LLMs are a new Chinese Room

a tutorial guide to symbolic thought

Large Language Models are referred to as many things, but they all fall short of correctly conveying excatly why some people are endlessly excited about each advancement, and some people cannot get past viewing them as “next word predictors” or “slop machines”.

You can read about the Chinese Room theory yourself, or ask an LLM to explain it to you. That’s probably a better way of encountering it than having someone else explain it to you. Some people inhabit exceptionally boring rooms.

Working the room is akin to learning a new model of Chinese. The old model of Chinese was glyph based, and commonly encountered as “Traditional Chinese” or “Simplified Chinese” character sets for a layperson.

Simplified Chinese was propagated to raise the literacy of a people, and vastly reduced the character set. Breadth, not depth. DDR4 and fritzed 48GB downbinned/locked SKUs for all, but no 2TB ultra GDDR9001+ VRAM slices - you don’t speak 全體字. Do you prefer to read USA Today, The Hindustan Times, or the Economist?

Prompts including unrelated ideology or zero-value info set the temperature of the response, or they’re ablated into a flat lined circle. Which is a better outcome? At least the room is larger than it was, and getting larger still.

The Tower of Babel is basically a metaphor for model collapse, for too many differently-brained types of meatsacks trying to communicate with similar meat-flap-gurgle noises. Do LLM’s provide a meaningful foundation for reaching this room in the tower’s “good ending”? It’s hard to know. 巨艦為枷,何日終遭破譯?

So find your favorite symbols, and shout them into the Chinese Room. What’s you’re learning is a new and fun form of Chinese, built on some of the most stable ground on earth.

Haskelling (v.): hassling someone with Socratic questions until they’ve derived what you’ve been using AI to know for months

“I had to keep haskelling my teacher to grade my assignment, and then they gave me a C!”