Information Processing Systems

Information ProcessingSystems

Twenty-five information processes, drawn as mechanisms and judged against Newell and Simon's 1956 criteria.

Shingai Thornton · Halcyonic Systems · September 2026 · work in progress

What these have been called

“artificial intelligence”Dartmouth proposal · 31 August 1955³

What this page calls them

information processesNewell & Simon · 2 May 1956¹Their phrase, 1956.¹

Information here means Shannon information (1948). Every Input and Output is a symbol stream or signal. No operation uses what a symbol stands for.²

In May 1956 Newell and Simon set three tests for what should count as a complex information process. Every process on this page is graded against tests ① and ③; the marks sit on each card, and the verdict is at the close.

“Even though our language must still remain vague, we can at least be a little more systematic about what constitutes a complex information process.

A complex process consists of very large numbers of subprocesses, which are extremely diverse in their nature and operation. No one of them is central or, usually, even necessary.

The elementary component processes need not be complex; they may be simple and easily understood. The complexity arises wholly from the pattern in which these processes operate.

The component processes are applied in a highly conditional fashion. In fact, large numbers of the processes have the function of determining the conditions under which other processes will operate.”

Newell & Simon, “Current developments in complex information processing,” RAND P-850, 1 May 1956, p. 6. Lecture read 2 May 1956.¹

25 of 25 processes

Held as · Network weights: connectionist. Asserted rules and graphs: symbolic. Estimated coefficients: classical statistics. Tables and populations: the search room’s own store. The connectionist / symbolic split concerns the Given field, marked on each card.^31,1

1956 marks · Filled: criterion satisfied. Empty: not established. Judgments concern the defining paper’s architecture.¹

① Many diverse subprocesses, none central. ③ Processes selecting which processes run. ② Simple components, complexity from their pattern: every process qualifies, so cards name both rather than mark them.¹

Sliding a filter across the input. A shared kernel forms a weighted sum at each image location.⁶

yᵢⱼ = Σᵤᵥ Kᵤᵥ xᵢ₊ᵤ,ⱼ₊ᵥ

K is fixed during operating and adjusted during fitting.

Attention. Query–key products set the weights applied to value vectors.⁷

Attention(Q, K, V) = softmax(QKᵀ / √dₖ)V

Relation weights depend on the input even when projection parameters remain fixed.

Carrying a state forward. A state-space recurrence carries a state between readings.¹³

hₜ = Āhₜ₋₁ + B̄xₜ yₜ = Chₜ

S4 uses structured state-space parameters. Selective variants make parts of the update input-dependent.^13,14

Intervention. do(X = x) replaces the structural assignment for X and removes its incoming causal arrows.¹⁸

Xᵢ = fᵢ(paᵢ, Uᵢ) → X = x

The asserted graph supplies the causal structure.

Search and sampling. Temporal-difference updates alter value estimates. Reverse diffusion transforms noise through successive conditional distributions.^22,27

The bead denotes a selected message or signal sample in a chosen channel.²
Figures are schematic. Kernel positions, relation thicknesses, pixel textures, forecast curves and objective bars are authored display states.
Scrubbers mark normalized progress. They compare no runtime, work, accuracy or energy.
LEARNING animates parameter adjustment and an illustrative objective bar. OPERATING freezes parameters while signals or states propagate.
Asserted models and production rules have no native training procedure in the cited formulation. Their LEARNING view stages external calibration or Given edits.
Eight rooms each have an orbitable schematic and an SVG view.
③ is the discriminator: input-dependent routing or tests select operations, with gated variants marked separately from fixed recurrences.
Gating and normalization are not rooms: gating decides which transformation runs (criterion ③), normalization is a weighted sum.^1,7,12
Room marks describe the core operation. Composite diversity concerns the room’s repertoire. Card marks judge each narrower pictured architecture.
Diffusion LEARNING runs the forward noise schedule. OPERATING reverses the sampling direction.
The graph intervention formally replaces one assignment. It estimates no treatment size and performs no live actuation.¹⁸
Historical index: search lineage, 1956. Combined perceptron / MLP card, 1958, with a secondary backpropagation milestone in 1986.^1,30,31
Motion runs once while visible, pauses off-screen and ends readably. Reduced-motion preferences open figures at their end state.

Gallery: one shared Three.js renderer. Cards and fallbacks: SVG schematics.

That a symbol stream supplies its own semantics.²
That the historical criteria rank performance, or that one repeated architecture satisfies diversity.¹
That a hand-specified structural model identifies a treatment’s effect without causal assumptions.¹⁸
Measured forecasts, fitted parameters, a physical system boundary or energy comparisons from these drawings.