Jan Lakoon, the head of Meta's AI division, just did a three-hour interview with Lex Fridman in which he covered robots, AGI, and so much more.

I went through the whole interview, I grabbed the bits that I thought were most interesting, and I encourage you to watch the whole video.

I'll drop a link to that in the description below.

In this first clip, Jan talks about how autoregressive LLMs, basically the Large Language Models that we all know today, like OpenAI's ChatGPT, won't get us to AGI.

Jan is a huge proponent of open source, but he doesn't think our current technology will allow us to achieve AGI.

Let's watch.

For a number of reasons.

The first is that there is a number of characteristics of intelligent behavior.

For example, the capacity to understand the world, understand the physical world, the ability to remember and retrieve things, persistent memory, the ability to reason, and the ability to plan.

Those are four essential characteristics of intelligent systems or entities, humans, animals.

LLMs can do none of those, or they can only do them in a very primitive way.

They don't really understand the physical world, they don't really have persistent memory, they can't really reason, and they certainly can't plan.

If you expect a system to become intelligent just without having the possibility of doing those things, you're making a mistake.

And then next, following up on the premise that our current technologies won't get us to AGI, Jan talks about the amount of data necessary to train Large Language Models as compared to what humans use to essentially train our own minds.

And he says that the amount of data used to train Large Language Models is actually not that much compared to what humans need.

And this is where we start to get into synthetic data.

And I've been preparing my thoughts about synthetic data quite a bit because the fact of the matter is, I don't think humans are going to produce enough data to actually reach AGI using Large Language Models, even if our current architecture could allow us to get there.

Synthetic data, meaning data that has been created by other artificial intelligence, is one of those necessary ingredients to reaching AGI.

Let's take a look at this video.

Those LLMs are trained on enormous amounts of text, basically the entirety of all publicly available text on the internet.

That's typically on the order of 10 to the 13 tokens.

Each token is typically two bytes.

That's two 10 to the 13 bytes as training data.

It would take you or me 170,000 years to just read through this at eight hours a day.

It seems like an enormous amount of knowledge that those systems can accumulate.

But then you realize it's really not that much data.

If you talk to developmental psychologists and they tell you a four-year-old has been awake for 16,000 hours in his or her life and the amount of information that has reached the visual cortex of that child in four years, it's about 10 to the 15 bytes.

And you can compute this by estimating that the optical nerve carry about 20 megabytes per second, roughly.

10 to the 15 bytes for a four-year-old versus two times 10 to the 13 bytes for 170,000 years worth of reading, what that tells you is that through sensory input, we see a lot more information than we do through language.

And that despite our intuition, most of what we learn and most of our knowledge is through our observation and interaction with the real world, not through language.

Everything that we learn in the first few years of life and certainly everything that animals learn has nothing to do with language.

And then again, continuing with the Large Language Models are not going to allow us to reach AGI.

He basically says that just using language is not enough to model the world, to have a world model.

Even though language is highly compressed, he describes how when people are talking and thinking about things, they're using much more than just language.

We have a world model inside our head and we're conceptualizing ideas and perceptions before we even convert those into language.

It's very interesting to think about.

But because of this, he says, we're not going to reach AGI with language alone.

We need some other kind of technology to help us augment the language part.

Let's look at that video.

It's a big debate among philosophers and also cognitive scientists, like whether intelligence needs to be grounded in reality.

I'm clearly in the camp that, yes, intelligence cannot appear without some grounding in some reality.

It doesn't need to be physical reality, it could be simulated, but the environment is just much richer than what you can express in language.

Language is a very approximate representation of our percepts and our mental models, right?

I mean, there's a lot of tasks that we accomplish where we manipulate a mental model of the situation at hand, and that has nothing to do with language.

And next, kind of proving the point that Large Language Models in their current form aren't going to reach AGI, Jan goes on to describe why Large Language Models are really good at some things like creative writing, like coding, passing the bar exam and other things like that, but really bad at other things like self-driving and the ability to simply pick up a cup.

And for some reason, Large Language Models are really good at some things and really bad at others.

Let's take a look.

You know, we have LLMs that can pass the bar exam, so they must be smart, but then they can't learn to drive in 20 hours like any 17-year-old.

They can't learn to clear out the dinner table and fill out the dishwasher like any 10-year-old can learn in one shot.

Why is that?

Like what are we missing?

What type of learning or reasoning architecture or whatever are we missing that basically prevent us from having level five self-driving cars and domestic robots?

And then finally, kind of closing the loop on AGI and what's possible with our current Large Language Model architecture, he goes on to talk about how kind of world model reasoning is out of reach for Large Language Models as they are today.

He continues to explain that when humans are talking about things and conceptualizing things in the real world, we're doing so much more than just speaking in our minds.

He's really doubling down on this point, which I find fascinating.

Let's take a look at this clip.

We're not going to be able to do this with the type of LLMs that we are working with today.

And there's a number of reasons for this.

But the main reason is the way LLMs are trained is that you take a piece of text, remove some of the words in that text, you mask them, you replace them by blank markers, and you train a genetic neural net to predict the words that are missing.

And if you build this neural net in a particular way so that it can only look at words that are to the left of the one it's trying to predict, then what you have is a system that basically is trained to predict the next word in a text, right?

