The Exchanges

Q to confirm or not, um, it's very much that Facebook ultimately is very much an AI company, meaning that there are certain, you know, companies where that use a little bit of AI on top on the side for features to do different things. But like at this stage, Uh, Facebook is completely an AI company, uh, and, and, and specifically a deep learning company. Is that, is that correct?

A Yeah, I mean, they're, uh, I mean, they're pretty much deep learning system in every single, uh, Facebook product, and they are very much at the core of them. I mean, the most obvious one is obviously Facebook and Instagram and the Facebook app and Instagram. Uh, they have AI at the core, right? The whole experience is driven By algorithm. And then we see, as I mentioned, the whole moderation is written by that. I think what's, uh, more recent and it's coming more to the surface is what we call AI experiences. So things that are driven by AI in a more visible way. So that's the kind of things you're going to see a lot more, uh, in, in the first coming future. And we'll talk about it today. So I think AI was a lot behind the scene. I think companies like Facebook or You know, Google, uh, AI has been behind the scene and really at the core of the, the engine of the product for many years now. But in the past year, I would say now you start seeing AI really at the forefront driving the experience as well.

AI assessment note: “Yeah, I mean, they're pretty much deep learning system in every single Facebook product”

Q Yep. Thank you. And maybe just like anchor, anchor this in people's minds, obviously not talking about anything, uh, uh, specific about, um, but Facebook, uh, the, the, the, the, the cost of like running those like inferences at scale, just, just like one go can be in the millions, right? Is that, is that correct?

A Well, I mean, not, not one utilization from one user, but often we talk like some, I think what has really increased lately has been also the training costs. Uh, actually, To be clear, the most costly thing we do in ML is still inference cost, because when you put a piece of content within Facebook, it's running hundreds of different ML based algorithm, and they all run on machine parallel, and it's using a huge number of machine, you know, when every time you post some things, right? So obviously it's not a million per post. Where it becomes millions is when you do training runs. So some of the training runs Uh, in the most advanced system that it comes from our company or other companies out there, uh, are starting to be extremely expensive. Yeah. Like you can look at one run in the, in the scale that you're mentioning and that's not sustainable, including for companies like, uh, like, uh, like Facebook.

AI assessment note: “Where it becomes millions is when you do training runs.”

Q and Uh, on, on, on, on Twitter. Um, so what, what's your, uh, sense of, um, where things are going in terms of AGI, so meaning, you know, general intelligence, uh, you know, what's, what's realistic, what's not realistic, what's the sort of like the, you know, from the, uh, ultimate, uh, experts and pros in the field, like what, what's your current sentiment about what's doable, what's not doable?

A Yeah, I mean, when I talk about that, I try to really have a balanced view, right? Because I'm concerned that people make claims out there that gives a, like, a distorted view of the reality, right? So I'm trying to communicate really three things. So the first thing is, uh, that, look, we are nowhere near, uh, human-level intelligence, right? So, um, the system that we're creating, I mean, I talk about Blendbot. It's a really fun system. It's really amazing what we are, but quickly as you interact with the system, you can realize, hey, this thing doesn't have a lot of common sense, right? Um, and so the system we have today are really limited, and I don't believe anybody has a good view as to when will match human intelligence, but it's not going to happen in the next decade, not going to happen in the next two or three decades. It's going to take much longer. How long, you know, I don't know, um, but let's not give the impression that it's around the corner. I think it will give me a disservice to everybody. So that's my first point. The second is I'm trying to have a bit, a bit on a crusade around this term AGI. I think it's very misleading term. Uh, I don't even know what it stands for. You know, it's, it actually stands for artificial general intelligence, but what people don't realize is that, you know, human intelligence is no, it's not general. It's actually very, very …

AI assessment note: “we are nowhere near, uh, human-level intelligence, right?”

Q Good. Uh, one question from Atakan. Um, more about the, I guess the, the, the, you know, most interesting techniques that you may be experimenting with. So specifically, so in which category are you expecting the new GAN like step forward in text and visual modalities? Is it going to be a reinforcement learning like few shots method for, uh, maybe, or unsupervised learning or something else?

A So I guess the one thing right now, and I've been, I started to be pretty vocal about it is, And yeah, and obviously has been pushing this for the past few years is that self-supervised learning is really showing promises everywhere, you know, so it's obviously that's the technique behind, uh, large language models, but, uh, our, you know, my team came up with new papers around vision. So we believe actually that these techniques will supersede the fields, you know, in pretty much every area. So self-supervision really is, is the way to go. It doesn't mean that it doesn't combine with some own supervision. But some pre-training based on self provision is really the future and in pretty much every area.

AI assessment note: “self-supervision really is, is the way to go.”

Q we talk about, you, you, you guys did some really interesting work or are partnering, uh, with third parties, as you mentioned, um, around the, the virus, around COVID. Um, and, uh, so in terms of, in particular, in terms of trying to detect what is, uh, legitimate information versus not, uh, and I read some stuff over at SimSearchNet and other things. Can, can you maybe expand on that?

