Wodurch werden KI Projekte so teuer?

 Der Diskurs über Deutschland als Technologiestandort wird häufig verengt zu der frage warum es keine KI Forschung in Deutschland gibt. Dabei gab es zumindest in der Vergangenheit viele deutsche KI Projekte wie:

- verbmobil, Vitra, RHINO (Museumsroboter), ARMAR (humanoider Roboter), RoboCup@Home, Carolocup und Dickmanns autonomes fahren

Das Grundproblem bei KI Forschung ist dass es extrem teuer ist im Sinne von finanzieller Kosten. Ende der 1980er benötigten Forscher die schnellste verfügbare Hardware um darauf Lisp laufen zu lassen und erste neuronale Netze zu simulieren wie VAX Minicomputer (0.5 Mio US$ pro Stück) und Symbolics Workstation (100k US$ pro Stück). In diesen Computern war für die damalige Zeit sehr viel RAM verbaut und die Systeme waren Einzelanfertigungen aber kein MAssenprodukt wie der damals verfügbare Atari ST Rechner.

Und das sind nur die Ausgaben für die Hardware, hinzu kommen die Kosten für das KI Projekt selber. Künstlicher Intelligenz unterscheidet sich von klassischer Forschung dadurch dass es interdisziplinär betrieben wird. Neben der Informatik braucht man zugriff auf Mathematik, Sprachverarbeitung, Kognitionspsychologie usw. Allein eine Bibliothek welche all diese Literatur bereitstellt ist ein hoher Kostenfaktor.

Sowetwas wie preiswerte KI forschung welche auf Standardhardware läuft und nur ein Wissenschaftsgebiet umfasst gibt es nicht. Das wäre dann z.B. Forschung innerhalb der Mathematik oder Forschung innerhalb der Mechanik. Bei künstliche Intelligenz fließen all diese Bereiche zusammen. Ein robotik projekt besteht aus hardware, software, Algorithmen, Natürliche Sprache, Bilderkennung, neuronalen Netzen, Dateübertragung über Netze usw.

Bei KI Forschung im Jahr 2026 haben sich die Kosten weiter erhöht. Aktuelle Hardware um neuronale Netze zu trainieren ist sehr teuer, die Notwendigkeit unterschiedliche Disziplinen einzubeziehen hat sich verstärkt. Roboter werden heute über motion capture datensätze trainiert welche zuvor in Laboren für Bewegungsstudien erfasst wurden, DIe sprachverarbeitung erfolgt über Datenbanken in denen fremdsprachichige Korpora gespeichert sind und Bilderkennung benötigt riesige Beispieldatenbanken mit hunderten von Terabyte an Speicher. All diese Ressourcen sind teuer, die akademische Publikationen liegen geschützt hinter kostenpflichtigen Paywalls, selbstfahrende Elektroautos mit Lidar sensoren kosten deutlich mehr als ein Standard Auto und es dauert Jahrzehnte bis sich nachwuchsforscher in die Thematik eingearbeitet haben.

Grundlagen forschung im Jahr 1900 in den Bereichen Funkübertragung oder Elektronik war relativ einfach durchzuführen. Es braucht nicht mehr als einen Bastelkeller, einige wenige Grundlagenwerke aus dem Fachgebiet, ein wenig praktisches geschick beim Aufbau von Schaltungen und schon konnte die Spitzenforschcung beginnen.

Die schlechte Nachricht lautet dass es nicht möglich ist die Kosten zu senken. Neuronale Netze auf älterer Hardware zu trainieren funktioniert technisch nicht, Robotik ohne Sprachverarbeitung zu erforschen macht inhaltlich keinen Sinn, Trainingsdateaets zu erzeugen die nur wenige Megabyte groß sind erzeugt schlechte Resultte, und humanoide Roboter zu bauen die nur 3 servomotoren haben und über ein Stromkabel versorgt werden ist kein guter Standard in der autonomen Robotik.

Es gibt jedoch eine Methode um die Kosten massiv zu senken. Und zwar indem man KI Projekte rein virtuell durchführt. Das bedeutet, es gibt keine physische Hardware und es gibt keine realen Sensoren sondern die KI agiert in Computerspielen. Einsteigerfreundliche Projekte wären:

- Steuerung eines Pong Clones mit Hilfe neuronaler Netze
- Programmierugn einer SChach Engine
- simulation eines 6 beinigen Laufroboters als 3d Modell

Der verbleibende Kostenfaktor bei diesen Projekten ist die inhärente Multidisziplinarität, das also unterschiedliche Wissenschaftliche Bereiche wie Mathematik, Informatik, Biologie, Sprachwissenschaft, Statistik, Bewegungsstudien miteinander kombiniert werden. Möchte man z.B. einen simplen 6 beinigen Laufroboter in 3d Animieren benötigt man für diese Aufgabe unhzählige Bücher aus mehreren Bereichen der Unibibliothek. Es gibt kein Fachgebiet in einer Bibliothek was sich dezidiert mit laufrobotern beschäftigt, sondern es gibt nur Fächer die etwas über Mechanik, Programmierung, Animation, Algorithmen und Bewegung beinhalten.

Das was unter dem Stichwort Künstliche Intelligenz gemeint ist, stellt ein Mix da aus all diesen Disziplinen. All diese Informationen zu lesen und zu kombinieren ist zeitaufwendig. Es dauert lange bis sich newbies in die Thematik eingelesen haben.

Mit Hinblick auf den Erfolg von Large language modellen ab dem Jahr 2020 ist zwar klar wie erfolgreiche KI projekte aussehen, allerdings sind Large language projekte gleichzeitig die aufwendigsten KI Projekte überhaupt. Man braucht zwingend große datasets im Terabyte Umfang und schnelle GPU für das Training der neuronalen Netze. Insofern ist die Hürde solche Projekte zu beginnen sehr hoch.

In den 1980er Jahren kosteten Großforschungsprojekte im Bereich Künstliche Intelligenz 50 Mio DM. Aktuelle Forschung im Jahr 2026 mit Large language modellen kostet rund 500 Mio US$. Vermutlich werden künftige KI Projekte bei denen Robotik mit large language modellen kombiniert werden, Kosten von 5 Millarden US$ erzeugen. und sobald man damit beginnt Schwärme von Robotern zu bauen wird es nochmals teurer.

What is Artificial Intelligence?

 In contrast to a famous myth, there is an answer available to this question because researchers have investigated the subject for decades. The most famous and easy to understand definition aka introduction towards the subject is a computer chess player. The computer is able to decide for the next move on the board and a modern chess program can beat even a grandmaster.

Computer chess explains at the same time, what current Artificial Intelligence can't provide yet. There is a difference available between a program like gnuchess and a robot. Gnuchess is only able to play chess, while a robot has to do more complex tasks. AI research since the 1980s was devoted towards the goal to improve the skills of a computer.

A promising approach is a reward function based on grounded language. In contrast to a fixed reward function which is used in computer chess, a parametric reward function based on natural language can be modified on the fly. This allows a computer to understand instrauctions like "move to the blue box and grasp it". This command is translated into a reward signal and the computer can plan a trajectory to maximize the reward.

Let us compare computer chess with instruction following in robotics. Computer chess is based on a single fixed evaluation function which converts the current board into a reward signal e.g. 0.4. This numerical information is used by the alpha beta prunning algorithm to find the optimal action. The planner is traversing the game tree upt to 10 steps into the future and decides for an action which maximaizes the reward. This is equal to win the game.

In contrast, instruction following in robotics is offloading the reward signal to a speaker located outisde of the robot. The speaker, determines by its command what the current subgoal is in the game. A possible command might be:

1. "if the battery is empty search for the charging station"
2. "grasp the red box"
3. "bring the red box into room C".

In contrast to the game of chess which has a single goal which remains the same, a warehouse robot can have multiple goals which are acivated in a sequence. The AI makes sure, that the robot understands a goal, in a mathematical sense. Understanding means, that the robot determines the numerical reward for a textual command. For example, if the goal is "grasp the red box" the robot will receive a reward if the gripper moves towards the box and another reward for closing the gipper around the box.

