KI Spitzenforschung leicht gemacht

 In online Foren wird leidenschaftlicher darüber debatiert ob ein Land oder ein Unternehmen womöglich den Anschluss verloren hat an die internationale Forschung auf dem Gebiet der Künstlichen Intelligenz und es werden möglichen Antworten debatiert um den Rückstand aufzuholen. Die unterschwellige Annahme lautet immer, dass Wissenschaft und Forschung etwas kompliziertes sei und man Paper schreiben müsste, Roboter bauen, Sprachmodelle deployen müsste. Wo also darum geht aktiv zu werden und etwas zu verändern.

Im folgenden wird eine These ausgebreitet die von diesem Mainstream diskurs erheblich abweicht und unterstellt dass Spitzenforschung etwas mit zuhören zu tun hat. Es geht nicht darum irgendwas zu erfinden sondern es geht darum das zur Kenntnis zu nehmen was schon existiert. All die highend Roboter, selbstfahrenden Autos, und automatischen Softwaregeneratoren gibt es längst. Man muss lediglich ein wenig Neugier entwickeln und fragen wo es dazu informationen gibt, wie es im detail funktioniert und wo man es ausprobieren kann.

Das Gegenteil von Neugierde ist Desinteresse, also eine Haltung bei der man ein thema ignoriert. Also gar nicht wissen will welche Fortschritte es auf dem Gebiet der Informatik gibt, kein interesse hat an selbst fahrenden Autos und keine Zeit hat sich anzuschauen wenn im Micromouse Wettbewerb die selbstgebauten Arduino Kleinroboter gegen eine Wand fahren.

Zuhören bedeutet dass man innehält, nichts sagt sondern dafür die Ohren spitzt um der Gegenseite RAum zu geben. Nicht das was man selber denkt oder sagt ist bedeutsam, sondern das was die Gegenseite formuliert muss sich entfalten. Um Zugang zu erhalten zur internationalen Spitzenforschung reicht es meist aus, Fragen zu stellen, die Gegenseite bitten etwas zu erklären und zu verstehen warum die Gegenseite sich damit beschäfigt.

Artcraft and the mythical man month

Software engineering is discussed in the literature since the 1960s. During that time the first software crisis was available. which means, that large programs including operating systems were recognized as key element in computing. On jan 1, 1970 the unix operating system was activated. Unix was an early example for a large scale software project. Since the 1990s the amount of codelines has grown dramatically, new programming tools and language were created.

The only thing what programmers can't do is to create more than 10 lines of code per day.[1] This is the Sound barrier. The only exception is if the source code was exported from a high level program, for example a GUI interface. Some advanced GUI tools can write the XML code for a picture on the screen, so in this special use case, the programmer can create much more than only 10 lines of code.

The limit is valid for real code programmed in C/C++, Java etc, because there is no program available which can export Java code. What is available instead are compiler which take source code as input and generates assembly code as output. Writing the source code is only possible by humans.

Some improvements were made by copy & pasting code from an existing project or by using software libraries. But even if code was copied and pasted, he needs to be modified, so in overall the former mentioned productivity of 10 lines of code per day, describes the reality.

The amount of written code line has prevented since the 1960s that software can be rewritten from scratch. Even if a programmer is familiar with Linux and knows how to program in C he is not able to program his own Linux kernel from scratch. Not because of copyright reasons but because of the limitation in productivity. The Linux kernel has around 40 Million lines of code. If someone likes to program a software with the same size, he will need many many days until the code is ready.

The only way to write large projects like an operating system or desktop applications is to split the work among programming teams, that the reason why version control systems like git have become popular, because they allow to maintain a large codebase with multiple programmers. This allows to realize closed source and open source projects both.

Suppose there is an invention available which generates software automatically by a computer program. instead of compiling C Code into an .exe file, such a fictional artificial intelligence is able to generate the c code by itself. Such an invention would revolutionize software development drastically, its the holy grail in programming.

Some attempts were made in the past to simplify software engineering, for example there are macro languages available like Visual basic, and there are high level interpreted language available like python. in both cases, the programmer can write an entire applications with a few lines of code which reduces the effort. But these tools haven't replaced programmers, somebody has to enter the source code. Creating a video game in Visual basic or in python will need lots of man hours.

