DEV Community

Cover image for How Norns Were Created: A British Programmer's Difficult Path Toward Artificial Life
Artem X
Artem X

Posted on • Edited on

How Norns Were Created: A British Programmer's Difficult Path Toward Artificial Life

During the current boom in AI, I would like to remember the development story of a British computer game from the mid-1990s, informally called "Tamagotchi on steroids."

Under the cute design of the characters and locations there was a research project in machine learning and neural networks, and its technical ideas still look unusual and interesting today.

It all began with a self-taught British programmer. This is the story of the development of the first Creatures. Let us begin.

If you find factual inaccuracies in the article, you can write in the comments or DM the author; I will definitely correct them. I am not sure how Habr calculates reading time, but at the very end I included a very large bibliography. You obviously do not have to read all of it, so keep that in mind.

Beat Zero: The Beginning

Steve Grand first sat down at a computer in 1977. It was a PDP-8, and he immediately "crashed" it by pressing DELETE and throwing himself out of the program into the operating system. In 1977 you could not just buy a PC in a shop, so he spent several months designing his own computer until, as he put it, "manna fell from heaven" in the form of the Nascom 1: a blazing 4 MHz of speed and an entire kilobyte of RAM for 199 pounds.

His first full program, a checkers program that learned to play, was written in hex codes and fit into 768 bytes.

In late 1986 and early 1987, Steve Grand worked at Logotron, a small British company publishing educational software. It was during that period that he read The Planiverse, a science-fiction book by programmer and mathematician A. K. Dewdney about a simulation of two-dimensional life.

The book impressed him so much that he proposed making educational software based on it, but the idea did not receive support.

In 1989, Logotron was acquired by the Longman publishing group, while the game division split off into a separate company. Legally it was registered on July 27, 1989 under the name Starclear Software.

Starclear later became Logotron Entertainment, under license to use the brand. In 1990 it began using the name Millennium, and in 1992 it was officially renamed Millennium Interactive. The educational side of Logotron went under Longman.

Grand later recalled that he ended up porting a side-scroller to PC for Millennium. Their expert said fast background scrolling on PC was impossible. Grand did not know it was impossible, so he did it. A few weeks later the game was ready, everyone was impressed, and more orders came in. That was how he unwillingly became a game programmer.

Millennium Interactive was based in the picturesque British village of Great Shelford. It was there that Steve Grand began work on the key projects of his career as a game developer and AI researcher.

Beat Zero and a Half: The Book That Started It All

To understand what Grand wanted to build in Creatures, you first need to read the book he himself called his main source of inspiration: The Planiverse: Computer Contact with a Two-Dimensional World by Canadian mathematician and computer scientist A. K. Dewdney, published in 1984. It is not fiction in the usual sense. It is science fiction written as an engineering report wrapped in a spiritual allegory.

Who was Dewdney?

Alexander Keewatin Dewdney (1941-2024) was a fairly well-known figure in the academic culture of the 1980s: a Canadian mathematician and computer scientist, professor at the University of Western Ontario.

In his first column in May 1984, Dewdney published the concept of Core War, a game in which two programs fight for survival in shared virtual memory. So the author of The Planiverse was not merely an armchair scholar. He actively popularized the idea that a program can be a form of life. That idea runs through all of his work.

Where did The Planiverse come from?

The book was not born as a literary project. In 1977, Dewdney became interested in the idea of a two-dimensional universe as a philosophical metaphor. In 1979, he published a small scientific monograph, Two-Dimensional Science and Technology, a serious academic analysis of how science and technology might work in a world with two spatial dimensions. In July 1980, Martin Gardner wrote a Scientific American column about that monograph, and the print run sold out in a few weeks.

After that, Dewdney received a stream of letters from readers around the world: physicists, engineers, biologists, chemists, each proposing ideas about how a two-dimensional universe could work.

In 1981, he published an expanded collection, A Symposium of Two-Dimensional Science and Technology, with the best of those ideas. By the time The Planiverse itself was written, the two-dimensional world had already become a collective scientific project involving several hundred people and roughly a decade of shared work. The 1984 book was the literary packaging of the result.

The main trick: everything is derived bottom-up.

I will not recount the plot here. The main reason to read the book is the obsessive technical elaboration of its world. Dewdney does not describe Arde as "an alien world where everything is different." He derives the properties of that world from its initial physical constants, and every decision is accompanied by a drawing or diagram.