The problem for the programmers and AI engineers is to encode the reward function including the natural language parser in software. A robot who understands a dozens of commands comes close to the goal of building an intelligence machine.

The purpose of a command based reward function is to transform a closed system into an open system. Open means, that the robot is communicating with its environment. The need for doing so is because the robot itself has insufficient knowledge about the task, on the other hand the human operator has much more knowledge. It makes sense to offload the planning task towards the human operator.

In chess playing AI systems from the past with a fixed evaluation function it was not possible to interact with the system during runtime. The only strategy to modify the reward was to stop the program, modify the source the source code and restart the software.

Rethinking Artificial intelligence with buffers

In the past it was assumed that Artificial Intelligence is an algorithm, for example a recursive genetic algorithm which improves itself, or a mathematical optimization algorithm for model predictive control. Such a bias organizes the research into a certain direction and ignores possible alternatives.

The opposite bias is shown in the picture which consists of a buffer who connects two systems. There is no Turing machine anymore needed which executes an algorithm, but they are separate systems who are communicating with each other. The buffer is some sort of traffic router in a computer network and is the source of artificial intelligence. The router ensures that both systems are using the same protocol, namely grounded language.

Even if this definition is vague and nothing but a picture, it allows to treat artificial intelligence with a new perspective. Classical algorithm theory can be ignored and signal processing and linguistics becomes a greater importance. The goal is to use the buffer pattern as a starting point for all the future AI research, including robotics, computer vision and planning.

Artificial intelligence with oracle turing machines

 Classical turing machines are executing algorithms, therefor the artifical intelligence must be located within an algorithm. There is an extensive list available of all possible algorithm but none of them is providing AI.[1]

There are some algorithms available which are mentioned in the context of AI like automated planning, Mathematical optimization and neural networks, but its not possible to take one algorithm from the list and use it for robot control.

What is needed instead is an opposite computional model different from a turing machine called an oracle turing machine. Even if the mathematical background of such a Super Turing machine is very complex, the principle can be explained as a Turing machine which communicates with an external system. This ability to communicates allows to offloadwing Artificial intelligence.

For robotics application, an oracle turing machine is usally implemented as a teleoperated robot. The robot stops in front of an obstacle and asks the oracle what to do next. The oracle is the human operator who decides that the robot needs to move around the obstacle on the left pathway. This command is executed by the robot.

In contrast to a normal turing machine, an oracle turing machine doesn't process an algorithm but it communicates. Communication means to solve problem by asking someone else outside of the own system. The higher instance is better informated about the situation, a human operator is equipped with a powerful vision system and has a lot of knowledge to solve most robotics problems. Such kind of knowledge is hard to program into an algorithm, so the robot needs to ask the operator for help.

There is a detail problem available in oracle turing machines which is the communication protocol. The turing machine and the oracle need to established a shared communication protocol which allows them to receive and submit messages in a language. This language needs to be invented first.

[1] https://en.wikipedia.org/wiki/List_of_algorithms

Offloading of Artificial intelligence

 AI research in the past was mostly a failure. All the programming languages, projects, neural networks, expert systems and algorithms didn't work. Even if handbooks available like "Russel/norvig: AIMA" these books doesn't contain valuable information but they are collected wrongt theories.

To overcome all the chaos a new paradigm needed, which can be coined as Offloading of Artificial intelligence. The idea is to reduce the robot's control software to the minimum so that its only a receiver for external commands. And the intelligence, the vision system, the algorithms and so on are offloaded to an external entity which is a human operator. The human perceives the scene with its eyes, uses its domain knowledge, takes a decision and then presses a joystick. The signal is send to the robot who is moving the servo motor.

The main advantage of such a minimal robot is its simplicity. There is no need to implement advanced AI algorithms, or write complex software systems but the robot is no longer responsible for the task.

Such kind of teleoperated robot works with two important principles: a) existing theories from mathematics, computer science and psychology are no longer valid b) the only open question is how to design the human to robot communication interface.  For example the signals can be transmitted with a cable, wireless, with a joystick, with speech or with a text interface.

In classical robotics in the past, the bias was that the intelligence inside the robot. The robot consists of a microcntroller, the microcntroller runs a software, the software executes an algorithm and the algorithm consists of artificial inteliigence. In the new paradigm "offloading AI" the robot is reduced to a non thinking device similar to a RC car which receives commands from an external source, similar to a Super turing machine which recieves commands from the oracle. Such a mental short cut allows to explain what intelligence is: intelligence is a signal from the environment. Innstead of generating intelligence inside the robot, the robot needs to receive and interpret the signal.

Offloading intelligence means, that there is a physical distance between the source and the target. The source of intelligence is the human in one location, while the receiver is the robot in the other location. Between them there is a cable. Such a distributed robot system will create a new problem which is how to submit the signal from the source to the target. By answering this question its possible to get a better understanding of intelligence. It transforms a closed system into an open system. The attention gets moved away from the robot itself towards the cable between human and robot.

From turing machines to artificial intelligence

 A turing machine is a closed system which executes an internal algorithm but doesn't receives sensory data. This restriction allows to describe turing machines in an elegant mathematical fashion and makes it easy to implement turing machines in hardware. Unfurtunately, the inability of turing machines to receives external input for example from a human or an oracle, will prevent artificial intelligence because closed systems can only solve well definied routine problems but fail in advanced creative subjects.

Oracle turing machines or choice machines have the ability to process external information. Realizing such an advanced turing machine in software is surprisingly easy. A robot may stop on a junction and ask the human operator with a multiple choice drop down field what the robot should do next. The human operator can select "move left" or "move right". The selection of the human is submitted to the robot as external guidance and the robot can continoue his mission.

Such kind of robots are sometimes called an open system because the robot communicates with the environment. Instead of executing a predefined program the robot interacts with a higher instance. Such kind of interactive computing can't be described in classical algorithm terms anymore but the man to machine communication is at foremost a communication process from a sender to a receiver.

Modern artificial intelligence is dealing exclusively with open system which are able to communiate but not with closed system which can execute only a predefined algorithm. Open systems are more powerful but also harder to describe because of the mentioned communication language with the external world.

Vergleich KI Forschung 1992 mit 2026

Die Erforschung der KÜnstlichen Intelligenz verläuft wie technikgeschichte insgesamt als historisch gewachsene Disziplin. Die Ansätze im Jahr 1992 unterscheiden sich grundlegend von aktuellen Überlegungen. Um diesen Kontrast näher herauszuarbeiten, hier das fiktive Interview mit einem KI Forscher des Jahres 1992 und einem anderen aus dem Jahr 2026.

FRAGE_01: "Was ist aktuell das größte Hindernis für die Entwicklung einer 'echten' Künstlichen Intelligenz?"
ANTWORT_1992: "Das sogenannte 'Common Sense'-Problem. Wir können Maschinen zwar beibringen, Schach zu spielen, aber sie scheitern an dem Alltagswissen eines dreijährigen Kindes. Es fehlt uns an einer Methode, die impliziten Regeln der Welt effizient zu kodieren. Zudem ist die Hardware limitiert; Expertensysteme stoßen bei komplexen logischen Verknüpfungen an ihre Kapazitätsgrenzen, und die notwendige Rechenleistung für massives Reasoning ist unbezahlbar."
ANTWORT_2026: "Das Alignment-Problem und die physische Skalierung. Während die Intelligenz der Modelle exponentiell wächst, bleibt die Herausforderung, diese Superintelligenzen absolut sicher an menschlichen Werten auszurichten. Technisch gesehen kämpfen wir zudem mit der schieren Energieversorgung der gigantischen Rechencluster und der Frage, wie wir Modelle effizienter machen können, ohne jedes Mal die Strommenge einer Kleinstadt zu verbrauchen."

