Back to the Future: The Story of Generative AI (Episode 1)
Welcome to our four-part blog series on the history of Generative Artificial Intelligence. Our journey through history will highlight the significant milestones and show how the entire concept of generative AI has fundamentally transformed with each step of development. From the early attempts of sketching probability distributions with pen and paper to today’s sophisticated algorithms that generate complex and creative content – each of the four steps marks a revolution, not just an update.
Why is the history of Generative AI so exciting? Because it demonstrates how each technological advancement has not only changed the methods but also the assumptions, usage, audience, and interaction with the models. What began as a tool for statistical analysis is today a creative partner capable of producing art, music, text, and much more.
Epoch 1 – Foundations
A well-kept secret: If you rearrange the letters of “Data Science”, you get “Statistics”. Just kidding. But it is true that the roots of data science go back to the 18th century. Back then, α, Θ, and other mathematical symbols had more of a mothball charm than venture capital appeal.
Mathematicians like Gauss, Bayes, and a number of clever Frenchmen recognized the value of counting early on. They counted, recounted, and compared the results – all by hand and very laboriously. Yet these methods are still relevant and proven today – a true evergreen!
With the invention and availability of electricity, a new era began. Data could now be processed and analyzed much more efficiently. The idea of an “electronic marble run” for data emerged – a system with switches and paths that triggered various actions based on data input, such as lighting a bulb or executing a function.
An early, actually functional form of Artificial Intelligence (AI) was born: algorithms based on observations and derived rules.
What Makes These Early Models Generative? Well, the “electronic marble run” could also be operated in reverse. Forward, it was a statistical model that assigned a category or value to an observation. For this, the model needed to have an idea of the data. Backward, however, high-probability examples of mushrooms, marbles, data – in other words, images or tables – could be generated through random draws. The generative capabilities of the models were often underestimated, as the forward function was the focus.
This methodology is called the Naïve Bayesian Classifier. “Naïve” here is not meant derogatorily but refers to simplifying assumptions that make modeling significantly easier. With naïve methods, one does not have to assume complex relationships between variables like mycelium, stem, and cap of a mushroom. One simply says: If the average quality of all three parts is good enough, then the mushroom is good.
Some of the first applications of these models included handwriting recognition (for example, at the post office, known as Optical Character Recognition or OCR), as well as spam filters and general text analyses up to this day.
This was the first insight into the foundations of generative Artificial Intelligence. In the next part of our series, we will dive into the world of neural networks and machine learning, which laid the foundation for modern AI systems. Stay curious and don’t miss the next milestone in the history of generative AI!
Don’t miss Part 2 of our blog series.
