Meta's Chief AI Scientist Yann LeCun Makes the Case for Open Source | On With Kara Swisher

27:13Are you worried about being outspent? To get me a smarter assistant doesn't seem to be a great business, but I don't know. I didn't take the job at Facebook when I was offered it in the early days, so don't ask me. But go ahead. Well, it's a long-term investment. I mean, you need the infrastructure to be able to run those AI assistants at reasonable speed for a growing number of people. As you said, there are 600 million people using Meta AI right now. By the way, there's another interesting number. The open source engine, LAMA, on top of which MetaEA is built, but which is open source, has been downloaded 650 million times.

27:56That's an astonishing number. I don't know who are all these people, by the way, but that's an astonishing number. There are 85 ,000 projects that have been derived from LAMA that are publicly available all open source, mostly in parts of the world. A lot of those projects are basically training Lama, for example, to speak a bunch of languages from Senegal or from India. So you don't think this money is ill-spent? No, I don't think so because there's going to be a very large population who will use those AI systems on a daily basis within a year or two and then growing. And then those systems are more useful if they're more powerful.

28:37And the more powerful they are, the more expensive they are computationally. So this investment is investment in infrastructure. In infrastructure, what's happening by private companies. Now, you said the concentration of proprietary AA models in the hands of just a few companies was a huge danger, obviously. There's also been critics of the open source model. They worry about bad actors could use them to spread misinformation, cyber warfare, bioterrorism. Talk about the difference. Does META have a role in preventing that happening, given you're handing these tools, these powerful tools, in an open source method?

29:14Okay, so this was a huge debate. It was. You know, just fairly recently, you know, early 2023, when we started distributing LAMA. The first LAMA was not open source. You had to ask permission and you had to show that you were a researcher. And it's because, you know, the legal landscape was uncertain. and we didn't know what people were going to do with it. So it wasn't open source. But then all of us at Meta received a lot of requests from industry saying, like, you have to open source the next version because this is going to create a whole industry. It's going to enable, like, a lot of, you know, startups and kind of new products and new things.

29:56And so we had a big internal discussion for several months internally, a weekly discussion, two hours, with 40 people from Mark Zuckerberg down. Okay, very serious discussions about this, about safety, about legal landscape, about all kinds of questions. And then at some point, the decision was made by Mark to say, okay, we're going to open source Lama 2, tell me how to do it. And that was done in kind of summer 2023. And since then, it's basically completely gem-started a whole industry. Why is it more safe than these proprietary models that are controlled by the companies? Because there are more eyeballs on it.

30:35And so there are more people kind of fine-tuning them for all kinds of things. And so there was a question as to, you know, maybe a lot of badly intentioned people will put their hands on it and then will use them for nefarious purpose. Well, Chinese researchers developed an AI model for military use with an older version of Mattis Llama model as a backbone. It's actually kind of a very kind of minor bad things, and he could have used one of the many excellent open source Chinese models. There's one called Quen. That's really good, which is on par with, you know, the best. So, I mean, the Chinese have good research, good engineers.

31:13They open source a lot of their own models. You know, this is not. So you don't think that's Meta's responsibility? You put it out there, the tools, and then what people do with it. No, it is to some extent, of course. So there is a big effort in the LAMA team, in the GNI organization, to red team all the systems that we put out so that we ensure that they, you know, are, at least when they come out, are minimally toxic and things like that, right, and mostly safe. That's a really important effort, actually. We even initially gave Lama to a bunch of hackers at DEF CON and sort of asked them, like, try to do something bad with it.

31:59And the result is we haven't been aware of anything really bad done with any of the models that we've been distributing over the last almost two years. Yet would be the word I would put that behind. Well, yet, but, you know, it would have happened already. I mean, there have been, you know, the public doesn't realize is because they think it just appeared with JGPT, but there have been LLMs, open-source LLMs, available for many years before that. And I don't know if you remember this, but when OpenAI came up with JGPT 2, they said, oh, we're not going to open-source it because, you know, it's very dangerous.

32:37So people could do really bad things. You know, they could flood the Internet with disinformation and blah, blah, blah. So we're not going to open-source it. I made fun of them because, I mean, And it was kind of ridiculous at the time. The capability of the system really was not that bad. And so, I mean, you have to accept the fact that those things have been available for several years and nothing really bad has happened. There was some, you know, a bit of worry that people would use this for disinformation, you know, in the run-up of the elections in the U.S. And all kinds of things like this, you know, cyber attacks and things.