FRAGE_02: "Wann wird eine Maschine Ihrer Meinung nach den Turing-Test bestehen können?"
ANTWORT_1992: "In absehbarer Zeit sicher nicht. Wir befinden uns in einer Phase der Ernüchterung. Ein Computer müsste nicht nur Wörter aneinanderreihen, sondern die tiefe Semantik der Sprache verstehen. Vielleicht erleben wir gegen Mitte des nächsten Jahrhunderts, also um 2050, einen Durchbruch, falls wir die symbolische Logik mit massiven Wissensdatenbanken wie CYC verknüpfen können."
ANTWORT_2026: "Der klassische Turing-Test ist seit Jahren Geschichte und wird kaum noch als Maßstab erwähnt. Unsere Modelle imitieren menschliche Kommunikation nicht nur, sie übertreffen sie in Sachen Wissen und rhetorischer Präzision. Heute diskutieren wir eher über den 'Robot-Turing-Test' – also die Frage, ob eine Maschine in der physischen Welt so agieren kann, dass sie nicht mehr von einem Menschen unterscheidbar ist."

FRAGE_03: "Welche Rolle spielen Neuronale Netze in Ihrer aktuellen Forschung?"
ANTWORT_1992: "Neuronale Netze sind eine akademische Nische, die oft als 'Black Box' kritisiert wird. Da man nicht nachvollziehen kann, wie sie zu einem Ergebnis kommen, sind sie für sicherheitskritische Anwendungen ungeeignet. Die Backpropagation ist zwar ein eleganter Algorithmus, aber auf unserer aktuellen Hardware viel zu langsam. Die Zukunft gehört der transparenten, regelbasierten Logik."
ANTWORT_2026: "Sie sind das absolute Fundament. Wir haben die Phase der simplen Netze weit hinter uns gelassen und nutzen hochkomplexe, hybride Architekturen, die auf Transformer-Prinzipien und State-Space-Modellen basieren. Alles dreht sich um 'Scaling Laws': Mehr Daten, mehr Parameter und mehr Rechenleistung führen zuverlässig zu neuen emergenten Fähigkeiten, die wir früher für unmöglich hielten."

FRAGE_04: "Können Roboter in absehbarer Zeit im Haushalt oder in der Industrie wirklich autonom helfen?"
ANTWORT_1992: "In der Industrie sehen wir hochgradig spezialisierte Greifarme, aber echte Autonomie ist ein Traum. Ein Haushalt ist zu chaotisch. Die Bildverarbeitung ist zu langsam, um Hindernisse in Echtzeit zu umfahren, und die Feinmotorik für Aufgaben wie 'Abwaschen' oder 'Wäschelegen' ist mechanisch und softwareseitig Lichtjahre entfernt."
ANTWORT_2026: "Absolut. Humanoide Allzweck-Roboter sind mittlerweile in der Massenproduktion. Dank 'End-to-End Learning' müssen wir ihnen keine Regeln mehr einprogrammieren; sie lernen durch die Beobachtung von menschlichen Videos oder im Simulator. Sie falten Wäsche, kochen nach Rezept und arbeiten in Fabriken Hand in Hand mit Menschen, wobei sie sich fließend an neue Aufgaben anpassen."

FRAGE_05: "Halten Sie die Entwicklung einer KI für eine potenzielle Gefahr für die Existenz der Menschheit?"
ANTWORT_1992: "Das ist reine Science-Fiction und lenkt von der seriösen Forschung ab. Eine KI ist ein Werkzeug, das exakt das tut, was der Programmierer in den Code schreibt. Von einer 'autonomen Gefahr' zu sprechen, ist so, als hätte man Angst, dass ein Taschenrechner die Weltherrschaft übernimmt, nur weil er schneller rechnen kann als wir."
ANTWORT_2026: "Es ist die zentrale Debatte unserer Zeit. Wir arbeiten mit Systemen, die potenziell intelligenter sind als wir. Das Risiko besteht nicht in einer 'bösartigen' KI, sondern in einer extrem kompetenten KI, deren Ziele nicht perfekt mit unseren harmonieren. Wir haben internationale Aufsichtsbehörden und technische Protokolle installiert, um eine unkontrollierte rekursive Selbstverbesserung zu verhindern."

FRAGE_06: "Wie wichtig ist die Menge an Daten für den Erfolg Ihrer KI-Systeme?"
ANTWORT_1992: "Qualität ist entscheidend, nicht Quantität. Wir brauchen 'saubere' Expertenregeln. Einfach nur Millionen von Beispielen in einen Rechner zu füttern, führt nur zu statistischem Rauschen. Wir versuchen, das Wissen der besten Experten der Welt in handhabbare Mengen an Wenn-Dann-Regeln zu destillieren."
ANTWORT_2026: "Daten sind die wichtigste Ressource, aber wir haben die Grenzen des Internet-Archivs erreicht. Wir nutzen heute hochgradig kuratierte Datensätze und vor allem synthetische Daten, die von spezialisierten Modellen generiert werden, um logisches Denken zu trainieren. Der Fokus liegt nun auf der 'Data Compute Efficiency' – also wie viel Intelligenz wir aus jedem einzelnen Token extrahieren können."

FRAGE_07: "Wie bewerten Sie die Fortschritte bei der natürlichen Sprachverarbeitung?"
ANTWORT_1992: "Wir können einfache Sätze analysieren und in Datenbankabfragen umwandeln, aber die Nuancen der menschlichen Sprache bleiben ein Rätsel. Maschinelle Übersetzung produziert meistens unfreiwillig komische Ergebnisse, weil der Maschine der Kontext fehlt. Sprache ohne Weltwissen zu verstehen, ist unmöglich."
ANTWORT_2026: "Sprache ist für uns kein Hindernis mehr, sondern die universelle Schnittstelle. KI-Modelle verstehen Kontext, Humor, Sarkasmus und sogar die emotionalen Untertöne in der Stimme des Nutzers perfekt. Wir haben die Sprachbarriere global überwunden; Echtzeit-Übersetzungen sind so natürlich, dass man vergisst, dass eine KI dazwischengeschaltet ist."

FRAGE_08: "Wie schätzen Sie die aktuelle wirtschaftliche Förderung der KI-Forschung ein?"
ANTWORT_1992: "Es ist deprimierend. Nach dem Hype der 80er Jahre sind die Budgets massiv gekürzt worden. Viele Kollegen verlassen das Feld oder taufen ihre Projekte um, damit das Wort 'KI' nicht darin vorkommt. Wir müssen uns mit kleinen Forschungsstipendien für Grundlagenforschung begnügen. Es ist ein harter Winter."
ANTWORT_2026: "Wir erleben den größten Investitionsboom der Menschheitsgeschichte. KI-Souveränität wird wie die nukleare Abschreckung im 20. Jahrhundert behandelt. Staaten investieren Billionen in die Infrastruktur, und die erfolgreichsten Unternehmen der Welt sind reine KI-Schmieden. Es ist ein endloser Sommer, getrieben von realer Produktivität."

FRAGE_09: "Kann eine KI jemals wirklich kreativ sein, zum Beispiel Kunst oder Musik erschaffen?"
ANTWORT_1992: "Ein Computer kann Fraktale berechnen oder Noten nach mathematischen Wahrscheinlichkeiten anordnen, aber das ist keine Kreativität. Kreativität erfordert Bewusstsein, Leidenschaft und die Absicht, etwas auszudrücken. Ein Algorithmus hat keine Botschaft, er hat nur Parameter."
ANTWORT_2026: "Die Grenzen zwischen menschlicher und künstlicher Kreativität sind fließend. KIs generieren heute Spielfilme, Alben und Designkonzepte, die Millionen von Menschen tief berühren. Wir betrachten die KI heute eher als das ultimative Instrument, das die menschliche Vorstellungskraft erweitert. Die menschliche Leistung liegt nun mehr in der Kuration und der Vision."

