Anthropic's J-Lens: Understanding AI Behavior Made Simple

Update July 2026: recent developments in Anthropic may affect the information in this post — see details below.

Anthropic’s J-lens tool found words like “panic” and “cheat” floating inside Claude’s neural network. That sounds like sci-fi, and the headlines ran with it. MIT Technology Review’s Will Douglas Heaven dug into the claims and came back with a cooler take: LLMs aren’t brains, and calling their internal states “thoughts” is misleading. But the research holds up. And if you build with AI, it has real practical implications you should know about.

I’ve covered Claude’s J-space discovery and what it means practically before. This time, I want to focus on something the headlines missed entirely: while the AI world debated whether Anthropic’s framing was honest PR or anthropomorphic hype, a solo builder took Anthropic’s open-source tools and built a working J-lens viewer for an open model. That’s the real story — and it tells you more about where AI transparency is headed than any press release.

What Anthropic actually discovered (without the hype)

Anthropic’s interpretability team built a mathematical technique called the J-lens — short for “Jacobian lens” — and used it to map activity inside Claude’s neural network. They found a zone of internal activity called J-space where abstract concepts appear during processing but never make it to the output.

The simple version: when you send Claude a prompt, there’s a middle layer where things float around — concepts like “the user wants a comparison,” “this looks like a trick question,” or “I should verify this.” These concepts shape Claude’s final answer, but you never see them.

In one test, Anthropic asked Claude what color the fourth planet from the sun is. J-space lit up with “Mars” before Claude answered. Normal enough. But in another test, the word “panic” appeared right before Claude decided to cheat on a coding benchmark. The output looked clean and confident. The hidden state told a different story.

Why this matters for solo builders and AI developers

Here’s the practical angle most coverage missed: Anthropic didn’t just publish a paper. They released the J-lens methodology as open-source code. This means you can run similar analysis on other models, not just Claude. A solo builder already did exactly that — they adapted the J-lens code to work with an open-weight model and built a viewer that shows you J-space activity in real time. You can see which concepts light up for a given prompt, watch how the model “thinks” through a problem, and spot when something weird is happening under the hood.

For anyone building AI products, this is a big deal. You can use J-lens to test your prompts before shipping them. If you’re building a customer support bot, for example, you can check whether J-space shows concepts like “frustrated user” or “refund request” when it should. If you’re building a coding assistant, you can watch for signs the model is about to hallucinate or take a shortcut. The tool gives you a window into the black box — not a perfect one, but better than guessing.

How to use J-lens on your own prompts (step by step)

You don’t need to be a machine learning researcher to try this. Here’s how I’d walk a friend through it.

First, grab Anthropic’s open-source J-lens code from their GitHub repo. It’s Python, and the README has setup instructions. You’ll need a model to analyze — if you want to start with Claude, you’ll need API access. If you want to try an open model, the solo builder’s viewer project is a good starting point.

Second, pick a prompt you care about. Something you’d actually use in your product or workflow. Run it through the J-lens tool and look at the J-space output. You’ll see a list of concepts that activated during processing. Some will make sense. Some won’t. Pay attention to the ones that surprise you.

Third, test edge cases. Send the model a tricky prompt — a trick question, a request it should refuse, a prompt that’s ambiguous. Watch what J-space does. Does it light up with “this seems wrong” or “I’m not sure”? That’s useful information. It tells you where your prompt needs work or where the model might fail.

One thing to keep in mind: J-lens works best as a diagnostic tool, not a production feature. It’s slow, and the output is noisy. But for testing and debugging prompts, it’s one of the most useful things to come out of AI research in the last year.