Then you can feed it a text, a prompt, and you can ask it to predict the next word.

It can never predict the next word exactly.

What it's going to do is produce a probability distribution of all the possible words in your dictionary.

In fact, it doesn't predict words, it predicts tokens that are kind of subword units.

It's easy to handle the uncertainty in the prediction there because there is only a finite number of possible words in the dictionary, and you can just compute the distribution over them.

Then what the system does is that it picks a word from that distribution.

Of course, there's a higher chance of picking words that have a higher probability within that distribution. So you sample from that distribution to actually produce a word.

And then you shift that word into the input.

That allows the system not to predict the second word.

And once you do this, you shift it into the input, etc.

That's called autoregressive prediction, which is why those LLMs should be called autoregressive LLMs.

There is a difference between this kind of process and a process by which before producing a word, when you talk, when you and I talk, you and I are bilinguals, we think about what we're going to say, and it's relatively independent of the language in which we're going to say.

When we talk about, I don't know, let's say a mathematical concept or something, the kind of thinking that we're doing and the answer that we're planning to produce is not linked to whether we're going to see it in French or Russian or English.

And next we continue our discussion about language as a world model.

Jan is in one camp that believes language alone cannot be a world model.

And I want to know what you think.

Do you agree or disagree?

When you see something like Sora, which uses transformers, which uses current architecture that is found in Large Language Models, it is pretty darn good at simulating the world.

But it's not perfect.

And if it's ever going to get there is still a question.

And he continues to hammer the point that Large Language Models are not enough to reach a world model.

But this actually reminds me of a couple things.

There's the movie Arrival.

And if you haven't seen it, it's a fantastic science fiction movie.

But the gist is aliens come down from space and they speak a different language and they have an expert linguist try to Figure out what that language is.

And over time, they Figure it out.

And the language itself is what the aliens are there to give humanity as kind of the great gift.

And all of a sudden, because this expert linguist now understands this new language, she starts to gain all of this power.

It's super interesting to think about that language actually shapes perception.

And that is something I believe in.

If we don't have a word for something, then most likely we can't describe it and maybe we can't even perceive it as it really is.

I believe language is incredibly important, but I'll probably defer to Jan thinking that language alone is not enough to have a world model.

You're saying your thinking is same in French as it is in English.

Well, it depends what kind of thinking, right?

If it's like producing puns, I get much better in French than English about that.

No, but is there an abstract representation of puns?

Is your humor an abstract representation?

Like when you tweet and your tweets are sometimes a little bit spicy, is there an abstract representation in your brain of a tweet before it maps onto English?

There is an abstract representation of imagining the reaction of a reader to that text.

You start with laughter

And then Figure out how to make that happen?

Or Figure out a reaction you want to cause and then Figure out how to say it so that it causes that reaction.

But that's really close to language.

But think about a mathematical concept or imagining something you want to build out of wood or something like this.

The kind of thinking you're doing is absolutely nothing to do with language, really.

It's not like you have necessarily an internal monologue in any particular language.

You're imagining mental models of the thing, right?

If I ask you to imagine what this water bottle will look like if I rotate it 90 degrees, that has nothing to do with language.

Clearly there is a more abstract level of representation in which we do most of our thinking and we plan what we're going to say.

If the output is uttered words as opposed to an output being muscle actions, we plan our answer before we produce it.

And LLMs don't do that.

They just produce one word after the other, instinctively, if you want.

And in this next clip, Jan talks about if we're able to have a world model based on prediction alone.

And again, the way that Large Language Models work is that it's essentially just predicting the next token in a sentence, the next word in a sentence.

And he actually says, yes, just based on prediction alone, that part of the technology, we will be able to achieve AGI.

We will be able to achieve a world model.

But just based on language prediction alone, he doesn't think so.

And this starts to get into simulation theory.

Do you need to understand every single atom?

Do you need to predict how every single atom is going to move to truly create a world model?

It turns out maybe not.

And this is actually something a Twitter user had replied to me, which I had never really thought about.

With simulation theory, I always thought that you had to predict every single atom, but that's not necessarily true.

You only have to predict enough that the perceiver, the person who is viewing this reality is convinced that it looks real.

Just enough, but not everything.

Let's take a look at this clip.

Can you build this, first of all, by prediction?

And the answer is probably yes.

Can you build it by predicting words?

And the answer is most probably no, because language is very poor in terms of weak or low bandwidth, if you want.

There's just not enough information there.

Building world models means observing the world and understanding why the world is evolving the way it is.

And then the extra component of a world model is something that can predict how the world is going to evolve as a consequence of an action you might take.

What model really is, here is my idea of the state of the world at time t, here is an action I might take, what is the predicted state of the world at time t plus one.

Now that state of the world does not need to represent everything about the world.

It just needs to represent enough that's relevant for this planning of the action, but not necessarily all the details.

Here is the problem.

You're not going to be able to do this with generative models.

A generative model is trained on video, and we've tried to do this for 10 years.

You take a video, show a system a piece of video.

And then ask it to predict the reminder of the video.

Basically predict what's going to happen, either one frame at a time or a group of frames at a time.

