The perfection in those examples makes me suspect that they are cherry-picked or part of the training data. Especially the handwritten text is not always clear and could reasonably be interpreted differently. I'd expect a machine-learning model to get at least some things wrong some of the time.
If I wanted to use this in an application, I'd definitely want to see some accuracy figures on validation data as well as a few failure cases to see whether the output remains reasonable even when it is wrong.
The examples are actually very simple compared to a lot of crazy stuff Mathpix can recognize so it's an honest representation of its capabilities. Mathpix is built for perfection because 99% isn't good enough.
> the handwritten text is not always clear and could reasonably be interpreted differently
Digital pen input contains more info than the resulting bitmap; strokes are lost while rasterizing.
That info was the reason how old devices were able to reliably recognize characters written by a stylus. It worked well even on prehistoric hardware, such as 16MHz CPU + 128 kB RAM in the first Palm PDA.
If I wanted to use this in an application, I'd definitely want to see some accuracy figures on validation data as well as a few failure cases to see whether the output remains reasonable even when it is wrong.