33:13None of that really has happened. It's still good to be worried about such things. Well, I mean, you have to be, you know, watchful and do what you can to prevent those things from happening. The point is, you know, you don't need any of those AI systems for disseminating disinformation, as Twitter has shown us. Okay. Good. There. Good. I like how you get your little digs in. I'm watching it very carefully. You did an Elon one, the secretive drama queens of open AI. I got that. So you also get a lot of flack online recently for saying that cultural institutions, libraries, foundations should make their content available for training by free and open AI foundation models like LAMA, presumably.

33:50You're responding to a new data set that Harvard released made up over a million books. But those are public domain works, not works by living authors, artists, academics. Talk about the concerns and the flack you got about these AO models vacuuming up all of our cultural knowledge from the creators, writers, researchers without getting any credit. I mean, internet companies are known for scraping. I think Walt called, I believe it was when it used to be called Facebook, rapacious information thieves. But he may have been talking about Google. So talk to me about that, the controversy that happened with that.

34:24Okay. outside of all of those kind of legal questions. If you have this vision that, you know, AI is going to be the repository of all human knowledge, then all human knowledge has to be available to train those models, right? And most of it is either not digitized or digitized but not available publicly. And it's not necessarily copyrighted material. It could be, you know, the entire content of the French National Library, a lot of which is digitized but not available for training. So I'm not, this is, I was not necessarily talking about copyrighted work in that case. It's more like, you know, if you are in, so I'm from, my family, my father's family is from Brittany, okay, the western part of France, right?

35:10The traditional language spoken there, which was spoken until my great-grandfather, is Breton. Breton is disappearing. There is something like 30 ,000 people speaking it on a daily basis, which is very small. If you want future LLMs to speak Breton, there needs to be enough training data in Breton. Where are you going to get that? You're going to have cultural nonprofits, you know, kind of collecting all the stuff that they have, maybe governments helping, things like that. And they're going to say, like, you know, use my data. Like, I want your system to speak Breton. Now, they may not want to just hand that data just like that to, you know, big companies on the West Coast of the U.S.

35:51But a future that I envision, this is not company policy, this is my view, is that the best way to get to that level is by kind of training an AI system, a common AI system, repository of human knowledge, in a distributed fashion. So that there would be several data centers around the world using local data to contribute to training a global system. You don't have to copy the data. Who runs that global system? Who writes Linux? Okay. Right. So that should exist for all of humanity. Yeah. I mean, who pays for Wikipedia? Right. I pay$7 a month, but go ahead. Good idea. Or the Internet Archive, right?

36:36Yeah. So for Linux, in the case, actually, Linux is mostly supported by employees of companies who tell them, you know, to actually distribute their contributions. And you can have kind of a similar system where, you know, everyone contributes to this kind of global model. That's AI for everybody else. Which is AI for, you know, LLMs in the short term. For things that aren't necessarily monetizable. Yeah. Well, you monetize on top of it, right? I mean, Linux, you don't pay for Linux. But if you buy a widget that, you know, runs Linux like an Android phone or a car that has, you know, Linux in its touchscreen, you pay for the widget that you buy.

37:17So it's going to be the same thing with AI. The basic foundation model is going to be open and free. It does feel like that it's a coalescing of small amount of powers running everything. It does at this point. And that vision is a lovely one, but it's not occurring. Well, it's, my opinion, it's actually inevitable. You've been in a public debate. You like to debate with other godfathers of AI. Your Turing Award co-winners, Jeffrey Hinton and I think it's Joshua Bengio. Yep. They both have been ringing alarm bells, warning about the potential dangers of AI. quite dramatically, I would say. They've called for stricter government regulation, oversight, including R &D.

37:54You've called their warnings complete BS. I don't think you missed words there. Talk to me about why that's complete BS. And one of the things you disagreed was one of the first attempts at AI regulation here in the U.S., California bill SB 1047. Hinton and Bengio both endorsed it. You lobbied against it. You wrote, regulating R &D would have been apocalyptic telephonic consequences on the AI system. Very dramatic of you, sir. You said the illusion of existential risk is being pushed by a handful of, quote, delusional think tanks. These two aren't delusional, I don't believe. Hinton just won the Nobel Prize for his work.