A few examples:

  • Physics. In a two-dimensional world, gravity, light, sound, and any radial forces decay as 1/r, not 1/r^2, because a circle around a point in 2D consists of two rays, not a sphere. This changes everything: atmosphere holds differently, distance behaves differently, and the physics of sound and optics is different.
  • Biology. A two-dimensional animal faces a fundamental problem: a through-gut would split it in half. Most of Arde's fauna avoids the problem by not having one at all; digestion works in portions. Yendred has a more advanced evolutionary solution: a zipper. The esophagus literally unzips to let a piece of food pass, then zips back up behind it. Yendred is also bilaterally symmetric, has four arms in total, and has its mouth between two eyes.
  • Architecture. All houses on Arde are underground; otherwise regular two-dimensional rivers and winds would destroy them. From above, an Ardean city looks like a one-dimensional anthill: nothing is visible except entrances through which inhabitants run back and forth. Nails are useless in 2D, because any rod through a wall would tear the wall in half, so everything is fastened with glue and tape.
  • Technology. Dewdney includes diagrams of 2D fishing boats, steam engines, rockets, animal nervous systems, newspapers, and musical instruments. Each drawing is accompanied by an explanation of why it must work that way. There is even a one-dimensional analogue of Go called Alak, described in enough detail that people later actually started playing it.
  • Sociology. Arde's society is also derived from the constraints of two-dimensionality. Roads, for example, are a special problem: any object on a road blocks motion entirely, because you cannot go around it. Transport is arranged differently as a result.

The book is packed with diagrams and drawings, and physically it almost looks like a technical manual with a novel attached.

Why exactly did this inspire Grand?

Dewdney does not describe the behavior of Arde's inhabitants; he derives it from the laws of their world. A two-dimensional animal does not "act two-dimensional" for show. It behaves that way because it could not otherwise exist under such physics and biology.

The same principle became the foundation of Creatures. Norns are not scripted to behave "as if alive." Their behavior is supposed to emerge from a simulation of biochemistry, neurons, and genes. The principle "do not describe behavior; describe mechanisms from which behavior follows" comes directly from there.

Grand repeated this thought many times later, in interviews and in his book Creation: Life and How to Make It: the difference between "seeming alive" and "being alive" is the difference between a script and a simulation.

Beat One: Robin Hood and a Custom Engine, 1990-1991

By that point Grand already had his own engine behind him: Microcosm, which he had written around 1979-1980 as a flexible rule-based system, originally for educational simulations on the BBC Micro. The goal was to escape the rigid structure of classical text adventures.

Technologically, this was not a game engine in the modern sense, but rather a framework of autonomous agents with an if-then rule base. At Millennium it was renamed Gulliver and treated as a technological advantage for a commercial product.

Development of The Adventures of Robin Hood started in July 1990, and the game was released in September 1991 on Amiga, Atari ST, and MS-DOS. Grand was the programmer. The isometric interface was made in the spirit of Populous.

At first the game was meant to be about cowboys. The team had gone partway down that road, then the mood changed, someone casually suggested Robin Hood, and Grand thought: fine, I will make that. The co-designer was Ian Saunter, one of Millennium's founders. Grand drew the graphics at first; later they were polished by artist Robin Chapman.

Technically, Robin Hood is already 80% proto-Creatures. It has 64 locations, about forty NPCs, each with its own set of AI rules, about 600 rules in the system with a projected 1500 by the final build, and 32 attributes for every sprite - attributes Grand described as making up its "soul": hunger, optimism, sympathies, personality type.

Grand considered one of the engine's strengths to be that all characters exist continuously. They are all there, doing something offscreen, unlike rooms that disappear when the player is not inside them.

Years later, Grand described the idea this way: in Rome and Robin Hood, there is no explicit plot. Every character simply has a set of rules for how to behave, and the plot emerges from interactions between autonomous characters and the player.

In other words, Robin Hood is already the same philosophical experiment as Creatures, but built on rules instead of biology.

