Hacker News Reader

Could there be a similar scheme for floats?

Also, is there a penalty for using float16s for CPU centric code?

Using fixed-point representation would allow using this compression method.