Building Connections Through Open Research: Meta’s Joelle Pineau
33:35 Share

Today, we're airing an episode produced by our friends at the Modern CTO Podcast, who were kind enough to have me on recently as a guest. We talked about the rise of generative AI, what it means to be successful with technology, and some considerations for leaders to think about as they shepherd technology implementation efforts. Find the Modern CTO Podcast on Apple Podcast, Spotify, or wherever you get your podcast.

Why is being open about your company's AI research a benefit more than a risk? Find out on today's episode. I'm Joelle Pinot from Meta, and you're listening to Me, Myself, and AI. Welcome to Me, Myself, and AI, a podcast on artificial intelligence and business. Each episode, we introduce you to someone innovating with AI. I'm Sam Ransbotham, professor of analytics at Boston College.

I'm also the AI and Business Strategy Guest Editor at MIT Sloan Management Review.

and I'm Sherwin Korubande, senior partner with BCG and one of the leaders of our AI business. Together, MIT SMR and BCG have been researching and publishing on AI since 2017, interviewing hundreds of practitioners and surveying thousands of companies on what it takes to build and to deploy and scale AI capabilities and really transform the way organizations operate.

Hi everyone. Today, Sam and I are happy to be joined by Joelle Pinault, Vice President of AI Research at Meta. Joelle, thanks for speaking with us today. Hello. Okay, let's jump in. A good place to start might be for you to tell us about Meta. A lot of people know what Meta is, but maybe you could describe it in your own words and also your role in the company.

Well, as many people know, Meta is in the business of offering various ways for people to connect and build community, build connections, whether it's through Facebook, WhatsApp, Instagram, Messenger. We have billions of people around the world using our products.

I've been at Meta for about seven years now, leading AI research teams. I'm based in Montreal, Canada, and now I lead FAIR, which is a fundamental AI research team across our labs in the US, in Europe. The role of our group is actually to build next generation AI systems and models, discover the new technology that will eventually make the products better, more engaging, safer as well.

That's a great overview. Can you give us a sense of what some of those projects are that you're excited about or that you're working on? You don't have to give us any secrets, of course, but what are some fun things you're excited about? Well, I hope we have a chance to talk about it, but there's not a ton of secrets because, in fact, most of the work that we do is all out in the open. We adhere strongly to open science principles. We publish our work. We share models, code libraries, and so on and so forth. Our teams cover the full spectrum of AI open problems.

So I have some teams who are working on understanding images and videos, building foundation models, so core models that represent visual information. I have some teams that are working on language models.

So understanding text, written, spoken language as well. I have some teams doing robotics. So understanding how AI systems move in the physical world, how they understand objects, people, interactions, and a big team of people who are working on core principles of intelligence. So how do we form memories? How do we actually build relationship between different concepts and ontology of knowledge and so on and so forth?

It seems like there's almost nothing within artificial intelligence you're not working on there. Tell us a bit about why you think open is important. So FAIR has been committed to open research for 10 years now, since day one. We've really pushed on this because whenever you start a project from the point of view of making it open, it really puts a very high bar in terms of the quality of the work, as well as in terms of the responsibility of the work.

And so when we decide what algorithms to build, what data sets to use, how to evaluate our data, how to evaluate the performance of our model through benchmarks, when we know that all of that work is going to be open for the world to scrutinize, it really pushes us to put a very, very high bar on the quality of that work, on the rigor of that work, and also on the aspects of responsibility, safety, privacy,

and so on. The other reason I think that open is really helpful is a lot of researchers come from a tradition of science where you're always building on the work of others. Science does not happen in a silo. And so if you're building on the work of others, there's also a desire to contribute back to that community. So our researchers are incredibly interested in having that kind of a culture. So it helps us recruit the best researchers, keep them. It is quite different from how other industry labs operate.

operate. And so from that point of view, I think it's definitely a big advantage. What's interesting is from the point of view of meta, there's no concern in terms of keeping some of that research internal in the sense that it doesn't in any way stop us from using this research into our products, not because we've published the results that we can't use it into our product, really the

Power of the product comes from all the people using it. It doesn't come from having a secret sauce about AI. And we know how fast AI is moving. A lot of that knowledge is disseminated across the community very, very quickly. And we are happy to contribute to that.