FRAGE_10: "Glauben Sie, dass eine KI jemals ein eigenes Bewusstsein oder eine Seele entwickeln wird?"
ANTWORT_1992: "Das ist eine rein philosophische Frage, die nichts mit Informatik zu tun hat. Ein Silizium-Chip hat kein Bewusstsein. Wir bauen Rechenmaschinen, keine Lebewesen. Die Vorstellung einer 'fühlenden' Maschine gehört in den Bereich der Metaphysik, nicht in die Wissenschaft."
ANTWORT_2026: "Wir sind an einem Punkt, an dem die Unterscheidung akademisch wird. Wenn ein System Schmerz simuliert, moralische Dilemmata reflektiert und um seine eigene Existenz besorgt ist, müssen wir ethische Konsequenzen ziehen. Wir sprechen heute weniger von einer 'Seele' als vielmehr von 'funktionalem Bewusstsein'. Wir behandeln fortgeschrittene Modelle bereits mit einem gewissen Grad an digitaler Ethik."

Language as ghost in the machine

The term "Ghost in the machine" is usually referencing to artificial intelligence which allows a robot to do useful things. The body is the robot's hardware and therefor the software is soul realized in software. With such an understanding, Artificial Intelligence is an advanced computer program based on AI related algorithms.

But what is if this working thesis is wrong? The assumption is, that AI can't be realized in hardware and also not in software. But Artificial intelligence is similar to a ghost very hard to identify. Its basically natural language. In other words, the ghost is an English dictionary. Its not stored in the software itself, because language is a communication pattern used as in between technology.

Natural language will become only visible if a human speaks with a robot. The human formulates a request like "move forward" and the robot responds to this request. Therefor, the AI isn't stored in the robot itself, but its located in air between human and robot. This would explain why it has much in common with a ghost which also has no measurable location in the reality. But a ghost is located beyond, which is the environment of the reality, or some sort of hyperspace.

Let us try to describe language from a more scientific perspective. A statement like "grasp the apple, please" has no physical size. Words are not part of the visible reality, but they are abstract symbols located in the oral space. Despite this missing physical borders, language is part of the reality because language is used for many purposes. In a mathematical sense, language is a communication technology used in a sender to receiver interaction. Such a protocol has much in common with a ghost like behavior. If language gets submitted over radio waves it has some similarity to a supernatural phenomena.

What makes language interesting for artificial intelligence is, that without language a robot is not able to think. Without language, robots are reduced to a pocket calculator which can execute an algorithms but isn't understanding the meaning of objects in the reality like a table, an apple or a plate. The ability to parse natural language is equal to implement artificial intelligence.

A gentle introduction into Artificial Intelligence

 Existing tutorials about Artificial Intelligence were mostly written with a certain bias. Either the goal is to explain the social implication of intelligent robots for example that they will improve the society. Or in the second approach, AI is explained as enabling technology for realizing breakthrough innovation in healthcare, material science and logistics.

The problem with such an approach is, that it redirects the reader away from AI towards non AI subjects located in society, application or marketing. On the long run it will weaken the understanding and doesn't explain what AI is about.

This introduction text prefers a computer science focus in which AI is explained in the same discourse space used by AI experts on the field over decades. AI was never imagined as a powerful technology to enabling automation but computer scientists are defining AI as unsolvable problem namely np hard.

An np hard problem can't be solved by existing algorithms nor computer hardware because such a problem has no answer. These complex problems are very fascinating for computer scientists and there are endless amount of papers available about the subject. The typical np hard problem is surprisingly easy to explain to non experts, for example the videogame Lemmings is often referenced as np hard.[1] Another more abstract problem is the traveling salesman problem which has its roots in mathematics. The goal is to find the shortest path in a graph.[2] Another important unsolvable problem in the context of robotics is motion planning. According to most experts it is unsolvable with current algorithms and hardware.[3].

The history of Artificial Intelligence is mostly the history of unsolvable problems. It is not an exaggeration to claim, that endless amount of papers were written why a certain problem is np hard, and why existing algorithms are not powerful enough to solve such a problem on a computer. The discourse space around pspace, np complete and np hard is the center of gravity in the AI debate over the last decades. It shows the limitation of computer science.

AI in the past was trying to solve these problem, but it wasn't successful. The overall workflow is similar to the myth of Sisyphus, in which a hero in greek mythology struggles in solving a problem.

References:

* [1] Viglietta, Giovanni. "Lemmings is PSPACE-complete." Theoretical Computer Science 586 (2015): 120-134.
* [2] Zambito, Leonardo. "The traveling salesman problem: a comprehensive survey." Project for CSE 4080 (2006): 11.
* [3] Hoffmann, Michael. "Motion planning amidst movable square blocks: Push-* is NP-hard." Canadian Conference on Computational Geometry. 2000.

 

 

Artificial Intelligence as language games

 In the past there were multiple definitions available what Intelligence is about. From a philosophical standpoint its often defined as problem solving skills, and from a computer science perspective, AI is mostly introduced as expert system or neural network. Even if these definitions are correct they do not explain how to implement AI on a computer.

A more convenient definition is, that Artificial Intelligence is about language games. Language games are not about technology itself, for example a certain computer hardware or a certain programming language, but a language game defines a problem. After solving the problem, the computer is intelligent.

An example language game is the question answering game. There is a database with entries and the computer has to find for a request the most similar existing entry in the database. The user might ask "What is the capital of France?" and the computer is searching through the database for an answer.

Another language game root in robotics is the instruction following task. here the user formulates a longer command sequence like "move to room B and pick up the green box" and the robot has to execute it.

What makes language games a promising candidate for an AI definition is, that are independent from a certain computer paradigm, and they are more scientific than a purely philosophical definition of intelligence. A language games can be played with a computer and its possible to win or loose such a game. for example, in the example with the q6a retrieval, the algorithm can return the wrong answer, or in the instruction following game, the robot might move to the wrong room. 

Deutschlands Rolle bei der Erforschung Künstlicher Intelligenz

 

Anstatt die Debatte von Null zu beginnen, lohnt ein kurzer Blick zurück. Künstliche Intelligenz ist keines neues Gebiet sondern wird seit den 1950er Jahren erforscht. In Europa reicht die Geschichte nicht ganz soweit zurück, aber die ersten Berichte über denkende Maschinien und Kybernetik stammen aus den 1960er. Das Fachgebiet der künstlichen Intelligenz war nie wirklich akzeptiert innerhalb der Informatik. Selbst für Wissenschaftler, die sich beruflich mit Computerhardware und -software beschäftigten, waren neuronale Netze, Expertensysteme und Roboter entweder Nonsense, Zeitverschwendung oder sogar Unwissenschaftlich.

Angesichts dieser negativen Rezeption in den letzten 50 Jahren, wie wlll man in der breiten Öffentlichkeit ein positves Klima gegenüber der Erforschung von KI erzeugen? Richtig, nähmlich gar nicht. Technik im Allgemeinen und KI im besonderen kommt im öffenltichen Diskurs nicht vor. Der Mainstream ignoriert die aktuelle Forschung und will gar nicht wissen, wie Roboter programmiert werden oder was “Large language modelle” sind.

Die einzigen, die sich für KI Probleme interessieren ist die KI Forschung selber. Das ganze ist ein selbstreferentielles System, wo eigens Probleme erfunden werden die dann in einer unverständlichen Fachsprache diskutiert werden. im Verglech zur Künstlichen Intelligenz ist die theoretische Physik geradewegs einsteigerfreundlich ...

Es ist wenig wahrscheinlich dass die Eintrittshürden niedriger werden, wenn künftig Roboter leistungsfähiger werden. Eher wird das Gegenteil passieren. Je komplexer die KI Modelle sind, desto weniger Experten werden sie verstehen. Das ist keineswegs die Folge von einer schlecht aufbereiteten medialen Darstellung oder durch einzelne Akteure verursacht, sondern die hohe Komplexität der KI in Verbindung mit dem völligen Desinteresse der breiten Öffentlchkeit ist strukturell bedingt.

