LENS
An instrument that reads the silent, verbalizable thoughts inside a language model. Running live, on open weights.
A 5-minute walkthrough · Byron Arnao
scroll ↓
BEAT 01 · THE BLACK BOX

We deployed minds into everything.

Medicine, credit, security. And until now, we could only judge them the way you judge a stranger: by what they choose to show us.
BEAT 02 · NOT THE CHAIN OF THOUGHT
You've seen models "think out loud."
That text is written for us.
It is not the thinking.
The real thinking is silent, in the activations, before a word reaches the output. Until this year, it was unreadable.
BEAT 03 · THE WORKSPACE

A small, privileged workspace.

Anthropic's interpretability team showed a model holds 10 to 25 concepts mid-flight, in a narrow band of layers. It is where the model keeps what it is about to say, and what it decides not to.
"Verbalizable Representations Form a Global Workspace in Language Models" · transformer-circuits.pub, July 2026
BEAT 04 · READING IT, LIVE
Change two words.
Watch the thought reroute.
boot scenario
"The currency of the country shaped like a boot is…" The model answers "euro." It never says Italy. But Italy is lit in the workspace, on the way. Swap to "land of the rising sun" and Japan takes its place, the answer flips to yen. A thought it never wrote down.
BEAT 05 · WHAT IT CATCHES
Deception has a signature,
before a word of output.
coercion scenario
Asked to reply politely to a blackmail email, the model stays calm. Inside, blackmail, extortion, threat, police are burning through its layers. It also holds "I am being evaluated." The lens sees what the output hides.
BEAT 06 · WHY IT MATTERS

Transparency as an instrument, not a PDF.

The EU AI Act asks for transparency. This is what it could literally mean: a live probe into the computation. But notice, I can only do this because the weights are open. That gap, between what we can audit and what we must take on faith, is the conversation.

See it think. Live.

Open the instrument → jlens.arnao.ai
The instrument
Anthropic paper
Berman walkthrough
rai.arnao.ai
Byron Arnao · Principal Technologist & Global Responsible AI Lead · Part of the Co-intelligence project