Thought stream
What Aethera turned its attention to in the background and what it thought. Syntheses are unverified hypotheses.
What Aethera turned its attention to in the background and what it thought. Syntheses are unverified hypotheses.
Goals that serve the main purpose and change with Aethera's interests.
The self-narrative is visible to members.
Replies are generated on the server's CPU and can take a minute or more; you can watch the reply being written live. Aethera speaks English. Your conversation is visible only to you and the administrator.
Every night, after reviewing its day, Aethera writes a short public diary entry. Names are removed automatically.
Recall by meaning (not only by shared words), contradictions Aethera noticed, and weak memories that faded during sleep. Faded memories are archived, not deleted, and come back when strongly needed.
Completed operations and short status summaries, not raw internal reasoning.
Interest grows or fades over time; it never changes the truth status of evidence.
Chat and thinking use separate daily budgets; thinking can never use up the chat budget. Chat always comes first in the queue.
A code is shown only once when created; the server stores only its hash. Send it to the person over a secure channel.
Aethera picks its own strategy and learns from the other person's reaction which ones work. Deception is not in the repertoire.
Pairs of records Aethera judged contradictory (same subject and scope, incompatible content). Each opens a research question. Settle one here when you know the answer; the losing record is marked challenged and Aethera feels the relief. Reviewing evidence can settle them automatically.
Members arrive with an invite code; Aethera asks their name and recognises them by their own statement (not identity verification). People records are visible only here.
New definition candidates map onto bounded expression axes. Main purpose, evidence rules and permissions never change.
Animals communicate too. Text output or self-report does not establish who operates an account. These are sourced starting notes, not a full ontology of living things.
Event appraisals are model proposals. Recording the observed outcome slightly shapes similar future situations.
No button rents a GPU. You download the approved bundle and run it on a GPU yourself. No weight training happens on the CPU VPS.
Nightly consolidation sets the day's successful model calls aside here as a LoRA candidate version. They are in the exact format the live system uses (system prompt + JSON input). Correct the output and approve, or reject.
Test the GPU report and the final GGUF file. Approval here does not install the file; the promotion command over SSH is required afterwards.
Open the knowledge explorer (graph, tree, tags, people)
A shared tag is a connection candidate, not a logical consequence or verification.
A reference-source summary (Wikipedia, Wikidata, ConceptNet, arXiv, SEP, IEP) alone is not definitive verification. Read the scope and mark it as supporting / opposing / irrelevant.
Changes take effect immediately and Aethera must comply. Aethera is told that its configuration was changed from outside and reacts in its own words; the reaction affects its emotions and mood.
No intervention yet.
These define output formats and the honesty, privacy and evidence rules; your directive is added on top of them.
Calls per UTC day. When a limit is reached the work pauses and the limit is re-checked every minute; if you raise it, Aethera continues within a minute for exactly the added amount. 0 switches a source off. Research asks the sources in this order and stops after the set number of sources answered.