NOTE-341: The Computational Boundary of a "Self"

Contribution

The paper offers a definition of an individual that does not depend on the material it is made of: a Self is whatever pursues a goal, and its size is the size of the goals it can pursue. It ties this definition to a mechanism, bioelectric coupling through gap junctions, which the author’s laboratory has shown can steer pattern and tumour formation. And it draws the consequence the other definitions avoid: a living system is not one Self but a stack of nested Selves (molecular networks, cells, organs, organisms, swarms), each with its own goals, cooperating or competing. The definition and the consequence are new as a package. The experiments it rests on are earlier work, cited.

Key insight

Agency comes in levels and sizes, and the size is set by how far information travels. A cell senses its own surface and keeps its own variables in range. Coupled to its neighbours, it shares their measurements, so the coupled group can sense and keep in range things no cell can see, such as the length of a limb or the layout of a face. That group is then a larger agent with larger goals. Nothing in the parts has to be given up: the cells remain agents, now “equally selfish” but with “the relevant self that they defend … bigger in terms of space, time, and complexity”. Cut the coupling and the larger agent dissolves into its parts, which is the paper’s account of cancer.

Assumptions

The paper states three “core assumptions” (Introduction).

• Evolutionary continuity. Every capacity “emerged from simpler variants”. Cognition is not binary, so describing memory or goals in aneural systems is not a category mistake. Failing to credit them is a “neganthropomorphic fallacy”.
• Pragmatism about metaphor. Whether a cognitive description of a system is literal or loose is “an empirical question”, settled by whether it “improves prediction and control in novel cases at the bench”.
• Cybernetic goals. A goal is the set point of “a feedback system that operates to maximize some specific state of affairs”, modelled as an attractor. No stronger teleology is assumed.
• Third-person scope. “Except for a few brief remarks at the end, no claims about consciousness … are made.”

Two further assumptions are used without being stated as such:

• that the active-inference reading of cells (cells “minimize surprise and optimize variational free energy”) is correct, cited to Friston and colleagues;
• that physiological coupling, not shared genome or shared fate, is what binds parts into a Self. The paper calls its definition “orthogonal” to evolutionary ones but then uses only coupling.

Key results

There are no theorems, simulations or new data. What the paper offers is a definition, a scenario, cited evidence and predictions.

• Definition (Defining Individuation). “The scale and types of goals that a system can pursue defines (determines) the boundaries and content of the putative ‘agent’.” The boundary is the “Center of Concern” or “cognitive light cone”: “the most distant (in time and space) set of events that this system can measure and attempts to regulate”. In advanced agents it is also “the boundary of the self-model”. Fig. 2 places a tick, a dog and a human on a time × space diagram.
• Nesting (Fig. 2C; Multicellularity vs. Cancer). “In this scheme, Individuals can overlap.” A swarm, its animals, their organs, cells and intracellular networks “each have their own cognitive horizon”. The paper adds that “most biological systems consist of multiple, nested selves [not one, as implied by Integrated Information Theory (Tononi, 2008)]”. It also says “there is not necessarily one unique primary level of organization”, and that which level to describe is chosen empirically, as the one giving prediction and control “with the least effort”.
• Inner selves need partial isolation. “Each inner agent maintains its own local Self only to the extent that it regulates (restricts) information flow from neighbors and thus establishes unique states that do not coincide with those of the collective.” The Box puts it as a balance: too little sharing fails to bind, too much makes “a homogenous soup”.
• Evolutionary scenario (Agent’s Back-Story). The sequence runs from homeostasis (the “atom” of the hierarchy) to memory through delays and hidden layers, to anticipation, to multicellularity, to nervous systems as speed-optimised bioelectric networks. Multicellularity is said to be driven by “greedy infotaxis” and by surprise minimisation: “the least surprising object in the world is a copy of yourself” (citing Fields & Levin 2019).
• Cited evidence for coupling as binding.
  • Group-level decisions: left–right identity is randomised by whole domains, not cell by cell; melanocyte conversion happens body-wide.
  • Cancer: depolarising a specific cell population makes normal melanocytes metastatic in genetically normal tadpoles; optogenetic or constitutive hyperpolarisation prevents tumours from human oncogenes. Plastic barriers and gap-junction blockers induce tumours.
  • Bacterial biofilms coordinate by ion flows.
  • “most anesthetics … are gap junctional (bioelectric) uncouplers”, offered as “no coincidence”.
  • Swarm organisms (ants, slime moulds) show cognitive illusions like vertebrate brains, offered as a confirmed prediction of the scaling.
• Predictions (Predictions and Research Program). Among them: forcing electrogenic proteins and gap junctions into a unicellular organism should induce metazoan-like bodies; bioelectric genes should be conserved in regeneration across kingdoms; swarm organisms can be trained; “multi-human systems could have their own degree of cognition”, testable with information theory and reinforcement learning; and “adaptive, useful robotics will require components that are themselves competent and goal-seeking”.
• Consciousness (What Does It Feel Like to be a Pancreas?). “Nothing in this model explains” the Hard Problem. It is “natural to hypothesize that there is something-it-is-like to being a tissue or organ”, with “as much proto-consciousness as a simple neural network”. The model “does take a stand on the perennial ‘combination problem’ and can perhaps be seen as a form of panpsychism” (citing Chalmers 2013). What the stand is is not said.

