Brilliant piece thank you. I particularly like the observation by LeCun of how 'world building' is going to be used by everyone in tech as they try to rebuild their moats. Looking forward to the snake oil traders jumping on that...
Yeah, given how fast "AGI" went from technical term to marketing copy, you can expect "world models" to follow the same path, with every company claiming the label to raise funding.
This is fantastic. Hugo, I likely don't even understand what's wrong with this question, but will attempt it anyway.
Do any of these approaches or problem spaces map onto representations of conceptual knowledge? I have in mind here, institutional knowledge, like the Constitution, or criminal code, or cultural etiquette. Things evolved apes "make up". Is it enough to encode it in language and depend on LLMs, or should there be a layer of actual machine understanding? Like representation + action for robotics?
Really glad you're interested, and there's nothing wrong with the question, it's actually pointing at something interesting.
The short version: LeCun's argument is about how machines learn the physical world. He thinks language is a terrible bottleneck for learning physics, because physics is continuous, high-dimensional, and most of what matters can't be expressed in words. You can't learn to catch a ball from text descriptions of catching.
But institutional knowledge is almost the opposite case. The Constitution is language. Criminal code is language. These things were literally invented in text and exist as text. So for that domain, LLMs are arguably working in the native representation already, not trying to squeeze physics through a language bottleneck.
Where it gets interesting is one layer deeper. Knowing the text of a law and understanding how that law actually operates in a society are two very different things. A law interacts with incentives, power structures, human psychology, cultural norms, all of which are grounded in the real world, not just in the document. So there might be a version of the same split: the "letter" is language-native and LLMs handle it well, but the "spirit" requires something closer to a world model of human behavior. Understanding why a constitutional principle works, what happens when you push it, what breaks if you remove it, that might need the kind of causal understanding that LLMs are not great at.
LeCun's real target isn't language per se, it's the generative training objective. He thinks predicting the next token teaches you correlations, not causes. If that's true, it applies to legal reasoning too, not just robotics. An LLM might predict what a court would likely rule, but not why that ruling follows from constitutional structure. Whether that distinction matters in practice is, I think, genuinely open.
My intuition is that the split, as you say, must exist. That the ultimate challenge is a causal epistemology that supports novel creation and use of world-model counterfactuals.
Brilliant piece thank you. I particularly like the observation by LeCun of how 'world building' is going to be used by everyone in tech as they try to rebuild their moats. Looking forward to the snake oil traders jumping on that...
Yeah, given how fast "AGI" went from technical term to marketing copy, you can expect "world models" to follow the same path, with every company claiming the label to raise funding.
This is fantastic. Hugo, I likely don't even understand what's wrong with this question, but will attempt it anyway.
Do any of these approaches or problem spaces map onto representations of conceptual knowledge? I have in mind here, institutional knowledge, like the Constitution, or criminal code, or cultural etiquette. Things evolved apes "make up". Is it enough to encode it in language and depend on LLMs, or should there be a layer of actual machine understanding? Like representation + action for robotics?
Really glad you're interested, and there's nothing wrong with the question, it's actually pointing at something interesting.
The short version: LeCun's argument is about how machines learn the physical world. He thinks language is a terrible bottleneck for learning physics, because physics is continuous, high-dimensional, and most of what matters can't be expressed in words. You can't learn to catch a ball from text descriptions of catching.
But institutional knowledge is almost the opposite case. The Constitution is language. Criminal code is language. These things were literally invented in text and exist as text. So for that domain, LLMs are arguably working in the native representation already, not trying to squeeze physics through a language bottleneck.
Where it gets interesting is one layer deeper. Knowing the text of a law and understanding how that law actually operates in a society are two very different things. A law interacts with incentives, power structures, human psychology, cultural norms, all of which are grounded in the real world, not just in the document. So there might be a version of the same split: the "letter" is language-native and LLMs handle it well, but the "spirit" requires something closer to a world model of human behavior. Understanding why a constitutional principle works, what happens when you push it, what breaks if you remove it, that might need the kind of causal understanding that LLMs are not great at.
LeCun's real target isn't language per se, it's the generative training objective. He thinks predicting the next token teaches you correlations, not causes. If that's true, it applies to legal reasoning too, not just robotics. An LLM might predict what a court would likely rule, but not why that ruling follows from constitutional structure. Whether that distinction matters in practice is, I think, genuinely open.
Fun question to sit with.
My intuition is that the split, as you say, must exist. That the ultimate challenge is a causal epistemology that supports novel creation and use of world-model counterfactuals.
Excellent article, pleasure to read.