How does the new super chatbot work, why it’s so much better than other voice assistants – and what its future might look like.?

Algorithms and chatbots are taking on more and more tasks and intervening in people’s behavior: For example, special artificial intelligence (AI) supports call center employees in their work by specifying how to formulate conversations. A novel chatbot for restaurants automatically recognizes phone orders and even allows fluent conversations between the caller and the algorithm – in the narrowly defined area of ​​​​food orders.

The interface between man and machine is therefore given special attention. This was also shown last week when the American start-up OpenAI made its new chatbot ChatGPT freely accessible on its website – and triggered worldwide hype.

Within five days, OpenAI boss Sam Altman reported more than one million registered users of ChatGPT. To put this in context: LinkedIn launched the German edition of its business network in early 2009 – and passed the one million mark after almost a year in December 2009.

Description of technical products to poems

The AI ​​behind ChatGPT – the acronym stands for Generative Pre-Trained Transformer – indeed provides very impressive answers. And on a wide range of questions, from descriptions of technical products to food recipes and the formulation of poems.

What’s more, all of this works not only in English but also in German. When asked “What’s the difference between a cheetah and a leopard?” the bot hesitates for just under a second – and then spits out a letter for letter: “The cheetah and the leopard are both big cats that belong to the big cat family.” are followed by three more sentences on the differences between the two cats, which could have come from an animal encyclopedia almost ready for printing.

OpenAI’s chatbot is based on GPT -3, the third version of a language model based on deep learning. The algorithms of the language model are trained using gigantic amounts of data. And so the reason for the success of ChatGPT for Hinrich Schütze, Professor of Computational Linguistics and Co-Director of the Center for Information and Language Processing at the Ludwig Maximilian University of Munich, is not a scientific breakthrough. “Rather, it shows how important the mass of data and computing power that OpenAI has at its disposal,” explains Schütze.

OpenAI was co-founded in 2015 by Tesla boss Elon Musk, among others, who has since retired from the board of directors. Another prominent investor is the US software company Microsoft.

Chief developer as a support employee

According to Schütze, the language models on which the chatbot is based are trained in a relatively primitive manner: “They are actually only supposed to predict the next word,” says the AI ​​and linguistics expert. “If you want to master this well, it requires factual knowledge and understanding of language.” In the case of ChatGPT, this works so well because the algorithm has been fed and trained with more and more data in recent years. OpenAI does not publish exact figures and details; However, the bot is said to contain all websites, the online encyclopedia Wikipedia and the entire Github developer network. Through all the training, the AI ​​behind the chatbot knows that “Paris” is the most plausible continuation of the sentence “The capital of France is…”, explains Schütze.

The combination of a powerful language model with gigantic amounts of data ensures that ChatGPT works so well across the board. “It’s as if the VW support employee – who answers customers’ questions about spare parts – was also the VW chief developer,” says Michael Witzenleiter, head of the Conversion Maker start-up from Offenburg in Baden-Württemberg. “The AI ​​has, so to speak, stored the knowledge of the world and can also provide answers outside of the actual question – that is what makes it so fascinating.”

Witzenleiter knows what he’s talking about: his start-up uses an interface for one of the OpenAI language models for its own service: Conversion Maker uses AI to generate its own texts for the websites of e-commerce providers. “We trained the AI ​​with our own data for this purpose,” says Witzenleiter.

But ChatGPT has another special feature: the AI ​​is already very good at recognizing the context of a question from the word order. In fact, when asked the follow-up question “Do they look different?”, ChatGPT knows that cheetah and leopard were asked before – and answers correctly in terms of grammar and content: “Yes, cheetahs and leopards look different.”

Recognizing the context as a unique selling proposition 

This understanding of the question context is a unique selling point of ChatGPT, emphasizes Sebastian Schubotz, chief developer of the Leipzig AI start-up, which published the gastro chatbot Foodcall in March of this year. “This is really new – most bots can’t do anything with statements like ‘this’ or ‘that’ alone,” says Schubotz, who, however, uses a different language model for Foodcall: “GPT-3.5 is very broad, so ChatGPT also works so good,” says the AI ​​expert. “But there are currently no individual adjustments – such as specializing in gastronomy as we do.”

This is one of the reasons why he sees the future use of ChatGPT primarily as a kind of extended search: “It could be something like the future of search,” says Schubotz. “So something that works smarter than Google does today .”


Please enter your comment!
Please enter your name here