According to Yann LeCun (@ylecun), choosing a batch size of 1 in machine learning training can be optimal depending on the definition of 'optimal' (source: @ylecun, July 11, 2025). This approach, known as online or stochastic gradient descent, allows models to update weights with every data point, leading to faster adaptability and potentially improved convergence in certain AI applications. For AI businesses, adopting smaller batch sizes can reduce memory requirements, enhance model responsiveness, and facilitate real-time AI deployments, especially in edge computing and personalized AI services (source: @ylecun).