Roland Busch

They can train them as long as they want.

They will not get to the level which we can use on the shop floor.

It will not work.

So they need... I strongly believe that the LLMs need specific, domain-specific, machine-specific data in order to really make a difference.

But then, if you do that, then you really can make a step up which is fundamentally higher.

I give you two examples which are maybe interesting.

One is...

When we talk about it was an optical inspection kind of task where we used an RLM and say, okay, show me the problem.

Hit rate was okayish, but not to a level we needed.

Then we start training the model with not so many data.

which is anyhow important because if you're manufacturing in a PPM level, guess what, how many mistakes you get a day.

But on those which you have, and then you create obviously some synthetic data around those, once you train the model, your hit rate goes up substantially.

It's much, much higher than if you use the next best and next best model.

Is there convergence to some extent?

But I do believe there's a certain ceiling where you can train as well as you want.

If you do not get specific training, then you have a problem.

Now, my second example is we, and this is now really, now I give you a little bit the spirit of how deep you have to go.

Take a, we have a manufacturer, Italian, they do robots, a crib in the box.

Crib in the box for any kind of parts.

And obviously you can train a robot to make this crib in the box.