Fully justified text in LaTeX documents

 

The LaTeX software is known as a high quality typesetting program. The reason why LaTeX generated documents are described as high quality is because they are looking the same like hand-typeset documents created 200 years ago. To understand LaTeX better we have to describe what classical hand typesetting is about.

Typessetting 200 years ago was at first the art of creating fully justified text. The interesting fact is, that this formatting style is only seldom described in the literature. MS-Word has a simple button to activate it and LaTeX is using this formatting style as default. So the user has no further explanation why this is needed.

Fully justified text means basically to adjust the words on a line so that the right and left edge are straight. In case of hand typesetting this art was very complicated to realize and the main reason why it takes so long until the metal characters are sorted on a page. The main trick for created adjusted text lines is to use glue between the words.

In the past, all the newspapers were typeset with this formatting style. What the manual typesetter were trained is to realize this unique shape. From a technical perspective it is possible to create flush left pages with manual typesetting as well. It would be even a bit more economical. But nobody has done so. A closer look into old newspapers will show that 100% of them were typesetted with fully justified text.

The idea of LaTeX is to imitate this formatting style. A standard LaTeX generated document will look like a newspaper which was manual typesetted in the past.

It is not very hard to guess what the opposite of fully justified text is. Flush left text is equal to missing typesetting. A flush left formatted text looks very different from what is used in newspapers in the past. The main reason why LaTeX generated texts are looking all the same is because of missing flush left formatting. The typical latex user assumes that it is prohibited to use a flush left formatting style.

Programming larger projects with Forth

 

The sad news is that it technical not possible to run existing c source code on a forth cpu. The reason is, that Forth cpus have no registers but only a minimal stack. And no c compiler in the world is able to generate the assembly instructions for such a machine. Even for mainstream CPUs like the 6510 it is hard to write c compiler because this CPU has also a low amount of registers. And in a forth cpu the situation is more dramatically.

So the only practical method to write software for the GA144 and other CPU is to hand code the software in Forth. Unfortunately, Forth is known as one of the hardest programming languages ever next to Dyalog APL. An easy way to learn the Forth language should be explained in the following blog post.

Instead of trying to understand the idea of a push down stack the more beginner friendly approach to write something in Forth is to use modular programming. Modular programming is a powerful software engineering method which is supported by most languages like Pascal, C, Python, C++ and of course Forth. The idea is that each file contains of 4-5 functions plus a handful of variables. The functions can access the variables and the module can be included by other modules. The concept is some sort of strip down version of a class.

Modular programming in Forth is a bit more complicated than doing the same in C. but it is not impossible in writing such programs. It is mostly a question of getting experience in writing stack based functions. The concept of modular programming is the same. It allows that the source code can grow to a size of 1000 lines of code and even more. And each single file has a limited size of not more than 100 lines of code. If the file is getting bigger, the programmer has to outsource some of the routines and creates some interfaces to communicate with other modules.

The interesting situation is that Forth based modular programming works well with existing Forth cpus. The only thing what a certain Forth system needs to support is the ability to execute a word, and also the ability to read and write to variables.

Understanding the 1990s in terms of AI

 

In case of normal compurer technology the 1990s and the year 2021 are not very different. In the 1990s wokstation computers were invented already which includes the Unix operating system. Video gaming consoles were widespread available and letter were created of course with a computer but not with a typewriter. So the only new thing today is, that the CRT monitores were replaced by flat screens and the computer is a bit faster. But this is not a revolution but only a normal minor improvements of the last 30 years.

In the context of Artificial Intelligence the situation is much more different. The situation in 1990s and the situation today is nearly the opposite. The people in the 1990s had only a rough understanding what AI is, and no practical demonstrations are available. Instead of describing what the situation today is let us take a closer look into the past.

The first thing to mention is, that in the 1990s the internet was in an early stage. It wasn't the dominant media which includes all the other media but existing media technology like the cinema, television, books and printed journals were used instead. What the people have learned about Artificial Intelligence were published in these media. That means, the understanding of robots and AI were influenced by movies and books about the subject. Robots played only a small role in all the existing media, and it was unclear if such technology can be realized soon. There was a big gap between robots in the movies and robots realized in reality. And this gap was projected into the future. Basically spoken AI was something not available in the 1990s. It was used as part of a science fiction plot in which robots take over the world or it was explained in a computer science book, that solving the np hard problem is not possible.