The open question is, if the artcraft project has found the answer to the productivity limit of human programming. The promise of a large language model combined with an agent workflow is, that a machine create the entire software project from scratch.  If this promise is valid, its equal to breaking the sound barrier in software engineering.


[1] Brooks, Frederick: The Mythical Man-Month, 1975

Update zur Wordcraft Software

 Das Thema Wordcraft wurde hier im Blog schon erwähnt. In sehr kurzer Zeit gab es viele Veränderungen, die im folgenden gewürdigt werden sollen. Zuerst einmal gibt es nach nur 1 Tag bereits eine neue Version. Das schnelle Entwicklungstempo wurde also beibehalten, zweitens gibt es in einem Online Forum eine Umfrage, worin Benutzer darüber abstimmen können welche Software als nächstes von einer KI nachgebaut werden soll. Ganz oben auf der Liste stehen Programme die häufig eingesetzt werden wie Outlook, Onenote aber auch Maya und Solidworks werden genannt.

Für diese Programme gibt es bisher Open Source alternativen, zum Beispiel kann man Outlook durch Evolution ersetzen. Scheinbar reicht das aber nicht, die Nutzer wünschen sich einen 1:1 Nachbau der ihren vertrauten Programme.

Ob das die Hersteller dieser Software lustig finden ist unklar. Ein Pressestatement von Microsoft gibt es nicht. Was es aber gibt ist ein Fakt und zwar gehört die github Plattfortm auf der die artcraft Software zum Download bereitsteht ausgerechnet Microsoft. Microsoft sieht also dabei zu wie auf ihrer eigenen Open Source Plattform Klone ihrer Standardsoftware wie Word, Excel und Powerpoint heruntergeladen werden.

Halten wir mal die Ist situation fest: seit 1 Monat wird über Künstliche Intelligenz die komplette Photoshop Suite plus die Microsoft Suite nachprogrammiert und weitere Anwenderprogramme stehen zur Diskussion. Ferner gibt es in kurzer Folge Verbesserungen dieser Programme. Offenbar wurde vor 1 Monat eine Art von Revolution gestartet, bei dem Large language modelle nicht nur dazu verwendet werden, kleine Codesnippet zu erzeugen die bei Stackoverflow als Antwort dienen sondern wo über Künstliche Intelligenz komplette Programme nachprogrammiert werden, in einer anderen Programmiersprache als das Original, also Rust anstatt C/C++.

Die meisten Beobachter machen den Fehler, die Thematik artcraft aus der Perspektive von rechtlichen oder ökonomischen Dimensionen zu betrachten. So wurde gefragt ob das Projekt ein Urheberrechtsproblem hat, oder welche ökonomischen Auswirkungen entstehen. Das ist aber nur ein Detailproblem, die eigentliche Frage dreht sich darum was das aus Sicht von Künstlicher Intelligenz bedeutet. Ist KI inzwischen so mächtig, dass Softwareprojekte die früher in Jahren erstellt wurden, in wenigen Wochen nachgebaut werden können?

Wordcraft comparison between version 0.7.0 and 0.14.0

 a) version 0.7.0 was released on Oct 14, 2026, need 89 Megabyte
b) version 0.14.0 was released on Nov 2, 2026, need 100 Megabyte

For a better understanding for the development process behind the AI generated code, two version of the wordcraft software should be compared which were released with an interval of 14 days. In Open source projects from the past, such a short period won't change the software drastically, because the amount of changes in a software is measured by the amount of commits. And before a commit can be made, the source code needs to be written which takes time.

In case of wordcraft, most of the commits were created by Claude Opus so the amount of commits is much higher. The open question is: does this affects the functionarlity of the software? Perhaps most commits are only random modification without real purpose. So there is a need to take a closer look into the executable file.

The older version a) is working reasonable well. The user can enter text, insert tables, export the document into the pdf format, and even an English only spell checking is available. The program has a lower quality than the original MS Word software, but it can be used to create documents and open existing .docx files.

The newer version b) has additional features:
- easier font selection in styles
- additional proofreading languages
- faster rendering of the text on the screen

Adding these features in around 2 weeks is only possible with Artificial Intelligence. No human development team can provide the code changes in such a short amount of time. If this commit frequency is valid also for the future, the wordcraft project will evolve into a production ready word processing software.

