Why Meta Wants To Build Artificial General Intelligence — With Joelle Pineau

0:00The head of Meta's AI Research Division joins us today to discuss the company's pursuit of human level artificial intelligence, the cutting edge of AI, why it's open sourcing, its large language, and plenty more than other than the company. interview the company is giving about its recent news. All that and more coming up right after this.

0:24You're used to hearing my voice on the world bringing you interviews from around the globe. And you hear me reporting environment and climate news. I'm Carolyn Buehler. And I'm Marco Werman. We're now with you hosting The World Together. More global journalism with a fresh new sound. Listen to The World on your local public radio station and wherever you find your podcasts.

0:51Welcome to Big Technology Podcast, a show for cool-headed, nuanced conversation of the tech world and beyond. Boy, do we have a show for you today. We're recording to the minute on Wednesday here, right before we drop this episode, because there's breaking news coming out of Meta, all about the moves that they're making with their AI division, their pursuit of human level intelligence. We have none other than Joelle Pino here to talk to us about it. She's the head of Meta's AI Research Division, formerly called FAIR, now I guess it's called Mare. Still FAIR, Fundamental AI Research, we love the name.

1:26Okay, well keep it FAIR, keep it running. We spoke actually in October 2022 before ChatGPT, so this is going to be a really cool moment to talk a little bit about where we've come from there and where we're going. Joelle, welcome to the show, great to see you. Thank you, Alex. Great to be here. So if you recall, in October 2022, when we spoke a couple times at the World Summit AI, it's kind of funny because the big storyline then was whether AI is sentient. And this was kind of a moment where all the big research houses had big large language model chatbots internal, and they hadn't released it yet.

2:05And it's kind of interesting how society starts to talk about a breakthrough, right? It sometimes goes in a weird direction before we're actually refocused on what matters. And now I think we are refocused on what matters. There's been much more talk beyond sentience in terms of the near-term viability of this technology. I'm curious just to start, what has surprised you in the research since that discussion? Not necessarily about, okay, we all know that it's now taken off and it's been hyped, but has there been anything that's made you sit back and be like, wow, we can actually do more than we thought we could a year or a year and a half ago.

2:42So many times this year, honestly. And it's great to think back to that point in time. I hope you didn't ask me for any very specific predictions, even for someone who's deeply in the space of AI, just predicting how this is unfolding continues to be full of surprises. I will say, you know, it's also been interesting. The faster we progress, the more we have a sense of how much more is left to do. And so though, you know, back you mentioned back in October 2022, we were worried about sentience and we hardly talk about it now. And yet we are so much further along on the map in terms of our ability to have models that deeply understand information and process multimodal data.

3:30So we're getting further along, and yet we worry about some of the more concrete problems. We've talked a lot this year about safety, for example, about how to make sure that we have models that are performing well, but also are aligning with the values of peoples and the needs of people, which I consider sort of a much more grounded problem that we can tackle with research. So that's, I think, the major change that I see. Significant progress, but that means we also have a much better view of what are the real problems we need to solve. Yeah, it's funny because back then we also had a discussion about whether we should be focusing on like the short term or the long term problems.

4:13And obviously, those are both worthy of attention. And it's kind of wild that the focus on the long term problems, it seems like, blew up OpenAI over a weekend and maybe it's been put back together now. But the talk from Meta now is actually focused on some of the more big ideas that people might have thought were more long term. But now it actually seems like it might be closer than we think, at least according to some of what we hear from OpenAI and others. So this is a quote from Mark Zuckerberg that just came out fairly recently. He says, as recently as last week, we've come to view that in order to build the products that we want to build, we need to build for general intelligence.

4:53So, I mean, Jan LeCun, in our discussions, I've been speaking with him since 2015, one of your colleagues, he's always talked about how the goal is building for artificial general intelligence. So when I saw Mark come out with that last week, I was like, yeah, yeah, that's been the focus for Meta. But all of a sudden, it almost feels like there's a more pragmatic or it feels more real now than it did before. Am I reading that right? What is leading us to now start to talk about this as something that's not pie in the sky 20, 30 years down the road, but something that might be achievable in the nearer term?

5:28Yeah. I mean, Jan and I and the team in FAIR have been talking in those terms for many years. It's been clear we've been putting in place sort of a portfolio of projects that are trying to build the building blocks towards general intelligence. In the last year, Mark, as well as many others, has taken a deeper interest in what's going on in AI. I think he was always aware of a lot of the good work we were doing, but didn't dig in quite as deeply. And in the last year, definitely has. And through a lot of conversations, I think has come to see how, in many ways, the path even to bringing AI to the products that people use and love from the company, the path to making those AI systems better goes through building general intelligence, not narrow intelligence.

