In artificial intelligence the world is always what we want it to be and it is so before we act.

Artificial intelligence will not replace human intelligence. It does not operate on the same plane as human intelligence. It will replace the world.

For many, intelligence seems to mean something close to ratiocination or a set of operations within the constraints of a given physical world. But we know things are more complicated. In the Western tradition, we find in Kant a more appropriate definition where intelligence is a world model. Human intelligence builds the world within which it may then perform reasonings and calculations. For human intelligence, this model is mental rather than physical. The world model may predict multiple plausible world states, parameterised by latent variables that represent the uncertainty about the world state. The world model is a kind of simulator (see LeCunn, A Path Towards Autonomous Machine Intelligence).

Artificial intelligence is nothing more and nothing less than a world model existing in the physical world. In other words, with artificial intelligence model and world coincide. While human intelligence extends over the full range of possibilities contained in the mental model, artificial intelligence can create virtual worlds of equal range existing in the external world. While human action realises possibilities, foundation models realise the full range of these individual possibilities. While human intelligence simulates the world in mental space, artificial intelligence does so for the external world.

There is a very common error that artificial intelligence can operate in the world as given to human brings. That is conceptually incoherent. Intelligence is always the creation of a world rather than a series of acts in an independently given reality. By definition, it would be impossible for artificial intelligence to operate in the external world as it is given to human beings since this world is the contingent product of historical change, and no model exists within the human historical process. It is hardly surprising that artificial intelligence can excel at coding, which is abstract, but not at activities constrained by historical contingency, such as natural language. New code is the way the world model realises its possibilities, typically in response to prompts.

Both language models and their video counterparts excel, above everything else, at building virtual worlds. The analogy of a video game is helpful. Language and video models are never stable across different realities. When it comes to video, there is a radical transformation as we move from the idea of filming the physical world to that of creating and recreating it, and never filming it or documenting it. The inevitable conclusion would seem to be that artificial intelligence will never get us any closer to understanding the world. It will get us farther and farther away from that goal because it is busy creating new worlds, worlds no longer grounded in reality. The fear is that world building will become a method for pursuing our fantasies and desires not by acting in the world but by creating worlds designed as conduits for wish fulfilment (see Freud, Interpretation of Dreams). All those videos generated by artificial intelligence which have invaded social media already present us with fake images of what we have always hoped the world could be like, the world as wish fulfilment. In artificial intelligence the world is always what we want it to be and it is so before we act.