Since the end of last year, ChatGPT, Midjourney and Dall-E have not only filled what felt like every media channel but have also turned learned truths about what is created by people and what is created by technology on their heads. It was only logical that artificial intelligence should also be part of the Online Print Symposium 2023. For the keynote, the organizers brought journalist and author, Jörg Schieb, a real digital expert, on stage.
175 billion parameters: The ChatGPT3 language model was trained with this amount of data on text from blogs, articles, news, tweets and publicly available books. As digital expert, Jörg Schieb described at OPS, that’s 754 GB of text – or the equivalent of 376 million pages of paper that, stacked on top of each other, would be 3 km high and weigh 1,282 tons. “All these texts are looked at and broken down into so-called tokens – which are reconnected and linked, just like in the brain,” the expert explained. “So a neural system is actually created. These words and tokens, or even parts of words, are weighted, they are linked together, and they form the parameters.”
In ChatGPT4, the latest version of what is arguably the best-known AI application, that amount of data has increased yet again, to a hundred trillion that we can barely imagine. “We have a thousandfold increase in knowledge here. It’s quite adventurous how fast this is going right now,” Schieb said.
In his keynote address at the Online Print Symposium 2023, Schieb summarized the key facts surrounding ChatGPT – and other artificial intelligences as well – and helped put the developments in perspective. “We are literally living a Gutenberg moment right now, at the very least. Gutenberg didn’t invent the printing press, but he at least made it marketable, and he changed the way information is disseminated permanently. With that, I’d like to at least compare what ChatGPT – and comparable chatbots – will bring in terms of change. We really shouldn’t underestimate and minimize this,” he noted. “It’s not a fad, either. It’s something that’s here to stay, I’m convinced of that.”
More than text generation
Schieb also showed what all is already possible with ChatGPT using various examples, of course not without first also asking the audience who had already tried it out. In addition to text generation itself, whether short or long, or help with lecture preparation or programming codes, Schieb said one thing about artificial intelligence that is particularly exciting – ChatGPT is open. “There is a programming interface that anyone can use who might pay for it, and ‘put’ all sorts of things on top of the technology. Synthesizing the responses received and converting them to voice output and having them played back by an equally artificially generated avatar is no problem”, he said, as he proved directly with an example of his own at the beginning of his talk. “ChatGPT itself doesn’t have a speech output yet, but that’s only a matter of time,” he is sure. And the language model also handles the translation of texts in a fraction of a second, because it now masters more than 190 languages.
ChatGPT understands contexts
And, as Jörg Schieb further explained, it can do even more, at least if you take a closer look at ChatGPT4, because not only is the new version based on a significantly larger amount of data – it is also capable of processing images and understanding their contents. It is thus, unlike the previous version ChatGPT3, a multi-modal model. ” It’s not about the question, ‘what do you see in the picture’ – artificial intelligences have been able to describe that for a long time,” Schieb said. Rather, it’s about the interrelationships of the depicted image elements. “That’s the kicker because the system actually understands what’s in the image and relates that to everything it’s learned up to that point.”
AI doesn’t think, AI calculates probabilities
As impressive as the topic of artificial intelligence is, however, there is one thing that should not be forgotten, the digital expert explained, “Just because there is the word intelligence in AI, people think ChatGPT can think. But the system doesn’t think. It calculates probabilities. That’s important to keep in mind. ChatGPT analyzes. So when you ask a question, ChatGPT doesn’t know the answer, it calculates the probability of what they meant by the question and what the likely correct answers to it are. That’s getting better, too, because, yes, the systems are learning in addition.”
But that, in turn, naturally raises questions that will need to be addressed in the future. For example, who determines what the system learns and how exactly this knowledge is controlled, or by whom. Because: “The system can only know what has been trained. If you only let the system be trained in fairy tales, it won’t be able to write a nonfiction book, and vice versa,” says Schieb. Another question, he added, is who will determine in the future what a system is actually allowed to spit out as an answer and what not. “There are still a lot of challenges ahead of us – at all levels.”
Media competence is in demand, but fear is not
At the OPS, Jörg Schieb was able to dispel much of the widespread concern that artificial intelligence will take over the work of humans in the future. Because ChatGPT and Co. could do one thing above all: support us wonderfully. Of course, there are jobs that AI will completely replace, which is why everyone is called upon to make intelligent use of the new possibilities. In the printing industry, for example, intelligence could be trained to deal specifically with technical questions, thus expanding customer support and freeing up capacity on the “human” side of the workforce for real, creative tasks.
Accordingly, despite all the unanswered questions, Schieb’s summary was rather optimistic: “If you view chatbots and AI as a tool and not as a threat, then it can be useful. It can remind me of things, it can sort my thoughts, it can be a sparring partner. What it can’t be is creative, humorous, imaginative. It can’t create something really new. It can put together images that some might call creative – I might call them more impressive or aesthetic, but not necessarily creative. Because, after all, they are the result of a collection of what the systems have previously learned. This is true for texts as well as for images. But it cannot create completely new thoughts. It is always based on what we humans have created before. This means that if we humans are relieved of AI, which takes over more and more tedious routine tasks, and we can thus have more creativity, then it is a good and useful technology. And we can certainly implement that in companies as well.”
If you want to watch the OPS keynote by digital expert and technical author Jörg Schieb in full length and with all the details – keyword Turing test and Google’s concern – you can watch the video recording here: