NOTE-339: Examining the Society of Mind

Contribution

A compact, mechanism-by-mechanism guide to The Society of Mind (LIT-429), aimed at “those interested in implementing Societies of Mind” (§3). Minsky presents those mechanisms “in fragments and at a variety of levels”. Singh gathers them into one catalogue. He adds the theory’s early history, including quotations from unpublished 1976 drafts that he had from Minsky and Papert. He also places the theory against the AI of 1986–2003 and says where each neighbour falls short of it.

Key insight

The Society of Mind is a theory of organisation, not of a mechanism. Its claim is that intelligence comes from the diversity of many specialised processes and from how they are managed, not from one principle. Singh quotes §30.8: “The trick is that there is no trick.” So the hard problems are the ones a uniform architecture hides. They are how agents that cannot understand each other’s representations still cooperate, and how managerial knowledge chooses among them. That is the test Singh puts to every later system.

Assumptions

• Exposition, not argument. The paper describes the theory as Minsky stated it and does not defend it against its critics. The four 1991 reviews (in the issue LIT-418 closes) appear only through Dyer’s connectionist alternatives (§4.1).
• The author’s position. Singh was Minsky’s doctoral student, “presently collaborating with Marvin Minsky to develop an architecture for commonsense thinking” (author note). The paper is openly sympathetic: “we predict that The Society of Mind will still be read decades from now”.
• Scope. “In this article we could only examine a small fraction of the full Society of Mind theory” (§5).

Key results

• History (§2).
  • The copy-demo robot of the late 1960s: “no single method ever worked well by itself” (quoting the book’s Postscript).
  • Newell’s 1962 “single personality, wandering over a goal net” is the view the theory opposed.
  • Frames (1974) included Fahlman’s essay on frames as “a packet of related facts and agencies”. Hewitt’s Actors (1976) was a message-passing contemporary that Minsky and Papert’s agents differed from.
  • The 1976 “Brazil” drafts: “The mind is a community of ‘agents’. Each has limited powers and can communicate only with certain others.”
  • The 1977 IJCAI paper names the theory “The Society of Minds” and says mental abilities “both ‘intellectual’ and ‘affective’ (and we ultimately reject the distinction)” emerge from “quasi-political hierarchies” of agents, critics and censors. The 1980 K-lines paper describes “Divisions” of agents.
• Agents (§§3.1–3.2). An agent is “any component of a cognitive process that is simple enough to understand”. An agency is a society of agents that does more. Mental activity “ultimately reduces to turning individual agents on and off”, and a “partial state of mind” is a subset of agent states.
  • K-lines are “the most common agent”. They chunk a problem-solving episode, false starts included, so that its configuration can be re-entered.
  • Nemes (polynemes, micronemes) represent. Nomes (isonomes, pronomes, paranomes) control.
• Combining agents (§3.3). Frames are built from bound pronomes. Frame-arrays share slots across viewpoints, and shared slots are “the ancestors of paranomes”. Transframes represent events, with origin, destination, actor, motive and instrument.
• Problem solving (§3.4). Difference-engines, after GPS, are “elevated to a central principle”. Censors and suppressors carry “negative expertise”, which may be “the bulk of what we know, yet remain invisible”. Humour is linked to it, citing the jokes memo. A B-brain watches the A-brain.
• Communication (§3.5).
  • Agents cannot share consistent symbol definitions. Minsky: “The fewer things an agent does, the less likely that what another agent does will correspond to any of those things.”
  • So agents communicate by arousal through K-lines, and by randomly connected bus lines whose meanings both sides learn (“first invented by Calvin Mooers”, with “low probability of collision”).
  • They use an internal language of “grammar-tactics” and their inverses, from the 1991 reply.
  • Paranomes, “the most common method”, work with “no active communication at all”.
  • “Thoughts themselves are ambiguous!” (§20.1).
• Growth (§3.6). Protospecialists. Predestined learning. Accumulating, uniframing, transframing and reformulating. Learning goals from attachment figures (“goal learning” as against “skill learning”). Mental managers and Papert’s Principle (§10.4). The Principle of Non-Compromise. Stages that teach each other.
• Since 1986 (§4).
  • Symbolic–connectionist hybrids: Dyer, Shastri’s SHRUTI, Maes.
  • Graphical models resemble neme networks. Ring-closing is “similar in some ways” to Pearl’s belief propagation. Graphical models lack nomes and procedural control knowledge.
  • Brooks’s Behavior Language and Hearn’s K-line Language are attempts at a Society of Mind programming language.
  • Case-based reasoning: derivational analogy “came directly from the K-line theory”.
  • Blackboards fail to scale past a few agents “huddled around a blackboard”.
  • Cyc’s microtheories resemble agencies, but Cyc knows little about cognitive processes.
  • Soar is the “opposite” in philosophy but similar in mechanism, as Minsky’s 1993 review of Newell found.
  • Multiagent systems lack the heterogeneous, reflective architecture the theory proposes.

