"We are on the threshold of the web of AI agents"

Stefan, we've had AI since 2022 ...

(laughs) It always seems that way, doesn't it? AI has been around for much longer - but it really has ushered in a new era: We can now talk to machines in our natural language. We learnt this with the release of ChatGPT.

We started with AI at 3pc ten years ago. Back then, it was all about semantic technologies: for example, our editorial assistant automatically linked personal names with the corresponding biography or technical terms with a glossary entry. That was a real leap! But of course it's not comparable with our current AI development projects, where we're working on AI agents, which is super exciting ...

Wait a minute, before you explain that to us: Has the AI got any smarter?

How you take it: When you type a question into ChatGPT, you are talking to an LLM, a Large Language Model. This recognises relationships between words and the structure of the language from a huge amount of training texts. This huge network of relationships acts as a kind of knowledge store that provides a response in natural language to an input in natural language.

This is fascinating and seems very human. However, the LLM cannot really understand anything. It cannot reflect or learn independently from mistakes. It is and remains a complex - but in this sense also stupid - machine.

But still, in the beginning ChatGPT "knew" much less than it does now and made more mistakes!

Yes, remember abstruse mistakes like recipes for "fish legs" at the very beginning ... (laughs) It's almost a shame that the AI model was constantly being improved.
When ChatGPT was launched in November 2022, it was like a public experiment: all users trained the models by asking questions. Our data was systematically tapped in order to use our feedback to continuously improve the quality of the answers ...

Aren't there always warnings that the data I enter will end up somewhere else anyway?

Yes, a very important strategic issue! Also for many customers: What compliance am I pursuing? Do I want to open the door and send my users' data to the USA or China when they talk to a bot on my website?

To put it bluntly, this can be prevented. You can now host open source LLMs yourself - provided you have the expertise and correspondingly powerful servers. In addition, German and European cloud providers are now also entering the market, offering AI models as a kind of ready-made solution. This means that GDPR-compliant AI applications can be developed and implemented quickly. We decide which solution is the right one together with our customers on a case-by-case basis.

You keep talking about "models", recently I even came across "multimodal model" - could you please explain what that means?

ChatGPT and the like, for example Gemini, Mistral or DeepSeek, are referred to as AI models. It is important to note that all of these current AI models are GPT models, which stands for Generative Pretrained Transformers. They are therefore pre-trained, can generate new content and transform answers into natural-sounding language.

Multimodal models can not only process written language, but also spoken texts or visual information from photos, videos or the live image on a monitor, for example. Cat, dog, peacock moth? A table? A multimodal model not only recognises the table on the screen, but also provides an overview of its content. The larger new models in particular are very cool!

But they can! Unfortunately, they can't make coffee yet ...

No, they still can't walk around, they lack physical representation. What's more, they can't yet plan or pursue their own goals. So Schwarzenegger's Terminator is still some way in the future.

But when the AI starts doing things independently online?

You have to take a critical look at that: Such questions are about control. Basically, we give up control when, for example, we send an AI agent to book our hotel instead of searching through travel portals and transferring money ourselves. This loss of control is just crying out for online fraud or targeted attacks on larger digital infrastructures - we shouldn't be careless with the new possibilities.

Are there no legal regulations to prevent this?

Yes, there are. It's no coincidence that the implementation of the new EU AI Regulation with different levels of protection is currently being hotly debated. Normally, however, you are not in the high-risk area. It is important to make it clear to users that the friendly chatbot is not a human, but an AI, and what that means. There are already sensible and practicable solutions, somewhat comparable to the banners for privacy settings.

Let's switch from theory to practice: You said at the beginning that 3pc is involved in exciting research projects - what are they about?

We are currently involved in three major projects: One is developing an AI-supported augmented reality application for the operation and maintenance of industrial plants. A second is using mixed reality tools to involve children and young people in municipal planning processes.

Our latest research project is currently in the starting blocks: together with partners such as Fraunhofer FOKUS, the German Research Centre for Artificial Intelligence DFKI and major application partners such as dpa, we are working on using generative AI to support editorial work. Very exciting!

Supporting editorial work - what does that mean?

The aim is not to make editors superfluous, but to simplify time-consuming tasks with AI assistants. Just think of the area of accessibility: a generative AI could take over image descriptions and ALT texts or even translations of articles into plain language.

To prepare articles, the assistant can quickly compile information from various sources, from international databases, articles, photo collections, tables, etc. - including image suggestions and, of course, all source references. Material that you can work with immediately.
Or it prepares data for interactive applications, for example before elections: Statistics, maps and overviews of all constituencies and corresponding texts on candidates can be presented in a new way.

But that sounds like saving jobs ...

No, I see it differently: it's about preparation or additional benefits. The application creates space for exactly what an AI will never replace, what people - and only people - can do: Finding out something new, being on the spot, doing journalistic work with good instinct and sensitivity.

AI can provide information quickly and in a new quality. It can enrich the basis for articles or contributions and create additional added value by transforming complex content and making it accessible to different target groups.

Aren't we adding to the already unmanageable jumble of news?

Not if we use the AI assistant to generate really good content. It then plays a similar role to influencers in social media: they also act as filters, selecting topics for us from the unmanageable mountain of news and presenting them.
With the AI assistant and very good content that conveys a lot of content, brings it to the fore and has charisma, we achieve the opposite of "spamming the world" - we turn our customers into influencers of the future.

Quality content instead of mass-produced goods - I'm excited! Nevertheless, I get thousands of hits with every Google query ...

I think that will soon be a thing of the past: Millions of people are already asking ChatGPT and co. every day to get an answer instead of drowning in Google hits.

And this is just the beginning of something new, it will be the next challenge for all traditional applications, for search engines, but also for every website: There will also be chatbots for your own internet presence as a matter of course. They'll just search my content and give users answers to their questions. Because that's much easier than searching for information in the menu.

Taking it one step further: in future, everyone will have a personal AI assistant on their mobile phone instead of having to visit websites themselves. Open the app, say what you're looking for, send the AI agent off and sit back.

And the classic website is a thing of the past?

No, because the AI has to get the information from somewhere. But user behaviour will change fundamentally. It's no coincidence that we are currently working with customers on AI agents for their own websites so that users can find content easily using voice search - that will be the future.

We are indeed facing a huge upheaval: what is currently emerging is a web of AI agents. That's where the journey is heading - we don't yet know where exactly.
An incredibly rapid development, a real challenge! It's impressive - but we're looking forward to shaping this digital future together with our customers.

How can AI make your day-to-day work easier?

In our AI workshops, we show you what artificial intelligence is and how it can support you in your day-to-day work.
Together, we will look at specific challenges facing your institution or company and find suitable AI solutions that will make your day-to-day work more efficient and easier.

Curious to find out more? Contact us - we will put together an AI workshop that suits your needs exactly.