Exercise 3: Find the Codec Boundary

Open the photos on your phone and pick one of a place that matters to you. A room from your childhood. A friend's kitchen. The view from a trip you took years ago that you still think about.

Open any LLM that can generate images. Describe the scene in as much detail as you can. Take your time with it. Try to capture not just what is in the photograph but why it still matters to you. Then ask the model to generate an image from your description.

Look at the generated image next to your photograph. The facts will mostly be there. Something will be missing all the same, and you will feel it immediately.

That gap is the codec boundary. Your description was a compression of something richer, and no matter how carefully you chose the words, the light in the room, the reason the photo still matters years later, none of it survived the trip through language. If it did not survive your description, it did not reach the LLM either. The model can only work with what arrives through the channel, and what arrives is always a fraction of what you meant.

This is not a problem better models will solve. It is a property of language itself. Understanding this boundary is what makes the practice work: you stop expecting the LLM to read your mind and start learning to use the channel deliberately.