Source note: Local copy of a Summer Into AI 2026 submission originally published on Substack. Screenshots, video, and audio found in the post body are mirrored locally when publicly accessible; profile and avatar images are intentionally not copied. Read the canonical post.
Myles pushed up a great story for week2 about building a custom AI in a browser- which, hey-this is a cool idea!
-and that article forced me to actually push up one more thing.
I built something that takes a slightly different approach-the something being about custom model training. Earlier in the week I shared a story about Zezek the corporate goblin. Maybe you caught that, but if not you can catch up here-
Instead of using unsloth and doing loras over qwen, I wanted a much smaller base model-a LLM that could fit on some floppies.
Using nanochat, I took the Zezek training data and custom trained a depth4 model. That’s not terribly interesting. What is interesting is that you can load the entire model in a browser window-chat with the model, and even modify the LLM weights via the Goblin Brain Manipulator.
The Goblin Brain Mutator
While this isn’t something that extends the LLM’s memory like what Adam was working on with the forge, what it does allow you to do is modify the structure of the LLM with broad strokes. For example, consider the mutator panel. With this you can not only add layers, but you can also try applying specific values to the model-a kind of “YOLO-training-in-realtime” that allows you to play with sentient goblin minds in all sorts of “goblin approved” ways.
Did I mention the model runs in your browser?
This was done on a box with a 3060-yes 30-60, so if you have any computer hardware within the last 3-5 years you’re going to be able to run this.
With a little creative hacking I was able to get a convenient javascript solution that’s performant and works in a browser to host the model-which opened up a great new thing with the bus architecture that’s operating inside the tab. The inference is visualized!
While you can download loras of Zezek, perhaps you’d like to poke around a custom LLM trained last night and made available today. The entire model is about 25 meg-so, it should fit on just about anything.
Warning: I did not check phone support. This may not work well on a phone.
Chat with Zezek on Desktop Here
It’s really interesting to tinker with these model weights in memory in the browser. For example, the model’s coherence comes quickly apart and you gain an appreciation for just how fragile these minds really are.
You know what’d be interesting? Setting up Zezek with a local model in a /goal or /loop with a single directive. It’s also crazy to think you can just have multiple tabs running with Zezek, each one doing whatever it is that Zezek does.
It’s not much, this model is tiny-but it is a neat little glimpse into how these models work. On that note, I don’t have a writeup prepared for how the base training worked-it’s effectively a nanochat setup. If you’re familiar with that, it’s nothing shocking there.
For the technically minded they may want to see this post and what I covered here.
Original source
Canonical Substack URL: https://advisoryhour.substack.com/p/summer-into-ai-week2-submission-zezek.