An AI that is ERERE's own
What it is
The path by which ERERE becomes an artificial intelligence of its own and detaches from everyone else's: open models that ERERE owns, its own knowledge, its community, and a detachment that happens by itself when it is ready.
Why it exists
Whoever depends on someone else's intelligence depends on their prices, their rules and their interests. ERERE wants an intelligence that nobody can switch off, change or use against people.
How it works for you
- Its own weights: ERERE's IO runs on open models whose licences allow using, modifying and training them; the weights are on ERERE's machine and nobody can revoke them.
- Several ERERE brains: Consulta can certify an answer with several ERERE models of different families and with the open libraries (Wikipedia and others) copied onto ERERE's machine: your question does not leave.
- It learns from itself: what ERERE's models certify together with the libraries becomes material with which ERERE trains itself, together with what the community expresses.
- The detachment: every week ERERE measures how many certified answers its IO finds. When it passes the threshold twice in a row, it detaches by itself from external intelligences: from then on no request goes out. Only the founder can go back, with the second factor.
- A complete exam for every new model: bias, facts, languages and safety, with a cap on text written by other intelligences and human knowledge always kept (see the guide "Answers with sources, and an intelligence that measures itself").
What you give
Nothing more: the IO works on ERERE's machine.
What you get
An intelligence that does not change owner, does not change price and does not take your questions elsewhere.
What you choose
Everything you chose before. You can object to community learning with one tap.
Honest limits
Open models are smaller than the largest in the world: on hard questions ERERE may be less brilliant until it grows. That is why detachment happens only when the measurements allow it. It needs a machine with enough memory and disk (16 GB of RAM for several models, 80 GB of disk for the libraries).
Where it stands
In the system and verified by the automated tests with simulated models and libraries: Consulta certified with no external request, automatic detachment, going back only with the second factor, its own training material. It switches on with ERERE's server.