This public understanding is very different from today's situation. Today the average user of the internet will see lots of videos with working robots including the source code for the software to recreate them from scratch. This includes walking robots, self driving cars and Mario AI autonomous bots. Such a library of existing software which includes textual tutorials wasn't available in the 1990s. In the past, it was difficult to find an entire book about the subject Artificial intelligence and even the advanced one have described only a very rough situation. Another problem in the 1990s was, that modern programming languages like python wasn't invented. That means it was not possible for a newbie to create a line following robot because for doing so he had to learn languages like C++ first. Basically spoken, creating such a robot was in the 1990s a multi million dollar project for a research lab at the university but not an amateur project for a weekend to learn something.

Perhaps these examples have shown, that since the 1990s many things have changed. The dominant new situation is, that the former np hard problem seems to be solved. That means the amount of papers in which mathematicians are able to proof that AI can't be realized within the next 1000 years has reduced drastically. It was replaced by the understanding, that each month a new robot is presented which is a bit more powerful than the previous one.

What would happen if a person from the 1990s gets teleported into today's time? In case of normal computing the situation would be relaxed. He can use his former knowledge about how to use a mouse and a keyboard to control a computer. He will find the same MS-DOS prompt and the same Unix command line and sending e-mails is the same like 30 years ago. But in one subject the teleported individual gets confused strongly which is the current state of AI. After sitting down in front of a computer and watching a youtube playlist he won't believe what he will see. The shown examples in robotics are so much different than what he knows that the person wouldn't understand it anymore. It would be simply too much to see a biped robot, self driving cars, autonomous drones or the poor hitchbot who is waiting on a park bench at night.

The development in robotics wasn't an abrupt event but a continuous flow of changes. But at the end, the year 2021 and 1990 have nothing in common in terms of AI. There is a huge gap between both years and not a single revolution but many of them have occurred. Most of today's AI technology like biped robots, deeplearning, and game playing AI wasn't available in the 1990s. It has entered the world without any warning and there is no sign that the development will stop someday. It seems that the subject of AI was the single driver which has changed the world into a futuristic one.

Modular programming with any programming language

In the past it was some kind of game to compare different languages against each other. C programmers are convinced that their language is the fastest one, python programmers emphasizes how easy it is to write code and Forth programmers are proud of low energy consumption for a single instruction. It is hard or even impossible to bridge these communities.

On the other other hand there is a unique element which have all the programming languages in common. This feature is more powerful than stack based computing, and easier to create than object oriented programs. The single useful feature is the ability to write programs in a modular fashion. The term modular programming is a bit uncommon and sounds outdated so it makes sense to explain it further.

A module is a file which contains of 5-8 procuedures and 5-8 variables. The procedures are allowed to manipulate the variables and the idea has much in common with a class. A module was used in programming before the advent of oop. The pascal language knows the unit statement, the c language can include prototype files, and fortran porgrammer can create modules as well. So we can say that modular programming is the single paradigm which is available in all the languages including forth which has also modular capabilities.

Modular programming looks not very interesting on the first look but it is the key element to write larger program. A larger programs contains of 1000 and more lines of code. Such projects can only be realized in a modular fashion because it allows to structure the program into logical chunks which are programmed independent from each other.

The perhaps most interesting situation is that with modular programming all the languages including forth and C++ are easy to master. What the programmer has to do is to follow strict the rules. That means, if the program gets bigger he has to create a new file and put the procedures into this file. Programming means to manage procedures and variables. This paradigm allows to solve any problem in software.

Let us take a closer look into some larger Forth projects at github. Instead of explaining how a stack based language works let us focus only on modular programming. What we can see in these projects is that some files are there and each file has the same structure. At top there are some variables initialized and in the bottom some functions are written down which are accessing to the variables. So the concept is the same what c programmers are using and java programmers if they are creating new files for a new class. And yes the principle makes sense because it allows to write longer programs which has more features.

The interesting situation is, that modular programming has no limit in the code size. Because newly created submodules can be included in other modules and at the end there are 50 and more files available which have each 100 lines of code. So the overall structure is highly hierarchical. It is less powerful than real object oriented programming but it comes close to the idea.

It sounds a bit trivial to expain this programming style in detail because c programmers are doing so since the 1970s and it is the most common coding style ever. On the other hand the amount of books how are describing the principle is low. Most C introductionary books doesn't even have a chapter about the subject of creating programs with more than a single sourcecode file. Similar to other languages the main focus of a tutorial is to explain the language itself, which includes the statements, but this knowledge doesn't allow to create useful software.

