Grant Harvey

And actually you've got to break it down to what's the desired outcome?

What are the inputs I need?

Then you can actually reimagine it from the ground up and have it ready to be improved.

Like, so that's a couple of them.

Like, if the data's a mess, forget about going any further.

If you just don't understand your process, you may need to break it down to an outcomes basis.

And then lastly, like, if someone's like, oh, I want this task to be automated, like, okay, here's an example of the task.

Like, tell me why it was a good task.

If they're like, because I like it.

That's not an evaluation metric.

Thumbs up, thumbs down, not evaluation.

You've got to break it down as much as you can.

So like ideally be like, oh yeah, this is a good task because it gives a score of 17 out of 21 on these 21 different classification models.

That is a much better, that is something you can then use to actually measure if something's done well and then start to bring in agents to perform it because you can evaluate them correctly.

Otherwise, yeah, again, it's like,

Thumbs up, thumbs down.

It's why AI and the enterprise has been such a disaster.

ChatGPT is amazing at solving 10,000 problems at once.

What enterprise solves 10,000 problems once?

No, enterprise solves one problem 10,000 times.