Rhysbot explores the whole pipeline: collect public material, review it, train models, measure them, and let people try the results. Its different chat modes use different models and different paths through that pipeline.
An existing Qwen3-30B-A3B model receives relevant excerpts from Rhys’s writing, How Everything Evolved, podcasts, X and curated notes with your question.
The excerpts are looked up at inference time. This mode does not change Qwen’s weights. In side-by-side comparison, the other Qwen answer receives no Rhys corpus and keeps its own conversation.
Open the default chat ↗Start from existing Qwen3.5-4B base weights. Continued pretraining (CPT) on Rhys’s corpus ran for 1,000 updates, then supervised fine-tuning (SFT) ran for 400 updates.
A saved LoRA adapter contains the learned changes. The checkpoint comparison runs that adaptation beside the same 4B base with its adapter disabled. It receives no retrieved corpus excerpts.
Choose “Trained 4B vs Qwen 4B base” ↗nanochat began with random weights: 4 layers, an 8,192-token vocabulary and a 512-token sequence length. It ran 2,000 pretraining updates, then a short 80-update SFT run.
This mode uses only that saved scratch checkpoint. No Qwen, retrieved passages or hidden fallback answer for it. Every valid reply shows its checkpoint run, weights hash and token counts.
It is a tiny learning experiment. Current samples are often incoherent; SFT logged NaNs from update 44. Finishing a training run is not evidence that it became a useful chatbot.
Choose “Pure nanochat Rhys” when available ↗The corpus is reviewed text with source and speaker metadata. It contains examples a model could learn from; it is not itself a model.
Inspect the corpusPredict the next token in text. Nanochat learns its weights from scratch. Qwen CPT adapts an already-trained model through a small trainable adapter.
Train on question/answer or other supervised examples. The initial SFT set contained 18,363 Anki question/answer examples and 723 public-writing continuations, rather than a broad set of Rhys’s first-person conversations or personal opinions. Those are dataset counts across train, validation and test; split membership does not prove an optimizer consumed every example.
Use saved weights to generate a new reply. Retrieval can supply extra passages at this stage. It gives the model context without another training run.
More corpus does not automatically mean new weights. The speaker-separated podcast and YouTube corpus was added after the original training runs. The frozen checkpoints have not magically learned that update. Fast chat can retrieve the newer indexed sources; a model needs an explicit new training run to learn from them.
A prompt can ask for a direct first-person reply as AI Rhys, grounded in relevant sources, with citations and clear uncertainty. That can improve conversational style without training new weights. Studying an Anki card does not establish Rhys’s belief, and a podcast guest’s statement does not become his opinion.
The synthetic voice is a separate speech model using Rhys’s authorized voice reference. It reads generated text; it does not give the language model new knowledge or turn a reply into a statement by the human Rhys.
These figures come from the saved training logs and held-out measurements. Lower loss means better performance on that measured objective; it does not establish factual accuracy, personal fidelity or good conversation.
