“Sundar Pichai speculates that the performance improvements of early large language models were more pronounced for coding tasks than for general language tasks, a signal Google may have missed.”