What's the Update?

Short and sweet, it's an update to Tudle's front end. Mobile users (100% of the userbase), will notice going forward that quickly tapping letter buttons no longer causes the browser to attempt to zoom in. It was quite the annoying behavior. I'd call it a bug, except for some reason mobile platforms consider double tap zoom a feature. Anyways, that behavior has gone the way of the dodo now.

What Progress has Been Made?

We'll order this section from least amount of work to most done...

Notes

While pretty much the simplest webapp a person could ever make, Notes is going to cost money to host and run, which is something we don't have right now. The site is predicted to cost $150 in August. This includes the Azure bill, domain, and various subscriptions (like Github) that I pay for to help me develop. For that reason alone, until we get some more sponsorship, start generating revenue, or I find a way to reduce cost without wasting a month of my life, I probably won't post SuperSlowNotes as an application anytime soon.

Flappy Dog

Turns out creating a flappy bird clone in JavaScript is pretty easy. Unfortunately, the hard bit is the art, something I'm not exactly an expert at. After reaching out to some artists and being ghosted, I took at stab at doing what the cool kids do and generated some art with AI. The journey started with ChatGPT. I quickly ran into a roadblock though when I realized it's quite good at generating one-off images, but very bad at staying consistent or drawing animation frames. Below is an image from that chat session.

Since then, I've been researching how to generate art more effectively. However, it left me at a pretty big standstill, so I decided to shelve the project and move over to Ceridwen for the time being.

Ceridwen

Two down, last to go.

Before we dive in let's review the concept of Ceridwen. The concept here is to have an AI LLM generate horoscopes as a character called Ceridwen, who is inspired by the mythological Welsh sorceress/goddess of the same name. The horoscope would be distributed on the site, as well as opt-in emails (for premium members, which everyone is right now). I don't really "Social Media", but I've also recognized creating Finstas (do people still use that term?) and updating posts there could also be a future feature.

Anyways, after reviewing membership costs for API access to LLMs, and enjoying the eyewatering billing process of cloud providers, I knew that Ceridwen was going to have to generate on my self-hosted hardware. This is all well and good, and I was able to successfully stand up a self-hosted LLM pipeline to generate the content sufficiently with only a couple of hours of work. The problem is... well... it was just kind of boring. Despite sometimes reading a horoscope myself, I didn't really have an attachment to the text that was being generated or see any reason why someone would read it each day, even if it was a "mystical sorceress" giving me my future.

So, I sat and thought for a while at the computer and brainstormed. That little while turned into an hour, then two, and then finally it was 1AM. Not long after that, two days later, I had the light bulb.

I let the business school training kick in, and I thought about who my target audience would be and what would make Ceridwen more appealing to them, and that's when I realized that I needed to actually create a character.

I recognized that there were some important aspects about the target audience (which I consider myself a part of):

1. They're highly educated people who spent copious amounts of time on the internet in their youths (if you don't get the joke go listen to Camp Counselors podcast).
2. They grew up in a time when fictional characters and concepts were being more defined, commercialized, and immersive. Think Club Penguin, Hello Kitty, Pokémon.
3. They form emotional attachment over time with these characters, grow nostalgia, and talk about them with friends.

So, I decided Ceridwen was going to go from an ethereal, mystic, descriptive text only figure, to an instantly recognizable character and personality that the audience could relate to. But unfortunately, that's when I circled back to some recuring problems we've seen already in this post. In order to do this effectively I need visual art, a very talented artist, and preferably lots of art content that can go with each post. Again, lacking an artist, I decided to try my hand a second time at generating art with AI.

The last stint of art generation was quite the journey. It started with remembering what little I knew about creating character art and then researching the workshop and brainstorm practices of the best artists and animation studios. Then I had to try and guide the LLM step by step into consistently generating the same character design over and over by stepping it through these concepts one by one. From wire frame to model reference sheets and character paint kits, I had the AI step through the artistic process like any animation artist would. Even then, as of today I still only get a consistent character generation image about 50% of the time. However, that's not even the biggest hurdle I stumbled across. I did most of this work with the agents hosted by the large LLM providers, not on my local hardware. Moving to my local machine and creating workflows there, I realized the hardware I have simply isn't up to the task, and that's a problem I've yet to solve.

Wrapping up, Ceridwen has morphed into what's referred to as a "Character IP". We've generated some art for her. I have a general idea of her personality (you can probably glean what it's like from her expressions), and I have no idea how I'm going to generate her content in a serialized fashion without taking out a second mortgage.

Below you can see some of the concept images used to build up Ceridwen.

Peace, Matt :)