Lwill ChatGPT replace Google soon? Or book authors and journalists? Maybe even a programmer? Iryna Gurevych, Head of the Chair for Computer Science at the TU Darmstadt and an expert for language models, sees ChatGPT on the right track in many areas – but there is still a long way to go before search engines or entire professions are replaced.
Although ChatGPT’s interface looks like a search engine, the artificial agent can do much more: It provides facts and definitions, can write computer programs and even hold conversations. It can even be used as a writing assistant – you can ask for a text for the opening speech of an event, for example, “and the model suggests what you typically say in such situations. And in many cases that looks amazingly real and human,” says Gurevych in the FAZ podcast “Artificial Intelligence”.
The model was specially optimized for dialogues. For this purpose, ChatGPT was trained with training data and reinforcement learning with human feedback. That’s why the results look like a human dialogue. In addition, it also learns the “better answers” for a given context, reports Gurevych. Nevertheless, many issues remain unanswered in the current version of the language model. The code of the software developed by the US company OpenAI is not visible, so that one can only speculate about the exact functionality. One thing is clear: ChatGPT can only reproduce information that was contained in its training data set.
“I asked ChatGPT which corpus it was trained on,” reports Gurevych. The model replied that a large number of publicly available documents such as websites, books and other texts were included, but they were only dated to 2021. The model is therefore less able to deal with temporal aspects and does not provide any information about current developments. Because ChatGPT says it has no access to search engines – in contrast to Facebook’s Blenderbot, for example. It will be exciting when different technologies are integrated in this area. “ChatGPT and Google with Lamda are complementary today, but may eventually merge in the future.” Lamda is Google’s AI chatbot.
Even longer texts, for example for books, in which complex relationships are explained, are currently still very difficult. However, the next language version GPT4, which is due to appear in spring 2023, should contain around 500 times more parameters – and thus also recognize more complex relationships.
ChatGPT describes itself: ChatGPT is a neural network model developed by OpenAI and used for text generation. It is specifically designed for use in chatbots and other applications where it is intended to have human-like conversations. The model was trained on the basis of GPT-3, one of the most powerful automatic text generation systems. It has the ability to generate human-like responses to questions and queries by accessing a variety of data sources and using that information to formulate the responses. One of the interesting things about ChatGPT is that it is able to adapt to the context of the conversation and reply accordingly. This means that it is able to respond to questions and requests in a way that is relevant and understandable to the user.
The episode is part of our podcast “Artificial Intelligence”. He explores the questions of what AI can do, where it is used, what it has already changed and what contribution it can make in the future. With Peter Buxmann and Holger Schmidt, the FAZ brought two proven AI experts on board for the podcast.