July 2026 Summaries

David McKay argues in his blog post that the current AI infrastructure overly relies on general-purpose models with extensive prompts for narrow tasks, a practice he believes is inefficient and unsustainable. He advocates for a shift towards smaller, continuously trained models tailored to specific tasks, utilizing existing data from production traffic. McKay criticizes the discontinuation of OpenAI's fine-tuning platform, suggesting that the capability to automatically and continuously fine-tune models should be a standard feature in AI infrastructure. He envisions an architecture where a central model delegates tasks to a fleet of specialized models, each retrained independently based on domain-specific needs. This approach, he contends, would improve efficiency and adaptability by decoupling model training schedules, contrasting with the current practice of retraining monolithic models at high costs. McKay emphasizes that the necessary technology for this transformation exists, but lacks implementation, and proposes a simple "auto-tune" feature as a potential solution.