Claims

id	claim	strength	support
C1	The Society of Mind grew from Minsky and Papert’s robot work and from frames, with agents descended from frames with attached procedures	moderate	quotations from Minsky’s own retrospective passages and the frames memo; the frame-as-ancestor link is Singh’s reading
C2	Minsky and Papert drafted a society-of-mind book by 1976 and abandoned it	moderate	quoted from the book’s Postscript; the “Brazil” drafts are cited as unpublished and were given to the author, so they cannot be checked
C3	Agents in the theory communicate mainly without shared symbols: by K-line arousal, learned bus codes and paranomes	strong as a description of the book	§3.5 quotes the book and the 1991 reply; it describes, it does not evaluate
C4	Derivational analogy in case-based reasoning “came directly from the K-line theory”	weak: asserted	cites Carbonell (1983) [24] but quotes nothing showing the lineage
C5	Graphical models lack nomes and procedural control, and will need them to scale to commonsense	weak: a prediction	§4.2, argued from the theory, not from any experiment
C6	No blackboard system has been built with hundreds of agents	weak: asserted	§4.5, no survey cited for the negative
C7	The Society of Mind theory has not been implemented, unlike Soar	moderate	§4.7; consistent with §4.3’s partial attempts (Brooks, Hearn)
C8	The theory was “ahead of its time” and the field’s fragmentation explains its neglect	weak: opinion	§5

Concepts

• Agent / agency. An agent is simple enough to understand. An agency is a society of agents that, seen from outside, may act as one agent.
• K-line. An agent that turns on a set of agents. It is often formed by chunking a problem-solving episode.
• Neme / nome. K-lines for representing (“data”) and for controlling (“control”), “analogous to the data and control lines in the design of a computer”.
• Polyneme / microneme. Arouse partial states across agencies (an “apple-polyneme”), or broadcast diffuse context.
• Pronome / isonome / paranome. Short-term role-binding; uniform operations across agencies; linked pronomes that keep parallel representations in step.
• Transframe. A frame for an event, with before and after states and their causes.
• Negative expertise. Knowledge of what not to do, held by censors (which act before the thought) and suppressors (which act before the action).
• Papert’s Principle. “Some of the most crucial steps in mental growth are based not simply on acquiring new skills, but on acquiring new administrative ways to use what one already knows.”
• Principle of Non-Compromise. A conflict between agents is a sign to reformulate with a third view, not to average.

Connections

• The 1986 book (LIT-429). This is the record’s second-hand map of it. Singh quotes its §§2.5, 6.12, 10.4, 20.1 and 30.8 and its Postscript.
• The 1991 reply (LIT-418). Singh quotes it on “agent” and “agency” and takes its internal-language mechanism from it. He names Dyer’s review from the same issue.
• The Emotion Machine (LIT-432). Singh’s footnotes 2–4 describe three things before publication: selector K-lines for “cognitive-emotional states”, panalogy, and the six levels (reactive, deliberative, reflective, self-reflective, self-conscious, self-ideals). The 2005 draft has all three. Its level names differ slightly (“instinctive” and “learned reactions”), and its sixth level is “self-conscious reflection”. The draft cites this paper in its Introduction.
• The earlier memos and papers, filed at the same time by another reader: the 1974 frames memo, the 1977 “Plain Talk” paper (which Singh shows already calls the theory “The Society of Minds”), the 1980 K-lines paper and the 1980 jokes memo. Singh’s reference [13] gives the jokes memo’s title as “Jokes and the Cognitive Unconscious”, MIT AI Lab Memo 603. The memo (November 1980) is titled “Jokes and their Relation to the Cognitive Unconscious”. The version Minsky posted is headed “Jokes and the Logic of the Cognitive Unconscious”.
• Superimposed coding and sparse memory. The bus scheme Singh describes (§3.5) uses random subsets of wires so that many symbols share a small bus with “a low probability of collision”. That is the regime of near-orthogonal sparse codes in which Kanerva’s sparse distributed memory works. Bricken and Pehlevan (LIT-269) read SDM’s intersections of Hamming balls as attention. NOTE-242 cites Minsky and Papert’s “Best Match Problem” from Perceptrons, not this programme. The parallel is mine; Singh does not draw it.
• Collective intelligence (LIT-046). Singh warns that “the societies of The Society of Mind should not be regarded as very much like human communities, for individual humans are ‘general purpose’, and individual agents are quite specialized” (§1). Pilgrim et al.’s framework is built on capable individuals. On Singh’s reading, the Society of Mind is not a case of it.

Bearing on the record

• THEORY-043 cites C3 in its architectural route: a criterion that marks a group down for loosely coupled units marks a society of mind down too. The paper is an exposition and supports no THEORY of its own.
• It is the source for this record’s knowledge of The Society of Mind while that book is unread. Anything the record says about the book’s mechanisms should name this paper as the source until the book is read.
• Boundary. No instruction for machine-learning practice. §4 compares the theory with AI architectures, including probabilistic graphical models and connectionist hybrids. That is history and design comparison, and no anthology topic holds it as practice. I did not tag it anthology-candidate.

Limitations

• The paper is an exposition by an insider, and it is uncritical by design. It reports no implementation, and its mechanism catalogue inherits the book’s vagueness. Its §4.3 says it has been hard to find “good ‘higher level’ abstractions for agencies”.
• Its history depends partly on private documents (the 1976 drafts) that a reader cannot see.
• It surveys only work that resembles the theory. It does not discuss work that argues against it, beyond Dyer’s alternatives.

Open questions

• Did any later system implement enough of the theory (K-lines, nomes, Critics) to test it? Singh’s own architecture work with Minsky would be the place to look. It is not in this record.
• Is the lineage from K-lines to derivational analogy documented by Carbonell, or only by the Minsky school?

Corrections

• none to a seeded skim (there was no seed)
• Reference titles. Reference [13] names the jokes memo “Jokes and the Cognitive Unconscious”. Its title is “Jokes and their Relation to the Cognitive Unconscious”, per the header of Minsky’s posted copy. Reference [14] names the 1991 article “A Response to Four Reviews of the Society of Mind”. The published title, as Elsevier gives it, is “Society of mind: A response to four reviews” (Crossref gives only “Society of mind”). Reference [34] gives no title for Singh and Minsky (2003).
• Keywords differ. The PDF gives “multiagent systems” as a keyword. The journal’s page gives “intelligent systems” in that place.