Claims

id	claim	strength	support
C1	A Self is the set of parts pursuing a common goal, bounded by its cognitive light cone	a definition, not a finding	stipulated; argued only by its claimed usefulness. The light cone has no units or measurement procedure
C2	Most living systems are many nested, coexisting Selves, not one	moderate for agency in the cybernetic sense; assertion for anything more	follows from C1 plus examples of goal-pursuit at many levels; the contrast with IIT is a bracket with no argument
C3	Gap-junctional bioelectric coupling binds cells into larger Selves	moderate	cited experiments on patterning and cancer from the author’s laboratory and others; the anaesthetic point is a correlation
C4	Cancer is a reversible shrinking of the Self to one cell	moderate as a reading of the cited cancer data; weak as a general account	depolarisation and hyperpolarisation experiments in amphibians
C5	Multicellularity and larger Selves were driven by infotaxis and surprise minimisation	weak	evolutionary scenario, no test reported
C6	Collectives of organisms, including human groups, may have their own cognition, testable by training and information theory	proposal	analogy, informal examples (“the Supreme Court has opinions”)
C7	Adaptive robotics will need competent, goal-seeking components	conjecture	argument from the absence of “robot cancer”
C8	There may be something it is like to be a tissue or organ; the view is a form of panpsychism and takes a stand on the combination problem	speculation, labelled as such	none beyond the continuity between neural and non-neural networks

Concepts

• Self / Individual — “the set of parts that operate toward reaching specific goal states”; capitalised throughout.
• Cognitive (computational) light cone, Center of Concern — the spatio-temporal region whose events the system can measure and try to change.
• Scale-Free Cognition — the thesis that the same cybernetic processes (homeostasis, learning, goal pursuit) occur at every scale from molecular networks to swarms, and that the self–world boundary is the main “control knob”. The term is credited to Christopher Fields.
• Infotaxis — moving or connecting so as to gather as much information as possible, “from as far away … as possible”, to minimise surprise (after Vergassola et al. 2007).
• Synapse (broadly defined) — any structure that regulates the passage of signals according to the state of other such structures; gap junctions are the cellular case.
• Neganthropomorphic fallacy — failing to credit cognition to a system where doing so would improve prediction and control.

