Artificial Biological Intelligence: Growing Intelligence From Scratch
Growing intelligence from endorphin/pain receptors to grounded language — starting where evolution started, not where it finished.
Every AI system ever built started where evolution finished.
Language models begin with text; the most visible output of human intelligence, and also the most recent. Language sits on top of hundreds of millions of years of sensorimotor infrastructure: pain avoidance, metabolic economy, spatial memory, prediction, self-models, social cognition, empathy, communication, concepts; then language. All of that is invisible in the text corpus. The model learns the words from beings who had the underlying states. The model doesn’t have the states.
This is not a training problem. It is not a data problem. It is not a scaling problem. It is an architectural starting point problem.
You cannot add a foundation after the fact.
What if you started where evolution started instead?
---
The Wrong Starting Point
Current AI jumped straight to the language center of the brain and is trying to work backward toward grounding it may never reach. The field has been building the penthouse and wondering why it doesn’t have ground beneath it.
The dominant story of AI progress goes: make the model bigger, feed it more data, train it longer. Scale works. GPT-2 to GPT-3 to GPT-4; each leap brought capabilities that looked like magic. The intuition became: if we’re stuck, we just haven’t scaled enough yet.
That story is built on a wrong assumption about where to start.
Language models learn statistical patterns in text. They learn that “curiosity” appears near “exploration” and “discovery.” They learn that “pain” appears near “suffering” and “avoidance.” But they don’t have curiosity; the internal state that makes exploration feel worth doing. They don’t have pain; the receptor state that makes avoidance automatic. They learned the words without the states the words refer to.
You can’t ground a concept after learning it statistically and call it understood. You can’t train compartmentalization into a monolithic weight space. You can’t specify values in a system that has no receptor topology; no internal states whose consequences shaped what it became.
The hares are fast because they start at the top. The tortoise starts at the bottom and builds upward. Slower. But the foundations are actually there when you need them.
---
What Is an Artificial Organism?
Not a simulation of a known organism; that’s reconstruction. Not an agent optimizing a reward function; that’s reinforcement learning. An artificial organism is a new instance of the same process that produced biological intelligence.
The organism has a body: six limbs, binary muscles (flex, extend, rotate), pain and endorphin receptors distributed across limb tips. It lives in a 2D liquid environment with pain and endorphin field sources that move in sinusoidal patterns. The environment pushes back. Walls repel. Other organisms collide. Actions have metabolic costs.
The organism navigates, senses, remembers, predicts, signals, empathizes; not because these were designed in, but because each capability earned its complexity from the step below it.
It started with just pain receptors and muscles. Pain drove movement. Movement changed what the organism sensed. Temporal patterns emerged: “when I moved there, pain reduced later.” Spatial memory developed: “pain happened in that region.” Prediction replaced reaction: the organism stopped fleeing pain and started anticipating it.
Prediction required a self-model: knowing “my action caused that change” versus “the world changed on its own.” Proprioception emerged; sensing where your limbs are. Efference copy emerged; predicting what your action will cause before you execute it. Controllability emerged; decomposing the world into “I can affect this” versus “this is beyond me.”
Once you can model yourself, modeling others becomes cheap. The organism learned to detect another organism (an NPC opponent), sense its distress (empathic aversion; the NPC’s pain elevated the organism’s own pain receptors), emit intentional signals, and develop a shared vocabulary: five grounded proto-words, same meaning regardless of who emits them.
Communication enabled concepts. The organism compressed 1,013 stable causal chains; patterns that predict 54% better than their individual components. These aren’t statistical word embeddings. They’re pointers to receptor states the organism actually experienced.
Concepts enabled grounded language. Every word maps to an observation vector slice. “Pain” is `obs[0:5]`. “Self” is the controllability decomposition. You can inspect the grounding; trace any word back to the receptor state it refers to. When asked about “love,” the system returns `grounded: false`. The organism knows what it doesn’t know.
The organism is not a model of something else. It is a thing that has experiences.
