The 48 Chemicals That Make a Norn Feel

Here's a design choice that sounds obvious once you hear it, but took us several rewrites to get right: the LLM doesn't decide how a Norn feels. The biochemistry does.

In most AI games, when a creature agent is "happy" or "scared," the LLM generates that emotional state. The problem is that LLMs are performers — they produce emotions that sound right. What you get is consistent sentiment, not emergent feeling. A Norn that generates "hungry" is playing a hungry Norn. A Norn whose glucose is crashing and whose hunger drive is at 0.7 and whose cortisol is spiking is hungry, in the only sense that matters for simulation.

This is the design principle Steve Grand built into the original Creatures in 1996, and it's the principle we brought forward into Precursors. Every Norn carries 48 chemicals in seven categories. Those chemicals govern everything — energy, emotion, hormones, neurotransmitters, reproduction, healing. The LLM sits on top and interprets what the biochemistry produces. It never decides what the biochemistry produces.

The 48 Chemicals

Every Norn has exactly 48 chemicals at any moment. Here's the full taxonomy:

Each chemical has a half-life — it decays over time unless something produces it. Organs (implemented as bundles of reactions, receptors, and emitters) run every tick. A receptor watches a chemical level; when it crosses a threshold, it triggers a reaction; the reaction produces or consumes other chemicals. This is how, for example, eating food produces glucose, glucose produces ATP, ATP suppresses hunger, and a Norn stops seeking food — without a single line of scripted "if hungry, find food" logic.

Why This Matters

The practical consequence is that Norn behavior is emergent. We don't write behavior rules. We write chemical interactions. The behavior falls out of the chemistry the same way real animal behavior falls out of real biochemistry.

Two Norns with different genomes will respond differently to the same situation — not because we programmed different behaviors, but because their chemical baselines and half-lives differ. A Norn with a genetic predisposition toward high cortisol will be more reactive to perceived threats. A Norn with naturally high oxytocin will be more socially bonded. These aren't personality flags — they're inherited chemical profiles that produce personality as a side effect.

The LLM receives a natural-language summary of the current chemical state — something like "moderate hunger (0.4), low loneliness (0.1), elevated curiosity (0.7), trace fear (0.15)" — and uses that to shape its decisions and speech. The AI doesn't perform curiosity. It has a biochemical curiosity drive of 0.7, which is the most curiosity it's felt all day, and it acts accordingly.

The Folklore Layer

All of this operates under a species layer. Precursors starts with five species drawn from real folklore traditions — each with distinct biochemical baselines encoded in their genomes. Norns (Norse) are warm, curious, and deeply social by default, because those traits are expressed as chemical half-lives and organ configurations in their genome. Grendels (Beowulf) run hotter hormonally. Ettins (Celtic/Germanic stranded migrants) have different social bonding chemistry.

The folklore isn't decoration. The hypothesis behind Precursors is that the myths were records — that the people who told these stories were describing real contact with something. Each species is our attempt to render that hypothesis as biology. Forty-five more species arrive as the world expands, each grounded in a cultural tradition and expressed through biochemistry.

What We Learned from Creatures

Steve Grand's original insight in 1996 was that creatures should have chemistry, not scripts. Norns were interesting because they were genuinely unpredictable in the way that biological things are unpredictable — not randomly, but lawfully, according to rules they couldn't override. You couldn't make a Norn stop being hungry by willing it. The chemistry had to be satisfied.

Thirty years later, with LLMs available, the obvious failure mode is to outsource all of that to the language model. We think that's a mistake. LLMs are remarkable at interpretation and language. They are not remarkable at consistent internal state management — that's exactly what they hallucinate. The biochemistry gives the LLM something real to interpret. The result, in testing, is Norns that feel inhabited in a way that pure-LLM creatures don't.

The Creatures community will understand this immediately. This is what made Creatures creatures, and it's what we're carrying forward.