Roland Busch

Remember, when you walk a plant and you see the whiteboard where a supervisor writes, I have a problem here, this is who we're working on, here I fixed it.

All that knowledge goes into this model.

So that you just say, okay, this is the pattern of a problem.

This is how the fix was.

The model knows it because we upload the data.

So it's a model which is trained.

This is why we talk about an industrial AI model, which is trained on the industrial data.

I mean, if it really runs across, but also very specific if it comes to certain machines, then the hit rate goes up from 60, 70% to the 95-ish, 80-ish plus, which is really then what you can do.

We don't do LLMs, so really large models which are trained on LLMs.

the whole knowledge in the world, this we don't do.

I mean, this is what we're using, we're using any kind of, and for specific tasks.

So we have some models which are very good in software, co-pilots and agents for software, some are good for... Now we are working on really genuine, newly product designs, not only just having the next code line, but really genuine designs, a completely different world.

And in some cases, we're working on co-pilots on the shop floor.

We talk about Microsoft, for example...

Are we having first use cases doing that?

So challenges, of course, that an industrial AI application doesn't accept hallucination.

I mean, you really have to be sure once it's in an agent out, it does what you want it to do.

We need the augmentation, absolutely.

Good point.

To the question, LLMs will be getting better and better, but I don't believe that these LLMs, if you do not train them really on specific industrial data, and this is where the augmentation comes from,