Wordcraft an AI generated software clone

 Open Source software is known since the mid 1980s. During this time, computer enthusiasts have reprogrammed Unix software in C. Many new programs and entire operating systems are available as open source and there are dedicated conferences of the FOSS community.

Open Source has many advantages over commercial software because Open Source programs are available as source code and they are distributed free of charge. The shared similarity to commercial software is, that human programmers are needed to create the code. Especially for larger projects like Office, or photo editing software this bottleneck is a serious problem.

Since a while there is a revolution available for the software industry triggered by large language models specialized for writing software. One of the newly created programs is Wordcraft which is Word clone and it is running on Windows, Mac and Linux. The original Word program consists of a large codebase and is only available for Windows, while the Wordcraft clone was written in Rust and runs even in the webbrowser.

It seems, that with the correct prompt and a recent AI tool its possible to clone software. A human programmer defines the goal, while the AI writes the code. This allows to bypass the slow development of human written software which is around 10 lines of code per day and programmer.[1]

Computers until the year 2022 were able to copy the software itself, a zip file which holds a software can be send from one computer to another in a minute. This allows to distribute open source software very fast. The bottleneck was to create new software. If someone likes to start a new open source project, he need a months upto years until the software was written and tested. The duration can be reduced by increasing the amount of programmers but its hard to coordinate these teams.

With the advent of vibe coding the former bottleneck can be solved partly. Similar to a neural networks which can generate photos from scratch, a large language model is useful to generate computer code from scratch. There is no need for a software team anymore, but software can be created by a single text prompt which gets executed in minutes.

Its pretty easy to predict the outcome of this technology. The amount of vibe coded software at github will grow exponentially. Every programmer who has in idea can program an entire software with the help of Artificial intelligence and publish the result online. Such kind of revolution will question existing commercial software and open source software both. The new boundary is between human written code and AI generated code.

Lets take a look into the project. The screenshot was taken in Linuxmint and shows the help screen. It gives the information who created the software, Claude Opus, and how many lines of code were needed in Rust, 

In a short text, a word document was created and converted into a pdf document. Also the word document was opened with Libreoffice. All the operations are working. It seems, that the wordcraft project is a very powerful clone of the original software.

Its interesting to take a look into the github page of wordcraft because it documents the rapid changelog. From the Linux kernel is known, that a huge amount of commits is added to the project, so a new kernel gets released every 2 months. In case of the wordcraft project the frequency is much higher. The version shown in th former screenshot is 0.7.0 from Oct 14, 2026. The current latest version is 0.13.2 with lots of improvements. 

The duration between both releases was only 2 weeks. No human programming team is able to improve the software in such short amount of time. The only explanation why the software is changing so fast, is:
a) the software version 1.0 was written already and the shown version at github is outdated
or b) the version 1.0 doesn't exist yet and an artificial intelligence is improving the software on a daily basis

Its not clear how the software was made. What is known for sure is that there is no other Wordprocessing available written in Rust with such a high amount of features. Calling the project a Word clone is an understatement.

Sources
[1] Brooks, Frederick P.:The Mythical Man-Month, 1975

Vocabulary game for a warehouse robot

Grounded language can be demonstrated with a vocabulary game. The game engien prsents a camera picture and a human has to enter the correct word. In case of a warehouse robot game, such a quiz consists of a BNF grammar which holds only the possible words, and a picture which contains all the possible objects and actions.

BNF Grammar

<word>    ::= <objects> | <actions>

<objects> ::= "wall" | "door" | "charging station" | "human obstacle" | "box" | "shelf" | "robot obstacle"

<actions> ::= "move straight" | "turn left" | "turn right" | "u-turn" | "stop" | "speed up" | "speed down" 

Picture

The interaction with the game is easy to explain. The random generator selects the picture with the "door" and the human has to enter the word "door" into the input field. The game engine recognizes that the answer is correct and rewards the human with a +1 score.

The harder to explain question is why such a vocabulary game is relevant in robotics. It has to do with the previously mentioned BNF grammar. The grammar provides a minimal state space for the game. Its not important what sort of camera is used by the robot and which programming language is available e.g. C++, but only the BNF grammar counts because it holds all the possible game assets. In the concrete case the BNF grammar is stored in a text file with 251 bytes and contains 14 different words. These words are providing the conceptual space for the robot.

Suppose a human operator submits a command sequence like:
1. Enter the door, bring the box the shelf
2. then move to charging station
3. if there is a human opstacle or a robot obstacle, stop!

