No matter which news portal you look at: ChatGPT and the artificial intelligence behind it are THE topic of the hour right now. Other AI tools, such as DALL-E, Whisper or Midjourney, are also attracting a lot of attention. So it’s high time to think about the extent to which this technology has what it takes to also support online print shops in their daily business.
Warning, spoiler: Just because it looks fancy and simple, artificial intelligence – particularly ChatGPT – is by no means that easy to implement without effort. Indeed, after the initial excitement about the new possibilities, sooner or later one also has to deal with issues such as credibility, data protection or labeling.
OpenAI: “Supergroup” of the tech industry
But let’s start from the beginning, because it is first necessary to understand what ChatGPT actually is and who is behind it. The chatbot was developed by OpenAI, a U.S. startup that was founded in December 2015 by a “supergroup” of the tech industry: Elon Musk was among the founders, as were Sam Altman, Greg Brockman and LinkedIn co-founder Reid Hoffmann. Also on board were well-known researchers such as Ilya Sutskever, Wojciech Zaremba and John Schulman. Together with a number of backers – including Amazon Web Services and major German investor Peter Thiel – the startup set out to become one of the leading research institutes in the field of artificial intelligence and to serve the general public with its research results without being influenced by financial matters. At least, that’s what the startup’s first official announcement says.
The research institute was originally a non-profit institution, but later formed a limited partnership with capped profits for investors – so it was no longer 100% non-profit. In the meantime, a non-profit foundation has been superimposed on the whole thing in order to adhere to the originally formulated values and goals and still be able to generate profits to a certain extent through the limited partnership business form. If you want to know more about it: The IT specialist Noel Lang has prepared a very good and detailed description of the development and milestones of OpenAI in a Youtube video.
Incidentally, Elon Musk left OpenAI in 2018 to devote more time to Tesla and SpaceX. However, this did not harm the start-up at all. The first products and models followed soon after: OpenAI Gym (2016), GPT (2018), GPT-2 (2019), GPT-3 (2020) and the image AI DALL-E (2021) had long since been launched when ChatGPT was made publicly available at the end of November last year and triggered a huge hype.
With Microsoft’s recent announcement of another “multi-billion-dollar” investment in OpenAI – some news portals such as Der Spiegel report a figure in the range of $10 billion – the further development of AI is likely to pick up speed once again. Not least because Microsoft plans to integrate OpenAI’s models into its end-user and enterprise customer products in the future. And the infrastructure of the Azure cloud service is also to be further expanded. This is understandable, after all Azure is already the exclusive cloud service on which OpenAI develops.
In addition, just a few days ago, the first information became known about what prices will presumably be called for the use of the AI in the future. Because ChatGPT will not remain permanently and free for all users as it was in the first weeks – the operation is then obviously simply much too cost-intensive.
What exactly is behind ChatGPT and how does it work?
ChatGPT stands for “Conversational Generative Pre-training Transformer” and is a natural language processing model developed by OpenAI, as the artificial intelligence explained to us in a self-experiment in exactly this formulation.
It is based on the GPT technology that we explained two years ago here on beyond-print.de (currently in version 3.5) and is, warning, to quote ChatGPT: “specialized in processing texts. ChatGPT can run applications like text generation, translation, automatic summarization, and many more. It can also be used in chatbots and virtual assistants to enable natural and engaging interactions with users.”
To put it more simply: ChatGPT is able to capture the content of textually transmitted questions, comments and commands and to give the (hopefully) correct answers based on a huge amount of data – in a naturally formulated language. Even enumerations, stories, poems (although the rhyming thing doesn’t work so well in German yet) or even programming codes can be output or explained by ChatGPT’s Artificial Intelligence, and that even without recognizing the formulations as computer-generated at first glance.
And yes, it works. Well even – at least if you made it to the input mask due to the massive overload of the servers. The progress compared to our first tests of GPT-3 more than two years ago is enormous – probably the hype is just because the results are really impressive. Even in German, although the development has its origin in English, of course.
ChatGPT is a fan of the written word
The AI is trained using previously provided data, primarily in text form. Or as it answers itself, “The information that ChatGPT draws on comes from a variety of sources, including online texts, books, articles, and other documents. The model is trained in advance with this data to [enable]it to understand and generate natural language.” And further, “ChatGPT mainly processes text files to build its database. These files may be in various formats, such as .txt, .csv, .json, or .xml. Some of these files may also contain structured data such as tables or lists. It is important to point out that ChatGPT is a natural language processing model, so the data being trained should be mostly text to perform well.” So ChatGPT likes to read.
Furthermore, a role in the construction of the database is also played by the so-called reinforcement learning by human feedback, as heise recently explained in an article the learning capabilities of ChatGPT vividly.
So can the chatbot also be useful for online printers?
First of all: It’s actually amazing what the AI spits out – and it’s fun to play around with it. Even the question about the most important trends in online printing is not left unanswered. Here, ChatGPT lists one thing in particular, in addition to web-to-print applications and sustainability: personalized printing – quite clearly, the chatbot has read the Beyond Print articles carefully. But all kidding aside. ChatGPT can be a powerful tool, and we don’t just mean as a question-and-answer game to pass the time.
Because ChatGPT, at least that’s how you can think of it, can also be used in a professional environment. For example, as support in on-site consulting in online stores, where chatbots are already in use today. But also in customer service, in marketing, when it comes to creating posting texts, or for building product pages and here especially for formulating descriptive texts. It is precisely in these areas that manually time-consuming tasks can be optimized through the use of artificial intelligence.
With the progress that the language model has made since its last version, applications such as these are no longer outlandish and are very much imaginable. ChatGPT can be integrated via API – but anyone who believes that things will continue just as trivially after that is mistaken.
Because ChatGPT, like every new “employee”, has to be trained, i.e. fed with the information relevant for the purpose and the resulting possible questions. And not only that, the AI must also be constantly monitored – and preferably by humans.
Trust is good, control is better
We wanted to know how ChatGPT ensures that the database and consequently the resulting answers are correct, and simply asked the AI itself. The result was interesting: “To ensure that the answers generated by ChatGPT are correct, the model is regularly monitored and checked by humans. There are also various measures that can be taken by OpenAI and other developers to ensure that the model is not misused to spread misinformation.
There are also techniques such as fact-checking and verification that can be used to ensure that the responses generated by ChatGPT are accurate. This includes using multiple sources to verify the accuracy of the information and using tools to check the credibility of websites and sources.
It is important to emphasize that the responsibility for verifying and validating the information generated by ChatGPT rests with the end user, as AI models, such as ChatGPT, can only be a support and aid in decision making and should not be considered a definitive source of truth.” The last point is especially interesting: because the responsibility and liability for the accuracy of the statements will probably ultimately be borne by the person who uses the AI for his purposes.
“Side effects” to deal with.
In addition: the better AI-supported text generation becomes, the more other topics and issues come into focus. For example, the question of data protection (where is the information stored?), the accuracy or manipulability of the database, or the obligation to label texts generated in this way.
It’s no wonder that some professions are already worried that in a few years they will be replaced by technologies such as ChatGPT. Or that, for example, the last hour has struck for homework or papers, since no one will be able to say for sure whether the content has not simply been compiled by ChatGPT – which should also cause journalists to worry about their future. Incidentally, there already seem to be the first apps here that can check whether a text was written by an AI or a human, as the example of 22-year-old US American Edward Tian shows.
But even global market leaders could be attacked by AI models. There are at least already voices that see a shake-up of Google’s market power coming.