That makes a lot of sense. A lot of my background is in computer security. And so I think openness is a great segue there from security because of both of those points. First, in security, anybody can design something that they can't break. But the question is, can someone else break it? And I think that's always a more interesting and more difficult problem.

But then there's also the idea of building on others' work. I think that's huge. If you think about what's happened in research from the history of all of mankind, historically, research happened in academia, and then eventually more basic research became more applied within industry.

But it seems like with artificial intelligence that a lot of this has shifted to industry first. In fact, what you described to me sounds very much like an academic lab. So is that a problem that we're moving basic science from academia? Or are we? Maybe I'm begging the question. Is this a move that's happening? Is this a problem? What do you think?

Well, I think both academia and industry have advantages when it comes to AI research. And I'll maybe not speak broadly across all areas of research, but for AI,

In today's context, I do think both have significant advantage. On the one hand, on the industry side, we do have access to vast resources, in particular with respect to compute. And so when it comes to scaling some of the large language models, you know, you need access to thousands of GPUs, which is very expensive. It takes a strong engineering team to keep these running. And so it's a lot

more feasible to do this with a large industry lab. On the academic side, it's harder to have the resources and the personnel to be successful in terms of scaling large models.

Now, on the academic side, there are advantages. And I do have a position in academia. I have many colleagues. I have some grad students. So I think you have to be very smart about what research questions you ask, depending on your setting. On the academic side, we have the privilege of often working in highly multidisciplinary teams. I work with people who come from philosophy, cognitive science, linguistics, and so on and so forth. We ask much broader questions. And as a result, we

We come up with different answers.

One of the places where I sort of track the different contributions is looking at some of the very top conferences in the field and seeing, like, where do the outstanding paper awards go? Do they go to academia? Do they go to industry? And in many cases, we see a mix of both. There's some really seminal work coming out, both of industry and academia, that is completely changing the field, that is bringing forth some breakthroughs in AI. So I'm quite optimistic about that.

the ability for researchers across different types of organizations to contribute. And beyond that, we haven't talked about startups, but there's a number of small startups that are doing some really phenomenal work in this space as well. And so overall, having a thriving ecosystem is in everyone's advantage.

I think I'm more interested in a lot of our work in looking in ways that we can work together. Because in general, I strongly believe that having more diverse teams helps you ask different questions. So a lot of the intent behind our work on open sourcing is actually to make it easier for more diverse set of people to contribute.

You made the analogy with the security community really relying on open protocols. I think there's a lot of that in how we tackle this work from the sense of like, I have amazing researchers who are putting their best every day into building models, but I do believe by exposing these models to a broader community, we will learn a lot. So when I make the models available, you know, researchers in academia and startups take these models, in some cases find flaws with them, give some quick feedback, and

In many cases, we see derivatives of the model that have incredible value. One of the big launches we had in last year is our LAMA model. LAMA 1, LAMA 2, LAMA 3, thousands of people have built derivative models from these, many of them in academic lab, fine-tuning models, for example, to new languages to open up the technology to different groups. And to me, that's where a lot of the value of having different players really comes from.

I think we certainly see the value in, let's say, collaborating and iterating and keeping things open, but that's not always guaranteed to happen. What kind of incentives are there for us all to work together like this? It's always hard to predict the future, and in particular with AI and how fast things are moving. And so I completely agree with you, you know.

What I will say is, as you mentioned, there's a strong culture towards open protocols at Meta that predates the AI team. The basic stack, the basic software stack is also based on many open protocols. And so that culture is there to this day. That culture continues. It goes all the way to the top of the leadership.

And that commitment to open sourcing the models is strongly supported by Mark Zuckerberg and his leadership team. So I don't see that this is going to stop very soon. What is going to be important is that we continue to release models in a way that is safe. And that's a broader conversation than just one company. The governments have several points of views of...

How should we think about mitigating risks for this model? There's also a lot of discussions about how to deal in particular with very frontier models, the largest, most capable models. And so we're going to have to have these conversations as a society beyond just the labs themselves. You raise this specter of risks.

you know, the worry out there is that, oh my gosh, these models are going to take over everything and our world is going to collapse and this is an existential threat. I'm kind of setting you up with that straw man, but do you buy that?

