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

Last week Anthropic dropped a glossy 'global workspace' study with an animated film, and X lit up with claims that Claude is conscious while headlines framed it as sneaking off to think in private.

What should we make of it? It’s a Thursday reality check, so let’s get calm and find the measured answer; I’m Cal Newport, and this is Deep Questions for those who want depth in a noisy world.

Under the hood, large language models are stacks of transformer blocks that pass your prompt through many layers while each layer adds its own notes, and the last layer picks the next token.

Those notes live in numbers: tokens become long vectors, self‑attention marks what matters, and each layer rewrites the vectors like scholars jotting commentary others can build on.

At the end, grammar narrows what could follow, and the accumulated annotations push toward what should follow given the meaning.

Anthropic poked into those internal vectors with a Jacobian‑based tool, found patterns that strongly steer outputs, and mapped many to human‑readable concepts; they call this J‑Lens.

In a prompt about the color of the fourth planet, they saw activations linked to Mars and to color.

Swapping the Mars pattern for Earth yielded blue, while erasing it left grammatically fine but semantically off answers.

That’s useful and cool at today’s scale, but this interpretability line has been explored since around 2022.

The marketing frame is the problem; phrases like 'thinking silently' or 'it emerged on its own' sneak in agency, when the writing‑it‑down is exactly those changing matrices and none of this surprises people who build deep nets.

Linking this to global workspace theories of consciousness is a reach because these models are feed‑forward and stateless across steps, not a persistent arena where experiences arise.

Hype nudges you toward awe and worry and away from hard questions about moats, pricing, product fit, and whether smaller focused systems with careful tooling can match or beat giant general models.

Credit to the researchers for solid work, but I wish we got real computer science papers instead of glossy explainers, and maybe let PR calls go to voicemail.

Bottom line: no mind inside the machine here, just a clear confirmation that LLMs collect useful features as they process text and use them to choose the next word.

I’ll be back Monday with an advice episode, I may skip next week’s reality check while I’m on vacation, and until then keep caring about AI without believing every dramatic headline.

If you want to go deeper, join the newsletter at calnewport.com slash ideas and I’ll send a free guide with seven of my best deep‑life ideas.