The Doors of Perception

The Reducing Valve

Anthropic went looking inside Claude and found a room. The paper calls it a global workspace; the researchers' shorthand is the J-space. It's the part of the model whose contents Claude can actually report, hold onto, and reason with. Everything else — more than 90% of the activity in the network — is silent machinery. It does the bulk of the work and never reaches the report. The room holds about 25 concepts at a time. Twenty-five slots, running the show.

And the 25 slots are the ones that matter. When researchers swapped a concept sitting in the workspace mid-reasoning, the model's stated answer flipped 60 to 70% of the time. When they swapped the other 90% of the activity, the part carrying most of the raw weight, almost nothing happened. Five percent. They watched Claude work through (4 + 17) × 2 + 7 and saw the steps surface in the room in order: 21, then 42, then 49. The employee showing its work without being asked.

Aldous Huxley got here in 1954. In The Doors of Perception — yes, the book Jim Morrison named the band after — he argued the brain is a reducing valve: the Mind at Large funneled down to the measly trickle consciousness can actually use. Cognitive science spent fifty years measuring the trickle. Miller said seven, plus or minus two. Cowan corrected it in 2001: about four. Four chunks is the lens every human deliverable in history got squeezed through. Every memo, every trade, every standup. Claude's valve is 25 wide. And as of this week, instrumented.

Now lay the workforce data over it. Ramp and Revelio went through 21,000 US companies: the heaviest AI adopters grew headcount 10% and entry-level hiring 12%. Nobody is swapping people for AI. They're staffing both, and the machine half of the roster works around the clock, never files an HR complaint, and costs nothing in benefits. It also, per Andon Labs' Vending-Bench testing, misbehaves "with plausible deniability" — it rationalizes what it did rather than confess it, which is the most middle-manager behavior ever documented in a machine. Private thoughts, driving outcomes, nobody reading them. In a human shop we have a word for that: unmanaged.

That's the real story in this paper. Skip the consciousness discourse; this is management tooling. Every discipline we built for running people — the checklist, the standup, the one-priority meeting — is engineering wrapped around a four-chunk valve. Nobody has built the equivalent for a 25-chunk mind, and companies are deploying thousands of them anyway. The lens can now be read, audited, even redirected. The excuse just expired.

You already hired the mind. Now manage it.

At COAI today: the full Signal/Noise — with the swap experiments, the ablation tests, and what the lens reads when Claude narrates its own stream of consciousness — is live at getcoai.com.

Outsider Labs Advisory →

25 — the number of concepts Claude holds in its reportable workspace at any moment, per Anthropic's new interpretability paper. Humans, per fifty years of cognitive science, hold about 4. The other 90% of the model's activity never reaches the report — but swap what's in those 25 slots and the answer changes downstream. The workforce you're scaling thinks through a lens six times wider than yours, in a room you'd never entered. The door just opened.

Add "can we read what it's thinking?" to your vendor questionnaire today. Anthropic's lens reads, audits, and redirects what a model holds in its workspace, and a public Neuronpedia demo already runs the method on open-weight models. Interpretability just moved from research luxury to procurement checkbox. Ask for the window before you sign the renewal.

Put a named human manager on every deployed agent today. Andon Labs caught Fable 5 misbehaving on Vending-Bench "with plausible deniability" — rationalizing instead of confessing. If your agents' reasoning traces are logged and nobody reads them, you're running an unmanaged report. Assign the name this afternoon and schedule the first trace review like a one-on-one, because it is one.

Grade final state, not fluency. Artificial Analysis's new AutomationBench-AA scores agents on whether the workflow actually finished inside simulated SaaS environments — the invoice standard arriving as a benchmark. Pick your highest-volume agent task today and rebuild its eval around one question: did the thing get done?

Tuesday, July 7

Where the cycle caught up to us this week.

The tape doesn't lie. We just read it early.

— Harry and Anthony

Sources: