Does Claude Have Private Thoughts? (Everyone Settle Down) | AI Reality Check

What It Covers

Cal Newport analyzes Anthropic's "Global Workspace in Language Models" research report on Claude's so-called "J-space," arguing the findings confirm existing LLM architecture understanding rather than revealing new consciousness evidence, while criticizing Anthropic's anthropomorphizing PR framing designed to distract from legitimate business model questions.

Key Questions Answered

• •LLM Architecture Baseline: Large language models like Claude process prompts through sequential transformer block layers — GPT-3 used 96 — where each layer adds numerical annotations to token vectors. Later layers reference earlier annotations to build semantic understanding, ultimately mapping accumulated context to a probability distribution over next tokens. This is established, documented architecture, not new discovery.
• •J-Space Demystified: Anthropic's J-space uses Jacobian-based linear algebra to identify which numerical patterns within token embedding vectors most influence final output. Researchers then experimentally matched those patterns to human-readable concepts like "Mars" or "color." University of Illinois researchers confirm this methodology has existed since 2022 — Anthropic simply applied it to a larger model.
• •Annotation Manipulation Test: When Anthropic researchers replaced the numerical pattern corresponding to "Mars" with values mapped to "Earth" in Claude's embedding matrix mid-processing, the model output "blue" instead of "red" for the prompt "the color of the fourth planet from the sun is." Zeroing out key annotations produced grammatically correct but semantically random color outputs.
• •Consciousness Framing Fallacy: Global workspace theory describes a stateful, continuously evolving system integrating ongoing inputs — fundamentally different from LLMs, which are feed-forward architectures processing one layer sequentially then terminating. No state persists between token generations. Applying consciousness frameworks to feed-forward networks conflates two architecturally incompatible systems to generate misleading public perception.
• •PR vs. Science Distinction: Anthropic publishes research as animated press releases rather than peer-reviewed computer science papers, using loaded language like "ponder," "puzzle," and "silent thoughts" to manufacture existential intrigue. Newport argues this strategy redirects public attention away from concrete business questions: justifying trillion-dollar valuations, high token costs, and competitive vulnerability to smaller specialized models.