Current workflow: Publish the actual transcript
Share creates a saved HTML snapshot with its own canonical URL and content-based metadata. Publishing requires confirmation; new shares are public, indexable and listed at /shared/.
can you speakConversation that can use what you paste
Auto mode uses supported tools and your sources first. It can ask a configured local model for open-ended conversation when those tools do not answer the request.
Automatic private learning
Eligible prose of at least 700 characters and 80 words becomes a saved source automatically while learning is on. Ask follow-up questions about it. Pasted statements are source material, not independently verified facts. This expands recall and word patterns. A separate neural network with 4,096,192 parameter capacity also trains on eligible text in the background, with validation before checkpoint promotion. The external local LLM is not retrained.
Chat accepts up to 100,000 characters. Private source storage is bounded at 50 documents and 2 MiB. Use Framework to pause learning or add a source explicitly.
Make a public conversation page
Click Share, review the transcript, and create a public link. Existing snapshots do not change when you continue chatting. You can revoke your published snapshots in the Share dialog. A real HTML file is saved for each snapshot with its own content-based title, description and canonical URL. Public snapshots use index,follow. Transcript links use ugc nofollow; directory and pagination links are followed. Previously link-only snapshots retain noindex,follow.
Anyone with the link can read the snapshot. New shares are public and indexable, listed with crawlable pagination at /shared/. Indexable pages have readable HTML, canonical links, descriptive metadata and sitemap entries; search engines decide whether to index them.
More deterministic mathematics
Try “gcd of 24 and 18”, “prime factors of 360” or “100 choose 50”. The tool catalogue also includes matrix operations, polynomial evaluation, correlation, quantiles, weighted means and fixed-rate calculations. Tools can be composed into workflows.
Review neural learning
Use “neural status” to inspect training, “neural next:” or “neural complete:” for language predictions, and “neural rollback” to restore the previous checkpoint. “Forget neural learning” removes neural weights and replay text; other memories remain separate. Training can improve some predictions while hurting others, so candidates are evaluated before use. This is not a guarantee of general intelligence.