Nilay Patel

We can automate even more of the things I mentioned earlier.

It seems like a lot of the frame for how you're thinking is Siemens has traditionally automated atoms.

Now you can automate bits.

Yes.

From, okay, you've decided how many units to produce, we'll produce them, to we're going to actually automate the deciding of how many units to produce.

Your agents, when you talk about industrial agents, so there's some line, something's gone wrong, a warning light goes off, and you say an agent will help you figure out what's wrong and potentially fix it by itself.

Is that based on an LLM?

Are you using one of the models from one of the big companies and it's just an LLM that you've trained to think about a line in that way?

Are these your models that you're training or are you augmenting models you're taking from OpenAI and Anthropic?

This is my fundamental question, and I've asked a lot of people this.

I'm very curious for your perspective because the domain is so much different.

I am not convinced that LLM technology as it exists today can...

Make the leap to do all of the things that people want it to do.

Right.

You see the gaps, even as you're saying an LLM on its own as it hallucinates enough to only be effective 60 to 70 percent of time, which is nowhere near good enough for all of the things people want it to do, especially at the labor replacement rates that some of these folks talk about.

Do you think it's good enough, or do you think it's the actual augmentation that makes the products that you build with it good?

And do you see LLM technology, the core technology, improving at a rate that might change your assessment of it?

All that data has to come from lots of different customers, right?

And you talk about Siemens as having all that data, but that data actually belongs to your customers.

Are they willing to let you aggregate so that you can develop the products at the scale that you're talking about?