So we can say that an individual programming language is less important. The fact that the for loop in forth works different from a for loop in python can be ignored. What the programmer needs to know instead is how to write programs distributed in units.

Safe artificial intelligence in the past

From a technical perspective it is very hard to prove that robotics isn't possible. It is even much harder to show, that AI is limited to certain level which can't be over jumped in the future. Also it is impossible to slow down the progress or reverse the development. There is no such thing like a frame to control the development of Artificial Intelligence. What is possible instead is to ignore the future and take a look into the past which is much easier to understand.

The major advantage of the 1990s compared to the situation in the now is that in the past robots weren't invented yet, also the amount of books about the topic was little. That means the 1990s were a time in which the problem of AI upraising wasn't there.

The 1990s were the decade before the upraising of the internet. The dominant media in this time was the television for the mass and the printed book for the educated scholars. The amount of information was small and no search engine were used. From the perspective of AI the situation was also in a very early stage. A robot like honda asimo who is able to walk upstairs wasn't invented. A company like Boston dynamics wasn't there and in the beginning of the 1990s the worlds strongest chess player was a human but not a computer. Also the subject of deeplearning wasn't invented. The only thing what was available were normal perceptrons with not more than 20 neurons. These systems were realized in software for the windows operating system but no useful application was known.

Self driving cars were also not invented in the 1990s because of many reasons. What was available instead were lot of movie related robots like the K.I.T.T. car, or the Data android in Star trek. And the perhaps most surprising fact is that the people in the 1990s were not convinced that Artificial intelligence wil become possible one day. A widepread believe was, that simple biped robots will be available in around 300 years in the future or they won't be possible at all. The underlying theoretical concept to prove this assumption was the np complete theorem. NP complete says basically, that all the AI problems can't be solved on a computer because of the large state space. A similar idea was introduced in the highlight report in the 1970s.

A widespread philosophical interpreation of Artificial Intelligence in the 1990s was, that since the beginning in the 1950s the AI researchers has promised many things but none of the goals were reach. So the rational understanding was, that AI is too complicated in general and it is not possible to build such machines. Nobody in the 1990s were able to disprove this assumption so it was equal to common shared knowledge.

So we can say, that the 1990s were the last real AI winter. AI Winter means that the subject was seen as impossible to realize and that the research towards the subject has stopped. What the computer scientists have done instead is to program normal software for example games, operating systems and very important network protocols for the upraising internet. That means, AI in the 1990s was an esoteric discipline not recognized very much.

Suppose a person from the year 2021 travels back into the 1990s and explains to the audience which sort of robots are available in only 30 years. He will say that biped robots are possible, that kitchen robots can be build, that self driving cars can be realized with neural networks and he will explain that the sourcecode for tetris playing AI Bots is distributed as open source to everybody. The audience won't believe any of these words. The audience will say, that it is impossible. If the time traveler would like to give the details and explain how these software can be realized the audience will leave the room because it is outside their horizon. It would be too much for a person of the 1990s to hear what reinforcement learning is about or that chess software can beat a human player.

The 1990s from the perspective of Artificial Intelligence

 

Describing the current situation in AI and robotics is nearly impossible. There are endless amount of projects and documentation available and at least once a week a new robot is shown at youtube which looks more human like than in the week before. Advanced topics like biped walking, speech understanding and playing of videogames are solved already or the chance is high that within the next 12 months such a success story will become visible and fiction and reality have merged to a complex media campaign. Is it not clear if a certain robot head is remote controlled or by an advanced algorithm or if a certain walking robot was able to do the steps in reality or it was drawn into the picture with animation technology.

If the current world is too complex it makes sense to increase the distance and observe something which has a clear frame around it. The 1990s are a great decade for describing artificial intelligence. The major advantage is that the amount of published books is known and that the level of technology in the 1990s was low. Biped robots weren't invented, chess machines were not able to beat the best human player (with a single exception invented by IBM) and most researchers were not sure if AI can be realized in the future.

Computer technology in the 1990s was in an early stage and most problems were located in slow hardware and bugs in the software.. Windows 95 was during that period a common operating system, and the internet was in the early beginning. This setup makes it easy to give a full overview about all robots and AI projects during this time.

Basically spoken, AI wasn't realized in the 1990s but it was a philosophical topic. The question was if experts in computer science think about how realistic AI will be in the future. That means 95% of the population wasn't informed about the subject at all, and the few computer experts who have used neural networks and expert system in reality were not able to demonstrate practical applications.