38:31Talk about that in particular. And by the way, Governor Newsom vetoed the bill, but is working with people like Stanford Professor Faye Faye to overhaul it. Talk about why you called it complete BS. You're very strong on this one. I'm very vocal about that, yes. Yeah. So Jeff and Yoshua are both good friends. We've been friends for decades. Yeah. You know, I did my postdoc in 1987, 88 with Jeff Hinton. So we've known each other for a very long time, for 40 years now. Same with Yoshua. I knew, I met him the first time. He was a master's student and I was a postdoc. So we've been kind of working together.

39:09We won this prize together because we work together at sort of reviving interest in what we now call deep learning, which is a root of a lot of AI technology today. So we agree on many things. We disagree on a few things, and that's one of them. The existential threat to the human race. The existential threat, yeah. Exactly. So Jeff... You're like, ah, no. They're like, oh, yeah, they're coming for us. I mean, Jeff believes that current LLMs have subjective experience. I completely disagree with this. I think he's completely wrong about that. We disagreed on technical things before. It was kind of less public.

39:46It was more kind of technical. But it's not the first time we disagree. I just think he's wrong. We're still good friends. Yoshua comes from a slightly different point of view. He's more worried. He's worried a little bit about this, but he's more worried about bad people doing bad things with AI systems. Yeah, I'm with him. Developing like bioweapons or chemical weapons or things like this. I think, frankly, those dangers have been formulated for several years now and have been incredibly inflated to the point of being kind of distorted so much that really they don't make any sense. Yes, delusional is the word you use.

40:27Well, I don't call them delusional. I call some of the other people who are more extreme and are pushing for regulation like SB 1047, yes, delusional. I mean, some people will tell you in the face, you know, a year ago, you ask them, like, how long is it going to take for AI to kill us all? And they say, like, five months. And obviously they were wrong. So this is what you're talking about. It's over AGI, artificial general intelligence, and how close we are. I would like you to explain it for people. When they hear it, they think about the plot of Terminator or iRobot or something like that. So Hinton and Bengio think the timeline for AGI could be more like five years and that we are not prepared.

41:08You said several years, if not a decade. You know, if you're wrong, you're going to be real wrong when it does kill us. So talk about why, you know, you'll be like, oh, we're not dead yet, and then we're dead. So talk about why you're not worried. So first of all, there's no question that at some point in the future, we're going to have AI systems that are smarter than us. Okay? It's going to happen. Is it five years, 10 years, 20 years? It's really hard to tell. In our kind of, or at least my personal vision of it, the earliest it could happen is about five years, six years, but probably more like 10 and probably longer because it's probably harder than we think, and it's almost always harder than we think.

41:50There's this history over, you know, the several decades of AI of people sort of completely underestimating how it is. And again, you know, we don't have domestic robots. We don't have, you know, level five side-driving cars. There's a lot of things that we don't know how to do with AI systems today. And so until we figure out kind of a new set of techniques to get there, we're not even on a path towards human-level intelligence. So once we, you know, a few years from now, once we have kind of a blueprint and some kind of believable demonstration that we might have a path towards human-level AI, I don't like to call it AGI because human intelligence is very specialized, actually.

42:31So we think we have general intelligence. We don't. So once we have a blueprint, we're going to have a good way to think about how to make it safe. It's kind of like, you know, if you kind of backpedal to the 1920s and someone is telling you, you know, in a few decades, We're going to be flying millions of people across the Atlantic and near the speed of sound. And someone would say, oh my God, are you going to make this safe? The turbojet was not invented yet. How can you make turbojets safe if you haven't invented the turbojet? We are in this situation today. So making AI safe means designing those AI systems in ways that are safe.

43:16But until we have it designed, we're not going to be able to make them safe. So the question makes no sense. You don't seem worried that AI would ever want to dominate humans. You've said this. You've said that current AI is dumber than a house cat. Whether AI is sent or not doesn't seem to matter if it feels real, right? And so how do you, if it's dumb or it doesn't want to dominate us or it doesn't want to kill us, what would be restrictions on AI and maybe AI R &D that you would seem reasonable, if any? I think if none is what you're saying to me. Well, none on R &D. I mean, clearly, if you want to put out a domestic robot and that robot can cook for you, you probably want to hardwire some rules so that when there's people around the robot and the robot has a knife in his hand, he's not going to shut his arm around or something.