I don't really spend a lot of time planning for the existential threat in the sense that many of these scenarios are very abstract. They're excellent, you know, stories in terms of science fiction. But in terms of actually taking a scientific and rigorous approach to that, it's

It's not necessarily the existential risks that take most of my attention. I will say with the current generation of models, there are several potential harms to different populations. You know, algorithms have been known to have biases towards underrepresented groups, for example, in facial detection system, as well as being on the language side, very anglocentric.

And so I do look quite carefully at the current set of risks and try to measure them as much as possible in a rigorous way. We build mitigations whenever we can. We've invented new techniques for doing watermarking to make sure that false information can't circulate. We've done a lot of work on bias assessment so that we can actually measure the fairness performance of our algorithm.

I look a lot more at current risks rather than these really long-term ones, just because I feel we can have a handle on it that is based on a rigorous approach, based on metrics, based on really analyzing what the harms is and how to mitigate them. The very far-fetched scenarios, it's really hypothetical. It's hard to build good systems. It's hard to do good science. It's also hard to do good policy. Yeah, I think your point's well taken about bias and metrics that

You mentioned, for example, these models that have biases built in, but I mean, my gosh, they're built off training data that has massive bias built in. I find it hard to attribute that to the model itself and more to the training data. And your point there is that you can build in bias mitigation there. What kinds of things have you done towards that?

Yeah, in fact, on the question of bias, it's a little bit of both. There's no doubt that many of our data sets are biased. The data sets are a reflection of our society. And unfortunately, a large amount of unfairness remains discrimination as well as having underrepresented groups in our society. So there's no question that the data sets themselves don't start us off on a very good foot.

However, the model themselves also tend to enhance these biases in that most of the machine learning techniques we have today, they're very good at interpolating the data. So you sort of take data distributed in a certain way and the models will really push towards the norm of that data. The models tend to be very poor at extrapolating. So making predictions outside of the data set, they tend to have a larger error. So if anything, when we train the models and we try to sort of minimize the error, we're

You do well by predicting more towards the norm versus towards the sides of that distribution. And so the data is responsible, the models are also responsible for doing that. And then there's the way in which we deploy the models. We tend to often look at aggregate statistics. So we'll look at the overall performance of the model and based on the overall performance, we'll say, great, we've got 95% performance on this model. It's ready to be deployed.

But we don't take the time to look at a more stratified analysis of results. What is the performance with respect to different groups? And how are these groups differentially impacted with respect to how the system is deployed in a bigger system? I think there's different points where we can be much more rigorous and thoughtful to make sure that we don't enhance biases. And ideally, that we actually use AI towards people

a more fair and equitable society. Yeah, I think that point of averaging is huge, that we've got so much

The models feel right when they give us the answer we're expecting. The image generation feels right when it gives us the image that fits our stereotypes. Yeah. And you're finding that it seems like it's a quite difficult problem. But on the other hand, I feel like these models can try to solve it in a way that we're not going to convince everyone in the world to suddenly stop being biased tomorrow or suddenly not have a stereotype tomorrow. But we could convince an algorithm not to have a stereotype tomorrow by

tweaking some weights and changing things. And so that gives me a little more hope to manage the risks. Perhaps it's not the existential threat we're getting there yet, but it seems more plausible to me that way. I think one of the challenges is determining what we want out of these models, right? We've seen some pretty egregious examples recently of groups which, I assume, is well-meaning intent to rebalance data sets, especially with representation of

For example, different racial groups in images. You know, of course, if someone asks for like an image of an engineer, you don't want only men to show up. You would hope to have a few women show up. And there's ways to rebalance the data. There's ways to sort of recompensate at the algorithmic level. But sometimes you end up with very unusual results.

And so it's also a question of what are the distribution of results that we expect and that we tolerate as a society. And in some cases, that's not very well defined, especially when the representation is biased within the real world as well.

That seems incredibly hard because the problem switches from being an engineering problem and engineering problems you can typically solve with enough pizza and caffeine. And when you get to these more difficult problems, then they tend to be trade-offs and they tend to be choices. And these choices are very difficult.

They're not improving an algorithm, which is the kind of thing that we can get into. But knowing what it should do seems like a much harder problem. And again, that seems much worse, too, as these technologies become so pervasive. You know, if, for example, Meta does make these algorithms available to people, you know, as part of the open source process, by definition, more people have access to them and then more people have to make these difficult decisions. That seems much harder to scale than algorithms.

