Alex C. Christou works across AI development, interpretability research and neurotechnology. This page explains the distinction behind the language example on the homepage: how difficult a sentence is to read, and how a language model processes it.
Readability and interpretability ask different questions
Readability concerns the effort a reader needs to understand a text. Sentence structure, unfamiliar vocabulary and qualifications can all affect that effort. Interpretability concerns the internal processes of a model. A clear output alone does not explain which representations or computations produced it.
These questions overlap when models analyse complex language, but they should be kept distinct. A model can produce a fluent summary without showing how it reached that summary. Equally, a diagram can make a sentence easier to discuss without revealing anything about a trained model.
A monetary-policy sentence
The Committee judges that a further withdrawal of monetary stimulus may be appropriate.
The word “may” qualifies the statement. It presents a possibility rather than a commitment. “Judges” identifies the statement as an assessment, while “further” relates the proposed withdrawal to an earlier change. These small choices matter when reading language that is deliberately precise and hedged.
The interactive example on the homepage lets readers select words and examine these relationships. Its connections are illustrative and hand-written. They are not measured attention weights, model activations or evidence of a model’s reasoning.
What would count as model evidence?
An empirical interpretability claim needs a specified model, an input, a method and results that can be examined or reproduced. It also needs a clear account of what a measurement supports and what it leaves uncertain. The homepage illustration supplies none of those measurements and should be read as an explanation of the research question.
This distinction is useful beyond monetary policy. When an AI system supports a business decision, readers need to distinguish the source evidence, the system’s output and any explanation offered for that output. Keeping those separate makes it easier to inspect a claim and recognise missing information.
Follow the work
For Alex’s current work and background, visit the main profile or Alex C. Christou on LinkedIn.
Alex also writes BCI, Decoded, a newsletter on the maths behind brain-computer interfaces and developments across neurotech.