Granite Switch: Building AI more like software (research.ibm.com)

IBM has announced significant advancements in AI model development with Granite Libraries and Project Granite Switch, aimed at making the construction of AI systems more modular and similar to traditional software engineering. These innovations come in response to the challenges enterprises face in accurately customizing AI models, which often require extensive retraining or complex prompts for adjustments. By introducing a modular approach, where independent adapter functions can enhance model capabilities without altering the entire architecture, IBM seeks to empower developers to customize and improve AI systems efficiently, akin to how software applications are built using interchangeable modules. Granite Libraries include task-specific adapters designed to perform functions such as relevance scoring and hallucination detection, significantly boosting model accuracy. For instance, using the requirement-check adapter function raised accuracy from 51% to 84% on benchmark tests. Meanwhile, Project Granite Switch allows for the seamless integration and activation of these adapter functions within existing model architectures, streamlining the inference process and eliminating the inefficiencies tied to switching tasks in traditional models. Together, these initiatives position IBM’s Granite 4.1 models as highly performant yet cost-effective alternatives to larger models, paving the way for more sustainable and adaptable AI solutions in enterprise environments.