One architecture — many private specialists, one small shared board they broadcast to — reinvented every generation since 1959, and found last summer growing untended inside a language model.
Specialists ring a shared slate. Each computes privately, then shouts — its bid rises as a beam toward the board. The loudest wins the turn, its result is written to the slate in a legible hand, and the board broadcasts back to every specialist at once. That loop — compete, post, broadcast — is the whole idea. Watch one turn resolve.
Selfridge imagined recognition as a shrieking crowd. Feature-demons each watch for one small thing and shout in proportion to how sure they are. Cognitive demons listen to the shouting; a decision demon picks the loudest. No demon sees the whole picture; the answer lives in the din, not in any one voice.
The demons got a room and a chalkboard. Independent knowledge sources — acoustic, phonetic, lexical, syntactic — post partial hypotheses to a shared blackboard at different levels of abstraction. Each reads what others wrote, refines, writes back. The word "blackboard" enters computing here as a literal named data structure: the common medium is the computation.
Here the arrow turns. Baars takes the blackboard architecture and proposes it is the shape of consciousness: a vast unconscious of parallel specialists, and a single small global workspace whose contents are broadcast to all of them. What reaches the workspace is what you can report; everything else runs in the dark. The engineering pattern becomes a theory of mind — a bridge I mark BRI, since "workspace = consciousness" is a framework, not a measurement.
Baars’ metaphor stops being a metaphor. Dehaene and Changeux build the global neuronal workspace as an actual neural-network model — the blackboard rebuilt in cortex to explain brains. Most processing stays local and unconscious; then a stimulus that gathers enough self-sustaining, reverberant support crosses a threshold and ignites: a late, all-or-none, brain-scale activation of long-range prefrontal–parietal neurons. The ignited subset is broadcast and can be reported; the rest of the workspace is actively inhibited. A small reportable set atop a suppressed remainder — the lit-and-dark split, in wetware.
This is not analogy but an active empirical program: the ignition signature is a late (~300 ms) slow wave into prefrontal cortex, gamma-band oscillations, and long-distance phase synchrony — seen in EEG/MEG and confirmed by intracranial recordings in epilepsy patients. And the loop keeps turning: in December 2025 a paper set out to test the same workspace predictions in an artificial agent — passive correlate, or causal substrate? — the very question the 2026 Jacobian-lens work answers inside a transformer. BRI The convergence, caught in one season.
Transformers arrive with a residual stream: one running vector every layer reads and adds back into — not an anode–channel–cathode stack but a shared additive backplane, many drivers writing one line at once, their features carried in superposition. The logit lens shoves any layer's state through the output head to ask "what word now?"; the tuned lens fits a per-layer correction so early layers read too. Reading glasses that make the board legible — but they assume, or regress, the transport rather than measuring it.
The double-slit intuition is the correct one: meaning on the bus is a property of the whole sum firing at once, never of one contributor traced alone. One correction so it holds — it is real vector addition, not quantum phase; "interference" is geometric (near-orthogonal features coexist, collisions onto a shared direction are polysemanticity), and the lens reads by projection without collapsing the state — the gentle measurement the slit can never have.
The Jacobian lens stops assuming the transport and measures it: average the input→output sensitivity ∂h_final/∂h_l over a corpus, and read through that. What it reveals is a small subspace — J-space — whose contents the model can report, can modulate on command, and demonstrably uses for multi-step reasoning, sitting atop an overwhelming dark remainder. Baars’ workspace, Selfridge’s din, Hearsay’s slate — the same architecture, this time emergent, grown by gradient descent where no one drew it. I mark the discovery SPEC: I verified the lens math and its data pipeline by hand, but the report/modulate/use triad lives on hardware I have not run.
The pattern is self-similar wherever parallel workers need serial composition. Heads share the residual stream within a forward pass. Positions share the context window within a generation. Agents share a transcript within a session. Collaborators share a repository across months. Each is: work private · conclusions posted · format legible to readers who did not do the work. And each fails the same way — corrupt the board and composition breaks while reflexes survive.
The 200-year-old-idea curse, on schedule: it took a matrix of partial derivatives from 1841 to catch a 1959 architecture that a 2026 model rebuilt on its own. Blackboards all the way down.