The last project I was working on was Control Room, a complete environment for novel creation end-to-end: from initial idea to final draft, utilising a team of agents that form a writing team. Basically, a writer’s room simulation as an application.

So, where is it, then?

The program is in an advanced state, and the state machine works – agents read and write issues for each other, you can work with agents in the editor portion of Control Room, it has a versioning system (a github-like on your local storage). What does not work is the closed loop. And not because the program can’t pull it off, but for very annoying reasons: Currently, models can follow 4, maybe 5 steps in a tool chain before they collapse and hallucinate. Output meant to be JSON are plain text, tool calls and receipts created in desperation because the model is overwhelmed. And that’s Claude and GPT. You can keep it very simple and run it on a SOTA model, but it will be wonky. What you could do is run the agents on Gemma or Qwen locally, and give them prompts in the Monaco editor like “Let’s continue this scene” or “Help me analyse this character arc”.

So, Control Room is stuck in a place where it’s “a VSCode for writers”. If that’s enough for you, download and use it. But it’s not what Control Room was meant to be, because the capabilities just aren’t there yet, and that’s frustrating, especially since I had hoped to make the teamwork run even on local models.

So, what does this mean?

It means Control Room is currently on ice, and I may or may not pick it up again later, when local models become more capable. Qwen3.6-27B is already a step in the right direction, but still too heavy to run it on a typical writer’s machine. Hell, even on an average gamer’s rig. You’d need decent VRAM to make t work properly (you better budget for 32k tokens for thinking alone), and the way KV cache works, future models need a better method to deal with context windows. Nvidia’s Nemotron series is an experiment with a hybrid architecture (with Mamba), and more efficient models must be high on every lab’s priority list to reduce inference costs. But we are not there yet.

What AM I up to right now?

I’m currently going down a rabbit hole! It’s a crazy adventure. I found an architecture that learns during inference and “grows weights”. That means this architecture’s models have true memory and continuity. They do not start from frozen weights, like transformer models. They remember you, what you talked about, and they get increasingly better at everything you do with it.

I’m still putting together the training data and will start a series about the project as soon as that’s done and the training starts. The first few entries will describe the model, the architecture it’s build on, and what I’ve seen during my first training sessions (I can already say with confidence that training this is not like transformer training at all: no throwing terabytes of internet crap at it in the hopes it sorts itself out). I’m creating the material myself, using nothing downloaded anywhere, and I’ll talk about the corpora and the curriculum when I get there, but this is actually pretty tough work. But I’m now working on the last part of the data (I’m creating material for the German, Japanese, and Chinese languages), and I hope to be done with it soon-ish. Maybe a week or two more.

My new project is called Ninereeds (yes, that’s the dragon from Terry Pratchett’s novel “The Colour of Magic”), and if you’re curious, check out the GitHub repo. But there’s nothing usable there just yet, so you’re not missing out by just waiting for me to post about it here in more detail.

Anyway, that’s it for now. This post was a heads-up: I’m not dead. I didn’t forget my website. I just had nothing worth saying to post here yet, but now I’m getting there.