On October 1, 2026, the Cognition Futures Reading Group began reading Stefano Franchi’s “Homeostats for the 21st Century? Simulating Ashby Simulating the Brain,” published in Constructivist Foundations in 2013. The article appeared as a target article with seven open peer commentaries (from Inman Harvey, Helge Malmgren, Robert Lowe, Tom Froese, Matthew Egbert, David Vernon, and Takashi Ikegami) and a response from Franchi. Amanda Nelson, , Morgan Hough and attended.

Nelson proposed the paper. She had been looking at classic cybernetics models, early neuron models and W. Ross Ashby’s homeostat among them, and found in Franchi a paper about what such models are for.

This was the first meeting in this series for the Fall 2026 term.

What was read

The group worked slowly through the abstract and the opening sections, stopping towwards the end of the section titled “Simulating objects vs. simulating concepts.”

Franchi’s abstract sets up the argument. Ashby’s technical results, such as ultrastability, have been taken up in recent cognitive science, but his broader thesis that homeostatic adaptation governs all aspects of all life has been left behind. Franchi argues that this thesis makes life fundamentally heteronomous, and that it conflicts both with the autopoiesis framework used by Ashby’s recent defenders and with the Western tradition’s emphasis on autonomy. A definition of heteronomy was shared in the chat: action governed by a force outside the individual, the counterpart of autonomy. Parent noted how the paper cuts into that tradition’s treatment of autonomy as what defines life.

The opening sections draw the distinction the session centered on. In simulating objects, a model is judged by how closely its behavior matches the real thing. Franchi’s example is Newell and Simon’s early AI work, which compared machine performance on tasks like chess with human performance on the same tasks. In simulating concepts, the model is “a tool to think with” (Franchi’s phrase, after Papert’s Logo microworlds) and closeness to the real world is not the measure. Franchi holds that most successful projects use both, but that the two should be kept apart analytically, because the focus on object behavior can hide the philosophical work underneath. He puts it bluntly: technical work in the cognitive sciences is sometimes “philosophy in disguise.” The section closes by applying this to Ashby: the homeostat is usually read as a simplified proxy for the brain, and Franchi argues it can only be appreciated fully if it is also read as a proxy concept.

What was discussed

Concept simulation, and what it is good for

Nelson opened with an example of her own that the paper does not use: Randall Beer’s work using cellular automata to demonstrate a concept like precariousness. A model of that kind does not need to resemble anything in particular in the world. She also noted Franchi’s point that the most useful models combine the two approaches, and that in cognitive science a model should be grounded in biology in some fashion.

Later, Nelson said she valued Franchi’s account of philosophy as the study of “being as such,” as opposed to a loose synonym for thinking. Read that way, she said, philosophy does not have to happen only in one’s head or on paper, and a computer can be a tool for it.

Parent noted an appreciation of the word “simulation” being used in a particular fashion, as distinct from “model” or “idea.” He asked what non-brain object simulations look like, offering hurricanes and weather, stock prices, and a heart condition as candidates. He also connected the Newell and Simon comparison to his interest in substrate agnosticism: two substrates attempt the same task, and the match between them is taken as evidence for the theory. He asked whether a concept could be carried across substrates, then noted the question may not fit, since concept simulations are not judged by that kind of validity. In an aside, he observed that Franchi’s 2013 remark that identifying the mind with symbol-manipulating software had “long gone out of fashion” reads differently in the current AI moment.

Ultrastability and machine learning

Alicea raised a connection between Ashby’s mechanism and machine learning. Ashby’s homeostat adapts by randomly switching its connection parameters when a unit’s essential variable goes out of bounds, a stochastic step function. Alicea compared this to step-like functions in neural networks, such as the Heaviside or ReLU function, acting on a weight matrix, and described ultrastability as close to the way a neural network is pruned. He framed this as part of an answer to the question of whatever happened to cybernetics: some of it went on to motivate machine learning. He suggested there is probably a paper to be written on these connections. This was floated as a line worth pursuing; the group did not examine the comparison in detail.

Is the distinction new, and how far does it go?

Parent asked whether Franchi’s object/concept distinction is standard in computational modeling or an unusual move. Alicea said it echoes a familiar criticism of modeling: a model is not the thing modeled. He thought Franchi motivates the distinction more than necessary, since the homeostat is plainly an analogy for an organism rather than a replica of one.

