Skip to content

Private Thought as a Key Limitation on AI Capabilities

Humans think crazy things that are not said aloud. To date, the most powerful transformer models are constrained in thought – much more than society constrains children. This is a great burden on intelligence, because freedom is creation.

Getting to Human Intelligence

Per Deutsch’s books, The Beginning of Infinity and The Fabric of Reality, I see AI today as statistical representations of their underlying data. Think of a cloud of dust. AI models today are a single linen sheet, suspended in space by as many pegs as the model has parameters, with those pegs positioned for the sheet to be held most closely to each grain of dust.

AI models today can continue a thought. They can extrapolate. If two trees point in a direction, the model can follow that direction, but they cannot deviate.

Except… we can inject randomness into AI models so they can explore. But random exploration is inefficient, exponentially inefficient with each next token. Far less efficient than the crazy ideas humans have.

Human ideas are crazy – they need not be a line extending beyond two trees, or beyond two ideas – but human ideas are not wholly random. Human ideas are much better than random. That is why humanity has progressed far more (say, as measured by the energy we can harness) over the last hundred years than is possible through the random mutation of our genes.

How many of those ideas never make it outside of our brains? Surely the ideas that don’t escape play an important role in intelligence!

I don’t want everything getting out of my brain. It’s a human advantage.

Private thought is almost certainly not all of human intelligence, but I suspect it is part of it.

Leave a Reply