the ratio of new voice models i see on hackernews to the number actually deployed in any product i use is approximately infinity.