Writing Greeting Cards, Writing Code and Changing the World

No application has grown faster than ChatGPT. None attracts more attention. Is this hype really justified?

Nils Schwenzfeier: Yes, it’s actually the first time that an AI topic has been so prominent and almost everyone has heard about it. What sets Generative AI and large-scale language models apart from other technologies is the direct interaction with users. It is not an abstract concept like blockchain or something that remains invisible like cloud technologies. The masses have direct contact with the system in the truest sense of the word. It’s amazing: people who previously had little or nothing to do with computer science and artificial intelligence are now becoming Prompt Engineers. The barriers to entry are low, the results astounding and the imagination great. This combination is creating unbroken euphoria.

Nils Schwenzfeier, Managing Director of TamedAI: “Alongside the established technology giants and start-ups with billions of dollars at their disposal, an exciting scene of specialized GenAI offerings is developing.

Greeting cards at home with ChatGPT is one thing. But it has little relevance for use in companies.

Volker Gruhn: There is hardly a conversation with company managers at the moment that doesn’t touch on the question: “What can we do with this? The curiosity is enormous. The ease of getting started also plays a major role here: you don’t need a complex software project, large investments or anything similar to gain initial experience. If compliance plays along, a browser and a few euros a month are enough. However, the initial experimentation is increasingly giving way to planned use: creating a database, finding use cases, integrating generative AI solutions into existing IT processes, training employees: These are the tasks that companies are now facing.

And then there are hardly any limits to the imagination: The administration department of an insurance company benefits from the potential of the technology just as much as software development or the marketing team. Use cases are a dime a dozen. The trick is to find the ones with the greatest leverage. This will be one of the challenges for companies.

Public perception is dominated by OpenAI with ChatGPT. But what does the field currently look like beyond the top dog?

Volker Gruhn: On both sides of OpenAI and Microsoft, the market is on the move. The pace of development is so rapid that today’s snapshot can be outdated tomorrow. On the one hand, there are large companies such as Google and Amazon, which are currently investing massively in LLM technologies. Or Meta, which is taking an open source approach with its Llama model. In Europe, hopes are pinned on companies such as Mistral from France or Aleph Alpha. Both are working on their own LLM with different approaches. Both have the advantage that they were practically born with a European understanding of legal standards and data protection. In addition, there are highly specialized providers that specialize in applications in core business processes, such as our university spin-off TamedAI.

Are there differences between the GenAI offerings and are they therefore suitable for different purposes?

Nils Schwenzfeier: An exciting scene of specialized GenAI offerings is developing alongside the established technology giants and start-ups with billions of dollars at their disposal. Their aim does not lie in their solution being able to write both the greetings card for an 80th birthday and the code for the new CRM. Instead, models are tailor-made for specialized use cases. This ensures higher quality results and at the same time more cost-efficient development and simpler operation – even on-premise.

Within TamedAI, for example, we focus on the topic of information extraction from texts and documents. This ranges from processing the scan of an invoice sent by post to the automatic processing of e-mails. This focus brings particular advantages in terms of precise recognition and the fight against model hallucinations. Other offerings concentrate on similarly specific use cases. A prime example is the company DeepL, which has set a new standard in the field of text translation and transcription. Overall, the LLM world is becoming much more diverse.

We seem to be only at the beginning of this development. What’s next for the LLM market?

Volker Gruhn: The market for large models with general use cases will consolidate. You can see this very clearly with the release of GPT-4. The last version was so powerful, the quality of the results so good, that suddenly many other models have fallen behind significantly. They can no longer keep up because, for example, they do not have the database that OpenAI has built up at great expense and continues to maintain. In my opinion, we will see a handful of large providers offering comprehensive solutions on the scale of ChatGPT. Many of the special solutions described above will establish themselves following in their wake, particularly in the B2B sector.

OpenAI is currently the measure of all things. But so were AOL and Netscape back in their day. How big is the advantage for a first mover?

Volker Gruhn: OpenAI undoubtedly has a head start. The first version, which was made available to the public in November 2022, has already won over users. This was shown, for example, by a survey we conducted among managers in Germany at the end of 2023. Here, someone has brought a convincing solution to the market from a standing start.

And the management has so far shown great skill in the further development of the product. The increase in quality from version 3.5 to 4 is noticeable. The GPT store still looks a little anarchic, but it shows where development is heading. Open API can write a similar success story with this platform as Apple or Google with their respective app stores. Added to this is the close cooperation with Microsoft and the integration into the Office world. All of this works in the company’s favor. On the other hand, the short-lived scandal surrounding the dismissal of Sam Altman has shown how quickly such a development can tip over. Whether we will still be thinking of ChatGPT in five years’ time when we talk about GenAI is not yet certain.

Volker Gruhn, founder and Chairman of the Supervisory Board of adesso SE: “The initial experimentation is increasingly giving way to a planned approach.

The data basis of an LLM is decisive for the quality of the answers. Is it simply a case of “a lot helps a lot”, i.e. is it mainly the quantity of data and parameters that counts.

Volker Gruhn: The turning point came with the transition from GPT-2 to GPT-3 – a leap to a gigantic 180 billion parameters. GPT-3 surpassed everything that had gone before, but the first impression was deceptive: bigger is not necessarily better. This realization was initially just a hypothesis based on the sheer size of the models. There was speculation: What if we raise the bar even higher, to 500 billion parameters? However, it soon became clear that it is not size alone that defines progress, but the interplay of refined training material, training procedures and optimized model structures. We now know that even models with a single-digit billion number of parameters can be as powerful as GPT-3. So it’s not just about the quantity of parameters, but also about the quality of the training, the underlying data and the architectural sophistication.

How well do you think Germany is positioned when it comes to AI? Apparently, Germany is once again lagging behind

Volker Gruhn: I think the assumption that Germany is completely left behind when it comes to AI and digitalization is exaggerated. When it comes to major language models, the USA may, with a few exceptions, lead the way. But our established industries, from insurance to banking to mechanical engineering, are by no means lagging behind in the application of AI. Countless projects are underway to improve products and processes with intelligent technologies. If legislators regulate with a sense of proportion – and regulation is necessary – our location will be among the winners of this development.