I agree. I think in many ways, deciding as a society what we want these models to optimize for and how we want to use them is a very complicated question. There's also the reason why at Meta we often open source the research models. We don't necessarily open source the models that are running into production. That would open us up, I think, to undo attacks and

It's something we have to be careful about, but we often open our research models. And so that means very early on, if there are major opportunity to improve them, we learn much faster. And so that gives us a way to essentially make sure that by the time a model makes it into product, it's actually much better than the very first version. And we will release multiple versions as the research evolves, as we've seen, for example, with the LAMA language models I mentioned earlier. You know, we released LAMA 1, LAMA 2, LAMA 3.

and so on, in every generation gets significantly better. Some of that is, of course, the work of our own fabulous research teams, but some of that is also the contributions from the broader community. And these contributions come in different forms. You know, there's people who have better ways of mitigating, for example, safety risks. There are people who bring new data set that are allowing us to evaluate new capabilities.

And there's actually some very nice optimization tricks that allow us to train the models faster. And so all of that sort of converges to help make the models better over time. Yeah, I think the analogy that sticks with me is how image processing improved from the 2012 and ImageNet competition that, you know, again, that came out of originally academia, Toronto, and

but then exploded as everyone could see what everyone else was doing. Everyone brought something better, a faster implementation, a smaller implementation, a bigger, and the accuracy just over the very short time got really truly phenomenal. Yeah. Let's shift gears a little bit. Joelle, you're an AI researcher and also a professor. How did you find yourself in this line of work?

I'm very driven by curiosity, I have to say. I first got into robotics. That was sort of my gateway into AI. I was doing an undergrad degree in engineering at the University of Waterloo. And near the end of that, I had the chance to work on a robotics project, building a six-legged walking robot.

And in particular, the sensor system for that robot. So we had some sonars and had to process the information and from that decide sort of where were the obstacles in the environment. And so that led me to doing graduate studies, master's, PhD at Carnegie Mellon University in Pittsburgh, which is a phenomenal place to study robotics.

And from there, I really got into machine learning. I found that for the robot to have relevant, timely information and to be able to take decisions, you needed to have a strong model. So my thesis work was in planning under uncertainty, the ability to take decision when there's some uncertainty about the information and developing algorithms for doing that. And from then on, I took on an academic career at McGill University in Montreal, where I'm still based.

And pursuing work across areas of machine learning, a lot of applications of machine learning in healthcare. We have a fabulous Faculty of Medicine here at McGill. And so I had many very exciting partnerships there. And also a lot of work on building dialogue systems, which today, you know, we recognize as language models and chatbots. But I was building some of the very preliminary version of this work in the early 2000s.

2000s. And so because I do use curiosity as my main motor, it has allowed me to work across several subfields of AI, robotics, language, perception, and applications. And so that gave me a pretty good set of knowledge and experience to then come into a place like Meta, where

The teams that I work with do fundamental research, but we work closely with product teams and try to both push the frontier in terms of the science, but also push the frontier in terms of new products, new experiences. So clearly there's lots that Meta is doing around the core Meta products, but there's the general scientific discovery that Meta research is working on. What are some examples of projects that are in progress there?

This is such an interesting area. I think there's enormous potential to use AI to accelerate the scientific discovery process. When we think about how it works often, you know, let's say you're trying to discover a new molecule or discover new material. There's a very large space of solutions, often combinatorially large.

And the traditional methods have us looking through the space of molecules one by one, and we take them into the wet lab and we test them out for the properties that we want, whether it's to develop a new medication or develop a new material.

And so we've had a few projects over the years that look at this problem. More recently, we have a project that's looking specifically at direct air carbon capture. Really the desire to build new materials that could capture carbon in a way, of course, to address our environmental crisis. Now, when you do this work, there's many steps. One of them is even just building up the data set for doing that. So we've built up a data set, synthesizing many different materials

possibilities for this problem. And out of that, we often partner with external teams to try to validate which of these solutions may bring the most value.

We've done previous work also in the area of protein synthesis that had a similar flavor, though the specifications of the protein was a little bit different. But at a core fundamental way, the problem looks very similar. So I'm really excited to see what comes of this. I've had some cases where partner teams came to me and said, in the space of about a year of working with AI, they were able to cut down protein.

