For sure , I actually just did it. I took a year of my own traces with grok, gemini, chatgpt, claude and trained a model not on the method itself, but on the inner workings of the whole closed loop and all my ai projects. It’s a Qwen 3.8 27b in GGUF, which was already decent at self-correcting and here’s the wild part: it reasons like a DeepSeek now. And not by my grading. I showed its traces to actual DeepSeek and asked what model it thought wrote them. Its answer was the V3 family with thinking mode, or an R1 distill. Which is pretty dope in itself. First time training a model so probably a lot of quirks I need to iron out. I’m still going through the training pairs before making the dataset and model public also aiming it at anyone building AI runtimes, especially folks doing steering in Rust and latent-space exploration.
The part I’m most excited about, taking that same GGUF and running it on the actual Niodoo runtime. Weights trained on a year of the loop’s own traces, steered by the engine that produced them. Full circle. Gonna be sick I’ll let you know how it goes.