LLMs AREN'T Models! Stop Treating Them Like Magic! #shorts

追加: 2026-05-06

[00:00:00]Companies are out actually trying to say they need to change what's called SR 11-7, which is the structure that helps you define what a model is and going through this process. In simple terms, a model has inputs and you're going to predict some sort of output and you have some sort of model error in there. And of course, this entire process is math and stats and statistical theory comes into play on how we test things and know that we estimated the right process. But there's an estimation part of this to get to some sort of prediction at the end of it and there's an error. Okay?

[00:00:27]LLMs don't do that, as many people know.

[00:00:30]They're not models. You can call it a model. It's more like an algorithm, but it's it's it's not a model. Um and it's also not deterministic in the sense that they keep adding in new information.

[00:00:39]It's still learning throughout this process. And even when you create and generate um a layer on top of an LM trained specifically to do a specific task, you're now introducing kind of variations of this. But there's no estimation and there's no exact answer typically, right? Now, we could use it for doing mathematics, but the LLM is not designed, operated, theoretically sound to do predictions of a very specific thing. It's not a model.

[00:01:03]Um and because of it's not a model, and I don't think you should use it as a way of black-box prediction magical model, um it should not be validated or reviewed by model risk management.

[00:01:16]And a lot of people I think are like, "What? You got to be kidding me." Like they can't believe I'm saying this. Um but it's not a model. It's mathematically, statistically, theoretically not a model. And unfortunately, only people with a background deep in math and stats and econometric tools will understand that.

[00:01:31]You have to understand that. Um if your MRM teams are out running and wanting to govern and process and review and validate use cases of LLMs, um I would say it's an overreach and an abuse of power in many instances here because like let's look and think about like how this is being done and why it's not a model in itself, okay? So, a model example here would be like I'll use some traditional examples. A model is like you put in a bunch of financial characteristics on a consumer and you predict out the probability they're going to default. Okay?

[00:02:01]Um another one would be like you put in a bunch of data to predict out like derivative pricing. It's another example. So, there's inputs that go into this. There's an estimation process and procedures. We're trying to minimize some sort of error at the end of it. And then at the we get some sort of prediction that we're going to get out of this and we're going to decision on.

[00:02:17]But the model itself has estimation um and it has a model error and it has some sort of prediction that comes out.

[00:02:23]That's what a model is. A perfect example of a non-model, right, is accounting.