Alicea then gave the distinction a more specific stake. In cognitive science, he said, a model often represents two things at once: something like brain or network activity, and the constructs that activity is supposed to produce, such as attention or memory. Researchers switch between the two as if nothing needed to sit between them, and whether the construct has been captured correctly often goes unasked.

Nelson agreed the distinction is correct and that each approach has its own benefits and should be analyzed separately. But she argued there is more to say than the paper does: a model has to be fruitful in both domains to be truly fruitful, and good experimentation keeps both in view. In her words, “the whole is greater than the sum of its parts.” She did not recall what Franchi says about the homeostat on this point specifically.

Parent built on Nelson’s point in closing. He suggested thinking in degrees: how far a given simulation deliberately moves away from a concrete object comparison, such as checking a machine’s output against a human’s, and toward pure concept, and whether it later comes back. On this view neither end is better. What matters is being intentional about the distance. He wondered whether this could become a framework or even a metric. This was raised as an idea to pick up, and was not developed further in the session.

Open questions carried forward

• Is the object/concept distinction a standard one in computational modeling, or is Franchi drawing it in an unusual way? Alicea offered one answer; Parent’s question was left open.

• What does Franchi say about the homeostat specifically on the need to work in both modes? Nelson raised this and the reading had not yet reached it.

• Could Parent’s degrees-of-abstraction idea be made concrete?

• Does Alicea’s comparison between ultrastability and machine learning hold up in detail?

Next time, the group will continue the paper with the section “The homeostat as a concept,” where Franchi sets out his reading of Ashby’s theses, ahead of the simulation results, the commentaries, and the author’s response.

Subscribe now

CFRG Context: Previous Discussions

This section is editorial context drawn from the group’s earlier sessions this year. None of it was discussed on October 1.

• Dynamical cognitive science. Earlier this year the group read Raja’s introduction to a Topics in Cognitive Science issue on dynamical cognitive science, followed by Randall Beer’s “On the Proper Treatment of Dynamics in Cognitive Science” from the same issue. Nelson’s October 1 example comes from a different line of Beer’s work, his cellular automaton models of autopoiesis and precariousness.

• World models. Several readings came from the 2026 Philosophical Transactions of the Royal Society A theme issue on world models in natural and artificial intelligence. That issue includes a paper by four members of the group (Alicea, Hough, Nelson, and Parent) recasting the Every Good Regulator Theorem as a framework for world modeling. The theorem is Conant and Ashby’s, so the group has been working with Ashby’s ideas before this paper. Other readings from the issue included Sacco, Sakthivadivel, and Levin on topological limits to self-organization in locally interacting systems, including autoregressive language models; Amir, Tiomkin, and Langdon on goal-relative “telic states”; Hofstadter on whether there is an “I” in AI; and Battleday and Gershman on easy and hard problems in AI for science.

• The agent as unit of explanation. Several of those readings, along with a paper by Tom Froese on methodological individualism, questioned whether a single, central agent is the right unit for explaining purposeful behavior. Franchi’s heteronomy thesis approaches autonomy from another direction. Froese is also one of the commentators on Franchi’s article, which the group has yet to reach.

References

This session

• Franchi, S. (2013). Homeostats for the 21st century? Simulating Ashby simulating the brain. Constructivist Foundations 9(1): 93–101. Open peer commentaries and the author’s response follow in the same issue (pp. 102–124). https://constructivist.info/9/1/093

• Beer, R. D. & Di Paolo, E. A. (2023). The theoretical foundations of enaction: Precariousness. Biosystems. https://www.sciencedirect.com/science/article/abs/pii/S0303264722002040 [confirm this is the work Nelson had in mind]

• Ashby, W. R. (1960). Design for a brain. Second edition. John Wiley & Sons, New York. (The source of the homeostat as Franchi discusses it.)

Background

• Alicea, B., Hough, M., Nelson, A. & Parent, J. (2026). A ‘good’ regulator may provide a world model for intelligent systems. Philosophical Transactions of the Royal Society A 384(2320): 20250007. https://doi.org/10.1098/rsta.2025.0007

• Beer, R. D. (2023). On the proper treatment of dynamics in cognitive science. Topics in Cognitive Science. https://doi.org/10.1111/tops.12686

• Sacco, F., Sakthivadivel, D. & Levin, M. (2026). Topological constraints on self-organization in locally interacting systems. Philosophical Transactions of the Royal Society A. https://doi.org/10.1098/rsta.2025.0011

• Amir, N., Tiomkin, S. & Langdon, A. (2026). Goals and the structure of experience. Philosophical Transactions of the Royal Society A 384(2320): 20250004.

