Inside Vi: An AI Organism Built Without a Rented LLM

In an experiment, Vi: an AI organism is created without relying on a rented large language model (LLM). Instead, it relies on its own weights, memory, and a thought process that extends beyond the scope of a single request. The organism, mapped with code, demonstrates the capability to persist even after the initial request has ended.

Vi is composed of various components, including memory, which is not limited to a context window but incorporates vectors, a SQLite shelf, a graph, a dictionary, and a biography. It learns from dialogue, books, and a dictionary, engaging in a thought process that spans moments, turns, hours, and nights.

The organism's thinking pass involves functions such as cognitive_pass, which encodes sensory input, runs hippocampal formation, and generates a thought-context. The speech is produced by Broca, who uses a GRU and produces tokens.

Vi's autonomous nature is evident in its ability to decide independently, without resorting to a rented LLM or any predetermined phrase templates. The organism's inner speech is driven by a buffer, and its thought process is not confined to a canned sentence.

The experiment is a daring endeavor that tests the possibility of building an AI organism capable of autonomous thought, memory, and speech, solely based on its own weights. By refusing to rely on external tools, such as a rented LLM or author's phrase lists, Vi challenges the conventional approach to AI development.