What a language model doesn't know it doesn't know

This is placeholder text so the site has something to show. Replace it with a real essay before launch.

A language model is very good at continuing a sentence. It is much less good at knowing when it should stop. The fluency that makes it useful is the same fluency that hides the edges of what it has learned.

Confidence is a style, not a measurement

When a person hesitates, we hear it. The voice drops, the sentence trails, the hands move. A model has no such tell. Every answer arrives in the same even tone, whether it is drawing on a thousand examples or none.

That matters because we read confidence as evidence. We always have. The brain uses fluency as a shortcut for truth: if something is easy to process, it feels more likely to be right.

Where attention goes

The interesting question is not whether a model is wrong, but where its attention went when it was. Often the answer is somewhere sensible and slightly to the side of the question that was actually asked.

We built machines that sound like they are paying attention. Learning to tell when they are is now our job.

More soon. For now, this is a card in the AI and Tech drawer, waiting for its real letter.