Keep CALM: New model design could fix high enterprise AI costs

Tencent AI and Tsinghua University have developed Continuous Autoregressive Language Models (CALM), a novel architecture that addresses the high computational costs of generative AI by predicting continuous vectors representing chunks of tokens rather than generating text token-by-token. This approach uses a high-fidelity autoencoder to compress multiple tokens into a single continuous vector, significantly reducing the number of generative steps and improving the performance-to-compute efficiency, making long-form AI analysis more feasible and cost-effective for enterprises.