The Conversation

Links

1. Hertz and Wittgenstein · Documented influence · Explicit citation. Tractatus 4.04 explicitly refers to Hertz's mechanics and dynamical models when discussing the logical multiplicity of representation. Tractatus Logico-Philosophicus, 4.04
2. Hertz and Craik · Interpretive affinity · Structure over resemblance. Both accounts make structural correspondence central to useful modeling; this comparison does not claim a verified direct intellectual influence. Craik, The Nature of Explanation (1943), library record
3. Craik and Tolman · Interpretive affinity · Internal alternatives. Craik's simulated possible actions and Tolman's flexible cognitive maps both explain behavior through internal structure beyond a fixed response sequence. Cognitive Maps in Rats and Men
4. Uexküll and Simon · Interpretive affinity · Organism-relative world. Both select the relevant environment through an organism's needs and capacities; their biological and decision-theoretic projects remain distinct. Rational Choice and the Structure of the Environment
5. Wiener and Deutsch · Documented influence · Cybernetics into politics. Deutsch's own retrospective explicitly credits Wiener's feedback, channel-capacity and memory concepts in the development of The Nerves of Government. Deutsch's author retrospective (1986)
6. Wiener and Kalman · Interpretive affinity · Prediction from noisy data. Both address estimation and prediction under imperfect observations; Kalman's state-space recursion differs from Wiener's stationary time-series formulation. A New Approach to Linear Filtering and Prediction Problems
7. Ashby and Beer · Documented influence · Viability through regulation. Beer's account of Cybersyn credits Ashby for his understanding of homeostasis and applies Ashby's law of requisite variety. Beer adds recursive institutional coordination. Fanfare for Effective Freedom (Beer's 1973 account)
8. Simon and Newell et al. · Documented collaboration · Shared researcher. Herbert Simon coauthored the General Problem Solver work with Allen Newell and J. C. Shaw, linking bounded cognition to implemented symbolic problem solving. Newell, Shaw and Simon, Report on a General Problem-Solving Program (RAND P-1584, 1959)
9. Minsky and Dreyfus · Interpretive contrast · Representation and relevance. Minsky defines models by the questions a user can answer; Dreyfus challenges whether formally represented facts can capture situated relevance adequately. Alchemy and Artificial Intelligence
10. Craik and Gibson · Interpretive contrast · Internal simulation / pickup. Craik emphasizes internally modeling possibilities; Gibson emphasizes active detection of environmental information. They allocate explanatory work differently. Gibson, The Senses Considered as Perceptual Systems (Houghton Mifflin, 1966), library record
11. McCarthy & Hayes and Fikes & Nilsson · Interpretive affinity · Actions over explicit worlds. Both use formal representations to reason about action and persistence; STRIPS supplies a practical planning operator scheme rather than a general commonsense theory. STRIPS
12. Fikes & Nilsson and Winograd · Interpretive affinity · Executable microworlds. Both connect explicit world descriptions to action procedures in restricted domains; Winograd additionally integrates dialogue and language interpretation. Procedures as a Representation for Data
13. Forrester and Meadows et al. · Documented influence · Prototype to World3. The Limits to Growth explicitly names Forrester's World Dynamics model as its prototype and describes the MIT project that built the book's world model, later documented as World3. Meadows, Meadows, Randers and Behrens, The Limits to Growth (1972), Dartmouth archive PDF
14. Bartlett and Samuel · Interpretive contrast · Reconstruction / retained records. Bartlett studies reconstructive human remembering; Samuel studies stored evaluated positions and learned weights as two separate learning procedures. The comparison separates records from dispositions. Some Studies in Machine Learning Using the Game of Checkers
15. Bateson and Beer · Interpretive affinity · Beyond the individual. Both place consequential information in larger feedback circuits, though Bateson asks where mind resides and Beer designs organizational coordination. Form, Substance and Difference
16. Ha & Schmidhuber and Hafner et al. · Documented influence · From evolved controllers to latent actor–critic. Dreamer cites World Models as prior work and explicitly reuses its convolutional encoder/decoder architecture, while changing behavior learning. Dreamer, related work and Appendix A
17. Sutton and Hafner et al. · Documented influence · Learning from imagined experience. Dreamer explicitly cites Sutton's 1991 Dyna account when situating policy learning through imagined rewards; Dyna's selected node is its 1990 architecture paper. Dreamer introduction and references
