result image I generated via Stable Diffusion in mere seconds, the more I am amazed with the stunning quality and complexity of the underlying technology: a perfect blend of Vladimir Putin and Donald Trump, which image could in my opinion be used 1-to-1 on the cover of Time Magazine. I really think it is that good.

I know a thing or two myself about computer image generation and manipulation, for example using software like Adobe Photoshop, Illustrator, Blender, etc. but it would be very hard if even possible at all to recreate the same image with only human skill and conventional image design / manipulation software.

The implications of these AI generative image techonologies are immense...

I'll come back to this topic more in-depth in future posts here on Hive, stay tuned.