And with the possibility of input by voice or keyboard—do employees have to be trained in the AI?
At the beginning, there’s initial training for the defined use cases in production. Plus, our employees have access to further training opportunities on the subject, including a learning path on prompt engineering. But they also learn how to use these tools in a creative way to try things out and see what works and what doesn’t.
In general, though, I believe prompting, or prompt engineering, is something you have to learn, so we’re considering whether we should offer training for it more widely throughout the company, and not just for selected IT and data professionals. It definitely helps to get more out of generative AI.
What teething troubles of AI or ChatGPT have you encountered so far?
Hallucinations are certainly a challenge. That was also a very delicate balancing act in direct customer interaction in the UK. You can largely rule out hallucinations by plausibility checks and the associated restrictions, but if you set the criteria too narrowly, the machine will tell you, “I can’t comment on that,” more often than you’d like. You have to be very careful and find the right balance. How to get a grip on hallucinations is perhaps the most important question to be solved at the moment, which is also at the center of AI research.
Will Mercedes-Benz only train its AI tools on its own data?
Yes. For example, if we want to explain our vehicles to customers visually, then this can only be done with our own training data. Incidentally, the training takes place exclusively in secure areas of these AI environments, so the data can’t be made public. There’s also some public data we can use for AI, but especially in the production environment, we rely on our own data.
Apart from Azure OpenAI Services in the production environment, what roles do other AI solutions play for Mercedes-Benz?
OpenAI is currently being portrayed in the media as a bit of an AI spearhead. And there’s a very good technical solution too, but we won’t limit ourselves to that. Of course, other companies have interesting solutions. We’re starting to look closely at open source alternatives. In addition to the large proprietary providers such as OpenAI, Microsoft or Google, we need to understand the open source alternatives.
I also believe we shouldn’t think of AI as an engine that stands somewhere unto itself. It needs to be deeply woven into our systems and processes. That’s why we require all our system partners to use AI elements in their environments. It must find its way into the entire system landscape, and it will.