But yeah, a large video model if you want.

The idea of doing this has been floating around for a long time, and at FAIR, colleagues and I have been trying to do this for about 10 years.

And you can't really do the same trick as with LLMs, because LLMs, as I said, you can't predict exactly which word is going to follow a sequence of words, but you can predict the distribution of the words.

If you go to video, what you would have to do is predict the distribution of all possible frames in a video, and we don't really know how to do that properly.

We do not know how to represent distributions over high dimensional continuous spaces in ways that are useful.

And there lies the main issue, and the reason we can do this is because the world is incredibly more complicated and richer in terms of information than text.

Text is discrete, video is high dimensional and continuous, a lot of details in this.

Okay, in this next clip, he breaks down why video prediction, basically frame by frame prediction, or even an entire video prediction, is really hard to do.

And this interview came out after Sora.

As you're listening to this, think about what Sora was able to achieve and compare it to what Jan is saying here.

If I take a video of this room, and the video is a camera panning around, there is no way I can predict everything that's going to be in the room as I pan around.

The system cannot predict what's going to be in the room as the camera is panning.

Maybe it's going to predict this is this is a room where there is a light and there is a wall and things like that, it can't predict what the painting of the world looks like or what the texture of the couch looks like, certainly not the texture of the carpet.

There's no way I can predict all those details.

One way possibly to handle this, which we've been working for a long time, is to have a model that has what's called a latent variable, and the latent variable is fed to a neural net.

And it's supposed to represent all the information about the world that you don't perceive yet, that you need to augment the system for the prediction to do a good job at predicting pixels, including the fine texture of the carpet and the couch and the painting on the wall.

That has been a complete failure.

And we've tried lots of things.

We tried just straight neural nets, we tried GANs, we tried VAEs, all kinds of regularized auto encoders.

We tried many things.

And in this next clip, he goes into the details about why video prediction doesn't work with our current architecture, our current technology.

And of course, Jan recently released JEPA, which is what he believes is going to be able to augment Large Language Models and transformers to be able to allow to actually have a world model with video.

Let's see what he says about that.

The reason this doesn't work, first of all, I have to tell you exactly what doesn't work because there is something else that does work.

The thing that does not work is training the system to learn representations of images by training it to reconstruct a good image from a corrupted version of it.

That's what doesn't work.

And we have a whole slew of techniques for this that are a variant of denoising auto encoders.

Something called MAE developed by some of my colleagues at FAIR, masked auto encoder.

It's basically like the LLMs or things like this where you train the system by corrupting text except you corrupt images.

You remove patches from it and you train a gigantic neural net to reconstruct.

The features you get are not good.

And you know they're not good because if you now train the same architecture but you train it supervised with label data, with textual descriptions of images, you do get good representations.

And the performance on recognition tasks is much better than if you do this self-supervised retraining.

The architecture of the encoder is good, but the fact that you train the system to reconstruct images does not lead it to produce to learn good generic features of images.

When you train in a self-supervised way.

Self-supervised by reconstruction.

Let's watch what he has to say about JEPA.

And again, JEPA was just released by Meta under Jan's leadership with his entire team.

And it's a very interesting project.

It gives the ability to augment Large Language Models with the ability to predict video.

And I'm still wrapping my mind around it.

I'm going to let him explain it.

The alternative is joint embedding.

Instead of training a system to encode the image

And then training it to reconstruct the full image from a corrupted version, you take the full image, you take the corrupted or transformed version, you run them both through encoders, which in general are identical, but not necessarily.

And then you train a predictor on top of those encoders to predict the representation of the full input from the representation of the corrupted one. So joint embedding, because you're taking the full input and the corrupted version, run them both through encoders, you get a joint.

And then you're saying, can I predict the representation of the full one from the representation of the corrupted one?

And I call this a JEPA.

That means joint embedding predictive architecture, because there's joint embedding and there is this predictor that predicts the representation of the good guy from the bad guy.

And the big question is, how do you train something like this?

And until five years ago or six years ago, we didn't have particularly good answers.

He's going to talk about what is the difference between Large Language Models and and why them working together might be the solution to reach AGI and a world model.

First of all, what's the difference with generative architectures like LLMs?

LLMs or vision systems that are trained by reconstruction generate the inputs, right?

They generate the original input that is non-corrupted, non-transformed.

You have to predict all the pixels and there is a huge amount of resources spent in the system to actually predict all those pixels, all the details.

In a JEPA, you're not trying to predict all the pixels, you're only trying to predict an abstract representation of the inputs.

And that's much easier in many ways.

What the JEPA system, when it's being trained, is trying to do is extract as much information as possible from the input, but yet only extract information that is relatively easily predictable.

There's a lot of things in the world that we cannot predict, like for example, if you have a self-driving car driving down the street, there may be trees around the road and it could be a windy day, so the leaves on the tree are kind of moving in kind of semi-chaotic random ways that you can't predict and you don't care.

What you want is your encoder to basically eliminate all those details.

We'll tell you there's moving leaves, but it's not going to keep the details of exactly what's going on.

When you do the prediction in representation space, you're not going to have to predict every single pixel of every leaf.