44:05So, you know, those are guardrails. So, the design of current AI systems, to some extent, is interestingly unsafe. You could say it this way. A lot of people I admit are going to hate me for saying this, but they're kind of hard to control. You basically have to train them to behave properly. What you want, and this is something I've proposed, is another type of architecture which I call objective-driven, where the AI system basically is there to fulfill an objective and cannot do anything but fulfill this objective, subject to a number of guardrails, which are just other objectives. and that will guarantee that whatever output the system produces, whatever action it takes, satisfy those guardrails and objectives and are safe.

44:56Now, the next question is how do we design those objectives? And a lot of people are saying, oh, we've never done this before. This is completely new. We're going to have to kind of invent a new science. No, actually, we're pretty familiar with this. It's called making laws. We do this with people. We establish laws, and the laws basically change the cost of taking actions, right? And so we've been shaping the behavior of people by making laws. And we're going to do the same for AI systems. The difference is that people can choose to not respect the law, whereas the AI system by construction will have to.

45:31Now, both these people, Hinton and Benji, endorsed a letter signed by current and former OpenAI employees calling employees that AI companies have the right to warn about serious risks by the technologies and ordinary whistleblowers wouldn't protect them. You didn't endorse it. At the same time, we've seen some regulation in the EU. They differentiate between high-risk AI systems and more general-purpose models. They have bans on certain applications that, quote, threaten citizens' rights, facial images. I suppose this robot who wants to knife you. What is the model here to make it safer, to make people – you're suggesting we wait and see when bad things happen before putting up guardrails.

46:10Let's wait until there's some murder happening or not. I can't tell. No, no, that's not what I'm suggesting. I mean, you know, measures like, you know, banning massive, you know, face recognition in public places, that's a good thing. Like, you know, nobody would really think that's a bad thing except if you are an authoritarian government. Some people think it's a great thing. Yeah, it already exists in some countries, actually. But that's a good thing, right? And there are measures like this that make complete sense, but they are at the product level.

46:45Also, changing the face of someone on some embarrassing video and stuff like that, I mean, it's kind of already legal, more or less. The fact that we have the tools to do it doesn't make it less illegal. There may be a need for specific rules against that, But I have no problem with that. You know, I have a problem with this idea that, you know, AI is intrinsically dangerous and you need to regulate R &D. And the reason I think it's counterproductive is, you know, in a future in which you would have those open source platforms I was talking about, which I think are necessary for things like democracy in the future, then those rules would be counterproductive.

47:26They would basically make open source too risky for any company to distribute. And so we kind of kill. So that these private companies will control everything. That's right. A small number of private companies on the west coast of the U.S. would control everything. Now, talk to any government outside the U.S. And tell them about this future where everyone's digital diet would be mediated by AI assistance. And tell them that this will come from three companies on the west coast of the U.S. And they say, like, that's completely unacceptable. Right. Like, this is the death of a democracy. Like, how will people get a diversity of opinions, right, if it all comes from three companies on the West Coast of the U.S.?

48:04We all have the same culture. We all speak the same language. Like, this is completely unacceptable. So what they want are open platforms that then can be fine-tuned for any culture, value system, center of interest, whatever, so that users around the world have a choice. They don't have to use like three assistants. They can use, you know, a lot of them. So you're worried about domination by OpenAI, Microsoft, Google, possibly Amazon. Anthropic. Anthropic, which is Amazon, really. So last two questions. You were awarded the 2024 VIN Future Prize. There's so many prizes in your area. I never get any prizes.

48:43For transformational contributions to deep learning. In your acceptance speech, you said AI does not learn like humans or animals, which take in a massive amount of visual observation from the physical world. But you've been working to make this happen. You've been talking about it a while. Where do you imagine it being in years? Will it be like humans or animals or where? Well, so, yeah, I mean, there is a point at which we're going to have systems that learn a little bit like humans and animals and can learn new skills and new tasks as efficiently as humans and animals, which is frankly astonishingly fast.