Beat Two: Rome AD92 and the Channel to Maxis, 1991-1992

In 1991-1992, Grand reused Gulliver for a second commercial game: Rome: Pathway to Power, released in Europe as Rome AD92. It used the same isometric Microcosm engine. The game spans the Roman Empire from the eruption of Vesuvius in 79 CE to 92 CE, and the player progresses from Roman slave to Caesar. Grand's wife Ann wrote the documentation, Richard Joseph wrote the music, and Saunter was again co-designer. The game was released in Europe by Millennium in 1992, and in the US by Maxis in 1993.

And here the most important part begins.

Grand later explained that after Rome AD92 failed, Maxis CEO Jeff Braun still wanted to work with him and asked for a proposal. Grand proposed Creatures, an idea he had wanted to make for years after being inspired by Dewdney's The Planiverse. Maxis was thinking in a similar direction, so the fit seemed natural.

In other words, the failure of Rome unexpectedly opened exactly the window Grand had been waiting for for six years. Maxis in 1991-1992 was the epicenter of a worldview shift in commercial game development toward simulations.

Grand later noted that in 1991 Will Wright was already at Maxis with a prototype of The Sims, and both of them had independently arrived at similar ideas.

Beat Three: The Project Is Born on a Motel Balcony, 1992-1993

In September 1992, Creatures was "officially" conceived: Grand made the first sketches on a motel balcony in the town of Winthrop, Washington.

On November 4, 1992, he wrote "A Mouse for Windows", a desktop pet running around icons. On November 16 came "Little Computer Ewoks", by analogy with Little Computer People from 1985.

On March 8, 1993, the document Small Furry Creatures: A Mythography appeared: for the first time with that working title borrowed from Douglas Adams, and for the first time with Journey, Grendels, Ettins, and Shee appearing in the notes. On June 1, 1993, we get the first surviving entries from Grand's programmer diary.

Meanwhile, life at Millennium continued as usual: the studio released James Pond 3 in 1993, began developing MediEvil in 1995, and kept making normal commercial games. From late 1992, Grand was officially working on Creatures, initially as a DOS project, but within Millennium he was still, for another year and a half, one person with a gigantic ambition.

Beat Four: Creatures Receive the Right to Live, 1993-1994

1993: the first crisis, a stylistic one.

By mid-1993, Grand had already been working on Creatures for a year as a solo developer inside Millennium, with Maxis expected as publisher.

But a quiet conflict began inside the Millennium team. Grand wanted to embed mythology - English or Norse - for the internal coherence of the world, while the team pulled the project toward a classical adventure game.

On December 12, 1993, Grand formally clarified the issue. He had never conceived Creatures as an adventure game or a game with a fixed plot. The idea was that players would create their own stories. But people interpreted his mythographic notes too literally and tried to turn Creatures into some kind of Norse adventure.

This was the first serious signal that the project was not understood inside the studio. Everyone expected an ordinary game from it.

March 1994: "Creatures 0", the first working prototype.

By early 1994, Grand assembled the first working prototype under the working title Small Furry Creatures. A version marked "Millennium version" was handed to Millennium Interactive on March 22, 1994. The prototype ran under DOS, not Windows.

There was also a "Maxis version", confirming that Maxis remained the expected publisher at least until that point. In the same March, Small Furry Creatures was shown publicly for the first time, in a preview in Megazone issue 37.

The prototype is now known in the community as Creatures 0. It survived and can be run under DOSBox. It already has a neural network, biochemistry, and the beginnings of learning, although the norns look "like chickens" and often get stuck.

April 1994: Maxis cancels the project.

And then disaster happens. Grand's diary preserves the entry: Maxis officially cancelled the project, although they were still interested in discussing SimSeaWorld later.

By the mention of the mysterious "M" in earlier entries, we can reconstruct that Maxis had been involved from at least March 1993 to April 1994, about a year.

In other words, the very publisher for whom Grand had started making Creatures after Jeff Braun's interest left the project. Apparently this was connected to internal turmoil at Maxis; Grand later wrote that some kind of internal upheaval had happened there.

This could have been the death of the project. No publisher, two years of work by one programmer, and the rest of the team not understanding what he was doing. In most studios, a project in that situation would quietly be shut down.

May-July 1994: Millennium makes a decision.

But something important happens next. Instead of shutting the project down, Millennium loosens the leash.

From May 3-10, 1994, Grand's diary documents a prototype "Secret Adventure Mode." Then, on July 6, a key decision is made that determines the project's future.