6:16And we've done a ton of work on AI on the platform over the last few years. That was what I would call more narrow specialized models. We can continue to do that, but the bigger step change are going to come through the more general model, building foundation models, building up to world models that essentially can capture a much richer version of the information. So I think that's what you're hearing from Mark. It's things that you've been hearing from Jan, myself, and others through the years. We're working together to connect these pieces together, both the research roadmap as well as the product roadmap, and make sure that we have the ability to connect these together.

6:59So the ability to have our research quickly diffuse in the best way possible through the product, and the ability to learn. The thing about general intelligence is you have to solve many different problems to have, you know, the ability to claim general intelligence. And fortunately, there are a lot of use cases across meta, across our family of products. And so that's giving us wonderful material with which to work. So why? So let's go back to, you know, that October 2022 discussion that we had before ChatGPT come out. Like the idea of me asking you this question about like, why is human level intelligence now in focus?

7:36I never would have asked it. It just didn't seem like it would be something that would be relevant to ask. But now it does seem more relevant and we're hearing it more and more in the discussion. So you mentioned world models, foundational models. But what about AI research now is allowing us to ask those questions? I think it's because, you know, the models are getting increasingly general. If you look at a model like CHAT-TPT, the Lima family of models that we've been releasing, you know, they started just as word prediction models. All they would do is take in sentences and predict what comes next.

8:15And what we're seeing is we can use them now through many other uses, whether it's to predict things that are not just words, but they're actually code. And some of that code is actually executable. Or you can predict, you know, the components of an image, and then you can plug in a diffusion model or other kind of synthesizer to realize the information. So what started as just language model has become much more general on its own, it gives us a path, it may not be the path, but it gives us at least a path to move towards general intelligence. And it's an exciting one. It's one that we're exploring.

8:54It doesn't mean that we've stopped exploring other paths towards general intelligence, but that is definitely the one that has proven to make the fastest progress. And what would you say the path is? The path is to essentially capture a lot of human information through this representation that we call language. And so the hypothesis that, you know, even things that are not necessarily text-based originally, if you describe them through these discrete tokens, and sometimes these discrete tokens are the words that we use to express, but sometimes discrete tokens are, for example, code numbers, essentially like chunks of images.

9:36And these discrete tokens are a path to representing all of human information. There was a study that came out last year, basically saying that these models can't generalize outside of their training set. And I think that was like a lot of the hype around these models were people saying that they were really able to have these capabilities that you you wouldn't expect emergent capabilities. And the study basically pushed back on it and was like, listen, they're not gonna generalize beyond their training set. And your evaluation of that study basically, you know, leads you to believe either A, if you believe that study, then you're a lot less optimistic about this wave.

10:20If you don't believe that study, you can really use your imagination and believe that what we're hitting on now, these foundational models that you talk about can lead us in directions that we never could have dreamed of. So I'm curious what your evaluation of that study is and how we should be thinking about this. I tend to really like be quite balanced on a lot of these questions. I think it's very easy to kind of, you know, pull opinions to one side or another. But the truth is like machine learning algorithms can generalize. That is a property of how we build these algorithms. Even the simplest, just linear models, they do linearize.

10:57They just linearize along a line. So, you know, the fact of the matter is, though, when you project that into a very, very high dimension, so some of these models have hundreds of billions of parameters, you have to think of like you're learning a function in that really high dimensional space. The directions in which you can generalize are so many that it's hard to know which are the good directions to generalize and which are the poor directions to generalize. the more data you have, the more that constrains that question. So I do believe they can generalize. I think they generalize relatively narrowly, or at least, you know, as long as you stay close, you get a good manifold of information.

11:42When you start to go really far afield from your data, because the dimensions are so large, you get all sorts of noise. So the advantage, and one of the reasons a lot of the progress has been through better and bigger data sets, bigger but also cleaner data, is because that really defines which parts of this really high dimensional space are the most interesting one. And when there's not a lot of data to populate that space, then the models tend to regurgitate the things that they've been trained on. So let's go back to that Zuckerberg quote that I read earlier. We've come to view that in order to build the products that we want to build, we need to build for general intelligence.

12:22Now, we talked a little bit about why that is now relevant, the path towards general intelligence. But now I'm kind of left with another question, which is, why does Meta need to build general intelligence in order to build the products that you want to build? I mean, yeah. I mean, just looking at like a couple of the AI products we've released this year, you know, one of them is the Meta AI Assistant. People who are in the US have been able to try this out on some of our platforms where you can essentially ask for questions and ask for assistance. In that case, you know, there's a sense that it has to understand a very large spectrum of information to be able to do well.

