Hmm.

And so that's the first step is understand like where are we? What is state-of-the-art within the field? And then the second step is what are the things that experts would say could make that better? And so that's where you take the benchmark and turn it into an understanding of like, huh, this is interesting.

It's constantly giving advice to go down this path, and it's not clearly looking at these data sources, or, or maybe its, um, its tone is wrong in these steps, right? This is where you have to bring the subject matter experts, in this case advisors, in, into the fold and say, "Tell us about this." Like literally looking at transcripts of interaction and annotating, right? So that's the second step.

And then you can build what we call public goods, which, um, for the, you know, the AI geeks, uh, following, uh, following this, that could be a model harness, which says, "We're gonna attach this to a model, and it will instruct it to do the following things," or, "It will point to the following data sets automatically," right? And then you do the process again.

You say, "Okay, well now let's take a look with the model in this harness and put it up against that same benchmark we used.

Did it get any better?" Usually you see some kind of improvement, but it's still not human level quality, right? So there's a version of that which is wash, rinse, repeat.

Along the way, you also have to be looking at people who are playing around with this, experimenting with this at a very, very small scale and saying, "Which parts of this would you rely on the technology to do versus where do you need the technology to say, 'Go talk to a human.'?" Uh, and, and how can you...

Yeah.