Allostat

Allostat is a possible route to general intelligence in which homeostatic imperatives drive behavior. It builds on the Homeostasis Theory of Cognition and Consciousness (HTCC) from my i3AI research, and in particular on one of its ideas: cognitive homeostatic variables, or hvars.

Many imperatives, no controller

An Allostat is a set of hvars. Each hvar is an agent that keeps one variable within bounds - energy, hydration, safety, progress toward a goal. At every step, each hvar proposes actions and says how urgent its need is, and a fixed motor layer combines those proposals into one action. Nothing sits above them deciding what matters most, so behavior comes from many imperatives competing for the same body. The name comes from allostasis - stability reached through change - because the regulators themselves adapt.

From deliberation to reflex

Each hvar starts slow and general: every decision is a deliberation with a large language model that can read the whole system, look at what happened, and rewrite the hvar's own code. Each session has two jobs - decide what to do now, and leave the hvar cheaper for next time. What it learns moves down a ladder: into judgements from a small, fast model for questions that need judgement, and into plain code for everything that is really arithmetic or geometry. Deliberation is kept for situations that are new or that went wrong.

In the first runs the ladder moves quickly. Within a few sessions, almost every decision is made by code the hvars wrote for themselves, and what brings deliberation back is a death the code cannot explain. Those reviews find problems no single regulator could see - a newborn losing every blade of grass to the parent it stands next to, one regulator's aversion quietly pushing against another's escape - and fix them in the code that all individuals share.

Motivation that grows

Hvars can create, change and retire other hvars, and add internal state of their own - a remembered map of where water was, a record of where wolves were last seen. The structure of motivation is something the system grows rather than something I design. Many individuals share one Allostat: every member of a species runs on the same set of hvars, each with its own state, and every life and every death is evidence the whole set learns from.

Where it is tested

In a predator-prey world, sheep need grass and water and wolves hunt them, and each species can run its own Allostat, learning against the other. In Crafter, an open-world survival and crafting game with a ladder of achievements, a player starts with hvars for its vital needs and a single one for progress; whether that one splits into goals of its own - wood, a table, a pickaxe - is one of the things a run will show.

The same design now watches a coding agent. Replaying a recorded working session one turn at a time, a single hvar for the quality of the agent's work - as the person directing it judges it - splits into regulators for particular kinds of shortfall: keeping to the scope that was asked for, testing on the real device before shipping, carrying work through instead of listing it, reporting in plain words. Each new regulator starts by watching, and has to earn the right to step in. A second prototype moves this into live work: a separate regulator watches a customer-support agent as it talks with a customer, takes guidance from a person about what good support means, and can interrupt a reply to correct it. Between conversations it can also learn: finished conversations become revised instructions for the support agent, adopted only if a paired rehearsal against the original instructions shows they do better.

Goals are regulators too, but short-lived ones: each request sets up a goal whose target is what was asked for, and the goal steps in if the agent stops short, then disappears once the target is reached. The system also sleeps. A survival regulator tracks how often its recent decisions were wrong, and when that pressure builds, a rhythm regulator puts the system to sleep, where it dreams: it reviews its mistakes, groups them by cause and fixes each cause once, in the code every goal shares. The measure that matters is precision - an intervention helps only when it says what the person would have said next, and every push they did not want costs them a correction of their own - and raising it is what the current runs work on.

Where it came from

Allostat grew out of a series of experiments with small, inexpensive judgement models as a substrate for large numbers of semi-intelligent actors - closer to biochemistry than to a single powerful reasoner. Those experiments moved from measuring the models, to fast-and-slow agent harnesses, to regulators built on homeostatic variables, and then to this: each regulator an agent in its own right.