AI models have more than they show

Work in progress

It was the summer of 2025 when I started going deep into AI. Up until then, they had been presented to me as next-word predictors. However, the more I used them, the more I noticed that there was much more to them. Next-word prediction just happens to be the way we ask them to produce output.

This belief was reinforced by Anthropic's interpretability study of these models, published on its Transformer Circuits blog.

On the Biology of a Large Language Model

Anthropic investigates the internal mechanisms used by Claude 3.5 Haiku through attribution graphs.

transformer-circuits.pub/2025/attribution-graphs/biology.html

July 7, 2026 · 11:34 PM

Today, Anthropic released new research showing that models have an internal thinking space, much like humans do.

This reinforces my view that:

  1. We need more people investigating the inner workings of these models.
  2. We need a better output paradigm than next-token prediction.