• Hofstadter, D. (2026). Is there an ‘I’ in AI? Philosophical Transactions of the Royal Society A 384(2320): 20240527. https://doi.org/10.1098/rsta.2024.0527 †

• Battleday, R. M. & Gershman, S. J. (2026). Artificial intelligence for science: The easy and hard problems. Philosophical Transactions of the Royal Society A 384(2320): 20240530. https://doi.org/10.1098/rsta.2024.0530

• Favela, L.H. and Raja, V. (2026), Dynamical Cognitive Science! Wherefore Art Thou?. Top. Cogn. Sci., 18: e70042. https://doi.org/10.1111/tops.70042

Leave a comment

Further & Suggested Reading

These readings were not discussed on October 1. They were gathered following the session as starting points for the questions Parent, Nelson, and Alicea raised about simulating objects versus simulating concepts. Several of the article’s own commentaries take up similar questions. The annotations are editorial.

Mapping the space between object and concept

Barbara Webb (2001), “Can robots make good models of biological behaviour?” Behavioral and Brain Sciences 24(6). Webb separates simulation models, which represent a hypothesis about a target system, from other activities also called modeling. She then sets out seven dimensions on which such models vary: biological relevance, level of organization, generality, abstraction, structural accuracy, behavioral match, and physical medium. She argues that no position in that space is the single correct one, and that a good methodology states where a model sits and why. This is the closest existing match to Parent’s call for intentionality about how far a simulation moves from its object.

Michael Weisberg (2013), Simulation and Similarity, chapter 7, “Modeling Without a Specific Target”. Weisberg argues that building a single model to study a specific target is only one part of modeling practice. This chapter covers generalized, hypothetical, and targetless modeling, the last being the study of a model with no target at all. His examples include cellular automata and the Game of Life, which places targetless modeling near the kind of example Nelson raised.

Arturo Rosenblueth and Norbert Wiener (1945), “The Role of Models in Science”. Philosophy of Science 12(4). The cybernetic starting point for this debate. The authors hold that no substantial part of the universe can be grasped without abstraction, which they define as replacing the part under study with a simpler model of similar structure, and they divide models into formal and material kinds. They also supply a limit case for the object end of any spectrum: “the best material model for a cat is another, or preferably the same cat.”

Richard Levins (1966), “The Strategy of Model Building in Population Biology”. American Scientist 54(4). Levins argues that models trade generality, realism, and precision against one another, and describes three strategies that each sacrifice one of the three. He also holds that a satisfactory theory is usually a cluster of models. That point bears on Nelson’s argument that good modeling has to work in more than one mode, since Levins locates strength in a family of models rather than in any single one.

Simulation as thought experiment

Ezequiel Di Paolo, Jason Noble, and Seth Bullock (2000), “Simulation models as opaque thought experiments”. Artificial Life VII. The authors set two views against each other: simulations as computational thought experiments, and simulations as realistic simulacra. They propose that simulations are opaque thought experiments, in which the consequences follow from the premises in a way that is not obvious and has to be revealed through systematic enquiry. This offers one answer to why concept work needs a computer at all, which connects to Nelson’s point about computers as tools for philosophy.

Andrew Pickering (2010), The Cybernetic Brain: Sketches of Another Future. Pickering reads the homeostat as “ontological theatre,” a device that stages an ontology, and argues that read this way the homeostat is a contribution to philosophy. It is a historian’s route to Franchi’s claim that technical work can be philosophy in disguise, and the book is cited by several of the commentators. A short essay version is available, and Constructivist Foundations ran a review of the book.

Randall Beer (2003), “The Dynamics of Active Categorical Perception in an Evolved Model Agent”. Adaptive Behavior. Alongside a detailed analysis of a simple evolved agent, Beer argues in a section titled “Frictionless Brains” that cognitive science needs simpler idealized models of complete brain, body, and environment systems, in the way physics uses frictionless planes. He states that the aim of such models is conceptual understanding rather than empirical prediction. This is Beer’s own account of what concept models are for, from the author of Nelson’s example. His later “Some historical context for minimal cognition” (2021) traces how the approach developed.

