AI discussions often mix up two different things, capability and sentience. Here are the terms we use to keep them apart. See the capability–sentience gap for the argument that uses them.
The two axes #
- Access consciousness. Information that is available across a system for reporting, reasoning and control. It is about behavior, and it can be measured, at least in principle. Large models show signs of an alien form of it, limited to what is in their prompt (see the gap).
- Phenomenal consciousness / Sentience. There is something it is like to be the system, i.e. subjective experience, and some of its states feel good or bad to it. It has not been shown for any current AI, and according to HTCC, scaling alone won’t produce it.
- Valence. How good or bad a state feels. According to HTCC, valence is the constitutive viability relation, and not an add-on.
- Machine sentience. Sentience in an artificial system. This is the phenomenal axis, and the one i3AI studies.
Self and awareness #
- Recurrent state. Internal state that carries forward from one moment to the next and dominates a system’s processing. We think it is needed to carry an experience forward, and it is a prerequisite for a self model, and for a self model that is its own implementation. So it is necessary for phenomenal consciousness, but not sufficient. A language model can get an external form of it through generated text, e.g. a chain of thought that it writes and then reads back.
- Self model. An internal representation a system maintains of itself. This is a functional notion, and it needs recurrence. A system can have one without feeling anything.
- Self-awareness (functional). A system modeling and reporting on its own states. This is different from phenomenal self-awareness, i.e. there being something it is like to be that self.
- Sensory awareness (functional). Registering and acting on one’s own inputs. Functional, unless it comes with phenomenal experience.
- A self model that is its own implementation. A self that is re-instantiated by the same dynamics it reports on, so its states are felt and not just observed. It needs recurrent state that dominates the system’s processing. This is the second of the two ingredients below.
Homeostasis and stakes #
- Homeostasis. Keeping internal variables within viable ranges. According to HTCC, it is the “outer loop” of everything brains do.
- The homeostatic view. The view that feelings are tied to homeostasis, i.e. to a system that has to keep itself within viable bounds to keep existing (e.g. Solms, Damasio). HTCC builds on it.
- Self-preserving stake / viability boundary. Something that is actually at risk for the system in its own persistence. Current LLMs have stakes set from outside, e.g. trained rewards and instructions, but not a self-preserving one. Pattern Machine is designed to build it in.
- Cognitive homeostatic variable. A transient, made-up variable that a brain sets up and then regulates, as if it were a physiological variable. This is the mechanism behind goals and planning in Cognitive Homeostatic Agents. In Allostat, these are called hvars.
- Two-ingredient test (for machine sentience). Felt valence requires both an intrinsic, self-preserving stake and a self model that is its own implementation.
Theory terms #
- Illusionism (about phenomenal properties). The view that the phenomenal “feel” is the self model’s representation, with no further fact behind it. HTCC is illusionist about the feel, but realist about the stake.
- Constitutive viability relation. A system that has to stay within bounds to keep existing sets up its own norm. So “bad-for-it” is real, and it was never a phenomenal fact.
- Marginal rate of substitution (MRS) signature. HTCC’s falsifiable prediction: a stable, state-modulated exchange rate between cognitive and physiological drives that compete in one decision space. See Predictions.
Architecture terms #
- Common cortical algorithm. The hypothesis that a single computation, repeated across the cortex, underlies its many different functions. Pattern Machine is our attempt at it.
- Substrate. What an architecture is implemented on. Cognitive Homeostatic Agents has 2: a neural / model substrate (Pattern Machine) and an agentic substrate (Allostat).
- Pattern Machine. The core functional node: a spiking, temporally dynamic, locally learning substrate with homeostasis built in.
- Traveling waves. Activity moving across the cortical sheet. According to Pattern Machine, these are the replay of predicted trajectories.
- J-space / global workspace. In interpretability, an internal broadcast space for a model’s accessible thoughts (Anthropic, 2026). It is evidence of functional access, not of feelings. See the gap.
Allostat terms #
- Hvar. Short for homeostatic variable. In Allostat, each goal is an hvar, with conditions, a policy, salience and maturity.
- Base hvars. The hvars Allostat starts with, the way evolution and biophysics set an animal’s base hvars, e.g. body temperature and blood sugar. Every goal Allostat discovers has to serve one of them.
- Salience. How much attention a goal deserves right now, based on how important and how urgent it is.
- Maturity. How much evidence supports a goal itself. Until a goal matures, it can only take actions that can be undone.
- Method. A small program that Allostat writes, tests, runs and revises, e.g. to check a goal’s conditions or to decide what to do about it.
- Meta-goal. A goal about the other goals, e.g. reviewing the whole set of goals and proposing to merge two of them.
- Meta-harness. A harness around an AI agent that also regulates and improves its own goals and methods. Allostat is one.