Who We Are我们是谁

Language is the world

AI can be trained at all because it is given a structure to fit.

Take Wittgenstein, one of the greatest philosophers of the twentieth century. The central question of the Tractatus is how a sentence states a fact. His answer was the logical picture, and from there he moved to the meaning of life. His account ended in failure, but his thinking lit up a road that was right.

The root of that question is what the content of knowledge actually is. Humans invented language in order to find an equivalent way of expressing the inner structure of the world. “The sum of the squares of the two legs of a right triangle equals the square of the hypotenuse” — the detail carried inside that sentence describes an intrinsic geometric property of the physical world. The early Wittgenstein said that language is the world. He meant that we represent the world with language: how you use language is what your world is.

y = f(x), where y is language, f is the rules of grammar, and x is the natural semantics we want to express — the physical world.

An LLM fits the high-dimensional mathematical structure inside language, y.

A world model, in the most original sense of the term, fits the high-dimensional mathematical structure inside the physical world, x.

Some structures of the world cannot be put into language at all unless you live through them: red, pitch, the feeling of joy.

The world does not have many dimensions — eleven at most. An LLM needs so many parameters in order to store more information, and to represent a subtler landscape.

Intelligence, in essence, is that high-dimensional mathematical structure itself.

Neural networks

Next-token prediction works because the drive inside it forces the network to carve out the correct landscape in semantics. RL on top of an LLM works because the drive inside it forces the network to correct itself toward solving the task better.

The main path to intelligence requires two things:

1. Self-supervision, to build the basic structure of the world and of language.

2. Reinforcement learning, to raise capability.

Whether the input is conditioned on action, and whether the output carries future and past world states, does not matter.

The question

What product do we want?

We want autonomy: an agent with self-awareness, initiative and judgment.

We want emotion: personality, desire, unpredictable tension.

We want possibility: resistance to alienation and oppression.

We want creation: more complexity, and stability outside the data distribution.

We want exploration: to let life bloom among the stars.

From the day Infimory was founded, our view of the product has been in tight alignment with our view of the world. This is our mission, our goal, and the starting point of every strategy we have.

The brain

We will build emotion, impulse and desire, following the basic needs of a living thing. From these come personality, logic, judgment and action. Physical AI will be more than a marionette; we will give it a soul.

Through world models and reinforcement learning — by rebuilding the objective, the model structure and the training paradigm — we will make physical AI create real value.