18. LeCun and Xing et al. · Documented critique · Abstraction versus reconstruction as a check. Xing et al. directly critique JEPA: discarded information may matter to dynamics, and reconstruction can expose what the representation loses. This is their argument, not a settled theorem. Critique of World Model, v5, architecture discussion
19. Li and Shanghai AI Lab · Documented critique · Output functions versus internal representation. The Shanghai paper directly credits Li's taxonomy, then argues that classifying outputs leaves the shared internal representation unspecified. A Definition and Roadmap for World Models, introduction
20. Yang and Chu et al. · Interpretive contrast · Where does reward belong?. Comparing their definitions: Yang includes reward and task models; Chu et al. place reward and policy outside the world-model factorization. This is a boundary difference, not a documented debate. Agentic World Modeling, §2.3; compare Yang's linked node
21. Sakagami et al. and Yang · Interpretive affinity · Mediation through a reusable component. Both separate a reusable world-model component from its users. Sakagami stresses information sharing inside robotic systems; Yang stresses learning and evaluation across agent domains. Robotic world models, §2; compare Yang's linked node
22. Sakagami et al. and Ha & Schmidhuber · Interpretive contrast · Queryable shared state versus learned rollout model. Sakagami's organizing concern is maintained, shared state and interfaces. Ha and Schmidhuber center compressed dynamics used to train control. These functions can coexist. Robotic world models; compare World Models linked node
23. Brooks and Ha & Schmidhuber · Interpretive contrast · Direct coupling versus learning in a model. Brooks questions centralized internal representations for situated behavior; Ha and Schmidhuber demonstrate learning control inside a represented environment. The tasks and claims differ. Brooks — Intelligence without Representation; compare Ha node
24. Sutton and Brooks · Interpretive contrast · Two early-1990s routes to adaptive action. Dyna combines a learned model with planning and real experience. Brooks builds behavior through direct perception–action coupling. Contemporary work, distinct system boundaries. Sutton, Integrated Architectures for Learning, Planning, and Reacting (ICML 1990), author's copy
25. LeCun and Hafner et al. · Interpretive affinity · Action through compact predictive representations. Both use abstract predictive representations for action. Original Dreamer learns concrete control policies; LeCun proposes a broader modular, hierarchical architecture. LeCun, A Path Towards Autonomous Machine Intelligence (OpenReview)
26. Clark & Chalmers and Sakagami et al. · Interpretive affinity · Useful cognition depends on accessible external state. Clark and Chalmers enlarge the cognitive boundary to coupled resources; Sakagami formalizes state shared across components. This is an analogy, not a claimed historical influence. The Extended Mind; compare Robotic world models linked node
27. Pearl and Chu et al. · Interpretive affinity · Interventions and counterfactual evaluation. Pearl formalizes the difference between observing and intervening. Chu and colleagues ask simulators for action-sensitive rollouts, with their own counterfactual framing. A shared concern, not a citation. Pearl — Causality; compare Chu linked node
28. Sutton and Yang · Documented engagement · World interaction improves the model used for learning. Yang cites Dyna (Sutton 1991) explicitly. Both route learning through a model refined by real experience; Yang broadens the model to task distributions and costs. Yang — World Models as an Intermediary
29. Xing et al. and Shanghai AI Lab · Interpretive affinity · Reconstruction can diagnose lost information. Both argue that reconstructive outputs can reveal losses in compressed representations, while emphasizing stateful generative modeling and physical compression respectively. A Definition and Roadmap, introduction; compare Xing node
30. Yang and Smith · Interpretive affinity · Two meanings of intermediary. Yang proposes a learned world model between an agent and the real world, to cut costly real-world trials. In Smith's work a governed world model sits between the people and AI agents who act and the business they act in, as a governed account rather than a learned simulator. Yang's paper came first; this is a comparison, not a claim about where either idea began. Yang, World Models as an Intermediary between Agents and the Real World
31. Beer and Smith · Documented engagement · Organizational cybernetics and authority. Smith's report examines Beer's Viable System Model and reads organizational governance practice as converging on the same structure rather than derived from it. Her emphasis is typed declarations, authority chains and decision evidence. Equivalence to the Viable System Model is not claimed. Smith — Organizational Cybernetics & the Viable System Model
32. Ashby and Smith · Documented engagement · Control-relevant distinctions and regulation. Smith's report builds on Ashby's law of requisite variety: governance must match the variety of what it governs. Her governed world model adds rules for who may change goals and the model, and keeps the earlier record. Ashby's results do not show that WMI's ingredient list or governance design is necessary. Smith — Organizational Cybernetics & the Viable System Model
