Interpretability
Also called mechanistic interpretability
Research into what is actually happening inside a model: which internal features and pathways produce a given answer.
Nobody programs what happens inside a neural network; it emerges from training. So even the people who build a model can't simply look up why it gave an answer. Interpretability researchers try to find out, by identifying the internal features a model uses, like a concept for "rabbit" or for "the user is upset", and tracing how they combine.
Results so far are striking and partial. Anthropic has shown that Claude plans some words ahead and that it can describe its own method inaccurately. The same researchers say their tools capture only a fraction of what a model computes, even on short prompts.
For example
Researchers found that when Claude writes a rhyming couplet, it picks the rhyming word for the end of the second line before it starts writing that line.
Put it to use
Where you'll meet interpretability in ChatGPT, Gemini and Claude, how to use it there, and the mix-up most people make.
Daily Bytes Plus is in the Android app.