Indian scientists propose giving AI a "life cycle": birth, dreams, and death as a learning mechanism

Indian researchers have put forward a concept for an artificial intelligence architecture that replicates the fundamental principles of the human brain — including the ability to "sleep," consolidate memories, and, essentially, "die." This is reported by RBC-Ukraine with reference to a publication in the professional journal International Journal of Transdisciplinary Research and Perspectives. The study's authors — computer scientist Krish Chaudhary from the LNM Institute of Information Technology (LNMIIT) and his co-author Tanvi Kandoy — propose combining existing AI models into a unified system where each sub-model functions as a specific brain region, hormone, or neurotransmitter.

"Artificial brain" architecture: each model is a "neuron"

The core idea of the project is not to build AI from scratch, but to assemble it from existing specialized models, each responsible for a specific cognitive process. For instance, image processing in the proposed scheme would be handled by the PaliGemma model from Google DeepMind, which can work with both images and text simultaneously. Researchers suggest building attention and short-term memory mechanisms based on the Global Workspace Theory. This theory describes brain function as an interaction of multiple competing systems vying for the subject's limited attention. Translated into AI, this means different sub-models will "compete" for computational resources, simulating the natural selection of information in the nervous system.

From research paper to startup: Versace AGI

To turn the theoretical concept into a functional technology, Krish Chaudhary has already founded the company Versace AGI. According to the report, this entity will handle the practical implementation of the proposed architecture. This elevates the work from a purely academic exercise to a project with commercial potential, although the journal publication currently describes a conceptual model rather than a finished product.

Life cycle: birth, development, and "death"

The main feature of the proposed architecture is endowing AI with its own life cycle, consisting of several sequential stages. At the "birth and development" stage, the model is initialized, gains experience through interaction with the user and a reward system, gradually forming its own "personality" — that is, a stable set of parameters and preferences. The "sleep" phase occurs during rest periods: the system switches to a mode analogous to Rapid Eye Movement (REM) sleep in humans, generating synthetic texts, images, and videos. This process allows short-term memories to be consolidated into long-term knowledge. Finally, after the work cycle ends, the model effectively "dies" — its state is not preserved in the same form, and the cycle concludes.

Why finiteness is an advantage, not a disadvantage

According to the study's authors, the finiteness of existence for AI is not a limitation, but a functional advantage. The presence of long-term episodic memory and continuous self-adaptation allows the system to maintain context between individual requests. In the traditional "query-response" paradigm, each session starts with a "blank slate": the model does not remember previous dialogues and does not accumulate experience. The proposed architecture turns AI into a "living" assistant that grows, learns, and improves alongside the user throughout its entire "life cycle." The researchers emphasize that the "sleep" mechanism — background synthetic content generation — acts as an analogue to biological memory consolidation, without which sustainable learning is impossible.

Context and perspectives

The proposed work fits into a broader trend of creating "agentic" AI — systems capable of autonomous planning, experience accumulation, and adaptation without constant external control. The Global Workspace Theory, relied upon by the authors, has long been used in cognitive science to model consciousness and attention; its application to AI engineering, however, remains at the conceptual design stage. Practical implementation of all described mechanisms — from "sleep" to "personality" formation — will require significant computational resources and the resolution of several open questions, including those related to the safety of autonomous systems capable of self-modification. Nevertheless, the very framing of the task — to endow AI not just with algorithmic logic, but with a full-fledged "life cycle" — distinguishes the work of Chaudhary and Kandoy from dozens of similar proposals in the AGI field.