AWS unveiled “Chain-of-Draft” (CoD), a new generative AI prompting technique designed to reduce inference costs and latency by forcing models to limit reasoning steps to five words or less. Based on research from Zoom AI, the method contrasts with the verbose “Chain-of-Thought” (CoT) approach by mimicking human mental notes, reportedly cutting token usage by up to 76% and latency by 79% in AWS Bedrock tests. The cloud giant demonstrated the efficiency gains using the “Red, Blue, and Green Balls” logic puzzle, noting that while CoD excels in structured reasoning, it may underperform in zero-shot scenarios or with smaller models under 3 billion parameters.