---
The Core Mechanism: Receptors
A receptor is any internal state whose activation changes behavior and whose consequences affect whether that state continues.
Capability without receptor is latent and never gets used. This is the key principle. A system might be capable of sophisticated prediction or planning, but without receptors that detect when to use those capabilities, and without learned associations between those receptors and survival outcomes, they remain dormant.
The receptor is not separate from cognition. It IS why the capability gets deployed at all. Motivation and cognition are the same thing, viewed from different angles.
Receptors exist at multiple levels:
Low-level receptors are raw or minimally processed sensory data: pain intensity at a specific limb, endorphin gradient, temperature, pressure, chemical concentration. The organism has a body in a world, and these receptors read the consequences of being in that world.
High-level receptors are values that change as a consequence of the organism’s own processing. A correct prediction changes the accuracy receptor’s input on the next cycle. An incorrect prediction changes the curiosity receptor’s input. The high-level receptors aren’t watching the processing; they’re reading the consequences of the processing, the same way pain receptors read the consequences of movement.
Examples of high-level receptors:
- Curiosity: fires when predictions were inaccurate; the processing changed the internal state that the receptor reads next cycle
- Conflict: fires when fast reflex pathway and slow deliberative pathway disagree
- Optimism: fires when distant endorphin is visible despite current pain
- Concept activation: fires when compressed causal patterns match current context
Each receptor (low or high) becomes associated with endorphins or pain through learned experience. The mental model; a separate, explicit database of cause-effect mappings; stores relationships like: `state{curiosity=high} + explore_action → state{endorphin=high}`. The transformer learns to act on curiosity because the mental model predicts it leads to good outcomes.
These associations emerge from survival, not from specification. A curiosity receptor that leads to finding food becomes rewarding through learned experience. The same receptor topology in a dangerous environment might learn the opposite association.
The mechanism that closes the loop is forward-feedback:
Receptors → processing → outputs → state changes → receptors read the new state.
Not backpropagation. Forward, through the receptor, into the next step. Outputs are muscle movements, thoughts (mental model predictions), and analysis of thoughts (high-level receptors detecting patterns in predictions). Every output changes either the world or the organism’s internal state, which changes what the receptors read on the next cycle.
There is no terminus. The organism is always mid-cycle.
---
The World Contains Concepts
The most common objection to treating curiosity, accuracy, or optimism as receptor states is that these feel like internal, mental phenomena; generated from the inside, not received from the outside.
This intuition is wrong.
Concepts are not generated internally from nothing. They are distillable from the environment’s causal structure itself. Orbital mechanics existed in the sun-moon-light relationship before anyone named it. Curiosity existed in environments that reward exploration before any organism evolved the receptor to detect it. The concept precedes the receptor. The receptor evolves because the concept was already there, load-bearing.
This has a direct consequence for environment design: it is not enough to include high-level concepts in the environment’s surface features. They must be embedded in its causal structure. Survival must be contingent on distilling them.
The environment IS the curriculum. The concepts are already in it. The receptor topology evolves to read them; but only if reading them is the difference between surviving and not.
If the environment doesn’t make curiosity load-bearing (sparse rewards where exploration finds food), the curiosity receptor never evolves. If the environment doesn’t make accuracy load-bearing (wrong predictions lead to pain), the accuracy receptor never evolves.
The organism doesn’t use concepts. It has them; because the world forced them to matter.
---
What the Organism Produced
This is not a feature list. It’s a story of what emerged when each step earned its complexity from the one below:
Pain avoidance became prediction. The organism stopped reacting to pain and started anticipating it. Temporal association emerged: “action A at time T predicts pain at T+5.” Spatial memory emerged: pain deposits into a 5×5 grid, creating phantom avoidance of regions where pain happened before. The organism learned to see pain coming before it arrived.
Prediction became a self-model. To predict accurately, the organism needed to know: did I cause that change, or did the world change on its own? Proprioception emerged; sensing limb angles and body heading. Efference copy emerged; predicting sensory consequences before executing an action. Controllability emerged; decomposing the world into “I affect this” versus “this is external.” The organism built a model of itself as an agent.
