Name: Brainstorm AI 2024: The Challenge of Defining AI Transparency
Uploaded: 2024-12-10
Duration: 13 min 25 s
Description: Manohar Paluri, Vice President, Engineering for Gen AI, Meta Interviewer: Jason Del Rey, Fortune

Manohar Paluri, Vice President, Engineering for Gen AI, Meta Interviewer: Jason Del Rey, Fortune  

Brainstorm AI 2024: The Challenge of Defining AI Transparency

Conferences

 Manohar Paluri, Vice President, Engineering for Gen AI, Meta Interviewer: Jason Del Rey, Fortune  

You're gonna help me with this pivot today after that performance. It's hard to match up, man. Yeah, we're gonna, we're gonna, we're gonna do our best here. So let, let's jump right in. Um Meta with Lama has a very distinct point of view on this topic. Can you explain why open your version of open? We won't get into the, is it really, really open? But um we could have those conversations outside but why your version of open and how do you think about the strength of open versus what some people see as the risks? It's a very often question that you get when you lead Lama and uh uh last year, when we faced this question, the answer wasn't super clear to us, but philosophically, uh we always believed open is the way to go all the way from open compute porch. And many of the things that meta has built uh Lama being the latest incarnation of it. I want to answer it in two ways. There is a particular way of thinking about A I development over the past four or five decades. Every time we overestimate what we can achieve with these models. And with the progress in A I, and we feel it's something to be enclosed and built upon and that's where the value is. But every time we hit a wall in unlocking General Intelligence, uh unlocking A I that actually need not be customized can be used out of the box. If you really look at the progress of A I today. In fact, even the 3.3 launch that we just did on Friday, there are multiple variations of those models uh and customizations of that that have already launched. And part of that reason is because it's open because people are able to build on top of it, customize it. So it's ridiculous for me to actually think of a world where we enclose A I and say you have an API for intelligence. Instead we, I genuinely believe the progress in A I over the coming decades is only going to happen if we actually go the open route where everybody can customize it, access it, um deploy various applications on top of it all the way from using an API but also customizing the weights. And that's where Lama uh vision stems from. And it actually has a tremendous amount of benefit for meta as well. So it has, it benefits the world, it benefits the developers, it benefits meta. So that's to me is the reason why we are open. So it's II I can understand seeing a A no brainer especially in the competitive landscape. And I and I get the, the speed to improvement and speed to innovation. Um We haven't really talked about the downsides and um you know, there, there are many out there that, you know, depending on your point of view or if you're a government official, you might see. Right. And so, you know, there's a report report in the fall that researchers linked uh to the Chinese military, we're building on top of an early version of L A. Um Again, I think the response was that's against our policy or our rules um still happening. So why should the world trusts meta with this open source approach with such powerful technology? When we know whether you know whatever your views of Chinese military are there, there could be actors out there who might not use it in the ways designed or intended. Um This particular one. so first and foremost, we should be very grounded on, on how we see the world uh with the data that we have and with the facts that we have, right, the competitive landscape of A I is pretty amazing in terms of the progress we have been able to make across the world. This is not a particular country. This is basically uh if you if you really look at the success of building the best models today, it has four components to it, this talent that you need to be able to build these models, there is compute which is flops that you put into these models, especially the flagship models are the largest large scale models that are actually used to distill and build very efficient models. And there is data, uh Lama at least has been built on publicly available data and license data. So unique data, the higher the quality, the better. And finally, the distribution which is aligning these models for various applications. If you have these four things, you are actually able to build the best technology and unlock value and the competitive nature is across uh the whole landscape across countries and continents. There is a huge uh competition in terms of building these best models. And to me taking a step back, if you want to play the long game of actually making sure this value accrues to everybody in the world. Again, you go back to how many people are building on top of it. How fast are you building on it and how responsibly and safely are you building it. And if you look at the operating systems that are available today, the safest operating systems are the open systems. So similarly, if you put out a model, there are thousands of people who are redeeming these models, there are thousands of customizations of these models that are much more efficient for hardware, much more safer for various applications. How is that going to happen if it's done by a single company or a single entity, it's going to only happen if everybody, all of the talent across the world are customizing this for applications and building a safer uh future and future that is more accessible to everyone. So that's, that's the real. So, so is the point that they, they're gonna always be use cases that we, we, we don't love but uh we're together, we're gonna build guard rails around that or am I, am I misinterpreting? I have a fundamental way of thinking about it, which is if you, if your router doesn't work, you actually don't go to the, you don't go to Linus Walters to actually complain about the router. You basically go to the manufacturer of the company to complain about the router. What is working. What is not similarly when I think about A I and these foundation models, that layer is not where you actually control or where you think about what is working or what is not what applications you actually enable on top of these models is where you have to have checks and balances. And to me that is the real place where you actually have significant value, both upside and downside. And we need to look at it from a responsibility and ethics point of view. That's why when you think about Lama, what we did is at a model level, it's so hard to compress all of the policies of the world into one model. It's not possible. So at a system level, when you build a system, you have this input and output guards, what does that enable you? That what enables you is that you can have the model but you have certain questions that you want to answer, you have certain answers that you want to curb. And that flexibility is what you need