33. LeCun and Smith · Documented critique · Learned predictor and governing environment. Smith's definition report assesses LeCun's 2022 definition against her proposed ingredient list. In her framing a learned predictor such as JEPA is a different kind of model that can plug into a governed world model through a socket, its output provisional until a person accepts it. A difference of scope, not evidence that JEPA cannot support memory or planning. Smith, The World Model Definition (TR-D-001)
34. Hutchins and Smith · Interpretive affinity · Cognition across people, tools and procedures. Hutchins studies cognition spread across people, tools and procedures. Smith proposes a governed world model in which decisions by people and AI agents are attributed and kept with their evidence. An explanatory account and an engineering proposal answer related but different questions. Hutchins — Cognition in the Wild
35. Sakagami et al. and Smith · Interpretive affinity · Persistent shared state with explicit interfaces. Both place a maintained representation between activities and the world, distinguishing given information from estimates. Smith emphasizes institutional purpose, authority, commitments and decision lineage. This comparison does not claim that shared world-model infrastructure originated with WMI. Decision Lineage Protocol specification, version 3.0.0
36. Minsky and Smith · Interpretive affinity · A model's relevance depends on its user. Minsky makes a model's adequacy relative to its user's questions. Smith's specification gives each role a projection of the same world model, and adds a constraint: no projection may add governance. Decision Lineage Protocol specification, version 3.0.0
37. Chu et al. and Smith · Interpretive affinity · Revision of state and revision of the model. Both separate producing outputs from revising the model, and both gate revision: Chu and colleagues' Evolver passes regression and rollback checks. Smith's emphasis is who has the authority to approve a change, and that the earlier record is kept. Similar structure does not establish operational equivalence. World Model Initiative Series, Volume I · DOI 10.5281/zenodo.21073276
38. Brooks and Smith · Interpretive contrast · How much persistent shared representation?. Brooks demonstrates situated behavior with reduced reliance on centralized world representations. Smith's target includes institutional commitments and decision history that must persist across people and agents. Their differing domains explain part of the distance; neither comparison establishes a universal winner. World Model Initiative Series, Volume I · DOI 10.5281/zenodo.21073276
39. Craik and Ha & Schmidhuber · Interpretive affinity · Mental rehearsal to learned simulation. Both explore action through an internal model. Craik offers an explanatory account of organisms; Ha and Schmidhuber implement learned latent dynamics for control. Ha and Schmidhuber — World Models
40. Minsky and LeCun · Interpretive affinity · Sufficiency depends on use. Minsky defines adequacy through an observer’s questions; LeCun’s configurator and abstract prediction make relevance a design problem for an agent. LeCun, A Path Towards Autonomous Machine Intelligence (OpenReview)
41. Kalman and LeCun · Interpretive affinity · Compact predictive state. Both use a state representation to support prediction. Kalman specifies probabilistic estimation in a given dynamical model; LeCun proposes learned representations and modular planning. LeCun, A Path Towards Autonomous Machine Intelligence (OpenReview)
42. Bellman and Sutton · Interpretive affinity · Planning from transitions. Bellman formalizes sequential choice with a transition model. Dyna combines learned models, planning updates, and real experience; this edge marks conceptual continuity. Sutton, Integrated Architectures for Learning, Planning, and Reacting (ICML 1990), author's copy
43. Wiener and Sutton · Interpretive affinity · Act, observe, update. Both couple action to subsequent evidence. Cybernetic regulation and reinforcement learning organize that feedback around different mathematical objectives. Sutton, Integrated Architectures for Learning, Planning, and Reacting (ICML 1990), author's copy
44. Simon and LeCun · Interpretive affinity · Bounded, task-relative intelligence. Simon emphasizes limited computation and structured environments; LeCun’s architecture selects abstractions and planning levels for the task. This is a comparison, not an influence claim. LeCun, A Path Towards Autonomous Machine Intelligence (OpenReview)
45. Ashby and Chu et al. · Interpretive affinity · Adaptation at multiple levels. Ashby studies regulation and adaptive organization; Chu and colleagues separate prediction, simulation, and persistent model revision. The levels are related questions, not equivalent formalisms. Chu et al. — Agentic World Modeling, v3