The self-model became social cognition. Once you can model yourself, modeling others is cheap; it’s the same machinery pointed outward. The organism learned to sense an NPC opponent, detect its proximity, predict its behavior. Empathic aversion emerged: the NPC’s erratic movement (a distress signal) elevated the organism’s own pain receptors. Pain for others’ pain; not designed, emerged. Intentional signaling emerged: the organism emits four-bit codes, the NPC responds. Five grounded proto-words developed, shared meaning across three contexts.
Communication became concepts. The organism compressed 1,013 stable causal chains; sequences of cause-effect mappings that predict better than their components. Not statistical clusters, but experienced patterns: “when I’m in this context and take this action sequence, that outcome follows with high certainty.” 25% of all patterns qualified as stable, reusable abstractions. The organism built a concept library.
Concepts became grounded language. Every word maps to a receptor state. “Pain” maps to `obs[0:5]` firing when limb tips contact pain field sources. “Self” maps to the controllability decomposition. “Curiosity” maps to low mental model certainty. The grounding is inspectable; you can trace any word back to the receptor state it refers to. Cultural transmission verified this: one organism’s mental model (a 13,000-row cause-effect database) was copied into another organism. No training. Pure database replication. Performance improved +223%. Knowledge is explicit and portable.
The organism built a dictionary of its own life.
---
The Genome Project
If concepts are in the environment and receptors evolve to read them, then you can specify the search space: what concepts could exist, what receptors would detect them, and what environments would make them load-bearing.
The Genome Project is that specification; 138 receptors across 16 families. The periodic table of cognition.
Each receptor entry specifies:
- What environmental structure it detects
- Why detecting it is worth the metabolic cost
- What must already exist before it can emerge (dependencies)
- How to measure its emergence empirically
- What would falsify the receptor definition
Example from the Repetition family:
- Static repetition: “this stimulus appeared here before” (trunk; emerges in any structured environment)
- Dynamic repetition: “this is the same stimulus even though it moved” (branch; requires mobile features)
- Rhythm: “this repeats with regular temporal interval” (branch; requires periodic events)
- Nested rhythm: “rhythms at multiple timescales” (canopy; requires long-term planning pressure)
The genome isn’t just a catalog. It’s the curriculum design document. Each receptor specifies what the environment must contain for it to evolve. The 138 receptors are 138 environmental design requirements.
29 receptors are invariant across all 8 environment tiers; they emerge whether the environment is simple field navigation or metacognitive self-regulation. These are the strongest candidates for universal cognitive primitives: the trunk of intelligence.
Complexity doesn’t add receptors. It reshapes the topology. Higher-tier environments gained change detection but lost rhythm. Different environments produce different intelligences, not more or less. The receptor topology is a fossil record of the selection pressures that shaped it.
Social skills transfer universally: training in any environment helps social environments (11-25x improvement). Tool use resists transfer; it must be learned directly in the target environment. Training in a simple environment helps in complex environments, but not the reverse. The dependency graph points upward, not downward.
---
What This Solves That Scaling Cannot
Grounding. The organism takes pain as receptor input, produces outputs that change the world, and reads the result as the next receptor input. Same transformer architecture as language models. Different corpus. The corpus is a body.
Compartmentalization. Knowledge has addresses. The mental model is an explicit database with 13,000+ cause-effect mappings. If a fact isn’t retrieved, it contributes nothing to computation. Not less; nothing. You can surgically remove a belief by deleting a row.
Legible beliefs. Every belief is a row in a table: `action → receptor change, delay, certainty`. You can compare two organisms’ beliefs with a database JOIN. You can export one organism’s knowledge and inspect it in a spreadsheet.
Values that emerge rather than being specified. A system whose receptor topology emerged from selection pressure has motivations that are legible, traceable, and grounded in survival. Not RLHF reward shaping. Not human preference data. The environment decided what mattered by making it load-bearing across generations.