13:03And as we incorporate more data and as we perfect this instant, the more it's going to have essentially world knowledge, the better it's going to be. Another example is for those who've been following our work on AR devices, the smart glasses that we released earlier this year also come now with an AI model also accessible mostly in the US at this time. There too, you know, you have essentially a more embodied version of this meta AI assistant that sees the world as you see it, that is able to take on some action. In this case, the actions are not just words. It can take pictures. It can provide information.

13:41It can record information. And so to be able to do well in a wide set of different tasks with a wide set of different people, different environments, you need to have to move towards more general intelligence. That's really where that connects the research work we're doing and already what we're seeing in terms of the applications that Meta is putting out there. Yeah. Now, let's say you do achieve this and you open source it. Is that kind of like the end? Like is human reaching human level intelligence or general intelligence kind of like the end of AI research? Or is there more to do after that happens?

14:23There is no end to this journey. I mean, I hope there's no end to this journey, right? Like, do we as adults sort of say, okay, I'm going to keep on growing my knowledge. And at some point in time, I don't know, for some of us, maybe 25, some of us, maybe, you know, 75, you decide like, okay, now I'm done. Like, I have reached where I am in terms of human intelligence. I don't think that's how it works for humans. The world is always evolving. There's always more to be curious about. And so I think that's the path that we are on with our AI algorithms. Similarly, they need to stay curious about the world that they evolve in.

15:01And over time, they need to figure out how to integrate that information and sort of rise to the challenge of the world that they're building. But because the environment is not static, I don't see us coming to an end. That's so interesting because it's always described as the finish line. And actually, There's people who would argue that there's no such thing as human level intelligence, that the second you hit that, you're basically left with super intelligence and game over. Yeah, I mean, I don't really ascribe to that scenario, I have to say. And the other nuance I will add to this, you know, often this notion of general intelligence is articulated in the context of like a single agent, a single uber-intelligent agent.

15:49And I don't think that's really where we will move towards either. There's clear evidence that as a species, humans, animals, we learn so much more through interactions and so much of our culture and our intelligence is derived from our ability to interact, to collaborate. So I think that's also going to be a super interesting door to open as we are on this journey to think about how do we build AI agents that are not just, you know, pushing for single entity intelligence, but are connected to a network of other intelligent agents, whether synthetic silicon agents or human agents. Well, it's so interesting because language, of course, like speaking of types of intelligence, language is only one type.

16:38Yes. And Jan and I spoke about this on a recent show, not so recent anymore, but your interactions in the world teach you so much that you never learn with language. Your understanding of gravity, for instance, is not something that you can implicitly understand from language. So are you doing research now to help Metta's research division figure out stuff beyond words and images? Absolutely. And I would say that may be one of the distinguishing factors compared to other research group out there. There's a strong belief that having AI agents that are deployed in the physical world, where the notion of embodiment is important, is something we should be pursuing.

17:18We have a research team that's dedicated to this. They do some work in robotics, in particular, because that's the best agents we have to consider physical embodiment, spatial constraints. It's not necessarily because Meta intends to commercialize robots. It's because by going through these essentially devices, we have a lot to learn about how to build AI models that live in the physical world. And in the work that we've done recently with the smart glasses, the models that proved to be useful for that use case came out of the work that this group was doing. People who were looking at robotics and devices, physical devices, living into the world and building AI models specialized for that was incredibly useful to inform the work going into the glasses.

18:05Of course, we also leverage the work we were doing on language and our LAMA family of models. But LAMA on its own doesn't make for the best assistant on glasses because it doesn't have enough of an understanding of the physical world, of images, and so on. Now, there are some people saying that the reason why Meta is now speaking about AGI is because OpenAI is speaking so much about AGI and other research houses are. And getting the talent to work on these projects is really difficult. This is something that Mark actually said in that Verge interview. I think that it's important to convey because a lot of the best researchers want to work on the more ambitious problems.

18:49So I got to ask you straight up, is the talk about AGI more of a recruiting thing? No. I mean, like, of course, we love to have great talent. And of course, this is a competitive market for talent. But we don't talk about anything just because someone else talks about it. Like, we genuinely are doing the work and we've been doing it for a number of years. There's no major shift in terms of our ambition to solve AI that's been inscribed in our mission and our goals for FAIR for many years now. Mark is talking about it now. I think he's excited about the work. It's wonderful to have his support to do it.

19:27but it doesn't necessarily fundamentally change the problems that we have to solve, the work that we're doing. I think there's also a sense that we are being more explicitly ambitious about this work, which goes along with some of our investments on the compute side, which are necessary to fuel that work. And so that's why it's coming out maybe more from what you're hearing from Mark. But I think if you go back and listen to what Jan or I or some of our other senior researchers have been saying for a number of years, there's not a departure there. Right. And briefly on the talent market, what does the talent market look like right now?

