You ever sit and just look at your multiple folders of abandoned projects? How long have you been backing said folders up from one drive to another, one pc to another?
For me it’s two decades. A few gigabytes I’ve been hauling about, essentially nothing more than reminders about how thoroughly incapable I am of actually following through on a plan without a massive effort of motivation, or some external pressure.
Let me start again… Hi, I’m Pander. I’ve recently been diagnosed with ADHD (finally) - in case you couldn’t tell. In the last few years since I learned I might have it, I’ve been finally actually making some progress on my backlog. I’ve managed to distil everything down to a handful of projects: Make a portfolio (this website) - A single place to put everything, something I can easily link to from Twitch, LinkedIn, Twitter, you name it. Keep a diary (this blog post) - Whether people care to read it or not, I mostly want to keep a log, put my chaotic thoughts in order, and have something to look back on and reference. Unify as many projects with promise as I can Learn Live2d and make an avatar Learn Blender and make Trackmania maps Put together a set of AI tools and build a modular AI system Make a multiplatform app to help create stories and craft worlds, with some way to maintain proper continuity across a shared universe Find a way to manage my time more effectively Take the AI modules and build my own LOCAL agent which can help me keep everything in order
You may have noticed that I keep mentioning AI - You aren’t imagining things. Ever since the first time that gpt-2 started to hit a couple of tech publications, I knew that the revolution was here. Yes, GPT-2, not 3, not 3.5, not ChatGPT, not whatever that megacorp ClosedAI are calling their latest model… We’re talking late 2018 to early 2019, I became OBSESSED with what the latest capability was, how much faster or more accurately the newest model performed.
Then I noticed some problems… I noticed a few of them. Nothing the AI was outputting felt REAL, uncanny images, extra fingers, malformed limbs, conversation flows that would devolve into GARBAGE after the first couple thousand characters. I instantly knew it was because the model had a limited attention window, lack of real mechanical understanding of the partial pieces of content it was cobbling together. I later learned this was called a context window, and the pieces were called tokens.
The companies started making the models bigger, more capable, able to pay attention to more tokens at once, with the ability to ‘intelligently’ trim the context window as conversations continued, in an attempt to maintain coherence. Then we got things like AI Dungeon, SillyTavern, Project folders, Retrieval Augmented Generation… And yet, none of it quite worked right, none of the attempts really understood all the issues at the same time.
This is what I want to work on, true memory storage and management, no more context limitations, no more hit and miss elimination of random tokens from the middle of a block of text. This is where the ASE comes in (AI Story Engine)
One of my first projects is to be a multiplatform app, built to run an open source LLM of the user’s choice (with some ethically sourced suggestions in the install process) This app will be able to ACTUALLY keep track of story events on a timeline, no more using an item in one scene, then having it in the inventory the next day. No more losing a weapon only to pull it stright back of its holster. I’ve been a gamer most of my life, and it seems painfully obvious to me that if a game that installs in under 1gb can keep an inventory system straight for the entire playthrough of a fully realised story, then it shouldn’t be rocket science for an AI to just hook into said system?
Maybe the AI industry doesn’t need more money? Maybe it just needs a system with actual logical rules? Why try and make the LLM itself so incredibly capable that it MIGHT handle these things, but can’t run on a high end gaming system? My PC can run 5-6 rpgs at the same time, and they all have coherent stories, but I can’t run a single LLM that fails to tell a single story? Please tell me I’m not the only person who sees the issue here?
I’ve spent the last 6 years researching the AI market and testing this tool and that tool, I know what the limitations are, what the capabilities are… Now I ‘just’ need to start putting things together. After a year of fighting with software stacks, development workflows, incompatible drivers, linux kernel errors, broken updates, and trying to get the NHS to help me deal with my brain - I think I’m ready to get things done.
Watch this space…