Quasar AI: the ecosystem’s own intelligent core
For more than six months we have been building our own intelligent system — the core, memory, knowledge, analysis tools and a software agent.
Why build our own
Calling somebody else's model over the network is the quickest path. But an external model does not guarantee a verifiable provenance for every statement and can change independently of our project: add to that a charge per request, and the fact that it knows about your task exactly as much as you told it.
We need something else: a system you can ask for grounds. So the core, the knowledge, the language analysis and the inference live inside the project.
How this differs from a wrapper around an external model
Quasar AI builds an answer with algorithms over knowledge taken from books, dictionaries and manuals, rather than by picking the most probable continuation of a text. That design has a price — it stays silent where it does not know — and it has a point: every statement has a provenance, and that provenance is never faked. If a source is unknown, it stays in the journal instead of turning into a confident citation.
Quasar Core
The computing core: parsing the request, finding knowledge, inference and verification of statements. The core thinks first — it analyses the request itself instead of handing it out to executors at random.
Quasar AI Agent
The working part of the system: planning a task, working with projects and software tools, writing and editing code. The agent knows the map of its own repository and builds a plan from the text of a task rather than from prepared templates.
Memory and knowledge
Knowledge lives as data, not wired into the code: an encyclopaedic cube of nearly two million articles, references for programming languages, dictionaries and grammars of two languages, parsed textbooks. Rebuild the source — rebuild the cube: otherwise the system answers from yesterday's knowledge and cannot tell you so.
How this relates to the products
The finished products of the ecosystem work on their own and do not depend on the research. But what is built in the core reaches them in time: the review of a trade, an explanation of a decision, a verifiable argument instead of a confident tone.
Where it stands and what comes next
The foundation, the language and a good part of the reasoning are closed: grammars of two languages and one sentence parser, the knowledge cubes, the logical inference module, multi-step reasoning, the planner and work with code. In progress: building an independent conclusion from several sources instead of retelling one, speech certified by the core, and checking a statement against several independent sources.
The Quasar AI page — how the system is built, the laws it works by and a roadmap of seven stages. No dates: this is research, not assembly from a drawing.