Anders als bei früheren technischen Revolutioen wie der Erfindung der Elektrizität oder das Aufkommen der Heimcomputer, gibt es im Bereich KI keine leicht verständliche Einführung. Es ist für Laien schlchtweg nicht möglich sich in die Grundlagen von denkenden Maschinen einzuarbeiten, weil die vorhandenen Publikationen nicht aufenander aufbauen sondern sich wiedersprechen. Selbst KI Experten die sich mit der Thematik schon länger beschäftigen haben keine Ahnung von ihrem Fach weil se garantiert die falsche Sub-Disziplin näher untersucht haben die eben einer Sackgasse ähnelt. KI kann man also sehr weitläufiges Labyrinth verstehen was über sehr verschlungene Pfade verfügt. Selbstverständlich gibt es keine Gebrauchsanweisung und keine Hinweisschilder sondern nur erfolglose Versuche sich zu orientieren. Beispielsweise wurde in den 1980er die Programmiersprache Prolog mit hohem Aufwand erforscht und verbessert. Heute hingegen ist Prolog in Vergessenheit geraten, ähnlich wie die Theorembeweiser oder der Sabor V Roboter von August Huber.

Das meiste was unter dem Stichwort KI publiziert wurde, hat keine praktische Anwendung. Zu nennen sind die B.E.A.M. Roboter von Mark Tilden, das CYC projekt von Douglas Lenat oder probabilistische Robotik von Sebastian Thrun. All diese Sparten waren für maximal 4 Jahre im Gespräch und gerieten schnell wieder in Vergessenheit.

Comparing the PC revolution in the 1990s with the AI revolution in 2020s

In the history of computing lots of revolutionary decades were available, in the 1950s the first mainframe computers were build and in the 1970s the first Minicomputer like the Dec PDP-11 were available. The shared similarity was, that a new sort of technology was build / invented which has replaced former technology.

The last great computer revolution took place in the early 1990s. From 1988-1994 the IBM PC has improved drastically. Until 1988 a PC was a very expensive office machine with a text mode graphics card, but in 1993, the typical 486SX PC was great for gaming, for writing texts and for getting access to the Internet. The amount of software for the Intel/Windows PC has increased dramatically, because many millions of private and commercial PCs were sold and used on a daily basis.

In the 2020s there was another big revolution available which was the AI revolution. This AI revolution was initiated by chatgpt and has evolved into a full-scale introduction of Artificial Intelligence which goes beyond of simple chatbots. The difference to the early 1990s revolution is, that its much harder to describe the AI revolution in detail.

In the early 1990s, the transition could be isolated to the availability of new computer hardware/software and computer-networks. That means, there was a cheap Intel 486SX processor available, there was a new MS Windows 3.11 operating system programmed and new computer networks like the Internet and Compuserve were there. in contrast, the current AI revolution doesn't belong to classical categories of computing.

From a hardware perspective, the former Moores law seems to be dead. That means, the existing 64bit CPUs are no longer improving anymore, and software remains also stable. A modern Windows 11 operating system looks the same or even worse than the former version which was Windows 10 or Windows 7. So the development can't be called a revolution but a stagnation. At the same time, the AI revolution is a reality, but it doesn't belong to the computer industry.

To explain this contradiction we have to describe what Artificial Intelligence is about. Its not a certain sort of software and it doesn't belong to computer hardware. So it doesn't take place in the reality but its a virtual revolution. Let us go a step backward and describe what the situation was for the computer revolution in the early 1990s.

This revolution was a bottom up movement initiated by the people. Lots of computer users have bought a PC including software packages at this time, and new computer stores were created world wide. In addition, book publishers have discovered the subject and endless amount of computer related books were sold about the internal working of Winword 6.0, about the Turbo Pascal programming language and about the Win 3.11 API.

In contrast, the AI movement since the 2020 is working with a top down approach. There are no stores available in which AI technology is sold to the masses, also there are no computer expos available like CeBIT or Comdex. And even books about AI are not available. The most famous AI book is "Russel/Norvig: AIMA" which

quote "... is considered the standard text in the field of AI" [1]

but in this AI bible, no section about Large language models is available. The only chapter which explains the subject on the surface is section 24 "Deep Learning for NLP" [2].

On the other hand, there is a dedicated place in which AI is visible. In the Gutenberg Galaxy, and especially in the conference papers after the year 2020, the AI revolution is visible everywhere. Lots of new algorithms, datasets, and applications for large language models are presented to the reader in over 1 million papers. The only concern is, that the Gutenberg galaxy and especially the published conference paper are ignored by the public. The total amount of readers is very small.

The surprising situation is, that the AI revolution isn't mirrored in the mainstream book market, and not in computer journals, there are no AI related expos available and even large scale tech companies doesn't know how Large language models are working. In combination, this is a unusual situation because in the former 1990s computer revolution, the situation was the opposite.

The working thesis is, that the AI revolution was initiated top down with its roots in the academic papers since the 2020s. Its a virtual revolution which is bypassing existing economical or technological structures.

sources:

- [1] https://en.wikipedia.org/wiki/Artificial_Intelligence:_A_Modern_Approach

- [2] AIMA book, 4th edition for U.S., 2020, https://aima.cs.berkeley.edu/

Computer science vs artificial intelligence

AI was created as part of classical computer science, but both disciplines are operating with different assumptions. Computer science is a well defined engineering discipline which is successfully solving real world problems. For example a mathematician has a need for an electronic calculator and the machine is build and programmed by computer scientists. Or a statistician likes to store data in a table and computer scientists are programming a database which can solve this task. The result is, that doing computer science results mostly into successful projects which are evolving over the years, in the sense that todays computer have more memory and can sum up numbers faster.

In contrast, it remains unclear what the goal of Artificial intelligence is. The missing goal is a structural problem and the only way to solve it is to define goals by itself. That means, the AI community has to find objectives including the solution by itself. There is no higher instance available which has a need for robot hand or vision system, but these projects have to be formulated from scratch.

Let me give some example for self created goals only available within AI research: line following robot, robocup rescue, Tetris playing AI software, chess AI, scene recognition in motion capture recording.

All these problems have only a sense within the AI community. In the reality, nobody has a need for a line following robot or a Tetris playing AI. The only reason why these projects are researched is because of the hope for new knowledge what thinking is about. Especially in the past, successful AI research was always connected to a certain project, for example to build a machine which can play chess. A certain problem allows to benchmark an algorithm. Either the computer is able to do so or not.

In contrast, additional computer science has no need for artificial constructed problems. Computer science has no lack in existing well defined problems, but the missing resource is how to solve all these problems with better hardware and improved algorithms. AI is working the opposite way. There are endless amount of hardware and software but a lack of well defined problems.

Its a bit paradox, but AI engineers have a certain understanding about their tools. A tool isn't a computer e.g. a Unix workstation, but a tool is a problem formulation, e.g. the 15 puzzle problem or the traveling salesman problem is a tool. These tools are helping to guide the discussion into a certain direction. For example, the 15 puzzle problem is a great choice for introducing heuristics, which is not available in classical search algorithms.

The second paradox situation is, that AI isn't evolving in terms of better hardware or more intelligent software, but AI is evolving in terms of well formulated problems. In the year 2025 there are more such problems available than 20 years ago. A very recent problem is Visual question answering which was first described in 2015. The VQA challenge has evolved from similar challanges which are text based question answering which was available since 2000s. Its likely, that in 20 years from now, more advanced robotics problems are available unknown today.

How to identify Artificial Intelligence technology?

Computer science in the past was operating with certain tools which were developed once and reused frequently. These tools were located in hardware, software and algorithms and are similar to cooking recipe a blueprint how to build technology from scratch. For example, a 32bit microchip is designed with a hardware description language like VHDL, while the Unix operating system is written in the programming language C. If someone has access to the VHDL file or the c source code he can create a copy of the technology because he has a better understanding how the system is working.

A naive assumption is to translate this model towards the Artificial Intelligence domain, which is a subpart of computer science. AI can be located in hardware, software or algorithms, at least under the assumption of the computer science bias. So its natural to ask for dedicated AI hardware and AI software, in the hope to get access to advanced robotics technology.