Such a command sequence is valid for the robot because it is based on words from the BNF grammar. The robot converts the command sequence into actions. This allows a high level human to machine interaction 

Rückblick auf GOFAI

Von 1960 bis 1990 beschäftigte sich die disziplin Künstliche Intelligenz mit dem Thema Oood old fashion AI (GOFAI). Der Begriff ist erst im Nachhinein für diese Zeitspanne vergeben worden, zur damaligen Zeit fehlte es an einer Selbstreflektion.

GOFAI meint im wesentlichen ein "closed system" zu programmieren, also ein intelligentes Computerprogram, was autonom Entscheidungen treffen kann, in die Zukunft plant und erläutert warum es eine bestimmte Entscheidung getroffen hat.

Beispiel Software projekte waren Mycin und Xcon, welche beide als Expertensysteme bezeichnet werden. Zum Höhepunkt der GOFAI Zeitspanne Ende der 1980er wurde versucht Expertensystem-Shells zu entwickeln, also Computerprogramme die für mehrere Bereiche eingesetzt werden können und wo der Nutzer Regeln und Fakten definiert, Auch das sehr bekannte Cyc Projekt von Douglas Lenat war eine solches Expertensystemshell.

Es war keineswegs ein Zufall, dass während der GOFAI Epoche das Ziel die Entwicklung von "closed systems" war. Ein solches Softwaresystem besteht aus Datenstrukturen und Algorithmen also die Bereiche die von der Informatik diskutiert werden. Auch in der Mainstream Informatik funktioniert Software nach diesem Muster. Es ist das Grundprinzip bei der Programmierung von Computern. In einer Programmiersprache wie C gibt es einen Header worin Datentype festgelegt werden, z.B. integer Variablen und strings, und dann gibt es einen Body worin Funktionen enthalten sind welche auf die Datenstrukturen zugreifen.

Gegen Ende der 1980er gab es mehrere kritische Stimmen an diesem Prinzip. Namentlich Hans Moravec und Rodney Brooks sind hier zu nennen welche beide Roboter entwickelte. Sie bemerkten bei der praktischen Programmierung von Robotern, dass GOFAI Systeme an ihre Grenzen stoßen. Es reicht offenbar nicht aus ein intelligentes Computerprogramm zu entwerfen.

Vielleicht ein praktisches Beispiel: angenommen die Roboterhardware funktioniert, und auf dem Microcontroller läuft ein Prolog interpreter der neuesten generation. Nach dem GOFAI Paradigma müsste der Roboter einsatzbereit sein. Problem ist dass der Roboter nichtmal im Stande ist einer schwarzen Linie auf dem Boden zu folgen. Sondern der Roboter wird exakt gar nichts machen, irgendeine Komponente scheint noch zu fehlen.

Auch andere Robotik-Entwickler bemerkten das ihre Roboter allein durch Expertensysteme nicht einsatzfähig waren und so wuchs die Kritik am GOFAI Ansatz. Es entstanden Theorien darüber was zusätzlich benötigt wird, diese Strömung wurde als bottom up robotics bezeichnet. Sie legt einen starken Fokus auf Sensor wahrnehmung und Motorsteuerung.

Der Wechsel von GOFAI hin zu bottom up robotics erfolgte ziemlich genau im Jahr 1990.

Pointing game for a warehouse robot

 A pointing game is an entry level language game to demonstrate grounded language. A formal BNF notation for a warehouse robot is:

<pointing-command> ::= "pointat " <target>

<target> ::= "charging_station"
           | "obstacle"
           | "box3"
           | "shelf_2"

The amount of verbs is only 1 and the amount of nouns is also very little. A command is always send from the speaker outside of the robot to the hearer which is the robot. The robot has to translate a command into an action.

From a technical perspective, such a game seems be trivial. The correct position of a word can be found in the internal database, so its a lookup request. at the same time, the game provides a powerful abstraction mechanism. it converts a complex domain, the warehouse robot, into a simple dialogue game. Even if the original game consists of hundreds of lines of code written in C++ the dialogue grammar for the pointing game remains simple. So the dialogue game acts as an API to get access to possible software routines of the robot.

The former mentioned BNF grammar acts as a user interface. It defines which commands are possible in the GUI. A human operator can select one of the commands from the BNF Grammar menu.