VIVI: An Attempt to Grow Cognitive Intelligence, Not Just Train It
The article by a developer on Habr introduces VIVI, a project aimed at building an AI system that develops like a living organism, accumulating experience and forming memory, rather than relying on huge pre-trained language models. The system handles contradictions, generalizes from observations, and shows curiosity by asking questions when knowledge is insufficient. It uses various memory types and explains its reasoning, all without an external large language model.
The author of the VIVI project describes an alternative approach to AI: instead of training on massive corpora and fine-tuning, the goal is to create a system that grows intelligence from experience, with memory structure as the basis of intelligence, not precomputed weights. Key capabilities include understanding text without rigid rules, parsing sentences into events with roles (who, what, when), having working, long-term, episodic, and semantic memory, resolving contradictions through experience, and forming generalizations. The system demonstrates this by learning that a dog is an animal but not a plant, weakening false knowledge rather than erasing it, and treating new objects like a lion with hypotheses based on similar examples. It also answers questions about past events, showing episodic memory and the ability to reason. Development follows stages from child to adult, aiming to reduce hardcoded logic and let the system make its own decisions. Main risks include architectural complexity and software debt; the author is planning next steps for deeper planning and curiosity.
Source: Habr — хаб ИИ —
original