DEV Community

Cover image for A Trilogy of LLMs, World Models, and People
Whetlan
Whetlan

Posted on AI-assisted

A Trilogy of LLMs, World Models, and People

In a Futurology interview, Yann LeCun suggested that LLMs might be one component of future AI systems because they’re good at encoding and producing text. What would the rest of the system need to do? Futurology interview, LLM discussion

Think of two people carrying a table through a doorway. One says, “A little more this way.” The other moves toward the window because they can see where their partner is pointing. Transcribe the conversation and you'll have “this way.” A reader can guess what it means. To know where the person pointed, somebody has to put that detail in the record.

That choice is part of why I think training data is never neutral. Text comes from people who had a reason to say or write something at a particular time. A text-only model can pick up quite a bit about rooms and furniture from what people write. If a table leg catches on the doorframe, though, the people carrying it can tip the table and find out whether that helps. A model working from the transcript learns the outcome when somebody writes it down. That's the sense in which I think of language as the innermost layer. The people in the room have already decided how much of the room to put into words.

A world model could learn what happens when the table turns by observing the move itself. In Yann LeCun's proposal for autonomous machine intelligence, a predictive world model would help a system plan what to do. The I-JEPA paper gives a concrete example of learning from images. It uses the features of visible regions in an image to predict features of regions kept out of view. Predicting how the table moves after somebody gives it a shove is a different task, involving a change over time and the action that caused it. LeCun's research proposal · The I-JEPA paper

The data still has a date. Suppose we give three models access to a physical world governed by the same laws. One gets the literature, mathematics, and measurement records of an ancient society. Another gets the books and instruments of a later period, and the last can use what's available today. I'd expect them to learn different things.

The books matter even when the models can make their own observations. Mathematics gives them ways to describe what they measure. The literature tells them how people at the time described things they thought worth recording. Put the same oscillating object in front of all three models. One might count the cycles. Another might also record the temperature and check whether changing it affects the count. The first model could get there on its own. It would have to come up with the question and find a way to make the measurements, given what it started with.

When I say the data has a date, I mean more than a timestamp on the recording. The models could all start collecting data on the same day and still inherit different ways of looking at it. Even the labels on a set of measurements tell you what the person who collected them thought might explain the result. If nobody recorded temperature, you can't go back to that record and separate a change in temperature from a change in the object. You can run the experiment again, assuming you know to ask and can control the conditions.

Some of that starting knowledge is built into the objects a model can study. Quartz could vibrate in the age of pendulum clocks. Making a quartz clock took a way to cut the crystal, make it oscillate in a circuit, and count the cycles. Once the clock existed, someone else could put it on a bench and take it apart. A model trained in that world would have a device to examine that an earlier model would have had to figure out how to build. The laws governing the crystal are the same. The contents of the room have changed. NIST's history of timekeeping

People grow up inside the same physical world a world model would learn in. We're also working with whatever knowledge is available at the time. And over the course of learning how to do something, we can change what the next person has to learn. A clockmaker can ask how much a clock gains or loses. Once people build satellite navigation, an engineer also needs to know how far apart the satellites' clocks are. A clock can keep a steady rate and still be out of step with the others. Synchronization becomes a scarce resource when people build something that depends on it. NIST on clocks and synchronization

Would a world with only a trained world model develop the way a world with people would? Assume the model can act as well as observe. I doubt the two worlds would follow the same path. The model might be very good at measuring a crystal's oscillations. I can see why it would want accurate clocks if it had already decided to build satellite navigation. I'm less sure what would give it a reason to build satellite navigation in the first place. That reason came from the way people wanted to move around the world and from the equipment they ended up relying on.

Give the system a score based on how much a single clock drifts, and recursive self-improvement gives it a way to keep working on that problem. It can change the circuit, repeat a measurement, or find a source of error the previous design ignored. As long as the score stays the same, each round still asks how to make that clock more accurate with the knowledge and physical resources the system can draw on. A different problem would start with a different observation. To make synchronization the goal of its next round of improvements, the system would need a reason to care about the difference between the two clocks.

If the model suspects that temperature affects clock drift, it can predict how the drift will change as the temperature rises, then repeat the measurement with only the temperature changed and check its prediction.

If the two worlds did start to diverge, what would show us that the model had found a problem of its own? Would it be enough to see it ask a question no human has asked? Or would we have to watch it change something in the world and find a variable no human had thought to measure?


Find me on StratCraft | GitHub

Top comments (1)

Collapse
 
alexcodebytes profile image
Oleksandr •

The distinction between learning the laws of the world and learning what is worth doing in that world is the most interesting part of this piece. A world model might discover how to make a clock increasingly accurate, but that doesn't automatically give it a reason to build a navigation system or care about synchronization. The objective shapes which discoveries become useful, and that objective doesn't necessarily emerge from better prediction.

I also like the point about inherited knowledge being embedded in physical objects, not just written records. A later model doesn't merely start with more information; it starts in a world already transformed by earlier discoveries. That makes the comparison between the three models especially compelling.

Perhaps the strongest evidence of genuinely independent problem-finding wouldn't be a question humans haven't asked, but a chain of actions: identifying an unexplained effect, designing an experiment to isolate it, and then changing its goals or methods because of what it discovers. The difficult question is how we distinguish that from a system simply following an objective we gave it in a less obvious way.