A Yes, that's obviously another thing that's quite important for us is to try to moderate the content on our platform, you know, in an open way, right? We are trying to make sure that people can have a free speech that they can discuss, they can discuss their views, but when we identify information that's completely misleading, you know, we try to, if you want to fingerprint it, and then, you know, figure out what claims it's made, and then figure out all the contents that's actually very similar to this. So we work often with fact-checking Organization. And then when they fly content that should be, uh, at least, you know, shown as a misleading, then we have this very advanced similarity algorithm that kind of like look at an embedding of the, uh, of the content itself. Again, it can be the image itself or a multimodal aspect of things, and then try to keep telling you all the other content is very similar so that, you know, we can have much faster action on our platform. And if you identify one piece of content, You can remove all the content that are very similar, even if it has been kind of modified temperate ways, or it's like the same version of the same claims.

AI assessment note: “we have this very advanced similarity algorithm that kind of like look at an embedding”

Q Thank you very much. Fascinating. Uh, so this sounds like it's quite extensible to other things. Um, do, do you, or can you, uh, also cover, uh, physician's handwritten notes, for example, to inform discovery?

A Yeah, we don't, so first, first on the extensible, it's not that it's really easy, right? Just think of it right now, we have just 15 bioinformations just focused on drugs, understanding of You know, life science data for drug discovery. So if you want to reproduce that in another domain, you will have to actually also have a lot of bringing domain knowledge and also bring the scientists. So, but we believe that the principle could be applied, especially the molecular exploration like this. Uh, and today, and to answer your second question, no, we're not, you know, you are not using at, uh, patient data at the moment. There's something you could do around the discovery, but it's not the primary, uh, data source.

AI assessment note: “no, we're not, you know, you are not using at, uh, patient data”

Q Right. Any, any, from a technical standpoint, any promising approach is that, uh, is that, is that the, Is that a brute force exercise around, uh, data and like the more videos you have, you know, the smarter the algorithms becomes or are there like different Ways you can do things? Any promising approach?

A I actually haven't seen the, uh, the actual error analysis at the end. So I don't know the, uh, I actually, I mean, there are multiple approaches you can do. You know, some is you can try to find the original content that has been modified. So that's one approach, right? You try to, uh, get, uh, you know, it's back to the bit of what I was mentioning earlier, which is we try to figure out a similarity to that content. The other is you have a, Pretty good understanding of the artifacts of the algorithm that are used today , right? So when a deep learning algorithm generates a deepfake, it leaves some kind of signature or has some kind of like side effects to the content you can try to identify. So, uh, there are a lot of, you know, aspects like this. It's pretty clear that the, the approach we will use will be on ensemble techniques, uh, and then several techniques that combine all these, you know, uh, understanding what's added before, Uh, looking at, uh, you know, understanding the algorithm. Also looking at the multimodality aspect, right? Because you also have, hey, who posted the video? What's the behavior? Uh, what's the context around it?

AI assessment note: “there are multiple approaches you can do. You know, some is you can try to find”

Q More technical question from Ofer. Um, how do you detect drifts and ongoing performance of your AI models?

A Oh, that's a really, really good question. Actually, I'll tell you my, my little, uh, uh, trick this one is that I have a very, very good data science team as part of my overall team, and they are actually creating systems. So my little joke I had in my organization, sometimes people who are good ML engineers or practitioner or AI scientists are not very good data scientists. Okay. And this is a very good data scientist question. And so for me is that I have actually a very strong, uh, data science team that builds actually tooling to detect, uh, this drift and you want to do it in live system, right? So you always try to connect the performance of your system. I mean, it's much more convenient to have offline system, but we always try to connect the offline performance to online performance and always try to link the performance of the UI system, like the change in model and to see the performance over time to your online data and, and to the live system. Very, very interesting problem. Very, very good idea and not easy at all to do. And another good reason to recruit a lot of good data scientists, even before you're recruiting your, uh, ML scientists and engineers.

AI assessment note: “data science team that builds actually tooling to detect this drift”

Q um, former speaker at Data Driven when he was founder of Wise.io and then, uh, I went to GE Digital and a professor at UC Berkeley. So on, on deepfakes, one danger is a blending of an individual image with, with one another. This has been shown to strongly sway preference in the context of political candidate selection, for example. Does Facebook ban the personalization of advertisement along this axis?

A Okay, I, I don't think I can answer, uh, That's a really tricky question. Okay. And I don't have all the policies in mind, you know, and I wouldn't venture to actually give. So you see, we have a very strong view on political ads, but there are limits, right? So as to what we can do, and some of them are banned. First of all, if the ad would say something that would be dangerous, uh, for people to follow or, uh, would be a threat to their health or such as things. So There are limits, um, and I think there were new statements also that are made around how much you can, uh, use, uh, uh, modified or, uh, uh, content for that, but I, I don't have all the detail, and I'm not sure I understood the question in the detail enough. Uh, happy to do that offline.

AI assessment note: “I don't think I can answer, uh, That's a really tricky question.”