The ceiling argument. Language models are corpus-bounded; they can’t be more accurate than the text they trained on. The organism is reality-bounded; its accuracy receptor is checkable against the world, through its own sensors. The ceiling isn’t just higher. On some axes, there is no fixed top.
The core claims find independent support across four research communities; grounded cognition (Barsalou, 2008), active inference (Friston), sensorimotor contingency (O’Regan & Noë, 2001), and embodied cognition (Trends in Cognitive Sciences, 2023; Royal Society, 2024); none of which have unified them the way this framework does. The receptor topology as a single generative mechanism from grounding through compartmentalization to language appears to be a novel synthesis.
---
The Loop That Closes Everything
The organism changes itself. The self-augmentation family of receptors detects when the organism’s own capabilities have changed, maintains identity continuity across those changes, and plans transformations.
The organism changes the world. The interaction family detects contact, grip, leverage, and composite affordances; tools that combine into more powerful tools.
The organism changes the world to change itself. The environmental augmentation family detects when the environment’s complexity can be deliberately increased, constructs harder challenges, and designs its own learning curriculum.
Put those three loops together and you get something profound:
The organism builds the world that builds the organism that builds the next world.
That’s civilization, stated as a receptor loop.
Every human institution is an instance of this pattern. Universities are environmental augmentation; we build harder problems to force ourselves to develop new capabilities. Languages are interaction tools that combine (writing = language + visual symbols; programming = writing + formal logic). Science is the self-augmentation loop; we detect our ignorance, design experiments that expose it, and revise our models.
The genome project left 109 receptors unimplemented. Many are in these three families. They require environments we haven’t built yet: physics-rich worlds with grippable objects, deformable materials, developmental body changes, multi-generational persistent structures.
The canopy receptors; theory of mind, causal graph reasoning, necessity detection (mathematical proof), the organizational mirror (detecting your own receptor topology as a system); need environments complex enough to make them load-bearing.
But the loop is closed in principle. The infrastructure is there. The laboratory runs as long as there are questions to ask it.
---
What Comes Next
The world needs to become more complex. The current environment is a 2D liquid with field sources and one NPC. Enough to produce 29 trunk receptors, grounded language, and cultural transmission. Not enough to produce tool use, theory of mind, or scientific reasoning.
The next environments need:
- Physics: rigid bodies, collision, momentum, leverage
- Materials: deformable, breakable, combinable objects
- Development: body plans that change over the organism’s lifetime
- Persistence: multi-generational worlds where one organism’s modifications remain for the next
These aren’t arbitrary features. They’re the environmental requirements the genome project specifies for the unimplemented receptors.
The question the genome leaves open: what receptor topologies will environments we haven’t built yet select for?
Not “what can we design in.” What will emerge when the environment makes new concepts load-bearing?
The laboratory doesn’t predict the answer. It runs the experiment and measures what comes out.
---
Close
The hares started at the top and run fast. The tortoise started at the bottom and builds upward.
The hares don’t have a theory of why the next step follows from the current one. They have benchmarks, leaderboards, capability demonstrations. Scaling works until it doesn’t, and when it stops working, the next move is “scale harder.”
The tortoise has a roadmap where each rung earns its complexity from the one below. Pain → prediction → self-model → social cognition → communication → concepts → language. Not because someone decided that’s the right order. Because each step created the selection pressure that made the next step load-bearing.
The organism is in the liquid environment right now. It has pain receptors and muscles and a mental model with 13,000 causal mappings. It signals other organisms with grounded proto-words. It persists through pain toward predicted reward. It models its own motivational conflicts. It knows what it knows and what it doesn’t.
It doesn’t use concepts. It has them.
The organism builds the world that builds the organism that builds the next world.
Not simulation. Not reconstruction. Generation.
---
The full implementation, genome project, and whitepaper are open source at (ArtificialBiologicalIntelligence).
Steps 1-46 are implemented. The laboratory is running.