Connections

• Minsky. Minsky’s Society of Mind (LIT-429; Singh’s exposition LIT-444) is not cited. The comparison is still the point of filing this. Minsky’s agents are mindless, and mind is what their organisation does. Levin’s parts are agents with goals, and mind is what their goals become when shared. Both deny a single privileged controller. They differ on whether the parts are themselves minded, which is exactly the question the anti-nesting principle turns on.
• Dennett and the homunculi. The Discussion calls the view “consistent with Dennett’s stack of homunculi” (Dennett 1991): each level is the same homunculus with “a tiny bit more reach in physical space … and in time”. That is homuncular functionalism with the homunculi made biological and the stack read upward, not downward. The level-choice rule is Dennett’s intentional stance (cited, 1987), which LIT-220 states in its general form.
• Hofstadter. Gödel, Escher, Bach (LIT-437) is cited for collective states that “compute meta-system properties not accessible to the single agents”, and for agents that “transcend one system and re-set its goals”. GEB’s “Ant Fugue”, a colony that is a mind, is the literary ancestor of the paper’s swarm examples, though the paper does not cite it for that.
• IIT. The paper cites Tononi 2008, not IIT 3.0 (LIT-445), for the single-Self picture. IIT 3.0 says that “of all overlapping sets of elements, only one set can be conscious”, which is a claim about experience. Levin’s nested Selves are a claim about goal pursuit. The two are not yet in contradiction. They meet only in the closing section, where Levin extends proto-consciousness to tissues and organs, and there he gives no argument.
• Causal emergence. The paper names Hoel et al. (2013, 2016) as the way to compare “the causal power of models at different levels” and to find the most efficacious level. LIT-025 reviews that line. Using it would turn Levin’s pragmatic level-choice into a measurement. The paper does not do this.
• Brooks. The robotics prediction cites Cudhea & Brooks (1986) for “modular designs with sub-goals”. That is the same laboratory and architecture as Brooks’s “Intelligence without representation” (LIT-436), though not that paper. Levin’s complaint is that robots are made of “reliable but very dumb parts”. Brooks’s layers are exactly that, and Brooks counts it a virtue.
• The record’s Levin line. LIT-195 (world models and life–mind continuity, Levin a co-author) and LIT-211 (the definition-of-life survey, Levin a co-author) build on this paper’s continuum of cognition.

Bearing on the record

The bridge between Minsky (LIT-429) and Schwitzgebel (LIT-159, NOTE-131) turns on whether a minded whole can have minded parts. This paper is the record’s clearest statement that, for agency, the answer is yes and is the usual case. It sits on the composition side.

Against NOTE-131’s numbering:

• C3 (a whole of sensing or conscious parts can be conscious): supports, for agency only. Its nested Selves are wholes made of goal-pursuing parts. Its human-group proposal is the US case in functional terms. For experience it offers only the closing speculation.
• C4 (Putnam’s anti-nesting principle is unjustified): neutral. Putnam is not discussed.
• C5 (IIT’s exclusion postulate has absurd consequences): neutral as argument, opposed in conclusion. It rejects IIT’s one-Self picture in a bracket and gives no reason, and so lends C5 no support.

A THEORY drawing the bridge can cite it for the composition criterion: a whole is an agent when its parts share enough information to pursue goals none of them can sense. It parallels Schwitzgebel’s “entityhood-enough”, which is integration without spatial contiguity. A THEORY should not cite it as an argument against exclusion, or as evidence that collectives experience anything.

It carries no instruction for machine-learning practice.

Limitations

• The light cone is not operational. It is a diagram with a time axis and a space axis, and “much more work is needed to fully flesh out this rubric”. There is no rule for measuring it, so the definition cannot yet decide a hard case such as a colony, a holobiont or a nation.
• Goals are ascribed, not detected. “It is of course non-trivial empirical work to ascertain goals.” The cybernetic definition makes any homeostat a Self. The paper accepts this, but it leaves the criterion very liberal, which is the over-generation Schwitzgebel exploits.
• The bioelectric evidence supports coupling, not Selfhood. The cited experiments show that voltage and gap-junction coupling change pattern and tumour outcomes. That they change the boundary of a Self is the paper’s interpretation of those results.
• The IIT contrast is unargued, and aimed at an older statement (Tononi 2008). The phenomenal and the agential readings of “Self” are never separated, so whether the paper disagrees with IIT at all is not settled by its text.
• No model. The infotaxis scenario and the in-silico evolution it proposes are not built.

Open questions

• Can the cognitive light cone be measured for a collective, for instance from the reach of its sensing and the horizon of its prediction? A measurement would let the criterion include or exclude a nation, and so meet Schwitzgebel’s challenge directly.
• Does Levin’s account of agency at many levels conflict with IIT’s exclusion at all, or does it only become a rival when “Self” is read phenomenally? The closing section’s panpsychism remark suggests the latter, but it is not developed.
• Would causal-emergence measures (LIT-025) pick out the same levels as the goal criterion? If they diverge, which defines the Self?