Artificial Intelligence in the 1990s was mostly a subject for movies and science fiction books. Lots of stories about intelligent robots were available in this period. Even it was not possible to build real robots it was within reach to imagine a future in which positronic brains and other technology allows humans in doing so.

The 1990s from the robotics perspective

 

... were great, because no such innovation was available. All the technology available today, wasn't invented in this period. What was common instead were normal computer technology plus some philosophical books about how to realize Artificial Intelligence in theory. It was even unclear if a chess computer would be able to win against the best human player. The general understanding about AI during the 1990s was, the same like it was formulated in the famous lighthill report. Basically spoken the idea was, that even some expensive AI projects were started at US universities none of them has resulted into something useful.

The only place in which robots in the 1990s were available was the cinema. In blockbuster movies and of course in the star trek series, many examples were shown. But again, none of these things were build by the researchers.

From today's perspective it is known, that the during the late 1980s, the Honda company has developed humanoid robots. But, during the 1990s the internet was not widespread available so even computer experts were not aware of it. It was common sense, that no one has tried to build walking machines because it is too complicated to realize. What was known in the 1990s were some examples for expert systems. Some of them were described in mainstream computer journals. Also it was known, that for playing chess or tictactoe some sort of Artificial Intelligence is needed. But it was unclear how exactly such technology can be realized.

From today's perspective the 1990s have much in common with the stone age. AI was something not invented yet and it was imagined that it will take 100 years or longer to build it. Only to get the figures, the plot of star trek TNG plays around 300 years in the future. And exactly this was the estimated duration until biped robots are build.

Perhaps one surprising insight is, that in the early 1990s robotic competition were not common. From a technical perspective such robots can be realized with 1990s technology easily. But at this time, no one has seen a need for doing so. What was common instead was to program smaller programs in prolog and lisp. For all home computers and PC a compiler was available and some books were available about the subject. What such prolog programs in the 1990s were able to do was nearly nothing. Some more advanced software was able to solve logic games, but most of the programs were created as hello world examples.

With today's knowledge it is possible to identify some advanced projects in the 1990s not mentioned yet. For example the MIT Leglab has built in the 1990s lots of walking machines. But again, the Internet wasn't invented so no one was aware of it. That means, even if they have build these machines and published papers about it, computer experts, and hobby programmers as well simply never recognized it. Or let me explain it the other way around. Suppose a time traveler would visit a larger university library outside of the U.S. in the 1990s and will read all the books He won't find a single piece of information about the MIT robots, the honda asimo project or any other advanced AI project from this time. Sure, if the imagined time traveler is visiting the MIT library and knows which paper he needs to read then he will find the information. But without such an advantage he stays completely clueless.

Symmetric typesetting with LaTeX

On the first look the LaTeX ecosystem looks highly complex. There are an endless amount of packages, tutorials and guidelines available. The surprising fact is that it is possible to reduce the idea behind latex to a single screenshot. What the LaTeX community tries to achive is shown in the top of the image. The typset paragraph looks similar to a poem. The headline is formatted with centering and the paragraph is fully justified. The justification was created by adjusting the glue between the words, plus minor adjustments of the microtype package.

In contrast, the picture at the bottom shows what the TeX community tries to prevent. In such a situation the symmetry is missing because everything was formatted with flush left.

The interesting situation is that this preference has nothing to do with the LaTeX software but with typography in general. So we have to understand why formatting something by centering it is perceived as beautiful. The idea is that the written word follows the principle of art. It is not only a book but it is a graphical creation. Or to be more specific, the reader of the text should get the impression. From an authors perspective it is trivial to format something with the centering mode.

Let us take a closer look into the shown paragraph. It is the same text, the same font and the software for rendering it was in both cases the latest version of the lualatex engine. The only difference is that in the second case the formatting style “flush left” was applied. Now the question to answer is which of the cases has a higher score in terms of typographic quality? The simple answer is that the top example gets the quality judgement “beautiful example for typesetting”, while the bottom case gets the judgment “low quality or even no typographic quality at all”.

This judgment is not based on any objective criteria but it is simple a preference for or against flush left. How can it be that the established latex software including the high quality Latin Modern Roman Font can create low quality typesetting? Because the untold assumption is that only centering text which includes the headline and the paragraph is beautiful and everything else is wrong.

The underlying reason why this rule is valid is because it takes more effort to create full justified text. If the text isn't created with an algorithm but with a hot metal printing press it will take endless amount of time to fully justify the text. What the typesetter has to do is determine the size of the words and calculate how many glue is needed. In contrast, putting the letters for the bottom example together can be realized much faster.