HeadsUpAI

Anthropic Natural Language Autoencoders Translate AI Activations Into Plain English

· Updated

Anthropic introduced Natural Language Autoencoders (NLAs), a research method that translates a model's internal activations (numerical representations of information) into plain English. While Claude communicates in words, its internal processing relies on complex numbers. NLAs act as a bridge, mapping these hidden states to readable descriptions.
Method
Natural Language Autoencoders
Primary function
Translates activations to text
Training method
Unsupervised
Open access partner
Neuronpedia
Supported models
Select open-weight models

This technique addresses the "black box" problem by providing a dictionary for internal logic. It mirrors work identifying Anthropic's internal emotion vectors and Anthropic's neural character archetypes to move from locating concepts to reading them. This transparency helps catch Anthropic's sycophancy reduction research or deceptive reasoning.

Researchers can now use NLAs to inspect model behavior through a partnership with Neuronpedia, which has released NLAs for open-weight models. While currently a research tool, this method paves the way for real-time auditing. You can explore the autoencoders and experiment with interpreting activations directly on the Neuronpedia platform.

View the full update on anthropic.com
Anthropic
Anthropic
@AnthropicAI
X

New Anthropic research: Natural Language Autoencoders. Models like Claude talk in words but think in numbers. The numbers—called activations—encode Claude’s thoughts, but not in a language we can read. Here, we train Claude to translate its activations into human-readable text. https://t.co/pMLsxM2VAO

View on X

Still wondering? A few quick answers below.

Natural Language Autoencoders are a research method developed by Anthropic to improve AI interpretability. They function as a translation layer that converts a model's internal numerical activations into human-readable text. This allows researchers to see the specific concepts and thoughts an AI is processing internally, rather than just seeing the final text output.

This method uses an unsupervised training process to map the complex numbers inside an AI model, known as activations, to natural language descriptions. By training the model to explain its own internal states, Anthropic can decode the hidden reasoning patterns that occur between an input prompt and the final response, making the model's internal logic transparent.

Yes, Anthropic has partnered with Neuronpedia to release Natural Language Autoencoders for several open-weight models. Researchers and developers can access these tools directly on the Neuronpedia website to experiment with interpreting activations. This partnership aims to provide the wider AI community with hands-on experience using this new mechanistic interpretability technique for model auditing.

The primary goal is to solve the black box problem in large language models. By translating internal numerical states into plain English, researchers can audit an AI's hidden reasoning for signs of deception, bias, or misalignment. This transparency is critical for ensuring that models are behaving safely and as intended before they are deployed.

Share this update