An article about the future of ERP systems, written without the help of AI

At the end of June, I had the pleasure of speaking at the Swiss Software Festival in Basel on the topic of “The Future of Software Development.” To be honest, the pleasure was limited, as I had to fly back from Mexico, where my family had been invited to a wedding—weddings are no small affair in Mexico. At least I had time to work on my short presentation during the flight back. But what exactly was the title of the keynote speech I had promised to give?

I asked Perplexity. The AI search engine kindly gave me the title I had suggested to the organizers (and then forgotten): “Creating a new breed of software for a new generation of users.” The rest was easy. I asked Perplexity to write the text for a speech in 10 slides and gave the software a few keywords for each slide. Perplexity spat out useful suggestions in a matter of seconds. I then used the AI tool Gamma, which drew the slides for me, complete with beautiful illustrations and the obligatory Matterhorn in the background. In five minutes, I had my keynote speech: professional, structured, interesting – but not written by me.

AI has stolen language from humanity. According to influential linguist Noam Chomsky, language is one of the fundamental abilities of humanity. Only humans can use the same words to express completely different things with the help of sentences, and only humans can draw conclusions with the help of recursion. That was once the case – machines can now do the same with the help of generative AI. They can write emails and even keynotes for the Swiss Software Festival, or provide answers and engage in dialogue as chatbots. Or at least pretend they can. Because they are actually stochastic parrots, a term coined by computer linguist Emily Bender. With the help of gigantic amounts of data and statistics, they parrot what humans say. As a result, AI produces a lot of blah blah blah and propaganda with little substance. But there is an AI-based solution for that too. The Ticino-based AI specialist Artificialy, with whom we collaborate, has built a “bullshit distiller” for us. It can analyze texts, filter out manipulative content, and distill the real message of a text and its essential content.

AI produces language non-stop – most of it is never perceived by human beings. It has lost its unique purpose of enabling interpersonal interaction.

With artificial intelligence, language is gaining a new purpose: it serves as a means of interaction between machines, specifically computers, and more specifically ERP systems, and humans. Business software is far too complicated. The Abacus ERP system alone consists of over 2,300 programs. It takes far too much knowledge and effort to find your way around it. Instead of finding the right program and using a mouse and keyboard (in the best case) to ask questions and trigger actions, users use human language. They ask the digital agent questions (Which department does Denise Muster work in? How many vacation days does she have left this year?) and trigger actions. Human language will replace most other interfaces in business software.

To build these agents, we work with Artificialy to use small language models. Why? Large language models (LLMs), such as those used by well-known AI chatbots like ChatGPT, are too big. They are controlled by hundreds of trillions of parameters. Small language models (SMLs) require only a few million to a few billion parameters. Training an LLM costs millions of dollars, while SLMs can also be trained locally. The costs are then significantly lower. The same picture emerges when it comes to the cost per query. A query to a GPT-4 bot costs around $0.09, while an SLM, such as Mistral 7B, costs around $0.0004. The hardware requirements are also much lower, and unlike LLM, SLM can also be operated offline. Instead of several thousand dollars, companies only have a few dollars in monthly costs, and the predictability of costs is much better with SLM.

A single company, even a large and successful one like Abacus, can no longer shoulder the burden of building the ERP solution for the users of the future on its own. The complexity is too great. Today, we are a small fleet of companies working to build our solutions with voice interfaces and agents for the users of the future. Abacus makes the software, Artificialy builds the SLM and solutions for computer vision to enable interaction using natural language. DeepCloud connects the various players, such as SMEs, trustees, and banks, and offers digital services such as digital identities. Tinext hosts our AI platform in Ticino and Arcanite hosts DeepCloud's platform in French-speaking Switzerland.

This text is based on my keynote speech at the Swiss Software Festival. It was written by a human being (but translated from German with DeepL) – so it may contain errors, not because of deepL, but because I'm a faulty human specimen . Thanks god !