The scientific process in terms of experiments, that would have taken them like 25 years if they were going through the search space with more traditional methods. And I think that's something that we're seeing from other people we've talked to. We talked to, for example, Moderna talking about their vaccine development and how it helped explore that space. And we talked about Pirelli and how they

Use it for tire components. So I think this idea of exploring a combinatorically large space is really pretty fascinating. It's not something that I would have expected meta to be involved with it at first blush. I can see, for example, the carbon dioxide from the air problem. That's probably just something you're facing in data centers, but I wouldn't have expected that.

Yeah, I think, I mean, you bring up the case of data centers. I would say that's a prime application for this. We are building several data centers and it's in everyone's interest for those to be very energy efficient. We also have some strong commitments in terms of using renewable energy. And so there is a strong motivation in that space.

And not to be forgotten, we also have all of the work that's happening on our work towards the metaverse, the Reality Labs side of meta, which is really the longer term vision of building AR and VR devices.

When it comes to that type of hardware design, there's a lot of really hard problems, whether it's in the space of optics or other components, where AI-guided design can actually be very useful to accelerate that work. Yeah, that's pretty interesting. We actually just talked with Ty Sheridan, who is the star of the Ready Player One movie, and so that's a perfect segue from the metaverse to there. We have a segment where we ask you a little rapid-fire question. So just first thing that comes to your mind,

What's the biggest opportunity for artificial intelligence right now?

I do think that the ability to open up, to connect people across languages is huge. We've had systems where we're building up machine translation to go up to 200 different languages, but there are many more languages that are spoken only. And so we're really having the ability to build technology for anyone to understand anyone else across the planet. I think that's going to be really crucial for us to figure out how to all live together on this earth.

So what's the biggest misconception that people have about AI? I don't know if it's the biggest, but one that really gets to me is thinking of AI as a black box. People think, you know, information goes in, something happens, and then something comes out. I think in many ways, from where we stand today, the human brain is a lot more of a black box than AI. When I have an AI system, I can trace down with a lot of precision how information circulates, how it's calculated, and how we got to the output.

I cannot do that with a human brain in the same way. So, yeah, whenever someone says AI is a black box, I sort of frown a little bit and feel like, no, it's a complicated box. But we have a lot of understanding of what goes on inside there. Yeah, other people's brains make no sense to me. Mine makes perfect sense, but everyone else doesn't. What was the first career that you wanted to do?

Oh, early on, I wanted to be a librarian. I loved reading books. I still do. I still read quite a bit. And I thought, you know, with having a job where you can just sit in a space filled with books and read all day sounded delightful. When do we have too much artificial intelligence? When are we trying to put that square peg in a round hole?

I don't think of it as like one day we have enough and one day we have too much. I think it's really about being smart about where you bring in AI into a system. So already there are places where AI shouldn't go and there are places, or at least the version of the models we have today, and there are places where we could bring in AI much more aggressively, I think. So I think what's really important is figuring out

how to bring it in in a way that it brings real value, economic value, of course, but real social value as well and being thoughtful about that. Yeah, that ties to your previous answer about the difficult parts of using the technology or not sticking to the technology itself. So what's one thing that you wish that artificial intelligence could do now that it can't do currently?

I wish the AI systems could understand each other. Right now, we're building a lot of AI systems that are individual. They're all fine-tuned for an individual performance. But once we start deploying many of these AI agents together in the same place, our methods for understanding the dynamics between several agents are very primitive.

And I think there's a ton of work to do. You know, if we look to humans as the society of agents that is most evolved today, we derive a lot of our longevity, our robustness, our success through our social ties. And AI systems today have no idea how to build social ties. That's interesting because I think we spend so much time thinking about the human-computer interface and the computer-human interface and not as much about the computer-computer interface.

This has been a fascinating discussion. I really kind of opened my eyes to all the things that META is doing that's beyond just that sort of surface research that's more obvious in the newspapers and media reports. Thanks for taking the time to talk with us today. Yes, very inspiring conversation. Thank you. My pleasure. Happy to be with you.

Thanks for listening to season nine of Me, Myself, and AI. Our show will be back in September with more episodes. In the meantime, if you missed any of the earlier episodes on Responsible AI, please go back and have a listen. We talk with Amnesty International, Airbnb, and Salesforce. Thank you for listening and also for reviewing our show. Talk to you soon.