Originally published as an Advisory Hour Substack post on 2026-08-16. This first-party site copy preserves the text, localizes public supporting media, and connects the record to its project trail.
Original post
"For me a stained glass window is a transparent partition between my heart and the heart of the world." // Marc Chagall
Unsloth let me build a copy of myself on a Qwen base. What this unlocks is a neat toy. I can chat with a shallow clone of myself. This is not a hot party trick. It won’t win anyone free tickets to the Cubs. It is, however, pretty neat to do and teaches unexpected lessons about the importance of prolific writing along the way.
I’ve tried using AI tools to try and figure out how I write before with varying degrees of success. If you read the LeadershipOS on my site, well, that’s an attempt at doing the “have AI figure out the principles” approach. I used AI to extract the prompt recipes that could be repeatable as well as key pillars of thought. These I include in my agentlab writeups in the LeadershipOS as a way to bring in my specific way of thinking to new agents I’m spinning up. The principles I work with are battle tested. More importantly: they work. I find my agents (Hermes and OpenClaw) perform better on the tasks I want done after they’ve taken time to read those articles.
This custom model project was one of those late night ideas. I’ve this growing mass of content on my website. It grows by the day. In day dreaming about what I might do with the site, I had a funny thought. What if I dumped the entire website as training material for a LLM. What would that do?
This article is a reflection of my first attempt at using my larger website corpus to create an Eric “V1”. I could always just talk to myself in the living room, but the dog and family start reaching fast for the telephone to call emergency services. Chatting with myself digitally feels like a safe bet. My inner monolog is so strong. The phonological loop likes to narrate my day at high verbosity. There’s a part of me that really wonders what happens when that loop is confronted with itself. Put enough mirrors together and its new physics. The gravity of this experiment hasn’t set in yet.
Where’s the demo? On a quiet machine sitting underground somewhere in the midwest. We can’t have too many Erics running around. Not yet, anyway.
The initial “v1” of the model wasn’t all that incredible. It’s a LLM. It was bound to the small amount of test data I provided. I’m sharing this reflection because in my learning I hope you’ll learn something, too. The exercise of building a LLM on myself has been like a gigantic light bulb getting flipped to the on position in the darkest of nights. It taught me something. It taught me something really unexpected.
I need to write more. A lot more. You probably do, too. More everything.
I write a lot. I’m a big fan of the morning pages of the Artist’s Way. I write a lot in and for places that I can’t put in the public. It’s just part of the NDA bound trade. I’d love to get all that content and make it LLM parquet magic. Unfortunately, data retention and corporate policies have an opposing view. So it goes.
What the LLM training has me realize is that all of that writing is a sad loss. I can’t use it to build a deeply robust model of my thinking or writing unless I start thinking of new ways to do so. I have intuited thoughts that I believe can overcome this.
I can do so by writing a lot more.
But looking over what I do have available, I think there’s a lot of other context that will be put into the pipeline. And I look forward to seeing what V1.1 brings. The main focus right now is in building a daily training process. What I learned building Zezek is that it is possible to build several models in a day. I already post on substack daily. What if my substack articles, which end up on my website… also found their way into a training data pipeline, and then this custom model of “me” was itself updated per day? It’s a kind of quasi continuous learning loop. It’s based on what I’m learning.
This could be a very interesting experiment. I asked the other Eric about this.
I like how the other Eric is suggesting my crazy idea is going to work. I’m going to need a lot more writing.
Stained glass Mirror Notebook
The Stained glass Mirror Notebook is a simple notebook that asks of a reflection to show itself in different light. Ruled Line Spiral Notebook - Stainglass Mirror Notebook
Stained glass notebook is a good concept for what writing is like. It’s a dazzling affair of ink. I’ve read many journals brought forward to the public eye. A lived life is not a polished statue. A life is multi-faceted, with sharp edges and unexpected turns. Life comes in shades of colors, and when viewed holistically then the full shape of life is revealed.
Connected work
- Custom Model Training Overview: Start with the practical framework for deciding when custom training is worth doing.
- Leadership OS: See the public operating principles that became part of the training corpus.
- Model Training: Browse released models, evaluation work, and local model experiments.