Vector-based AI reasoning could slash compute costs, researchers claim

Scientists have developed an AI system that reasons using vector representations rather than traditional transformer architectures. They report that this approach can run up to 11 times more cheaply than a leading OpenAI model, potentially marking a shift toward a post-transformer era in artificial intelligence. The method aims to reduce the computational burden of advanced reasoning tasks.

EXPANDED:

The reported system departs from the transformer architecture that underpins most contemporary large language models, instead performing reasoning through vector representations. This design choice targets the computational overhead associated with advanced reasoning tasks, which typically require substantial processing power.