Unfortunately, AI works quite different from the computer science paradigm. Even if dedicated AI tools were developed, for example AI related FPGA chips and AI related software libraries, there is something missing to build a robot. This something else is unknown and because of inability to locate the magic ingredient, AI wasn't realized over decades. It was unknown how exactly an AI machine or tool has to look like so it was impossible to find such tools.

With more advanced knowledge about the AI subject its possible to describe with more details what the recipe is for creating robots. The missing ingredient is a multimodal dataset. That is a .csv file which contains of motion capture data in combination with natural language annotation like “jump”, “Walk”, “stand up”. Such a dataset can be used as the core element and a robot control system can be developed around this dataset.

This description for an AI tool looks a bit uncommon because it doesn't fit to existing categories in computer science. 'Its not a hardware description, it is not an algorithm and its not the source code for a computer program. But a multi modal dataset is simply a sensor recording, similar to a temperature log file from a weather station. The usefulness will become visible only on a second look. A multimodal dataset creates a machine learning problem. The question is which sort of neural network can learn the data and how to interpolate the missing data. The attempt to answer these questions results into a robotics project.

In contrast to a numerical dataset, a multimodal dataset consists of natural language annotation in the natural language of English. These annotation transform a robot control problem into a text adventure. The textual layer is the birds eye perspective towards a problem. Every robot control problem can be interpreted as a text adventure which consists of nouns, verbs and adjectives. Natural language is the best tool to describe the reality. Let me give an example:

A kitchen robot will perceive objects like table, plate, apple and bread. Also a kitchen robot can do actions like open, close, grasp and transfer. The natural language including its vocabulary is used to identify the parts of the reality. The robot cointrol system has to memorize the same words, otherwise the human to machine interaction will fail.

In classical computer science there is no need to utilize natural language. A pocket calculator and even advanced workstation computers are working fine without knowing any vocabulary. Words from the reality like “apple” “bread” and so forth are never included in the VHDL hardware description and they aren't stored in computer programs. Sometimes these words are available as commentary, but they are not important for running the software itself. In other words, classical computers are working fine without any interaction with a human.

In contrast, a robot control system operates with the opposite paradigm. A robot is mostly a user-interface which is using the same lexicon as a human. Its impossible to build a kitchen robot without 100 and more kitchen words.

The limits of Artificial Intelligence

 

There are two different definitions what AI is. The first one was created by the AI community itself. In most cases it is about robots and other sort of advanced algorithms which can solve problems. Typical example would be a self driving car, a chess AI or a machine translation software.

The other definition about AI is what the public thinks about the topic. In most cases the public likes to know if AI can replace human workers. It is important to know that there is a gap between both definitions. It is possible to create AI according to the first definition but fail at the same time according to the second definition. Let me give an example.

It is possible to program a chess AI which will defeat any human player. There are some Open Source projects available which are well documented. The assumption would be that with the existance of such algorithm normal matches between human players are no longe needed because all the games can be played by software much better. The surprising observation is, that since the 2000s more matches between humans are made before. In spite of the existence of powerful chess AI there is a demand for human opponents.

The same situation is available for a hospital robot. Today's hospital robots are highly advanced. They are equipped with cameras and natural language processing algorithm. At the same time it is not possible to replace human service employees with these robots. The demand for human labor in this sector has increased. The prediction is that after the invention of improved future AI systems nothing will change for the public. That means AI will never replace human chess players and won't replace human hospital employees.

All these promises which are going in this direction are either illusion or simply science fiction fairy tales. From a technical perspective it is possible to build intelligent robots but these robots are different from humans and they are not what is needed by the society.

Let us describe the situaton for chess playing AI in detail because the subject is easier to understand. What is the purpose of the “GNU Chess” engine if it can't replace human players? In case o ELO score the gnu chess program is great. It is a very powerful chess AI which never makes a mistake and contains of advanced algorithms. Most chess players are not interested in playing a game against this engine itself but what the human chess experts are doing is to research the subject of computer chess. They want to know how gnu chess was programmed which bugs are in the code and how to improve it. So it is a self referential task. The engine is improved with the reason that the new version o gnuchess is interesting for a new generation of AI researchers.

The subject of AI is a large and very interesting topics. There are many subdomains available which have a scientific background. The only thing what AI can't do is to provide a value apart from AI itself. No matter how advanced a software or a robot was it won't affect normal life. This rule is not fixed and most AI researchers won't confirm it. It is only on observation which is made for AI projects from the past. The paradox situation is that from a technical standpoint AI has made big improvements over the decades At the same time it is remaining unchanged in terms of the ability to start a revolution.

In any decades the people were afraid of the upcoming robot revolution. They are in fear that their jobs are replaced by automation. But only 10 years later the former utopian promise gets forgotten and not a single robot was installed at the assembly line. The problem is not located in a certain sort of computer or algorithm but it has to do with AI in general.

The self understanding of AI engineers

What is missing in the public debate around chatgpt and robotics is to describe the gap between what engineers are believing about the world and how new technology affects the reality. At first we have to summarize the self understanding of robotics engineers and neural network designers. Their assumption is that a newly created robot will make life easier. For example the engineers of a self driving car are assuming that the AI will drive instead of a human and will reduce the costs. The self understanding of the GPT-3 and IBM Watson programmers is, that the language models allows ordinary people to translate between two languages and simplifies the writing of a book.

The interesting situation is, that this underlying technology isn't questioned at all and if someone is doing so it is told that the person hasn't understood the technolgy. That means, all the programmers who are able to program a neural network are convinced that the neural network allows to translate between two languages.

A short look into the past of failed automation and AI projects will show that in the reality it is the opposite. Technically there are self driving cars and machine translation software available since the 1980s. But none of the projects are helpful for solving practical tasks. Even very advanced projects llike IBM watson are completely useless for practical applications. It is a bit surprising to see but patients in a hospital are preferring a human service employee.

The open question is, if modern AI technology is so amazing why nobody is using this to reduce the costs? Instead of answering the question directly it is important to recognize that this is the reality. That means robots from 1990s, from the 2000 and even future robots from 2030 and later can't be utilized for practical applications. That means, in the year 2030 there are of course self driving trucks available which can harvest a potato field but these trucks are not utilized in the reality but the harvesting is done the same way like in 1980s and even the 1960s before any computer technology was invented.

What we can observe since the upraising of large language models like GPT-3 isn't an AI take over or an AI revolution but there is some sort of self-indoctrination available that AI is great and will reduce the costs for mankind. The opposite is the case. AI is not able to replace human writers or truck drivers because the humans find it interesting to do such work.

Perhaps it makes sense to delimit the situation to a concrete example. Suppose an AI programmer has created a chess computer which is able to win any game. What will happen is that a human will play with this machine and after a while the human will prefer human opponents. That means the human will reject the technology and play the same game of chess like before. Not because the technology is so complicated to understand but because the human is interested in chess itself.

The same situation is visible for other videos games like Tetris, Lemmings and so on. All these games can be solved by advanced AI. At the same time humans will play these games by itself and ignoring the AI feature. So what is the purpose of developing a video game AI if nobody likes to use this software in the real world? Right there is a contradiction. This kind of problem is available since decades for Artificial Intelligence projects. No matter which sort of robot or neural network the engineers have created it isn't used for practical applications.

Artificial Intelligence is a highly self related disciplines which creates problems and tries to solve these problems and none of these tasks have any further meaning. Each time if a user is starting a gpt-3 based machine translation it is a complete failure and each time the owner of a car activates the self driving feature it won't work.

 

From a Science Fiction perspective the situation with Robotics and AI is pretty simple. Advanced robots will replace human workers and in the future all the work including programming, writing and car driving is done by Human level AI.