49:19Like we can't reproduce this with machines, right? We have companies like Tesla and others have hundreds of thousands or millions of hours of cars being driven by people. They could use this to train AI systems, which they do. They're still not as good as humans. We can't buy a car that actually drives itself or a robotaxi unless we cheat. Waymo can do it, but there's a lot of tricks to it. And again, we can't buy a domestic robot because we can't make them smart enough. The reason for this is very simple. As I said before, we train LLMs and chatbots on all the publicly available text and some more.

49:59That's about 20 trillion words, right? A four-year-old has seen essentially the same amount of data visually than the biggest LLM has seen through text. That text would take any of us several hundred thousand years to read through. Okay, so what that tells you is that we're never going to get to human-level AI by just training on text. We have to train on sensory input, which is basically an unlimited supply. 16 ,000 hours of video is 30 minutes of YouTube uploads. Okay? We have way more video data than we know what to do with. So the big challenge for the next few years in AI to make progress to the next level is get systems to understand how the world works by basically watching the world go by, watching video, and then interacting in the world.

50:50And this is not sold, but, you know, there's a good chance that progress will be made, like significant progress will be made over the next five years, which is why you see all of those companies starting to build human-based robots. They can't make them smart enough yet, but they're counting on the fact that AI is going to make sufficient progress over the next five years, that by the time that those things can be sold in the public, that they, you know, the AI will be powerful enough. Right. Now I'm getting the glasses. I understand what you're up to now, finally. I actually believe in a four-year-old more than I believe in most of Silicon Valley, I'll be honest with you.

51:23I met people like you, as I was saying. This is my very last question and very quick because we've got to go. Who would like this? It's going to change learning. It's going to change this. It's going to make it everything better. Everyone's going to get along. And as you cite all the time, and I respect you for that, is there's hate. There's dysfunction. There's loneliness. Self-esteem among girls. Danger to people who are often in danger. controlled by billionaires of our government. Why do I trust you this time? Me? You. Just you. Okay. I'm not a billionaire. What? I'm not a billionaire. That's not the first thing.

52:03I'm doing okay, though. I'm guessing you are. Okay. I'm first and foremost a scientist. and I would not sort of, you know, be able to look at myself in the mirror unless I had some level of integrity, scientific integrity at least. I might be wrong. So you can trust that I'm not lying to you and that I'm not, you know, motivated by nefarious motives like, you know, greed or something like this. But I might be wrong. I might very well be wrong. In fact, that's kind of the whole process of science is that you have to accept the fact that you might be wrong and elaborating the correct ideas comes from the collision of multiple ideas and people who disagree.

52:58But look at the evidence. So we look at the evidence from the people who said that AI was going to destroy society because we're going to be inundated with disinformation or generated hate speech or things like this. We're just not seeing this at all. We're not seeing it. We've not seen it. I mean, people produce hate speech. People produce disinformation. And they try to disseminate it, you know, every way they can. A lot of people are trying to disseminate hate speech on Facebook. And it's against the content policy at Facebook to do this. Now, the best protection we have against this is AI systems.

53:39We couldn't do this in 2017, for example. 2017 AI technology was not good enough to allow Facebook and Instagram to detect hate speech in every language in the world. And what happened in between is progress in AI. So AI is not the tool that people use to produce hate speech or disinformation or whatever. It's actually the best countermeasure against it. So what you need is just more powerful AI in the hands of the good guys than in the hands of the bad guys. I'm worried about the bad guys, but that's a great answer. Thank you so much. Thank you. I really appreciate it.

54:23On with Kara Swisher is produced by Christian Castro-Russell, Kateri Yoakum, Jolie Myers, Megan Burney, and Kaylin Lynch. Nishat Kerwa is Vox Media's executive producer of audio. Special thanks to Corinne Ruff and Kate Furby. Our engineers are Rick Kwan, Fernando Arruda, and Aaliyah Jackson. And our theme music is by Trackademics. If you're already following the show, you get a free pair of Meta Glasses. If not, watch out for that stabby robot. Go wherever you listen to podcasts, search for On with Kara Swisher, and hit follow. Thanks for listening to On with Kara Swisher from New York Magazine, the Vox Media Podcast Network, and us.

54:59We'll be back on Monday with more.