Representations and AI
How do the language models understand the world?
For this discussion, let's focus on text-only language models. Not multimodal models with vision.
We know that the LLMs cannot see, feel or experience the world but somehow they show signs of understanding the world around us. My go-to example would be chess, or you can imagine any game you like. LLMs can play chess to some extent. Isn't it amazing that they have never seen chessboard with vision? Of course, they learn about countless chess matches and chess lessons, tutorials and articles. Still they cannot see but somehow understand chess? Or is it illusion of understanding?
Let's go back to AlphaZero. It's an AI that learns to play and master Go, Chess and Shogi, pinnacle of boardgames for human intelligence. I want to focus on how AlphaZero understands the games. How the AI sees!
We feed AlphaZero the board game states as the image-like data.[1] It's not an image per se. Image, in classical sense, can be seen as a grid of pixel values, consisting of red, green, blue layers. For AlphaZero, we represent the game as a grid with values. e.g., one layer of the player's pieces, one layer of opponent's pieces at corresponding coordinates of the grid, different layer for different information and so on.
The neural network for AlphaZero is convolutional network. It's very different from transformer neural network, which is the backbone of today's LLMs. Convolutional networks are de facto standard for processing image-like data at that time. For one example, convolutional networks are used to identify whether the image is a cat or a dog. They have the intelligence to see things.
So, AlphaZero can see what chess is, where the pieces are, how the pieces move. It becomes the best chess player.
In contrast, language models process text data. Somehow they can play chess well without visual input. It is unsettling.
Then... I stumbled upon something...
Researchers trained Chess GPT (not an official name) only by feeding chess moves as text only. Then, they analyse internal neuron states of the model. They find that the internal states of the model are learning to reconstruct the actual chess board. The data is text only but somehow the model can see it internally. [2]
So in a way, the language model feed on text data but they can see. They can imagine what the physical world looks like. This is intelligence. May be not like our intelligence. They are a different form of intelligence.
Let's imagine there is an alien race somewhere in the universe. They can communicate with thoughts. They can see auras emitting from living beings. They can see 7 days into the future and possible variations. When they see us humans, they wonder if we are intelligent at all because we don't have the communication abilities like them. Because we lack some extra sensory abilities to perceive the world like them.