Description is not understanding
A language model is trained on what we have said about the world. It is a spectacular compression of human description — and description turns out to carry a startling amount of the structure underneath. But it is still downstream of the thing itself. Ask it what happens next in a physical system, or across a supply chain, or in a society under stress, and you are asking it to interpolate between accounts rather than to run the situation forward.
World models are the attempt to close that gap: systems that carry state, that can be rolled forward, that can answer "what would happen if" rather than "what has been written about". Simulation, in other words, with a learned model of the dynamics instead of a hand-written one.
The part I actually care about
Most of the excitement here is about capability — robotics, planning, agents that do not fall over when reality diverges from the prompt. Fine. The question that keeps me up is a different one.
We are visibly short on empathy, and I think the shortage is a failure of perception rather than of morality.
People are not, in the main, indifferent to each other. They simply cannot see the dependency. The chain between a decision here and a consequence eight thousand kilometres away is too long, too abstract and too slow to feel like anything. So it does not.
Two routes, and I know which one I want
One route to fixing that is to change the human. Merge with the machine, widen the input channel, become the thing that can hold a planetary system in mind. The cyborg path. It may well happen, and I am not reflexively against it — but it solves the problem by rebuilding the person, which is an enormous thing to get wrong.
The other route is to make the dependency visible. If a world model can simulate the system we are all embedded in well enough to be honest, then the thing we currently have to take on faith — that we depend on each other, continuously and concretely — becomes something you can look at. Not an argument. A picture, with you in it.
That is a much less dramatic technology than a brain interface. It is also, I suspect, the one more likely to make people behave better, because it does not ask anyone to become someone else first. It just stops the interdependence from being invisible.
I do not know whether we get there. I am extremely curious, which is most of why I keep writing about it.