Does Claude Have an Internal Reasoning Workspace?

TL;DR

Anthropic reports that Claude contains a small internal workspace called JSpace, where concepts can appear before they are expressed and can influence later reasoning. Editing or removing these internal representations changed answers and impaired complex tasks, but the findings do not establish phenomenal consciousness or prove that Claude has subjective experiences like a human.

Transcript

Right now, your brain is doing way more work than you feel. It is keeping your body balanced, controlling your breathing, moving your eyes across this line, turning shapes on a screen into language, and deciding whether what I'm saying is worth your time. But almost none of that is visible to you. Most of the system runs silently in the background.... Read More

Key Insights

• JSpace is a small internal workspace identified within Claude that contains less than one-tenth of the model's internal activity and holds only dozens of concepts at once. Despite its limited size, it becomes important when tasks require deliberate reasoning rather than routine language processing.
• The Jacobian lens maps vocabulary words to directions in Claude's residual stream. Stronger activity in one of these directions makes the corresponding word more likely to appear later, allowing researchers to interpret internal concepts that are active before or without being written in Claude's response.
• Claude can internally represent judgments that never appear in its output. JSpace reportedly surfaced error when Claude examined buggy code, biological function from raw protein letters, injection and fake when search results contained a prompt attack, and ordered intermediate concepts during multi-step mathematics.
• JSpace has a causal relationship with some reported answers, according to intervention tests. Replacing the internal soccer pattern with an equally strong rugby pattern changed Claude's answer to rugby, indicating that the model's report was reading from the altered internal representation rather than merely displaying an independent decision.
• Claude can maintain one internal focus while producing unrelated visible text. When instructed to copy a sentence about painting while silently focusing on citrus fruits, it copied only the sentence, while internal patterns associated with orange, fruits, thinking, and imagery appeared inside JSpace.
• Intermediate concepts in JSpace can guide subsequent reasoning steps. For a question about the legs of an animal that spins webs, spider appeared internally, and replacing it with ant changed the answer from eight to six, showing that later computation used the workspace's contents.
• A shared concept in JSpace can support several different tasks. Replacing France with China caused answers about capital, official language, continent, and currency to change together, suggesting that different systems accessed one reusable country representation instead of relying only on separate task-specific facts.
• JSpace is not required for every language capability. Ablation preserved fluency, grammar, sentiment classification, simple textual fact retrieval, and some multiple-choice answers, but multi-step reasoning fell almost to zero, while summarization and rhyming poetry deteriorated substantially.
• More videos from this channel:
  • How Do Realistic AI Robots Change Global Power?
  • How Does DSpark Make AI Inference Faster?
  • How AI Is Shaping Music, Film, and Earth Tech
• More from this channel: AI Revolution

Questions & Answers

Q: What is JSpace inside Claude?

JSpace is the name Anthropic researchers gave to a small region of Claude's internal representations that appears to function as a shared workspace. It accounts for less than one-tenth of the model's internal activity and can hold only dozens of concepts simultaneously. Concepts placed there can become available for reasoning, reporting, and use by different parts of the network.

Q: How does the Jacobian lens read Claude's internal concepts?

The Jacobian lens, also called J lens, calculates a direction in Claude's residual stream for each word in the model's vocabulary. The residual stream is the main channel that transformer layers read from and write to. When activity strengthens along a word's direction, that word becomes more likely later, enabling researchers to interpret concepts active inside JSpace.

Q: Does JSpace prove that Claude is conscious?

JSpace does not prove that Claude has humanlike consciousness or subjective experience. The findings concern accessible internal content that can influence reasoning and be reported, which resembles the access-conscious activity described in global workspace theory. The transcript explicitly distinguishes this functional property from phenomenal consciousness, meaning whether anything is actually experienced from the inside.

Q: How did researchers test whether JSpace causes Claude's answers?

Researchers first observed soccer in JSpace before Claude named soccer as the sport it was silently considering. They then removed that internal pattern and replaced it with an equally strong rugby pattern while leaving the rest of the model unchanged. Claude answered rugby, suggesting that its final report depended on the edited workspace rather than an unrelated decision elsewhere.

Q: Can Claude think about something without saying it?

Claude can hold a requested concept internally while producing different visible text, according to the reported tests. When asked to copy a sentence about painting while silently focusing on citrus fruits, Claude output only the painting sentence. Meanwhile, JSpace contained patterns associated with orange, fruits, thinking, and imagery, separating internal focus from the words shown to the user.

Q: Does Claude perform silent intermediate calculations in JSpace?

The reported experiments indicate that some intermediate calculations appear in JSpace without being written. While Claude copied an unrelated sentence and mentally evaluated 3 squared minus 2, nine appeared internally first and seven appeared in later layers. In another task, spider emerged as an intermediate concept before Claude answered that a web-spinning animal has eight legs.

Q: What happens when researchers remove JSpace from Claude?

Removing, or ablating, JSpace did not make Claude produce nonsense. The model retained fluent speech, grammar, sentiment classification, simple fact retrieval from text, and some multiple-choice performance. However, multi-step reasoning dropped almost to zero, summarization became worse than that of a much smaller complete Claude model, and rhyming poetry also deteriorated substantially.

Q: How does JSpace separate automatic language processing from reasoning?

A Spanish-language experiment illustrated the distinction. After researchers replaced the Spanish concept in JSpace with French, Claude identified the language as French and selected Victor Hugo instead of Gabriel Garcia Marquez when asked for a writer in that language. Yet it still continued the original text in fluent Spanish, suggesting automatic continuation operated separately from reasoning about language identity.

Summary & Key Takeaways

• Anthropic researchers identified JSpace, a small region of Claude's internal representations that appears to make selected concepts available for reasoning and reporting. Using the Jacobian lens, they associated internal activation directions with vocabulary words, allowing them to detect concepts that Claude had not written or otherwise revealed in its visible response.

• Intervention experiments suggest JSpace is functionally involved in Claude's outputs. Replacing an internal soccer representation with rugby changed the reported sport, while substituting ant for spider changed an answer from eight legs to six. Injected concepts were also recognized, and silently requested thoughts appeared internally without entering the visible output.

• Removing JSpace preserved fluent language, grammar, sentiment classification, simple fact retrieval, and some multiple-choice performance, but severely damaged multi-step reasoning, summarization, and rhyming poetry. These results support a distinction between automatic processing and deliberate reasoning, while stopping short of demonstrating subjective experience or humanlike phenomenal consciousness in Claude.