Even that I don't think involved much of a "eureka" moment, but mostly incremental improvements due to the environment changing. A lot of neural-net techniques from the '80s and '90s that didn't work that well at the time, now work well because we have sufficient data and computation power (especially via GPUs) to make them work. There have been improvements as well of course, but they're evolutionary improvements, while the big breakthrough was big data + big compute.