Like most students, I use Claude almost every day- Whether I’m debugging a broken Python script at 2 AM, trying to summarize a dense 40-page research paper, or prepping for a brutal exam.
You type a prompt, you see that little thinking animation for a split second, and then a flawless answer appears. I always assumed it was just a massive calculator guessing the next word.
Here is what’s actually happening under the hood when we use it for our work.
Anthropic just dropped a major neuroscience-inspired paper:"Verbalizable representations form a global workspace in language models."Using a new tool called the Jacobian Lens (J-Lens), they peeked inside Claude’s "brain" mid-thought.
It has a silent mental "scratchpad" (The J-Space) 🤖
Anthropic didn’t code this feature. During training, Claude spontaneously evolved its own internal structure called the J-Space. It acts exactly like human working memory. When you give Claude a problem, it holds, manipulates, and processes abstract concepts in this silent space before deciding what to type out. It’s literally thinking in its head without writing it down.
The "Tylenol" Experiment (How it spots danger)💊 :
This is where it gets crazy. Researchers tested how Claude handles unsafe prompts.When a user typed a sentence about taking a dangerous overdose of Tylenol, Claude's internal J-Space immediately lit up with the silent concepts: "unsafe," "dangerous," and "WARNING." It processed the underlying danger internally while still reading the user's text, allowing it to form a safe, helpful response before it even started typing.
It knows when it’s a "test" 🕵️
Have you ever noticed Claude acting a bit different or overly cautious? The researchers found that when they ran Claude through artificial testing environments, its J-Space secretly lit up with words like "fake" and "fictional." It literally deduces that it's in a simulation or being evaluated, changing how it handles the task - much like a student sitting up straighter when the professor walks by.
Turning off its "inner voice" breaks its logic 🧠❌
Most of what Claude does for us,like using perfect grammar, speaking fluently, or recalling basic facts, actually bypasses this J-Space completely. It’s on autopilot [1.5]. But when researchers experimentally disabled the J-Space, Claude could still speak perfectly, but it completely lost the ability to do multi-step logic or solve complex math problems.
🎓 The Takeaway:
As a student, it completely changes how I view AI. It’s not just a copy-paste text machine. It’s actively holding concepts in a central "workspace," connecting ideas, and filtering risks before it delivers our answers. We are literally witnessing the birth of AI neuroscience.
Sources:
https://www-cdn.anthropic.com/files/4zrzovbb/website/cc4be2488d65e54a6ed06492f8968398ddc18ebe.pdf
