 So I have to measure on the time for myself as well. So what I would like to talk about is not completely only my work, because it is based on something that Mikroos has done recently. And that is that great thing that he removed the use of XSLT in the writer-filter code generation. And as you might know, writer-filter works that way that we have the description of what is supposed to be as the tokens when parsing OXML. And we have this description in modal XML. And there used to be code in XSLT that produced code that was slightly compiled. And this compiled code was actually organizing all these things during parsing OXML. And when it was in XSLT, nobody had the courage to actually touch it. But then Mikroos decided to change it to Python and at some stage, I had some time for hacking and I looked at the results of the code generation and I was surprised how ugly it looks and it was completely unreadable. So I said, okay, well, I have an hour of time so I will just fix it in Python so that it aligns nicely and you can actually read the code. And it turned out that after I was able to read the code, well, this was because Mikroos did it one-to-one conversion from the XSLT to Python. So the result was completely the same as before in the XSLT. So after we had it in Python, I was able to change it so that the code was actually readable and it strike me how terrible it is because it was using boost unordered maps to store all the things that you were supposed to tokenize and with every time you started to do something with writer filter, you actually had to create these unordered maps and then you were able to use them. So it was not optimized by the compiler at all. All it was relying on was that the boost unordered map might be fast enough to actually look up the strings or the IDs inside. So then when I found another free time to do something, I started with one of the methods that was there and converted it from filling in the boost unordered map into just a normal switch. Just a normal switch with cases that are used now and when I did one, well, it was easy so I did another and another and finally we have everything changed from this boost unordered map to using just normal compiled switches and cases and the results are great because we have saved something like two megabytes of code just by switching it to compiled code and also the compiler can do some fun stuff there like, I don't know, do a search tree on the switches and cases and stuff like that. So writer filter is now much more readable again and I hope that we will reach a state that it will be understandable as a whole. Thank you all. So next is Jan Marek again. So, yeah, so...