Foundation model

yeah well and that's the trick that we've learned that i'm sure is true here and he talks about it in his post that like what makes this possible is knowing what the objectives are and what the cons what the bounds are of what good looks like like you still have to know like what is a what is an acceptable system and what is an acceptable outcome And so like, you know, spot instances replacing your whole fleet is an unacceptable performance optimization because it drives the reliability part of that concern too low.

And so you can't just be I feed the God model my objective and type slash loop and hope, you know, it's like you're going to need there's going to be more sophisticated than that.

Paul, what do you think?

find this fascinating.

Exactly.

And that's why some of the most interesting work is happening in startups that are building these large foundation models for robotics.

Companies like Covariant and Osaro are both working on this.

Covariant's whole pitch is that they've trained a model on millions of pick and place examples from different robots, and then they can deploy that model to a new robot with just a few demonstrations.

So a large group of scholars at Stanford, um, and it was actually led by one of my co-founders, Percy Liang, came together and-

Who coined foundation models

... who coined the term foundation model.

We wrote this paper, uh, where that term came from, called "Opportunities and Risks of Foundation Model." And during that process, really the thing that I started to think deeply about was here is a model that is fundamentally new in our ecosystem.

So-

So, so on that, um, foundation models being at the center, you have this excellent essay where you mentioned that, uh, to this, uh, Claude has their own constitution, OpenAI has their own model.spec, and they don't have clear separation of power.

And they don't have some kind of em-- self-enforcing, um, measures if they go rogue.

What do you think is falling short? Like, what exactly separation of power in this kind of s- AI being in the central-- a foundation model being in the center would mean?

So you might need a diversity of models.

Mm.

That's like one possibility, is the agents doing the monitoring come from different foundation models than the ones doing the government work.

Or it could be something about the way they're, you know, how hosted-

Let's talk a bit about, like, the current developments, um, in, in this area.

We experienced recently, not just in biometrics, but other than biometrics, a lot of increase in works using set supervised and foundation models everywhere.

And I saw that one of your recent works also emphasizes using set supervised and foundation model-based approaches.

Why are these so attractive for generalized morphing attack detection?

So foundation models are pre-trained on a massive amount of data, which provides a very good starting, uh, point, uh, for all the other downstream tasks, and one of these tasks is morphing attack detection.

Uh, so in practice, um, we could see that, uh, foundation models provide very good image features, uh, to perform morphing attack detection and other related tasks.

That is one thing.

The other thing is, um, the avoidance of overfitting.