Rarely has an IT application caused as much excitement as the text robot ChatGPT. While universities, schools, or editorial offices are more concerned about the new tool, companies all over the world are considering how to use the clever bot sensibly for their business. Also in the ERP environment, ChatGPT and other AI systems from OpenAI offer interesting possibilities for use.
ChatGPT has been a hit from the very beginning: within five days of its launch on November 30, 2022, one million users registered. For comparison: Instagram took two and a half months, and Spotify took five months to reach this number of users. Since then, everyone in Germany has been talking about the digital chatbox that independently and quickly writes texts for all situations.
About “Eliza” – Mother of all Chatbots
Language robots are anything but new. In the 1960s, Joseph Weizenbaum developed “Eliza,” an early precursor to modern chatbots. The virtual psychotherapist was able to conduct a simple therapy session by asking the user simple follow-up questions. Sixty years later, language robots and digital assistants are part of everyday virtual life. Alexa, Google Assistant and Siri are at the forefront. And almost everyone had probably the pleasure of interacting with a chatbot that answers questions about products or services while shopping online or booking a hotel.
What makes ChatGPT so special?
ChatGPT is different. It delves much deeper into the meaning of linguistic inputs and can therefore answer complex questions. But what makes it so much better than previous language systems? Primarily, the huge amount of data it is fed: to date, around 300 billion words and 570 GB of data from books, web texts, social media, and other sources have been entered into the system. The training works according to the principle of reinforcement learning, meaning that answers are “rewarded” with positive or negative feedback. At the beginning, the whole process is still monitored by humans, but later the program optimizes itself with the so-called “Proximal Policy Optimization” (PPO). The result is a language model with around 175 billion parameters.
ChatGPT can understand sentences with up to 1,500 words, it can translate and master multiple languages. In German, grammar and punctuation are still lacking, but even these small errors will probably be ironed out soon. Another drawback is the system’s current level of knowledge: so far, ChatGPT has only been trained on data up until the end of 2021. The program answers questions about the conflict in Ukraine or the 2022 World Cup very eloquently, but the answers are either pure speculation or completely fabricated.
ChatGPT as a virtual service employee?
Even though ChatGPT is not yet suitable for use in companies due to linguistic shortcomings and limited knowledge base, the potential is there. Soon, the smart chatbot could help quickly and easily respond to customer inquiries. To do this, it would be connected to other applications in the company, such as a CRM or ERP system, via API (Application Programming Interface). The bot could then, for example, be able to assist a customer in finding a solution, such as in the case of a malfunctioning machine. A task that currently requires a human service employee.
In the future, ChatGPT could also write search engine optimized texts and descriptions for websites, product descriptions for online shops, or individualized manuals. The same goes for social media postings, emails, or blog posts. Even with complex topics, a human may still need to make adjustments, but the basic structure of such texts could soon come from the pen of the smart language robot.
Program code via bot
Another AI from OpenAI, OpenAI Codex, is even more exciting for software and ERP manufacturers. Based on the GPT-3.5 algorithm, which also trains ChatGPT, the system can translate natural language into programming code. Users could thus generate source code for simple apps themselves via command and thereby relieve the development department. This would free up capacity for them to concentrate on more complex issues.
The automatic generation of program code via voice command represents a further development of low-code programming. Low-code or no-code approaches already allow users to put together various building blocks and logics via an intuitive graphical user interface and thus create simple apps without programming knowledge. Such platforms are also increasingly used in adapting and extending enterprise software, particularly ERP systems.
If you take the story further, OpenAI Codex or similar AI models could greatly simplify and accelerate the customizing of ERP systems in the not-too-distant future. For example, by allowing the user to independently insert check steps when recording customer orders or add additional fields to the customer master data via voice command.
The crux of this quite interesting scenario is that in order to generate code, an AI such as OpenAI Codex needs access to many code examples. There are already ways to refine the model with custom frameworks and data structures, but in general, low-code platforms and ERP systems are proprietary software. Both the data model of the ERP solution and the (partially proprietary) frameworks and libraries used would have to be made known to the system or retrained. Such application scenarios are therefore still in their infancy.
AI with a future
In summary, it can be said that ChatGPT, due to its outstanding and exceptionally comprehensive results, definitely has the potential to significantly improve intelligent chatbots and provide basic content structures for complex elaborations. The development of program code through language could even mean a small revolution for software and ERP manufacturers. This is certainly still a thing of the future, but there is no doubt that artificial intelligences like ChatGPT and OpenAI Codex will increasingly make their way into the software world, and therefore also into the ERP world.
ChatGPT Overview
ChatGPT (Generative Pre-trained Transformer) is a dialogue-based chatbot that uses artificial intelligence to “understand” human language and generate responses that are practically indistinguishable from human responses. For text processing, ChatGPT uses the AI algorithm GPT-3.5, a natural language generation model that has also been optimized by human trainers using the learning methods of reinforcement learning and supervised learning. In the course of training, the chatbot was fed with countless data from books, websites, blog entries, studies or social media posts to learn how people use language. In order to improve text creation and the underlying model, ChatGPT’s AI is continuously trained through human feedback.
The Makers
The maker of ChatGPT is OpenAI. The company was founded in 2015 as a non-profit and independent research institute – one of the founders was then Tesla CEO Elon Musk. The organization’s goal was to research artificial intelligence on an open-source basis, develop new models and insights, and share them with other institutions and the public. But the non-profit image of OpenAI did not last long: in 2019, the OpenAI organization was restructured and now belongs to a parent company that follows a so-called “capped profit” model and is therefore also allowed to make profits. Microsoft has now invested around a billion dollars in OpenAI and plans to integrate ChatGPT into its search engine Bing.
In addition to ChatGPT, OpenAI has also developed some other artificial intelligences. One of them is the drawing program DALL-E 2. The name is a combination of WALL-E and the Spanish artist Salvador Dalí. The program combines image recognition and computational linguistics to create photorealistic images based on text inputs.
