Manohar Paluri, Vice President, Engineering for Gen AI, Meta Interviewer: Jason Del Rey, Fortune
Category
🤖
TechTranscript
00:00You're going to help me with this pivot today after that performance. It's hard to match up man. Yeah, we're going to we're going to do our best here. So let's jump right in.
00:11Metta with the llama has a very distinct point of view on this topic. Can you explain?
00:18Why open your version of open? We won't get into the is it really, really open, but we could have those conversations outside. But why your version of open? And how do you think about the strengths of open versus what some people see as the risks?
00:39So very often question that you get when you lead llama and, uh.
00:45Last year when we face this question, the answer wasn't super clear to us. But philosophically, we always believed open is the way to go all the way from open compute by torch and many of the things that meta has built and llama being the latest incarnation of it. I want to answer it in two ways.
01:04There is a particular way of thinking about 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.
01:21But every time we hit a wall in unlocking general intelligence unlocking.
01:27AI that actually need not be customized can be used out of the box.
01:33If you really look at the progress of a I today.
01:36In fact, even the three dot three launch that we just did on Friday. There are multiple variations of those models on 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.
01:59And 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.
02:14Deploy various applications on top of it all the way from using an API, but also customizing the weights and that's where Lama vision stems from and it actually has a tremendous amount of benefit from it as well. So it benefits the world. It benefits the developers. It benefits meta. That's to me is the reason why we are open so it's and I could understand seeing a no brainers, especially in the competitive landscape and I and I get the speed to improvement and speed to innovation.
02:44We haven't really talked about the downsides and there are many out there that depending on your point of view or if you're a government official, you might see right and so there's a report report in the fall that researchers linked to the Chinese military were building on top of an early version of Lama.
03:03Again, I think the response was that's against our policy or our rules still happening, so why should the world trust?
03:17Meta 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.
03:34This particular one. So first and foremost, we should be very grounded on on how we see the world with the data that we have and with the facts that we have right.
03:46The competitive landscape of AI is pretty amazing in terms of the progress. We have been able to make across the world. This is not a particular country is basically if you if you really look at the success of building the best models today. It has four components to it.
04:01There's 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 or the largest large scale models that are actually used to distill and build very efficient models. There is data.
04:17Lama at least has been built on publicly available data and license data so unique data, the higher the quality, the better.
04:23And finally, the distribution, which is aligning these models for various applications. If you have these four things, you're actually able to build the best technology and unlock value.
04:32And the competitive nature is across the whole landscape across countries and continents. There is a huge competition in terms of building this best models and to me taking a step back. If you want to play the long game of.
04:46Actually, 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?
04:57And 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 red teaming these models. There are thousands of customizations of these models that are much more efficient for hardware, much more safer for various applications.
05:13How 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 various applications and building a safer future and future that is more accessible to everyone. So that's that's the real problem.
05:29Is the point that they're going to always be use cases that we don't love, but we're together. We're going to build guardrails around that or my misinterpreting.
05:41I 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.
05:47You don't go to Linus to all this 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 AI and this foundation models that layer is not where you actually control or where you think about.
06:02Uh, 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 llama. 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.
06:27So 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 a 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're going to deploy these models and this is how we think about safety and responsibility.
06:53Let's shift gears real quick. I want to talk about some of the consumer applications of llama and meta AI. So many I started you would know the exact timeline started showing up in all our apps Facebook Instagram. What's app right in the search bar.
07:10What have you learned from you know? I think what some people saw is just slap slapping that on there and seeing what happened. What is the feedback been and then I want to talk about the scale of that use case.
07:23So we've been blown away by the adoption of meta. 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're what we're trying to do is llama is our engine that enables meta which is our universal assistant.
07:41But it also enables business eyes wearables which is a red band made of glasses that you can wear and you can invoke meta idea assistant and with hand have a hands free intelligence experience creator eyes and many of the other areas 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.
08:02So the goal for us with this foundation models is to enable all of this 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're trying to achieve building the future of human connection and the technology needed for it and llama becomes that centerpiece and meta is our universal assistant experience. So it's been pretty amazing to see how meta gets used across all of this.
08:24Yeah, how is do you I mean because I've.
08:27I've struggled to figure out what is the right you use case when I'm looking for. I don't know an old high school friend or something and that's probably not the right use case, but it's in that same search bar right and so what are the use cases that have been most common or maybe surprising so many of the so I kind of have 2 or 3 that come to my mind because I always believe.
