GenAI Shift: Focus Moves From Model Training to Real-World Applications

Large language models are maturing, with researchers noting that performance gains from training on new data are beginning to plateau. Absent a major architectural breakthrough akin to the 2017 transformer paper, significant leaps in base model capability are considered unlikely in the near term. As a result, the industry's focus has shifted toward building practical GenAI applications — such as agentic coding tools, RAG systems, and voice assistants — rather than advancing the models themselves. Simultaneously, companies are investing in infrastructure optimisation, including specialised inference chips and KV cache improvements, to make existing models faster and more cost-efficient. This application-driven phase is seen as an opportunity for builders and product developers to create value on top of current AI capabilities.