And that not only is a lot simpler, but also it allows the system to essentially learn an abstract representation of the world where what can be modeled and predicted is preserved and the rest is viewed as noise and eliminated by the encoder.

It kind of lifts the level of abstraction of the representation.

If you think about this, this is something we do absolutely all the time.

Whenever we describe a phenomenon, we describe it at a particular level of abstraction and we don't always describe every natural phenomenon in terms of quantum field theory, that would be impossible.

We have multiple levels of abstraction to describe what happens in the world, starting from quantum field theory to atomic theory and molecules and chemistry, all the way up to concrete objects in the real world and things like that.

We can't model everything at the lowest level.

That's what the idea of JEPA is really about, learn abstract representation in a self-supervised manner and you can do it hierarchically as well.

So that I think is an essential component of an intelligent system and in language we can get away without doing this because language is already to some level abstract and already has eliminated a lot of information that is not predictable.

We can get away without doing the joint embedding, without lifting the abstraction level and by directly predicting words.

In this next clip, Jan talks about hierarchical planning and what that essentially means is There is a ton of planning that needs to happen for basic actions that we take every single day. So to describe hierarchical planning, let's use an example.

Let's say I want to drive to the store.

First I have to get up and walk to my car.

How do I do that?

Well first I have to actually stand up.

Well how do I do that?

Well my brain has to tell my legs to use certain muscles to push me up.

And then how do I do that?

It's this recursive algorithm that is required, and every single step from something as basic as just walking to the car is actually thousands, or really just infinite steps if you really think about it.

How do you model that?

How do you actually model that with current architecture?

That's what he starts to talk about here.

Let's watch.

Hierarchical planning is absolutely necessary if you want to plan complex actions.

If I want to go from, let's say from New York to Paris, this is the example I use all the time and I'm sitting in my office at NYU.

My objective that I need to minimize is my distance to Paris.

At a high level, a very abstract representation of my location, I would have to decompose this into two sub-goals.

First one is go to the airport.

Second one is catch a plane to Paris.

Okay, so my sub-goal is now going to the airport.

My objective function is my distance to the airport.

How do I go to the airport?

Well I have to go in the street and hail a taxi, which you can do in New York.

Okay, now I have another sub-goal.

How do I go down on the street?

Well that means going to the elevator, going down the elevator, walk out the street.

How do I go to the elevator?

I have to stand up for my chair, open the door in my office, go to the elevator, push the button.

How do I get up from my chair?

Like you can imagine going down all the way down to basically what amounts to millisecond by millisecond muscle control.

Obviously, you're not going to plan your entire trip from New York to Paris in terms of millisecond by millisecond muscle control.

First, that would be incredibly expensive, but it will also be completely impossible because you don't know all the conditions of what's going to happen, how long it's going to take to catch a taxi or to go to the airport with traffic.

I mean, you would have to know exactly the condition of everything to be able to do this planning and you don't have the information.

You have to do this hierarchical planning so that you can start acting.

And then sort of replanning as you go.

Nobody really knows how to do this in AI.

Nobody knows how to train a system to learn the appropriate multiple levels of representation so that hierarchical planning works.

Lex is going to press Jan about scaling because what we've seen with the ability to scale Large Language Models is truly amazing.

We've gone from GPT one, two, three, now we're at four, almost at five, and the increase in parameters, the increase in GPU processing power has unleashed incredible capabilities from these Large Language Models.

You have Sora and Sora approved.

If you increase compute, you can get much better results.

Lex really tries to press Jan about, hey, if we keep scaling up, can't we reach super intelligence with our current technologies?

And he also asked what Alan Turing would think about current AI.

And if you're not familiar with Alan Turing, he created the Turing test amongst many other things and the Turing test is simply a test to see if artificial intelligence is truly intelligent.

And all it really needs to do is convince a human that it is human.

And with our current technology, that's very possible, especially if you just do it over text.

But even with voice, it's possible.

Jan says something really funny about it.

I'm not going to spoil that joke.

Take a look.

Does it make sense to you that they're able to form enough of a representation of the world to be damn convincing, essentially passing the original Turing test with flying colors?

Yeah, we're fooled by their fluency, right?

We just assume that if a system is fluent in manipulating language, then it has all the characteristics of human intelligence.

But that impression is false.

We're really fooled by it.

What do you think Alan Turing would say?

Without understanding anything, just hanging out with it?

Alan Turing would decide that the Turing test is a really bad test.

This is what the AI community has decided many years ago, that the Turing test was a really bad test of intelligence.

Next Jan discusses hallucinations.

Why do they happen?

Why are they such a big limitation for current Large Language Models?

And it is a huge limitation.

It's probably one of the biggest limitations to it being widely adopted in the corporate setting.

There are a lot of techniques to prevent hallucinations, but let's find out why they actually occur.

Because of the autoregressive prediction, every time an algorithm produces a token or word, there is some level of probability for that word to take you out of the set of reasonable answers.

And if you assume, which is a very strong assumption, that the probability of such error is that those errors are independent across a sequence of tokens being produced.

What that means is that every time you produce a token, the probability that you stay within the set of correct answer decreases and it decreases exponentially.

If there's a non-zero probability of making a mistake, which there appears to be, then there's going to be a kind of drift.

And that drift is exponential.