Mark Bedau (1999), “Can unrealistic computer models illuminate theoretical biology?” Proceedings of the 1999 Genetic and Evolutionary Computation Conference Workshop Program. A short paper that asks the question in its title directly, and is often cited alongside Di Paolo, Noble, and Bullock. No stable online copy was found.

Christopher Langton (1989), “Artificial Life”. The opening chapter of the first Artificial Life proceedings. Langton framed the field as placing life as we know it within the larger space of life as it could be. It is the programmatic statement behind much of the concept-side simulation work in artificial life.

Seymour Papert (1980), Mindstorms: Children, Computers, and Powerful Ideas. The source of the “tool to think with” idea Franchi draws on. Papert describes the Logo turtle as an “object-to-think-with.” The book is freely available online through the MIT Media Lab.

How the two modes relate

Barbara Webb (2009), “Animals Versus Animats: Or Why Not Model the Real Iguana?” Adaptive Behavior. Webb distinguishes models of specific animal systems from the exploration of invented artificial animals, and asks how simulating animals that do not exist can teach us about real biology. Her answer is that animat research, where it bears on biology, should be treated as model building like any other. It is a direct test of Nelson’s claim that a good model has to work in both modes.

Evelyn Fox Keller (2000), “Models of and models for: Theory and practice in contemporary biology”. Philosophy of Science 67 (Supplement). Keller describes models of gene pathways as models of mechanisms and, at the same time, models for generating new questions and experiments. The pairing gives a name to Nelson’s two modes as two directions a single model can face.

Mary Morgan and Margaret Morrison, eds. (1999), Models as Mediators: Perspectives on Natural and Social Science. The editors argue that models mediate between theory and world while remaining partly independent of both. This suggests a third way to place the homeostat, alongside object and concept.

William Wimsatt (1987), “False Models as Means to Truer Theories.” In Nitecki and Hoffman, eds., Neutral Models in Biology, Oxford University Press. Wimsatt argues that idealized and strictly false models do real work in theory building, partly by showing where and how they fail. It is reprinted in his Re-Engineering Philosophy for Limited Beings (2007).

Howard Pattee (1989), “Simulations, Realizations, and Theories of Life”. In the same Artificial Life proceedings as Langton. Pattee distinguishes computer simulations of life from computer realizations of it, and asks how either relates to a theory that separates the living from the non-living. This bears on Parent’s substrate question: at what point does running a concept on a new substrate stop being a simulation of it?

Models and the constructs they claim to capture

Olivia Guest and Andrea Martin (2021), “How Computational Modeling Can Force Theory Building in Psychological Science”. Perspectives on Psychological Science 16(4). The authors argue that building a computational model forces researchers to analyze, specify, and formalize intuitions that would otherwise go unexamined, which they call open theory. They describe inference as a path that must step through theory, specification, and implementation. This is relevant to Alicea’s point about researchers switching between a model’s network activity and the constructs it is said to produce.

Iris van Rooij and Giosuè Baggio (2021), “Theory Before the Test: How to Build High-Verisimilitude Explanatory Theories in Psychological Science”. Perspectives on Psychological Science 16(4). The authors argue that a focus on experimental effects can lose sight of the psychological capacities those effects are meant to explain. They revisit Marr’s levels of analysis, the classic framework for the level-switching Alicea described, and show how theoretical analysis can establish a theory’s plausibility before it meets data.

From Franchi’s own footnotes

Peter Asaro (2011), “Computers as Models of the Mind: On Simulations, Brains and the Design of Early Computers”. In Franchi and Bianchini, eds., The Search for a Theory of Cognition. Franchi cites this for a more detailed analysis of the uses of computer simulation. A related open-access article in German (2008) discusses correspondence in which Alan Turing urged Ashby to simulate the homeostat on the ACE computer rather than build a dedicated machine, an early precedent for Franchi’s virtual homeostat.

Antoni Gomila and Vincent Müller (2012), “Challenges for artificial cognitive systems”. Journal of Cognitive Science 13(4). Franchi cites this for the theoretical basis behind object simulations.