08:54What you use is the best way to actually think about the product imagine is one of the experiences that you can use for met in Meta. I so you can actually imagine anything and recently like 3 months ago. We had a birthday cake that I designed with my 4 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 like OK. This is pretty awesome and then you're building and then you're cooking had. Yes, no, I didn't cook it. I had some let's cook it.
09:24I wanted the daughter to be happy on the birthday.
09:28So there was one you can actually invoke Meta in a group. So if you think about how you.
09:34Use these models and interact with assistance. You're all you're naturally in a conversation in a group and you can actually invoke Meta to ask a particular latest score. So you tell me what's up group in what's up group. So I'm an Indian cricket huge cricket fan.
09:49We're always talking about what's the latest score in stuff. This is a group with my buddies and then we can ask my eye for the latest course. It will give you the latest information.
09:58We can actually have some really fun roasting sessions for the groups on for the people on the group. So these are things that are now on me. You can open and you can definitely do it. Yeah, I see an editor who would gladly roast me right now.
10:12So these are some exact so you want social plus this intelligence come together at the heart of it and you're already on what's up. You're already on Instagram. You're already on Facebook. So basically, we're bringing meta and all of these eyes in that place. So interactions with friends interactions with businesses is natural and intelligent.
10:30And so I won't go through all the way. We had a long back and forth kind of insightful and at the end I told me that perhaps more than half of the 600,000,000 might be not explicit, but it might be a little bit more than half of the 600,000,000 might be not explicit, but it might be a little bit more than half of the 600,000,000 might be not explicit, but it might be a little bit more than half of the 600,000,000 might be not explicit, but it might be a little bit more than half of the 600,000,000 might be not explicit, but it might be a little bit more than half of the 600,000,000 might be not explicit, but it might be a little bit more than half of the 600,000,000 might be not explicit, but it might be a little bit more than half of the 600,000,000 might be not explicit, but it might be a little bit more than half of the 600,000,000 might be not
11:00engagements where I know I'm engaging AI but maybe a meta AI is working behind-the-scenes on a result, on a search result. And
11:08I was just curious is is that accurate? And
11:13because I know, all the companies are very proud of these active user numbers. And so
11:19yeah, I'm curious for a take on what, whether, whether that sounds about right. That maybe half of that number is
11:25sort of not explicit uses but AI working behind the scenes.
11:29Let me answer that.
11:31So there are two answers to it.
11:33One is technology, when it works, should be seamless.
11:38So when you're actually doing a search query or when you're actually asking a particular
11:41answer, you're actually trying to do a thing, whether it's on our app, any of the surfaces,
11:47whether it's WhatsApp, Facebook, you're not necessarily trying to say, I want to use search,
11:51I want to use Meta.ai, I want to use this.
11:53You're trying to get a particular job done.
11:55You're trying to solve a particular problem.
11:56From that context, we're trying to make sure Meta.ai can understand what your query is
12:01and if it requires a language model or if it requires a real-time query or if it requires
12:05a friend graph access, we are able to do that seamlessly for you so you can actually get
12:09the job done in a much better way than before.
12:12That's how we think about it.
12:13And going back to the question that you're asking, it's really hard to break down which
12:18particular intent was search only, which particular intent was accessing a friend graph or looking
12:23at a particular substring or asking the language model for latest information and so on.
12:29So all of this together is what we're trying to unlock with Meta.ai.
12:32I was going to say, when you said two different parts, I thought that was going to be a really
12:36long non-answer, but we're really saying we should just trust Meta.ai's answer, maybe.
12:42Look, the technology is definitely improving day by day, and that's one of the other things.
12:48If you look at the launches we have done over the last year, we basically did five launches
12:51all the way from Llama3 family to the 3.3.
12:54And if you look at where we started the beginning of the year and towards the end of the year,
12:57we have more than 600 million variants of Llamas that people have built on, the 600
13:02million monthly actives that we talked about.
13:04And Meta.ai was in the fledgling part.
13:06We just started thinking about Meta.ai.
13:08So going from a product experience that doesn't exist to a product experience that's used
13:12by these many people, it's pretty amazing.
13:13I think we're all out of time.
13:15Thank you so much, Mano.
13:17Thank you for having me.