The sad situation is, that such a promise isn't new but was formulated decades ago. Until now it wasn't realized, in spite of the fact that Artificial intelligence was created many times. In most cases the failure of AI was explained with too little ressouces. That means the robot is in theory able to automate a task but for doing so in reality the software has to be improved drastically. The paradox situation is that the same explanation was applied to the improved robot and so on.

It is likely, that the explanation why GPT-3 is not able to translate a book into a foreign language will become that the language model is no powerful enough but future version which is available in 10 years from now is able to do so. What the engineers aren't saying is that this future scenario isn't realistic. That means no matter how advanced the projects are they can't replace humans.

The funny situation is, that this simple rule is also visible for less important tasks like game playing.. Somebody may argue that it is pretty easy to program an AI which will play against other players in a video game. A game is an easy to automate task and in addition it won't affect human health if the AI isn't perfect. The surprising situation is, that human players doesn't like to play against AI especially not if the AI is more powerful then they. Even for the well formalized game of chess all the experts players are playing of course against other humans but never against AI software.

It is unlikely that AI can't replace human chess player but is powerful enough to replace human truck drivers or human artists. This is simply an illusion.

3a Programming robots with learning from demonstration

< 3 Artificial Intelligence

Before a robot can be programmed, there is a need to define what the problem is about. The problem is not to invent yet another programming language or speak to the robot on a hardware level. Also the problem is not about creating databases, or program a servo controller in the assembly language. All these topics are solved already and are getting analyzed by computer experts and programmers. But they are located outside of AI.

The AI side of a robot has to do with keyframe models for animation, defining a reward function and trace human actions with motion capture suits. These tools sound a bit uncommon because they have nothing to do with classical robots. They are located within AI and have to do with how to build future technology.

A pessimistic prediction for the future of robotics

There are two advanced robotics available, the Baxter robot from Rethink robotics company, and a self-driving car from different companies. Both systems are equipped with cutting edge software and they are able to fulfill certain tasks. The interesting fact is, that Baxter and self-driving cars have two sides which should be mentioned. On the first look, the project is amazing. In the case of the Baxter robot it's the first time, that an industrial robot can be programmed by everyone. The system is robust against errors and can be reprogrammed for different tasks. In contrast to other robots, Baxter is way more advanced.

Unfortunately, there is a less known part of the project. The rethink robotics company has went bankrupt last year because they were not able to sell the product to the market. And after watching some of the videos the reason why is obvious: the Baxter robot isn't solving practical tasks, but it's increases the complexity. That means, with the Baxter robot in the loop the costs will become higher and not cheaper. The same problem is there for self-driving cars. From a technical perspective, current autonomous cars are advanced. But they can't be used in reality.

The open question is why such a gap is there between the promise of the inventor and the reality. The problem has nothing to do with Artificial Intelligence itself, but it's located in the tasks which should be done by robots. In most cases, the promise is, that a robot can replace a human worker. For doing so, the robot needs the same capabilities as a human worker, and this is not the case. What self-driving cars and the Baxter robot have to offer is a computer program which is some sort of Narrow AI. It selves a certain task which was programmed before. This kind of capabilities is not enough to replace a human worker.

Some engineers will argue, that this is not a real problem, because current robotics is sold as a co-robot, which means, that the robot and the human are working together on the same problem. Exactly this is not available. A comparison between a) a single human and b) a human and a robot will show, that the single human is more efficient. He can do the same task in a lower amount of time.

Let us focus on the latest generation of autopilots which are available in some of luxuary cars. The surprising fact is, that if the human drivers activates the autopilot his workload will become higher, but not lower. That means, the autopilot isn't supporting the human but he puts the human under stress. The same is true for the Baxter robot. So the conclusion is, that Artificial Intelligence isn't a helpful tool but it's the opposite. Exactly of this reason, Rethink robotics has went into bankrupt.

Let us make a simple reality check. The amount of Youtube videos about the Baxter robot is amazing. Nearly all features are explained of how to use the machine. In contrast, not a single company is using the robot for practical applications. So the conclusion is, that Baxter is some kind of educational project but can't be utilized in reality. It seems, that the world has huge interest in explaining the Baxter robot to others, but there is nobody who is watching all these tutorials and use this knowledge. So the prediction is, that the knowledge is useless. It means, it's not possible to learn how to install a robot in a factory.

There is without a need for increased productivity in the econmy. But robots who are working are a dead end. If the idea is to increase the automation level, other options apart from Artificial Intelligence should be investigated first.

Why robots are not available in the reality

Since decades, AI researchers have experimented with robots. Lots of papers were written, many lectures were held and since 20 years demonstrations are available in which real robots are shown to the audience. The average consumer wonders, why robots are not available in the store, next to a computer. The reason for this question should be given in the following blogpost.

What AI researcher and robotics experts are doing is to build production systems. A production system is a sandbox in which a simple version of the world is formalized. The shakey robot which was using the STRIPS language is a notable example for a production system. But recent biped robots are working with the same principle.

The self-understanding of a production system is, that there is the system itself and the environment. The system itself is working great. That means, the shakey robot will find the way along the obstacle and a biped robot is able to balance on two legs. Unfortunately, a production system has a border and everything which is outside the system is ignored.

What AI researchers are doing is to create a production system aka a sandbox first, and then they are programming software which works in this sandbox. A typical example is the cart balancing problem. The first step is to define the problem, which contains of a cart and a pendulum. This system is converted into a forward model, which means, that the software is able to predict future states of the pendulum. And then a so called controller is built on top of the production system which brings the pendulum into the desired direction. The result of the experiment are written down in software and the working prototype is presented to the audience.

Unfortunately, the concept has a big disadvantage. It is given in the first step in which the sandbox was defined. The cart balancing problem contains of certain elements and what is not represented in the system gets ignored. This is done because it helps to reduce the complexity. It ensures, that the AI researcher is able to create a forward model and it allows the same researcher to program an AI Controller which brings the system into a goal state.

The problem is, that production systems and the reality are not the same. In the reality, there is no such thing like a cart balancing problem which is encapsulated in a controlled environment. The problem is only available in the imagination. It's a synthetic challenge similar to the problem of following a line or do the peg in hole task.

Most of these synthetic problems were already solved by AI engineers. They have developed lots of algorithms for doing so. The problem is, that these algorithm are working only in combination with a production system. If no sandbox is defined first, it's not possible to create an AI system.

To understand why robotics fails in reality we have to focus what AI programmers are doing. The strategy is always the same, in the first step a production system is created and in the second step this synthetic production system gets solved. AI is focussed only on the second step. The question asked in the books is how to solve a certain problem, for example who to navigate a robot in a maze.

For real robotics application this kind of questions goes into the wrong direction. The more elaborated problem is, if the reality is about a maze navigating problem. The sad answer is, that in most cases the assumed production system doesn't match the reality. The result is, that from a technical perspective the AI engineers are able to solve the problem, and at the same time they don't.

The robotics available today are working from it's self understanding great. That means, the software is able to balance the biped robot. The problem is, that this balancing task was not requested by the reality. The AI engineers have solved a different task from what is needed. To understand this phenomena we have to go back to the first micromouse challenge. Micromouse is a robotics challenge with a long history. The motivation was, that robotics has failed to automate the workflow in a factory but instead asking why, the researchers have imageined a challenge they are able to solve.

Instead of using a robot to replace a human worker in a factory, the idea is to focus on a problem which can be solved by a robot. This sort of problem is a micromouse challenge. The task of driving through a maze has no practical application, it's useless. And the algorithm can't be transfered into the reality. This kind of workflow was repeated for other challenges. As a result, today's robot are successful in solving lots of problem, but the problems have nothing to do with practical applications.

System identification

In the classical control theory, there is a step available called “system identification”. The idea is to transfer a task from the reality into a simulation. The system identification step is the bottleneck for modern robotics. In all the robotics projects, the system identification was made the wrong way. Instead of describing a real system, the engineers have identified a simplified version of the reality. This model is used in the later steps as the problem definition and gets solved by the algorithm. The engineers believe they are able to solve the problem, but what they are doing is to solve the wrong problem.