20:04Is there a real scarcity in the type of people that can do this type of work? And what is it like recruiting against fast growing, and especially in terms of valuation, competitors like OpenAI, Anthropic, et cetera? Yeah, it's always been a very competitive market. I would say going back to about 2016, 2017. Since then, I don't really remember a year where it was like an easy, slow market in AI. And so it continues to be. One of the things that has changed in the last year or so is mostly on the startup scene, I would say. You know, three years ago, we didn't feel much competition with the startup scene.

20:47Now we do a lot more. I tend to view this as relatively positively, to be honest with you. And that's one of the reasons we open source our work. We genuinely believe that more people working on this is good. And so when we open source our work, we get to leverage the creativity of a greater number of people. And there's many more than we can hire. So I think that the very, very top talent that can train these models continues to be incredibly valuable to Meta as well as to other organizations. And fortunately, there's also a good pipeline of students. You know, I do have an affiliation with Mila, the Montreal Institute for Learning Algorithms.

21:25There are hundreds of amazing grad students coming out of that institute as well as others. We've set up some joint PhD programs in some cases so that these students have an opportunity to come work at least part time or through internships with us. And so we're both, you know, sharing with them the work that we do as well as having an ability for us to see whether they're a good fit for our work. So I feel like we have a great talent pipeline, but it continues to be a competitive market. Got to ask you about open source. Brad Smith from Microsoft has talked about how open AI is the most open.

22:02And I'm kind of curious from your position, are they living up to that open name? Is there real open sourcing there? And what is the state of open source? I mean, why is meta open sourcing outside of like, I mean, from like a, you know, meta is a business. So from a business perspective, why open source? Yeah, there's different levels of open sourcing, right? I do think, you know, having an AI model where you provide an API is sort of one layer of that, which is something that OpenAI has done. But there's a lot more that goes on. And so, you know, from just providing an API, you can make available the code that was used to train it.

22:39You can make available the trained model weights, which enables someone to run the models. And then there's a number of other artifacts that come across from this. We've been focusing on making available like model cards that give a better understanding and transparency about our models, good use guides, tools for safety and so on. So there's like a whole ecosystem of artifacts. I think the purists would say like everything has to be to be out there in an open way. So we even have some people who are coming from the software open source community who feel like we're not living up to the full view.

23:17Again, there's a continuum on this. It is clear that Meta has taken a much more open view than other big players in this space. And in particular, we've been releasing some of our code and model weights for some of our larger models, including LAMA. it comes from a lot of deep discussions in doing that. And so I think there may be a misperception that we would do this without any process or reflection, and it's a little bit of a religion. That's really not the case. We have quite a thoughtful process that's been put in place. You have to remember, we've been doing open sourcing work for 10 years since the first day of this organization.

24:04So we've built up a lot of muscle of how to do that in a responsible way. We do it in consultation with a wide set of people who have deep understanding of safety, ethics, and so on, who get brought into the process. And what's been wonderful to see in the last year is as the conversation has been moving and as the models have been going better and getting bigger, we've invested a lot more into being thoughtful about our release process. And so I would say now we have a much more mature process than we did a year and a half ago that involves a much more diverse group of stakeholders. We have a really rigorous process in terms of measuring the risks of these models across different categories of risk.

24:50So it's been exciting to see how much our commitment to open sourcing has driven us to innovate. And we've open sourced a lot of those innovations on the safety side. I think the Purple Llama tools is an example of that, which we released in December. And so it's been great to see that. I do hear a lot of people who are concerned about open sourcing. and I have many conversations with them, including other large organizations. My worry about closing the doors down now is that the models are only getting better. And so if we don't release them now, we really miss an opportunity to develop the muscle we need to make these models safer.

25:34And I don't think today's model are the ones that are going to you know, bring to the front the hardest questions. These models are yet to be trained. Is there anything standout that you've seen being built on top of the open source Lama model that Meta has put out there? Anything standout in terms of like a cool product that you've seen and anything concerning that you can talk about? There's definitely been dozens of product that are coming out of that. How about naming? Yeah, do you want to name one? Let me take an example, our segment anything model, which is a little bit different than our LAMA model, but I think has been the one that has been just incredibly impactful in terms of people quickly building on it.

26:19Our segment anything model is one where you take an image and it gives you a detailed segmentation of that. of that. We released it back in April, including a lot of tools and data to go along with it. And within days, we had people who had built up applications essentially for conservation applications. So being able to track down some species who may be endangered, using that to follow them. We had people use it for the treatment of medical images. So segmenting cells from some and these images. And it's been wonderful to see that explosion of work. On the language side, we also saw many people build up all sorts of different tools.