It is difficult to escape the huge hype surrounding Generative AI and ChatGPT. The public debate alternates between utopia ("AI will save the world and everyone will be rich!") and dystopia ("We will all die, or at least be unemployed"). The current media cacophony is too great and too many different interests are being pursued. There is a lot of belief and assertion and very little in the way of facts and evidence. Almost everyone has an opinion, an idea and a lot of questions about AI.
Generative AI and neural learning methods are very old.The first work on the subject was published i 1943 - even before the Von Neumann computer. In the 1990s, these methods were already flourishing: today, neural networks are installed as control software in almost every electronic device, e.g. in heating systems or car washes.
In 2016, the ALPHA-GO system once again made headlines around the world, but the impact was rather modest. ChatGPT then appeared on the scene in November 2022.
GPT stands foGenerative PreTrained Transformer. These are systems that not only classify data, but also generate "data", are (pre-)trained with a huge amount of different data and use the Transformer architecture. The training of these systems is very complexcosts several million dollars per training run, consumes a lot of electricity and therefore also has a gigantic carbon footprint.
The subsequent use of these pre-trained systems is very cheap in comparison and consumes very few resources. Emily Bender described this AI approachin in 2021 as "Stochastic Parrots", as in the end, based on the data used, only the probability is learned with which a word follows a next word depending on a certain context, the prompt.
This is how long texts are created step by step. We are all familiar with this from autocomplete, e.g. from old Nokia cell phones or WhatsApp.
The approaches are comparable, only with Generative AI they are thousands of times larger. It makes no difference whether the character string is Hebrew, French, German or JAVA. Systems such as ChatGPT are language models, operate on the basis of data and probabilities and have no real understanding of the world.
A little reminder of the statistics lecture: "Correlation does not necessarily mean causality." The fact that these systems have no understanding of the world is shown, for example, by the fact that people often have six or three fingers in AI-generated images. A mistake that only small children make when drawing.
Opportunities, limits and human perception of generative AI
However, anyone who has ever entered a prompt into such a system can hardly escape the fascination of this technology. However, it is difficult to convince interested parties, customers and partners of the limitations of this technology. We humans have a cognitive bias here - we look for faces, patterns and intelligence everywhere. This is innate to us. That's why we often fall for people with maximum self-confidence and minimum competence. It's similar with generative AI: great presentation and impressive results, but sometimes they can be completely wrong.
In manufacturing, such as mechanical engineering, the approaches of generative AI are currently still failing. For example, who wants to sit in an airplane that they know has been creatively maintained using Generative AI?
At Empolis, we combine the capabilities of language models in our AI-based Empolis Intelligent Views Platform®. With our knowledge model, we can guarantee the correctness of the answers in the technical domains and thus rule out a creative but incorrect answer with certainty. Errors are usually very fatal in production and maintenance and can have huge consequences. In addition, we will integrate the capabilities of Large Language Models (LLM) into the products Empolis Service Express®, Empolis Knowledge Express® and Empolis Content Express®.
Three use cases
We recently held a two-day hackathon in Würzburg, where Empolis experts demonstrated the feasibility of the ideas and realized the first prototype implementations.
This resulted in three use cases:
-
The "Empolis Knowledge Buddy", a chat-based assistant that provides precise and factually correct answers to questions based on company knowledge, asks queries and suggests solutions with verifiable sources.
-
The second use case, "Click, click, new sales contract", speeds up the quotation process by automatically identifying relevant sections from previous quotations.
-
The third use case, "There it is, the knowledge graph", deals with simplifying the creation and maintenance of knowledge graphs. Here, LLMs were used to extract facts from existing documents and chats with experts and provide tools to control, monitor and post-process these facts.
We have many new, self-proclaimed AI experts everywhere who are currently operating on the market with great promises. As one of the AI pioneers in Germany and the first spin-off of the German Research Center for Artificial Intelligence (DFKI), we have decades of expertise in the practical application of diverse AI technologies, confirmed by successful customers and partners. Our primary goal is to offer practical and reliable products. We invite all customers and interested parties to participate in the development of these solutions or to realize their own AI projects with Empolis.
We are working very intensively on the topic of Generative AI/ChatGPT and at the Empolis Exchange Summit 2023 on September 28 (online event) we will present the first extensions of our products Empolis Service Express®, Empolis Knowledge Express® and the Empolis Intelligent Views Platform®, which we are currently testing with selected customers and partners. You can register for the online event free of charge HERE.
We will be presenting generative AI enhancements for Empolis Content Express® at this year's tekom.
