From text to a candidate

Eligible user text is checked, cleaned of common copied interface clutter, chunked and queued privately; with shared learning enabled, it also contributes to a separate shared queue. The worker updates embeddings and network weights with gradient steps. The neural architecture now includes a trainable residual hidden layer. Replay mixes recent text with samples spread across older retained sources. Separate new-text, replay and base-language checks must pass before promotion.

Inspect before assuming improvement

neural status
neural next: the river
neural complete: the river

Status reports queue state, revision and candidate metrics. Next-word probability is not factual confidence. The 16-word-context network ranks authored wording and can produce labelled drafts; broader conversation uses the separate enabled local LLM.

Pretrained start and optional local teacher

The bundled seed starts with 3,239 active tokens and 54,000 inherited training updates, not random weights. The new residual layer is initialized to preserve its predictions. Optional operator-run Ollama teaching drafts language examples for review; an exact digest is required before queuing. Synthetic lessons are excluded from factual retrieval, and validation can still reject candidate weights. No ordinary model reply is automatically recycled into training.

Control the learner

pause learning
resume learning
neural rollback
forget neural learning

Rollback requires a retained previous checkpoint. Reset clears neural samples and weights, while source documents and chat transcripts remain separate. The 4.13-million-parameter capacity is not a count of learned facts. Private checkpoint storage is approximately 16.5 MB per version before database overhead.

Learning for everyone without publishing a chat

Shared learning is on by default. The chat header shows the current choice and offers Turn off; Options has the same setting. Previously saved off choices remain off. Future eligible user text becomes globally available as unverified source material and enters a separate shared neural training queue. Earlier private messages and assistant replies are not imported. Secret and personal-data filters are conservative exclusions, not a guarantee of anonymity. Do not submit private material while contribution is enabled.

shared learning status
shared learning off
shared learning withdraw
shared knowledge: your specific topic

Turning it off stops future contributions. Withdrawal removes retained contributions and clears the shared checkpoint for retraining from remaining text; other users may retain earlier replies. Private models never train from shared weights, so removing shared contributions does not require rewriting private models. Shared predictions are a fallback when no private checkpoint exists. The worker uses the same validation gates for both profiles. At most 256 shared chunks, 64 per contributor, are retained; each is at most 4,000 characters. Filters and quotas may skip messages. External LLM weights are unchanged.