It's like errors accumulate.

The probability that an answer would be nonsensical increases exponentially with the number of tokens.

This next clip is very interesting.

Yann talks about why he doesn't like reinforcement learning, or at least why reinforcement learning is very good, but only for a very narrow use case.

Let's watch.

Why do you still hate reinforcement learning?

I don't hate reinforcement learning and I think it should not be abandoned completely, but I think its use should be minimized because it's incredibly inefficient in terms of samples.

The proper way to train a system is to first have it learn good representations of the world and world models from mostly observation, maybe a little bit of interactions.

There's two things.

You can use, if you've learned a world model, you can use the world model to plan a sequence of actions to arrive at a particular objective.

You don't need RL unless the way you measure whether you succeed might be inexact.

Your idea of whether you were going to fall from your bike might be wrong, or whether the person you're fighting with MMA was going to do something.

And then do something else.

There's two ways you can be wrong.

Either your objective function does not reflect the actual objective function you want to optimize, or your world model is inaccurate.

The prediction you were making about what was going to happen in the world is inaccurate.

If you want to adjust your world model while you are operating in the world or your objective function, that is basically in the realm of RL.

This is what RL deals with to some extent, right?

Adjust your world model.

And the way to adjust your world model, even in advance, is to explore parts of the space where your world model, where you know that your world model is inaccurate.

That's called curiosity, basically, or play.

When you play, you kind of explore parts of the state space that you don't want to do in for real because it might be dangerous, but you can adjust your world model without killing yourself, basically.

That's what you want to use RL for.

When it comes time to learning a particular task, you already have all the good representations.

You already have your world model, but you need to adjust it for the situation at hand.

That's when you use RL.

To the part that I'm truly excited about.

Jan talks about why open source is so important.

He references Gemini 1.5's wokeness.

He references LLM biases and why open source models are the answer to all of it.

And I tend to agree with much of what he says.

Let's take a look.

A lot of people have been very critical of the recently released Google's Gemini 1.5 for essentially, in my words, I could say super woke, woke in the negative connotation of that word.

There's some almost hilariously absurd things that it does, like it modifies history, generating images of a black George Washington, or perhaps more seriously, something that you commented on Twitter, which is refusing to comment on or generate images of, or even descriptions of Tiananmen Square or the Tank Man, one of the most sort of legendary protest images in history.

And of course, these images are highly censored by the Chinese government.

And therefore, everybody started asking questions of what is the process of designing these LLMs?

What is the role of censorship in these?

You commented on Twitter saying that open source is the answer.

Can you explain?

I actually made that comment on just about every social network I can, and I've made that point multiple times in various forums.

Here's my point of view on this.

People can complain that AI systems are biased, and they generally are biased by the distribution of the training data that they've been trained on that reflects biases in society, and that is potentially offensive to some people or potentially not.

And some techniques to de-bias then become offensive to some people because of historical incorrectness and things like that.

You can ask the question, you can ask two questions.

The first question is, is it possible to produce an AI system that is not biased?

And the answer is absolutely not.

And it's not because of technological challenges, although there are technological challenges to that.

It's because bias is in the eye of the beholder, different ideas about what constitutes bias for a lot of things.

I mean, there are facts that are indisputable, but there are a lot of opinions or things that can be expressed in different ways.

You cannot have an unbiased system that's just an impossibility.

What's the answer to this?

And the answer is the same answer that we found in liberal democracy about the press.

The press needs to be free and diverse.

We have free speech for a good reason.

It's because we don't want all of our information to come from a unique source, because that's opposite to the whole idea of democracy and progress of ideas and even science.

In science, people have to argue for different opinions, and science makes progress when people disagree and they come up with an answer and a consensus forms.

And it's true in all democracies around the world.

There is a future is already happening where every single one of our interaction with the digital world will be mediated by AI systems.

We're going to have smart glasses.

You can already buy them from Meta, the Ray-Ban Meta, where you can talk to them and they are connected with an LLM and you can get answers on any question you have.

Or you can be looking at a monument and there is a camera in the system that in the glasses, you can ask it like, what can you tell me about this building or this monument?

You can be looking at a menu in a foreign language and it will translate it for you or you can do real-time translation if you speak different languages.

A lot of our interaction with the digital world are going to be mediated by those systems in the near future.

Increasingly, the search engines that we're going to use are not going to be search engines.

They're going to be dialogue systems where you just ask a question and it will answer.

And then point you to perhaps an appropriate reference for it.

But here is the thing.

We cannot afford those systems to come from a handful of companies on the west coast of the US because those systems will constitute the repository of all human knowledge and we cannot have that be controlled by a small number of people.

It has to be diverse.

For the same reason, the press has to be diverse.

How do we get a diverse set of AI assistants?

It's very expensive and difficult to train a base model, a base LLM at the moment.

In the future, it might be something different, but at the moment, that's an LLM.

Only a few companies can do this properly.

If some of those top systems are open source, anybody can use them.

We can fine tune them.

If we put in place some systems that allows any group of people, whether they are individual citizens, groups of citizens, government organizations, NGOs, companies, whatever, to take those open source systems, AI systems, and fine tune them for their own purpose on their own data, then we're going to have a very large diversity of different AI systems that are specialized for all of those things.

