Generative Computing and the Future of AI Software Development

LLMs have become remarkably capable, yet building reliable systems with them remains difficult. Unlike conventional software made of modular, independently testable components with well-defined interfaces, an LLM diffuses all of its capabilities across billions of undifferentiated parameters, offering no natural way to assemble, isolate, or verify a unit of behavior. I will present generative computing, an approach developed at IBM Research that aims to bring the rigor and modularity of software engineering to LLM-based systems. Rather than steering a monolithic model through hand-written prompts, generative computing builds AI from composable, testable parts with defined inputs and outputs, realized through modular adapters that can be trained, tested, and dynamically activated at inference time, and exposed to developers as ordinary code via the open-source Mellea library. I will discuss the system architecture that makes this efficient and show evidence that small modular components can match or exceed prompting on tasks like constraint checking and hallucination detection at lower cost, pointing toward a way to recover software's discipline without sacrificing the generative flexibility that makes LLMs valuable in the first place.