Ekong Ikpe released details on June 21, 2026, regarding Kitana, an experimental system designed to replace traditional token prediction with dictionary traversal. Unlike standard language models that rely on pattern matching and probabilistic guessing, Kitana functions as a cognitive system that prioritizes structured semantic networks. The core architecture uses a dictionary as a grounding layer, starting from basic symbols like A–Z and 0–9 to build toward spelling, grammar, and complex meaning. This approach treats language understanding as a process of dynamic construction rather than the retrieval of stored facts or probabilistic predictions.

The system organizes knowledge through a graph, or "tank," structure where every word maintains defined connections to others. This method aims to address the instability of meaning often observed in LLMs, which the developer argues arises because current architectures scale parameters and data without grounding meaning in structure. Kitana learns through a progressive curriculum similar to human schooling, where reasoning emerges from tracing relationships step by step.

The project is in an early stage and faces technical challenges inherent in human language, such as slang, ambiguity, contradictions, and exceptions. The developer is currently testing the limits of this structured approach to determine how effectively it handles linguistic messiness. Initial observations suggest that the system consistently prioritizes referencing formal definitions over statistical guessing, potentially offering a path where intelligence emerges from foundational structure.

Ekong Ikpe released details on June 21, 2026, regarding Kitana, an experimental system designed to replace traditional token prediction with dictionary traversal. Unlike standard language models that rely on pattern matching and probabilistic guessing, Kitana functions as a cognitive system that prioritizes structured semantic networks. The core architecture uses a dictionary as a grounding layer, starting from basic symbols like A–Z and 0–9 to build toward spelling, grammar, and complex meaning. This approach treats language understanding as a process of dynamic construction rather than the retrieval of stored facts or probabilistic predictions.

The system organizes knowledge through a graph, or "tank," structure where every word maintains defined connections to others. This method aims to address the instability of meaning often observed in LLMs, which the developer argues arises because current architectures scale parameters and data without grounding meaning in structure. Kitana learns through a progressive curriculum similar to human schooling, where reasoning emerges from tracing relationships step by step.

The project is in an early stage and faces technical challenges inherent in human language, such as slang, ambiguity, contradictions, and exceptions. The developer is currently testing the limits of this structured approach to determine how effectively it handles linguistic messiness. Initial observations suggest that the system consistently prioritizes referencing formal definitions over statistical guessing, potentially offering a path where intelligence emerges from foundational structure.