❮❮prev

Random train of thought - ironically (or appropriately) enough derived from bouncing some philosophy around with ChatGPT:

I've long realized that the Turing Test makes humanity biased towards systems that can communicate in human text. If a new intelligence comes to us that can't communicate in English (which for better or worse is the modern "lingua franca") many folks won't be able to recognize it as intelligent.

But this reflects an assumption that basic intelligence would precede the ability to speak. You see this in Asimov's story "Robbie": the nanny robot is a delightful companion, physically graceful, emotionally expressive, but mute... and the new fangled talking robot has so much machinery to deal with the language that it sits immobile, taking up a whole room.

But here's the thing: human language isn't just the test for LLM - it's the power source! There is so much intelligence imbued in the training corpus; I hold to the idea that what's most important is the relations and interactions and thus the context things make for each other, and the (shifting) mathematical landscapes that an LLM chat traverses is so rich with those that it powers the novel problem solving abilities we see.

So language isn't just the measure of smarts, it's the cognitive structure the smarts are built out of.