Nicholas Andresen

Dancers don't rehearse choreography by thinking next I will contract the left quadriceps while rotating the hip flexor 17 degrees.

They just dance.

What's happening is something more like high-dimensional pattern recognition.

Intuition's that fire before language catches up, a felt sense that doesn't decompose into words.

Language often comes after the real work, if it comes at all.

Something similar happens inside models.

As a model computes its next token, information passes through the network as several thousand-dimensional vectors, encoding uncertainty, alternatives, the half-formed feeling that something might be wrong.

But chain of thought forces the model to squeeze all of that into text, one token after another, an information compression ratio of over 1,000 to 1.

Much of the information is lost.

So researchers asked, what if we skip chain of thought?

What if models could think in their native format, vectors morphing into vectors, activations flowing through layers, all in some high-dimensional space that was never meant for human consumption and only emit English when they're done?

Projects like Coconut and Huggin are exploring exactly this.

Researchers call it vneuralese.

Neuralese isn't like thinkish.

Thinkish is compressed and strange, but it's still language that you could, in principle, parse with enough effort.

Neuralese isn't a language at all.

The reasoning happens in continuous internal latent states, in thinking that is never words.

You can't read Neuralese.

You can only watch what goes in and what comes out.

The good news.