to give at the application level. And then the policies that are more regional cultural geopolitical can kick in and say, OK, this is how we are going to deploy these models and this is how we think about safety and responsibility. Let's shift gears real quick. I want to talk about some of the consumer applications of wa and meta A I so um meta A I started, you, you would know the exact timeline started showing up in all our apps, Facebook, Instagram, whatsapp, right in the search bar. Um What have you learned from? You know, I think what some people saw is just, you know, slap, slapping that on there and see seeing what happened like what, what is the feedback been? And then I wanna talk about the scale of, of that use case. So we've been blown away by the adoption of me. We are close to 600 million monthly actives at this point across all of the surfaces that you talked about. But at the heart of what we are, what we are trying to do is Lama is our engine that enables me, I I which is a universal assistant, but it also enables business A is wearables A I which is a Rayan meta glass that you can wear and you can invoke me I I, the assistant and with have a hands free intelligence experience uh created A I and many of the other A I that people can customize towards their liking, whether it's a particular fan following they have on a particular character and so on and so forth. So the goal for us with this foundation models is to enable all of these product experiences. So humans, human connection is stronger, human agent connection is stronger. And that's kind of like the at the heart of what we are trying to achieve building the future of human connection and the technology needed for it. And Lama becomes the centerpiece and med is our universal assistant experience. So it's been pretty amazing to see how med I gets used across all of this. How do you, I mean, because I, I've struggled to figure out what is the right U use case when I'm, you know, looking for, I don't know, an old high school friend or something. Um And that's probably not the right use case, but it's in that same search bar, right? And so um what, what are the use cases that have been most common or, or maybe surprising. So, so many of the uh so II I kind of have two or three that come to my mind because I always believe what you use is the best way to actually think about the product. Uh imagine is one of the experiences that you can use for in Medi. I, so you can actually imagine anything. And recently like three months ago, we had a birthday cake that I designed with my four year old daughter. You can imagine it. You can add a princess, you can add an Eiffel Tower, You can make it a strawberry or chocolate flavor. And then you have this custom cake and you're like, OK, this is pretty awesome. And then your building and then your cookie had. Yes. No, I didn't cook it. I had some, let's cook it. I wanted the daughter to be happy on birthday. Uh uh So there was one, you can actually invoke my I I in a group. Uh So if you think about how you uh uh use these models and interact with our assistants, you're all, you're, you're naturally in a conversation in a group and you can actually uh invoke me. I I to ask a particular latest score. So you're talking about a whatsapp group in whatsapp group. Uh So I'm an Indian cricket huge cricket fan. We're always talking about what's the latest score stuff? This is this group with, with uh my buddies and then we can ask me for the latest score. It will give you the latest information we can actually have uh some really fun roasting sessions uh for the groups on uh for the people on the group. So these are things that are now on me. You could open uh and you could definitely do it. Yeah, I see an editor who would gladly roast me right now. But these are some exact. So you want social plus this intelligence come together at the heart of it and you're already on whatsapp, you're already on Instagram, you're already on Facebook. So basically, we're bringing me I I and all of these A I in that place. So interactions with friends, interactions with businesses is natural and intelligent. Um I was curious about that 600 million number that I keep hearing about active users. So I asked meta A I like, can you like explain like what is an active user mean or what is an engaged user? And so I won't go through all the, we, we had a long back and forth kind of, you know, insightful. And at the end, uh meta, I told me that perhaps more than half of the 600 million might be not explicit uh engagements where I know I'm engaging A I, but maybe a meta A I is working behind the scenes on a result uh on a search result. And um I was just curious is, is that accurate? And um because I, I know, you know, all the companies are very proud of these active user numbers. And so Um Yeah, I'm, I'm curious, I'm curious for your take on what, whether, whether that sounds about right? That maybe half of that number is sort of not explicit uses, but A I working behind the scenes, let me answer that. Uh So the two answers to it, one is uh technology when it works should be seamless. So when you are actually doing a search query or when you're actually asking a particular answer, you're actually trying to do a thing, whether it's on our app, any of the surfaces, whether it's whatsapp Facebook, you're not necessarily trying to say I want to use search, I want to use Medii, I want to use this. You're trying to get a particular job done. You're trying to solve a particular problem from that context. We're trying to make sure Medi I can understand what your query is and if it requires a language model or if it requires a real time query or if it requires a frame graph access, we are able to do that seamlessly for you. So you can actually get the job done in a much better way than before. That's how we think about it. And going back to the question that you are asking, uh it's really hard to break down which particular intent was search only which particular intent was accessing a frame graph or looking at a particular subs string or asking the language model for latest information and so on. So all of this together is what we are trying to unlock with me. I was going to say when you said two different parts, I thought, I thought that was going to be a really wrong non answer. But we're really saying we should just trust meta A I's answer. Maybe we're, we're, we're actually, the technology is definitely improving day by day. And that's one of the other things if you look at the launches, we have done over the last year, we basically did five launches all the way from Lama three family to 3.3. And if you look at where we started the beginning of the year and towards the end of the year, we have more than 600 million variants of a Lama that people have built on the 600 million monthly actives that we talked about. And me, I was in the fledgling uh part, we just started thinking about me. I, I, so going from uh product experience that doesn't exist to a product experience that's used by these many people. It's pretty amazing. I think we're all at the time. Thank you so much, Mano. Thank you. Thank you for having me.