Timeline of AI history

- 1950-1990 computer generation #1 to #4
- 1990-2000 AI winter
- 2000-2010 Deep learning hype
- since 2010 AI hype
- since 2023 chatgpt available

Classical computing consists of computer generation #1 to #4 which went from 1950-1990. The last and most important fourth computer generation was from 1980 until 1990 and included the homecomputer, internet protocols and the PC including graphical operating systems. The period from 1990 until 2000 can be called the last AI winter. During this decade the prediction for AI and robotics was pessimistic, it was unclear how to build and program such devices. Since 2000 more optimism was available which can be summarized with the buzzword deep learning. Deeplearning means to create a dataset and search in the data for patterns. Since around 2010, Artificial intelligence has become a mainstream topic. New innovation like the Jeapardy AI bot were created and new sorts of robots were available. Since around 2023 the AI hype has entered the mass market and the large language model became available.

Roughly spoken, the history of computing can be divided into 2 periods: 1950-1990 which was equal to classical computing and from 1990-today which consists of Artificial Intelligence after a slow start for the first decade.

 

Transition from 4th to 5th computer generation

4th generation is classical computing which consists of software and hardware, but 5th generation is about future AI based computing which isn't available today. So there is a transition available which should be explored in detail.

Classical computing is well understood because it has evolved over decades. The first computers were created in the 1940 with vacuum tubes, while later and mower powerful computers were based on transistors and microprocessors. Even if the advancement was huge the underlying perspective remains the same. The goal was always to improve the hardware and write more efficient software. In the 1980s this has produced advanced workstations which were connected to a world wide internet.

The more interesting and seldom discussed problem is how to make these powerful workstations from the 4th generation intelligent so that the computer can control a robot. The answer can be derived from the history of technology which took place from 1990 until 2020. After a slow start in AI during the 1990s which is called an AI winter a major breakthrough was available during the 2000s which was called Deep learning hype. Deep learning means basically to use existing neural networks which were available in the 1990s but train them on larger datasets with a deeper layer structure. Such training was possible because of the moores law which has provided since 2000 faster GPU and CPU technology.

Deep learning alone doesn't explain how to make computer smart but they are providing an important puzzle piece. The prestep before a deep learning model can be trained is to prepare a dataset. During dataset preparation a problem has to be encoded in a tabular format. This allows to convert a low specified problem into a high specified problem. Possible exammples created in the 2000s were motion capture datasets, OCR recognition datasets and even question answer datasets. The existinence of these datasets was a major milestone in AI development, because it allows to use computers to solve real world problems.

Datasets are important for deep learning because they are used to determine the score of a neural network. A certai trained network is able to reproduce the dataset with a score from 0 to 100%, for example to determine who many numbers the neural network has recognnied with OCR correctly. Such a score allows to compare different architectures and different training algorithms next to each other which is equal to transform the former AI alchemy into a mathematical science discipline.

The logical next step after the deep learning was to natural language processing. The idea was to introduce natural language to annotate datasets and use neural networks for chatbot purposes. This development took place from 2010 until 2020 and resulted into large language models which are available since 2023.

Let us summarize the development in decades:

- 1990-2000 AI winter with no progress at all
- 2000-2010 Deep learning hype with a focus on datasets
- 2010-2020 dataset based natural language processing
- 2020-today Large language models

This short overview explains, that Artificial Intigelligence wasn't created by randomly discovering a new algorithm, but AI was the result of a long term research effort which includes 4th computer generation, deep learning, natural language procsssing and large language models.

If we want to describe the devolopment as a single concept, than it would be the shifting focus from algorithms towards datasets.Classical computing until 1990 was influenced by the algorithm ideology. The goal was always to build new hardware and run a piece of software on this hardware as fast as possible. This attempt was realized with high efficiency language like C in combination with CISC architecture realized in microprocessors. unfurtunately, even the most advanced 4th generation computer's can't provide artificial intelligence. Even super computers built in 1990 were not able to beat human chess player and they were not powerful enough to control biped robots.

To solve the gap something more powerful than only algorithms was needed which is a dataset. A dataset is a problem formulation, its some sort of riddle stored in a table. Datasets allow to benchmark a system. Its some sort of quiz which has to be solved by a computer. Such kind of problem oriented computing allows to create artificial intelligence by selecing certain datasets. For example to control a robot the following datasets might be useful: object detection dataset, visual question answering dataset, instruction following datasets. If a neural network is able to solve all these benchmarks, the neural network is able to control a robot.

One interesting insight of modern AI research is, that the discovery has nothing to do with computing anymore. A modern robot build in 2025 doesn't contain of advanced hardware nor it runs advanced software. But all the components might be developed in the 1980s. That means, a microcontroller from the 1980s combined with the C language from the 1980s are more than powerful to create an advanced robot. What has changed is the problem addressed by the robot. The mentioned instruction following dataset wasn't available in the 1980s. A machine readable problem formulation is the key improvement.