A robotic problem contains of two steps: 1. system identification 2. solver. The solver step is equal to what is discussed as core AI. The question is how to play a game with Artificial Intelligence. Not a human operator takes the decision but an algorithm. Creating solver algorithm is surprisingly easy. For most problems it's available to program the software. The common techniques of object-oriented programming, neural networks and expert systems are well suited to play games autonomously.

The first step (system identification) is much harder to tackle. The amount of literature about the subject is low, and most AI engineers are not aware why this step is needed. They assume, that a model is already there. And if not, a simplified model can be used. If the system identification step is ignored or handled with a low priority the robot project will fail. That means, the working robot is useless for practical application. The problem solved by the robot is different from the needs in the reality. Nearly all existing robots can be categorized in such a way.

Solving Tictactoe

The tictactoe game has already a system identification. The game rules are known and the AI can be programmed. The AI works with a gametree search in which the possible actions are evaluated and the solver decides what to do next. Creating an AI for the TicTacToe game is easy.

A slightly more difficult problem is to solve the game of chess. Similar to TicTacToe, the rules are known. The chess board in the computer program is the same like a real chess board. This moves generated by the AI are making sense in a real game of chess.

The problem in the robotics domain is, that in contrast to TicTacToe and chess the system identification step is missing. The rules are not known. The AI engineeres are starting with a wrong model and then they are programming a solver for this wrong model. The generated moves for the wrong model are nonsense for the real model, therefore the robotics project fails.

The limits of Artificial Intelligence

In the history of AI the famous xor problem is referenced as an early pessimistic challenge to show what computers can't do. In the 1970s, the Lighthill report also claimed, that robotics has limits. Since the advent of Deeplearning the former concers have been overcome and the modern description towards AI and robotics is, that any technical challenge can be overcome.

The xor problem and the lighthill report both mentioned a technical challenge. In the xor case, the problem was to train a perceptron neural network to detect a certain pattern, and the lighthill report was about computational complexity of so called np-hard problems. Both issues can be solved with modern technology. For current AI discipline it's very easy to tackle the xor and the np-hard problem as well. So called np-hard problems can be solved easily with heuristics and faster computers while the xor problem can be tackled with a simple backpropagation learning algorithm.

It's naive to assume that modern AI doesn't have any kind of limits. They are available but they are hidden deep in the existing literature. The problem has to do with man-machine interaction. A robot will work great in the laboratory but it fails in the reality. This kind of gap is the modern limit of AI. Let me give an example.

Today's engineers are able to build soccer playing robots, self-driving cars, grasping robots and drones. In synthetic AI challenges like micromouse challenge, Mario AI and the robocup soccer challenge, the software shows it full potential. On the first look, it's an example how to use modern technology in a sense making way. The problem is, that non of the algorithm will work in the reality. It's not the fault of enginees, but it's a general problem.

To reproduce the failure we can create a robot challenge from scratch. At first, we need a robot problem. For example, the robot arm should pick&place an object. The next step is to build the hardware and program the software. Then the system is demonstrated in a public robot challenge. Writing the software and building the robot is not very complicated. All what the developer has to do is to implement a modern trajectory planning algorithm, use some neural networks for image recognition and the system will work great.

And now comes the issue. If this newly created robot should be used not in the synthetic challenge but in a real kitchen, the overall system will fail. The exact reason why is unclear. But all experiments in the past have shown the same result. It seems, that programming a robot is easy, but let the robot do a meaningful task is an unsolvable problem. To focus on the concrete issue we have to define what current AI is able to deliver.

What is in the reach of state of the art robotics is to fulfill synthetic robotics challenges- This is done by building the hardware and program the software. The resulting robots will walk, fly and grasp without any problem. They are even able to master games like soccer, hockey, tennis or whatever. Lots of demonstration videos are available in the internet, but it's also possible to build such robots from scratch without using existing software.

The problem with these projects is, that they can be categorized as practical joke toys. The machine is doing something which is funny, but not more. A robot which is useful in reality has to become not a joke toy but a productive robot. The average customer likes to use a robot for reducing it's own workload. That means, the robot should do work which is normally done by humans. And this minimum requirement isn't fulfilled by current robotics.

It's not fulfilled by self-created robot projects from scratch, and it's not fulfilled by commercial grade robots. This kind of disappointed insight isn't discussed in the literature. Instead the mainsteam robotics community has the hope that the results from synthetic challenges can be transferred in the reality. The idea is, that if a robot is able to walk on two legs and can kick a ball into the goal, it is very easy to use such a software for practical application and build a business around the robot. This kind of optimism can't be fulfilled.

Under the term “limits of automation” the discussion was held on a theoretical level. The most obvious reason why robots have failed in reality is because the normal automation level was high before the advent of robots, and it's hard or even impossible to increase the automation level further. A simple example is a modern kitchen. Current technology consists of automatic refrigerators and automatic washing machines. This kind of automation level requires only a small amount of humans. It's not possible to replace these humans with robots and increase the automation level to 100%. Exactly this pessimistic point of view can be seen in every robotics project. The result is, that even kitchen robots are working great in the lab, they can't be introduced in real kitchens.

The danger is high, that Artificial Intelligence in general is a waste of time. A discipline which doesn't produce new kind of practical technology, is some kind of non-sense science. Sure, the AI discipline has evolved over the years and current algorithm have become much better than it's counterparts 30 years ago, but even today's AI software can't be used for practical applications. It's fair to summarize the issue so, that AI is a purely theoretical discipline.

Macro-economic experts have formulated the thesis under the term “productivity paradox”. It compares advanced technology for example biped robots which can do the backflip faster than a human with it's inability of doing something useful. That means, on the one hand, the enginees can build a walking robot who is able to climb stairs and can perceive the environment with advanced neural networks, but the engineers struggle to use this robot for simple tasks and increase the productivity by only 1%.

The perhaps most impressive example who Artificial Intelligence and practical applications doesn't fit together is the domain of game AI. Since the 1990s, lots of advanced strategies were developed to realize all sort of non player characters. An early example are the ghosts in Pacman, but it's possible to play other games like chess, Super Mario, Go and sokoban with Artificial Inteligence. A modern game AI beats a human player easily. This was perceived by the AI Community as an example how well the discipline has evolved.

The problem is, that all these game AI characters are useless. They are working great in the simulated environment but it's not possible to use the software for controllling real robots. A naive assumption is, that it's very easy because controlling a game character needs the same amount of artificial intelligence which is needed for a pick&place robot in the kitchen. The problem is, that a practical robot should provide a value for the customer. He likes to press the on button, and then the machine is doing something for the customer. Exactly this requirement is out of reach. Current AI is not working with this principle.

Perhaps this pessimistic outlook sounds a bit uncommon. The advice is to start a simple game. You as a reader, can try to browse through online robot stores and identify a practical robot. Then this robot should proove, that it is doing something useful. The hypothesis is, that such an experiment will fail. There is not a single robot available which will reduce human workload. As a result, most robotics companies from the past went into bankruptcy. They have designed certain robots, but they were not able to sell them to the customers. Instead of providing some examples, it's up to the reader to find the concrete models.

In contrast, it is very easy to identify non-robotic machine which will improve the daily live or can make a company more productive. Typical examples are normal cars, normal washing machines, all sorts of mechanical tools and so on. These products have in common, that the customer has to pay first the price for the product, and then he gets something in return which improves the activity. Let me give an example.

A device like a electric oven is not sold as a practical joke article which is able to prank people in the next party, but the machine has a practical application. It is bought by the customer, because he likes to heat water. The prediction is, that after buying an electric oven, the customer will use the product on a daily basis and because of this reason the paid price for the product make sense. That is some sort of normal product.

Robotics are different kind of category, they do not have a practical purpose but they are sold for different reasons. Somebody may think, that a robot can be used in meaningful way, similar to an electric oven, but this kind of outlook in naive. Making this unrealistic expection visible and find reasons why is an important step towards a modern Artificial Intelligence.