Right.

I'll tell you, I talked to the French government quite a bit and the French government will not accept that the digital diet of all their citizens be controlled by three companies on the west coast of the U.S. That's just not acceptable.

It's a danger to democracy, regardless of how well-intentioned those companies are.

And it's also a danger to local culture, to values, to language.

I was talking with the founder of Infosys in India.

He's funding a project to fine tune Lama2, the open source model produced by Meta, so that Lama2 speaks all 22 official languages in India.

I mean, you can't have any of this unless you have open source platforms.

With open source platforms, you can have AI systems that are not only diverse in terms of political opinions or things of that type, but in terms of language, culture, value systems, political opinions, technical abilities in various domains.

And you can have an industry, an ecosystem of companies that fine tune those open source systems for vertical applications in industry, right?

You have, I don't know, a publisher has thousands of books, and they want to build a system that allows a customer to just ask a question about the content of any of their books.

You need to train on their proprietary data.

You have a company, we have one within Meta, it's called Metamate.

And it's basically an LLM that can answer any question about internal stuff about the company.

Very useful.

A lot of companies want this, right?

A lot of companies want this not just for their employees, but also for their customers, to take care of their customers.

The only way you're going to have an AI industry, the only way you're going to have AI systems that are not uniquely biased is if you have open source platforms on top of which any group can build virtualized systems.

The direction of inevitable direction of history is that the vast majority of AI systems will be built on top of open source platforms.

And on the topic of open source, Lex asked him, okay, if it's open source, how do you actually run a business based on open source?

What are the economics of open source?

And this has been proven over the years.

We have a lot of very successful open source projects that have made a lot of money.

Many of the standards that we use on the internet today, many of the standards that we use with databases, with code architecture, everything, a lot of that is open source.

And there are a lot of benefits with open source and a lot of companies have made a lot of money with open source.

Just look at Android for Google, just as an example off the top of my head.

Let's see what Jan says about the economics of Meta's open source contributions.

You have several business models, right?

The business model that Meta is built around is your first service.

And the financing of that service is either through ads or through business customers.

For example, if you have an LLM that can help a mom and pop pizza place by talking to their customers through WhatsApp.

The customers can order a pizza and the system will just ask them like, what topping do you want or what size, blah, blah, blah.

The business will pay for that.

And otherwise if it's a system that is on the more kind of classical services, it can be ad supported or the several models.

But the point is, if you have a big enough potential customer base and you need to build that system anyway for them, it doesn't hurt you to actually distribute it in open source.

And then Lex continues to press them about the open source economics.

Why doesn't another company just come along, take that open source project and build competition?

That probably will happen and that's probably good.

But at the end of the day, Meta is still Meta.

They have an entire multi-billion human user base to sell their products and services to.

Let's see what he says.

The bet is more, we already have a huge user base and customer base, right?

It's going to be useful to them.

Whatever we offer them is going to be useful and there is a way to derive revenue from this.

It doesn't hurt that we provide that system or the base model, right, the foundation model in open source for others to build applications on top of it too.

If those applications turn out to be useful for our customers, we can just buy it from them.

It could be that they will improve the platform.

In fact, we see this already.

I mean, there is literally millions of downloads of LLaMA 2 and thousands of people who have provided ideas about how to make it better.

This clearly accelerates progress to make the system available to a wide community of people and there is literally thousands of businesses who are building applications with it.

Meta's ability to derive revenue from this technology is not impaired by the distribution of it, of base models in open source.

We continue on the whole biases front.

Many employees at tech companies tend to lean left and that is a point that Lex makes and I tend to agree with. And so aren't those biases going to be built into Large Language Models?

And again, Jan points to open source being the answer.

I don't think the issue has to do with the political leaning of the people designing those systems.

It has to do with the acceptability or political leanings of their customer base or audience.

A big company cannot afford to offend too many people.

They're going to make sure that whatever product they put out is safe, whatever that means.

It's very possible to overdo it.

And it's also very possible to, it's impossible to do it properly for everyone.

You're not going to satisfy everyone.

That's what I said before.

You cannot have a system that is unbiased, that is perceived as unbiased by everyone.

It's going to be you push it in one way, one set of people are going to see it as biased

And then you push it the other way and another set of people is going to see it as biased.

And then in addition to this, there's the issue of if you push the system, perhaps you go too far in one direction, it's going to be non-factual, right?

You're going to have black Nazi soldiers in the image.

Yeah, so we should mention image generation of black Nazi soldiers, which is not factually accurate.

Right.

And can be offensive for some people as well, right?

It's going to be impossible to kind of produce systems that are unbiased for everyone.

The only solution that I see is diversity.

And diversity in the full meaning of that word, diversity in every possible way.

And continuing on the open source front, now they're going to talk about free speech, guardrails on AI, censorship, biases and everything.

I love this conversation.

And hearing Jan frame open source in this light makes a ton of sense to me.

Let's watch.

I mean, there are some limits to what the same way there are limits to free speech, there has to be some limit to the kind of stuff that those systems might be authorized to produce some guardrails.

I mean, that's one thing I've been interested in, which is in the type of architecture that before, where the output of a system is a result of an inference to satisfy an objective.

