Hacker News Reader

Isn’t this just the no free lunch theorem? Should you be expecting a more general framework to beat an specially trained algorithm?

Another concern is generality: just because it performs well on this dataset does not mean it will perform well on another.

View on HN

It really doesn't matter that the model will probably not perform well on another dataset, as it was built for a specific task.

It's also about flexibility: If Google ML doesn't provide you with a way to train their algorithms specifically for your use case it won't help you that they work well for generic text recognition tasks.

Reply on news.ycombinator.com