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

July 16, 2026

AI Summary

5 min read

Anthropic released a research report last week claiming to have found evidence that Claude has developed a "global workspace" — a hidden internal space where it "thinks about a concept without writing it down." The tech press ran with headlines like "Claude has carved out its own space to ponder" and "Anthropic found a hidden space where Claude puzzles over concepts." On X, the reaction was predictable: "Claude, my friends, by all counts, is a conscious entity."

Cal Newport is not buying it.

How Large Language Models Actually Work

To understand why this research is being oversold, Newport walks through the basic architecture of a large language model. Under the hood, an LLM is a sequence of transformer blocks arranged in layers — GPT-3 had 96 of them, newer models likely have more. When you submit a prompt, it gets broken into tokens (words or parts of words), and each token gets embedded into a high-dimensional vector — essentially a long list of numbers that captures the token's meaning.

Continue reading the full summary in the app — free to try.

Read Full Summary →

Free • No credit card required

Listen to Audio Summary Open in App

Never miss an episode of Deep Questions with Cal Newport

Get every new episode summarized in your inbox — free, ~5 minutes to read.

No spam. Unsubscribe anytime.

What you'll learn

1 (00:00) **Anthropic's New Report and the Hype** - Cal introduces the controversial Anthropic report and the breathless reactions on X and in tech media, including claims that Claude is a conscious entity.
2 (01:39) **The Report's Own Description** - Cal reads directly from the Anthropic paper's introduction, which describes finding a "J Space" of internal neural patterns that act like silent thoughts.
3 (03:14) **High-Level Tutorial: How LLMs Actually Work** - Cal begins a detailed explanation of the transformer architecture to demystify the report's claims.
4 (06:09) **Deeper Mechanics: Tokens, Embeddings, and Annotations** - Cal explains how prompts are broken into tokens, embedded as long vectors of numbers, and then annotated as they pass through layers.
5 (10:44) **Two Key Processes: Syntax and Semantics** - Cal breaks down what the layers are actually doing: learning syntax (grammatically valid next words) and semantics (which word makes sense contextually).
6 (13:36) **Returning to the Anthropic Paper: The J Space Method** - Cal explains the mathematical tool (Jacobian) used to decode the numerical annotations into human-interpretable concepts.
7 (15:53) **Concrete Example: "The color of the fourth planet from the sun is..."** - Cal shows a specific example from the paper where the J Space revealed annotations for "Mars" and "color," leading to the output "red."