That objective can include guardrails and we can put guardrails in open source systems.

I mean, if we eventually have systems that are built with this blueprint, we can put guardrails in those systems that guarantee that there is sort of a minimum set of guardrails that make the system non-dangerous and non-toxic, et cetera basic things that everybody would agree on.

And then the fine tuning that people will add or the additional guardrails that people will add will kind of cater to their community, whatever it is.

Next, if you wanted a little preview of when LLaMA three is coming and what it's going to be about, this is the clip for you.

Let's watch.

There's going to be like various versions of LLaMA that are improvements of previous LLaMAs, bigger, better, multimodal, things like that.

And then in future generations, systems that are capable of planning that really understand how the world works.

Maybe are trained from video.

They have some world model, maybe capable of the type of reasoning and planning I was talking about earlier.

Like, how long is that going to take?

Like when is the research that is going in that direction going to sort of feed into the product line, if you want, of LLaMA?

I don't know.

I can't tell you.

And there is a few breakthroughs that we have to basically go through before we can get there.

You'll be able to monitor our progress because we publish our research, right?

Last week we published the VJEPA work, which is sort of a first step towards training systems from video.

And then the next step is going to be world models based on kind of this type of idea, training from video.

There's similar work at DeepMind also, and taking place, people, and also at UC Berkeley on world models from video.

A lot of people are working on this.

I think a lot of good ideas are appearing.

My bet is that those systems are going to be JEPA-like, they're not going to be generative models.

And we'll see what the future will tell.

There's really good work at a gentleman called Daniel Haffner, who is now at DeepMind, who has worked on kind of models of this type that learn representations

And then use them for planning or learning tasks by reinforcement training.

Going back a few topics, they're going to talk about the power and efficiency in the human mind to process all of this data and what we get from it compared to Large Language Models.

And it's really not even a comparison.

Large language models require a ton of data, a ton of processing power, a ton of energy to train and to use versus a human brain, which is incredibly efficient.

We're still far in terms of compute power from what we would need to match the compute power of the human brain.

This may occur in the next couple of decades, but we're still some ways away.

And certainly in terms of power efficiency, we're really far.

A lot of progress to make in hardware.

Right now, a lot of the progress is a bit coming from silicon technology, but a lot of it coming from architectural innovation and quite a bit coming from more efficient ways of implementing the architectures that have become popular, basically a combination of transformers and components.

There's still some ways to go until we're going to saturate, we're going to have to come up with new principles, new fabrication technology, new basic components, perhaps on sort of different principles than those classical digital CMOS.

And next, they're going to start talking about AGI.

When is it coming?

And Jan has made this point over and over again.

A lot of people think AGI is just going to be some switch that is flipped one day, and it's just going to be this inflection point that AGI just has this quote unquote hard takeoff.

But he doesn't think that's the case.

He thinks it's going to be progressive.

Let's see what he has to say.

First of all, it's not going to be an event, right?

The idea somehow, which is popularized by science fiction and Hollywood, that somehow somebody is going to discover the secret, the secret to AGI or human level AI or AMI, whatever you want to call it, and then turn on a machine and then we have AGI.

That's just not going to happen.

It's not going to be an event.

It's going to be gradual progress.

Are we going to have systems that can learn from video how the world works and learn good representation?

Yeah, before we get them to the scale and performance that we observe in humans, it's going to take quite a while.

Are we going to get systems that can have large amount of associative memory so they can remember stuff?

Yeah, but same.

It's not going to happen tomorrow.

I mean, there is some basic techniques that need to be developed.

We have a lot of them, but to get this to work together with a full system is another story.

Are we going to have systems that can reason and plan perhaps along the lines of objective driven AI architectures that I described before?

Yeah, but before we get this to work properly, it's going to take a while.

And before we get all those things to work together

And then, on top of this, we have systems that can learn like hierarchical planning, hierarchical representation.

Systems that can be configured for a lot of different situations at hand, the way the human brain can.

All of this is going to take at least a decade, and probably much more, because there are a lot of problems that we're not seeing right now.

We have not encountered, and so we don't know if there is an easy solution within this framework.

It's not just around the corner.

I mean, I've been hearing people for the last 12, 15 years claiming that AGI is just around the corner and being systematically wrong.

And I knew they were wrong when they were saying it.

I called their bullshit.

And next, he talks about AI doomerism and what that actually means and good AI versus bad AI and how that whole thing is going to play out.

Something that I'm very interested in.

And I made a video about it recently because it is such an interesting topic to see the wide spectrum of people's beliefs about AGI and AI doom.

Okay, so AI doomers imagine all kinds of catastrophe scenarios of how AI could escape or control and basically kill us all.

And that relies on a whole bunch of assumptions that are mostly false.

The first assumption is that the emergence of superintelligence could be an event.

That at some point we're going to Figure out the secret and we'll turn on a machine that is super intelligent.

And because we've never done it before, it's going to take over the world and kill us all.

That is false.

It's not going to be an event.

We're going to have systems that are like as smart as a cat, have all the characteristics of human level intelligence, but their level of intelligence would be like a cat or a parrot maybe or something.

And then we're going to walk our way up to kind of make those things more intelligent.