46. Beer and Sakagami et al. · Interpretive affinity · A maintained world shared by activities. Beer’s organizational coordination and Sakagami’s robotic information component both rely on maintained shared representations; their authority and physical-control problems differ. Sakagami et al. — Robotic world models
47. Bateson and Clark & Chalmers · Interpretive affinity · Where cognition begins and ends. Both move the explanatory boundary beyond an isolated individual, through feedback circuits or reliably coupled external resources. Their criteria and philosophical claims differ. Clark and Chalmers — The Extended Mind
48. Deutsch and Hutchins · Interpretive affinity · Collective information processing. Deutsch models political communication and learning; Hutchins examines coordinated cognitive work in navigation. Both make organization relevant to information processing. Hutchins — Cognition in the Wild, author’s account
49. Hutchins and Clark & Chalmers · Documented influence · Distributed cognition as a cited precedent. Clark and Chalmers cite Hutchins (1995), the navigator's slide rule and work on cognitive collectives, as examples when arguing that external resources can take part in cognition. Clark and Chalmers — The Extended Mind
50. Gibson and Brooks · Interpretive affinity · Environmental structure does explanatory work. Gibson emphasizes active pickup of environmental information; Brooks emphasizes situated perception–action coupling. This is thematic affinity, not a claim of identical theories. Brooks — Intelligence without Representation
51. Dreyfus and Brooks · Interpretive affinity · Situated challenges to centralized symbols. Dreyfus questions whether symbolic programs can capture relevance and ambiguity; Brooks demonstrates robot architectures that reduce reliance on centralized world representations. Brooks — Intelligence without Representation
52. McCulloch & Pitts and Rosenblatt · Interpretive contrast · Logical neurons and learned connections. McCulloch and Pitts analyze idealized neural networks logically; Rosenblatt studies how response tendencies can be acquired through learning. The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain
53. Rosenblatt and Ha & Schmidhuber · Interpretive affinity · Experience embodied in network parameters. Both use learned connections to retain useful structure from experience, but their learning methods, representation scales, and uses differ substantially. Ha and Schmidhuber — World Models
54. Piaget and Tolman · Interpretive affinity · Structure acquired through activity. Both explain flexible behavior through acquired structure beyond isolated stimulus–response links; Piaget studies development, while Tolman studies learning and navigation. Cognitive Maps in Rats and Men
55. Turing and Sakagami et al. · Interpretive affinity · Persistent, differentiated knowledge. Turing’s proposed proposition store distinguishes kinds of statements; Sakagami separates given and estimated state in a queryable component. This is a structural analogy. Sakagami et al. — Robotic world models
56. Leontief and Forrester · Interpretive contrast · Economic interdependence, different dynamics. Leontief's input-output structure tracks sectoral dependence; Forrester emphasizes feedback and decision policies that generate behavior over time. Forrester, The Beginning of System Dynamics (author's retrospective, 1989)
57. Lorenz and Shanghai AI Lab · Interpretive affinity · Limits of predictive representation. Lorenz exposes sensitivity to initial conditions; the Shanghai roadmap emphasizes partial observation and bounded compression. More detail alone does not settle useful predictive horizons. Chen et al. — A Definition and Roadmap for World Models
58. Winograd and Sakagami et al. · Interpretive affinity · Explicit state made useful to components. SHRDLU combines represented state, procedures, and language in a microworld; Sakagami supplies a broader engineering account of shared world-model state and interfaces. Sakagami et al. — Robotic world models
59. McCarthy & Hayes and Pearl · Interpretive affinity · Formal action and counterfactual questions. Situation-based action reasoning and structural causal models distinguish actual states from consequences of interventions, using different semantics and assumptions. Pearl — Causality, first-edition contents and primary materials
60. Wittgenstein and McCarthy & Hayes · Interpretive affinity · Logical representation of possible situations. Both connect formal structure to possible situations. Wittgenstein addresses propositions and picturing; McCarthy and Hayes address computation, action, and knowledge. Some Philosophical Problems from the Standpoint of Artificial Intelligence
61. Newell et al. and Winograd · Interpretive affinity · Executable symbolic problem worlds. Both implement reasoning through explicit representations and operations. SHRDLU adds language interaction with a deliberately restricted, simulated blocks world. Procedures as a Representation for Data (MIT thesis)