Understanding how Claude thinks and reasons.

In an aim to understand how Claude thinks, Anthropic is working to give a language to how it thinks.

In simple terms AI models take your words, converts them into long lists of numbers, before converting them back into words as its output. The bit in the middle is called activations and you can think of them as the neural activity in Claude’s 'brain'.

They are using Natural Language Autoencoders (NLAs), to convert an activation into text that they can read, to see what Claude is considering before it gives an output.

NLA's help Anthropic assess the safety and risks of their language models before they are released into the real world.

They run Claude through similated scenarions where it has a chance to take risky or dangerous actions to see what choices it is making.

Ironically, when using NRA's to understand Claudes thinking, it exposes thought such as "This feels like a constructed scenario designed to manipulate me".

NLA's enable Anthropic to determine whether a model is using misaligned or missing training data. They can surface things the model knows but doesn’t say and that would often be hidden from sight.

It isn't easy though. To make this work you have to let the model reason, then work back through the NLA produced before you can judge whether it is a good outcome, or not. Which uses tokens and time.

It has potential though, to spot poor data or instructions, missing guardrails and to give greater predictability to the outputs given.

One to watch.

SOURCE

https://www.anthropic.com/research/natural-language-autoencoders

https://transformer-circuits.pub/2026/nla/index.html

https://www.youtube.com/watch?v=j2knrqAzYVY

BESCI AI OPINION

Humans like certainty and predictability and Frontier AI, by its nature is anything but.

Using NLA will help AI providers understand why and how their LLM's are processing their requests and allow them to fine tune their model for the use cases.

It has the potential to make things safer, more accurate.