And as we make them more intelligent, we're also going to put some guardrails in them and learn how to kind of put some guardrails so they behave properly.

And we're not going to do this with just one, it's not going to be one effort, but it's going to be lots of different people doing this.

And some of them are going to succeed at making intelligent systems that are controllable and safe and have the right guardrails.

And if some other goes rogue, then we can use the good ones to go against the rogue ones.

It's going to be my smart AI police against your rogue AI.

It's not going to be like we're going to be exposed to like a single rogue AI that's going to kill us all.

That's just not happening.

There is another fallacy, which is the fact that because the system is intelligent, it necessarily wants to take over.

And there are several arguments that make people scared of this, which I think are completely false as well.

One of them is in nature, it seems to be that the more intelligent species are the ones that end up dominating the others, extinguishing the others, sometimes by design, sometimes just by mistake.

There is sort of thinking by which you say, well, if AI systems are more intelligent than us, surely they're going to eliminate us, if not by design, simply because they don't care about us.

And that's just preposterous for a number of reasons.

The first reason is they're not going to be a species, they're not going to be a species that competes with us.

They're not going to have the desire to dominate because the desire to dominate is something that has to be hardwired into an intelligent system.

It is hardwired in humans.

It is hardwired in baboons, in chimpanzees, in wolves, not in orangutans.

The species in which this desire to dominate or submit or attain status in other ways is specific to social species.

Non-social species like orangutans don't have it, right?

And they are as smart as we are, almost.

And then next, they tackle the comparison between AI, AGI, and nuclear systems.

And there's been a lot of comparisons between nuclear and AI, and I don't necessarily agree that's an apples to apples comparison, and Jan tends to agree, and he explains why.

But not only that, he explains why AI is essentially going to be our filter to the rest of the internet and to other AI models.

An AI model is not going to be able to directly communicate with us.

We're not even going to see it.

He actually brings up the analogy of a spam filter for email.

We get a ton of spam, and most of which we never see and we never need to see.

It's interesting to think about, but it's also a little worrying to think that an AI is going to be my filter to all of my information diet.

So that AI system designed by Vladimir Putin or whatever, or his minions, is going to be trying to talk to every American to convince them to vote for whoever pleases Putin or rile people up against each other.

They're not going to be talking to you.

They're going to be talking to your AI assistant, which is going to be as smart as they are.

That AI, because as I said, in the future, every single one of your interactions with the digital world will be mediated by your AI assistant.

The first thing you're going to ask is, is this a scam?

Is this thing telling me the truth?

It's not even going to be able to get to you because it's only going to talk to your AI assistant.

Your AI assistant is not even going to, it's going to be like a spam filter.

You're not even seeing the email, the spam email, right?

It's automatically put in a folder that you never see.

It's going to be the same thing.

And next, he talks about robots, and I've been making a lot of videos about robots lately, so let's see what Jan has to say about robots.

The next decade, I think, is going to be really interesting in robots.

The emergence of the robotics industry has been in the waiting for 10, 20 years without really emerging, other than for like kind of pre-programmed behavior and stuff like that.

And the main issue is, again, the Moravec paradox, like how do we get the system to understand how the world works and kind of plan actions?

We can do it for really specialized tasks.

The way Boston Dynamics goes about it is basically with a lot of handcrafted dynamical models and careful planning in advance, which is very classical robotics with a lot of innovation, a little bit of perception, but it's still not like they can't build a domestic robot.

We're still some distance away from completely autonomous level five driving, and we're certainly very far away from having level five autonomous driving by a system that can train itself by driving 20 hours like any 17 year old.

A lot of the people working on robotic hardware at the moment are betting or banking on the fact that AI is going to make sufficient progress towards that.

And last, Lex asks a good question.

What gives you hope about humanity over the next few decades?

And Jan gives some good answers there.

I really love hearing this.

I love ending it on kind of a positive note.

We can make humanity smarter with AI.

I mean, AI basically will amplify human intelligence.

It's as if every one of us will have a staff of smart AI assistants that might be smarter than us.

They'll do our bidding, perhaps execute tasks in ways that are much better than we could do ourselves because they'd be smarter than us.

It's like everyone would be the boss of a staff of super smart virtual people.

We shouldn't feel threatened by this any more than we should feel threatened by being the manager of a group of people, some of whom are more intelligent than us.

I certainly have a lot of experience with this.

That's actually a wonderful thing.

Having machines that are smarter than us that assist us in all of our tasks or daily lives, whether it's professional or personal, I think would be an absolutely wonderful thing because intelligence is the most, is the commodity that is most in demand.

That's really what, I mean, all the mistakes that humanity makes is because of lack of intelligence, really, or lack of knowledge, which is related.

Making people smarter, which just can only be better.

I mean, for the same reason that public education is a good thing and books are a good thing and the internet is also a good thing intrinsically.

And even social networks are a good thing.

If you run them properly, it's difficult, but you know, you can help the communication of information and knowledge and the transmission of knowledge.

AI is going to make humanity smarter.

That's it.

I encourage you to check out the video in full.

It is very long.

Hopefully this cut down version helped.

If you liked this video, please consider giving a like and subscribe and I'll see you in the next one.