Show HN: Topas-DSPL – A 15M param AI that solves hard reasoning tasks(ARC-AGI-2) (github.com)

Bitterbot AI has unveiled the TOPAS-DSPL, a streamlined 15 million parameter AI model designed to tackle complex reasoning tasks effectively. This innovation distills the full TOPAS Neuro-Symbolic architecture by focusing on the Dual-Stream Programmatic Learner (DSPL) and its Bicameral Latent Space framework. Remarkably, the model achieves a 24% solve rate on the ARC evaluation set, demonstrating its robust reasoning capabilities without relying on the auxiliary memory systems and symbolic aids typically utilized in such architectures. The significance of TOPAS-DSPL lies in its ability to address a critical issue in standard recursive models—compositional drift—by isolating algorithmic rules and execution states into separate streams. The Logic Core acts as the CPU that dictates instructions based on demonstrations, while the Canvas Core executes these commands in a manner reminiscent of spatial processing. This separation not only enhances the efficacy of the reasoning process but also allows dynamic reprogramming at each recursive step, enabling efficient learning with fewer parameters. With its advanced optimizations and modularity, TOPAS-DSPL is well-suited for applications that require robust reasoning while minimizing resource use, enhancing accessibility for researchers and developers in the AI/ML domain.