Michael Hayward, one of Millennium's founders, decides that the game needs more time to mature and tells Grand to take it out of the schedule. Five months are added to the schedule, along with a list of things to redesign:

  • More graphical decoration.
  • More facial expressions.
  • A new norn look, more monkey-like.
  • A unique norn for each player.
  • A full code rewrite.
  • The ability for norns to travel from one computer to another.
  • The ability to talk to the player.

This is essentially the moment when Millennium decides that the project is not just another game in the catalog, but an R&D task with its own economics. Taking a project off the release schedule was a nontrivial decision in the 1990s. It meant: we are ready to invest without understanding when it will pay back.

On October 10, 1994, another important detail appears: plans for DDE support are finalized. This laid the foundation for external tools to communicate with Creatures, which would later turn into COB objects and the modding community.

November 1994: the Warner demo, and six days later, Cyberlife.

Here comes the resolution. On November 14, 1994, Michael Hayward first demonstrated a Creatures prototype to publisher Warner Interactive. Warner's reaction was strong: they compared the breadth of the audience for Creatures to the effect VisiCalc, the first spreadsheet, had once had.

That may be the best compliment a technological product can receive from a publisher: not comparison with another game, but with a category-defining tool that changed personal computers.

On November 20, 1994, exactly six days after the Warner demo, Millennium created a separate subsidiary, Cyberlife, specifically to build products around Grand's artificial-life concepts. The result was the Creatures series.

From that date, Grand ceased being "a lone programmer inside Millennium" and became the head of his own R&D unit. The team around the project started to grow.

The timeline can be summarized like this:

  • Maxis cancels the project in April 1994, depriving Creatures of its publisher and justification for existing.
  • Hayward personally decides not to close the project, but instead to increase the time budget and redesign it.
  • The extra five months of prototype polishing pay off: by November, the result is something Warner Interactive compares with VisiCalc.
  • Warner's strong reaction and the promise of a serious publishing contract, a one-million-pound advance signed in mid-1995, give economic justification for the separate unit that will develop the game.

Beat Five: The Birth of the Visual Style, 1994-1995

The first strategic decision concerns how the game should look at all. The March 1994 Creatures 0 prototype used pure pixel art, ordinary for a DOS game of the time.

The final product needed a stronger visual style, and the solution Grand found was radical.

The idea: build the world physically.

On January 23, 1995, Grand had the idea not to draw Albia, but literally build it as a background model, then scan and digitize it. Project artist Mark Rafter agreed a few days later and began work.

For the PC industry of 1995 this was an unusual move. Doom had already been out for two years, Quake was one year away, artists were massively moving to digital painting, and the idea of building scenery by hand like a museum exhibit sounded almost antique.

But Grand had his own logic: he wanted Albia to feel not like a drawn background, but like a real place. Dollhouse physicality gives the image a mass that flat illustration cannot reach by definition.

Execution: a team of museum model makers.

Mark Rafter drew the model designs. Physical execution was assigned to Complete Fabrications, a local Cambridge studio of model makers that had previously worked for museums. The total model budget was about 15,000 pounds. The finished model filled three large glass display cases, each worth about 5,000 pounds.

The scale was genuinely museum-like. The ocean scene with the Statue of Nornity measured 4 by 3 feet and 5 feet high, roughly 1.2 x 0.9 x 1.5 meters. The forest scene was 6 by 3 feet with the same height, roughly 1.8 x 0.9 x 1.5 meters.

These were not tabletop miniatures, but full demonstration dioramas with miniature interior objects, environmental props, and even electrics. Garlands were embedded in the model to control lighting, because a day-night cycle was originally planned for Creatures. In the final game the cycle remained technically, but was removed visually; it would return in Creatures 2.

Some locations, such as the desert island and small waterfall, were not part of the model and were drawn separately later.

Digitization and post-processing.

The finished dioramas were photographed with a digital camera, then reworked for the technical limitations of PCs of the time. Jason Riley and Colin Swinbourne handled retouching, adapting the image to the palette and resolution of the era.

This pipeline - physical diorama, digital photography, manual retouching - gave Creatures its recognizable painterly 2D style: the feeling that you are looking at a scene through the glass of a museum display case, with slightly blurred physical depth in the light and shadows.

Digital graphics in 1996 could not have achieved this effect alone: the palette was too flat and the outlines too hard.

Meanwhile, everything else is being decided.

While the model is being built, two other important processes are underway. By mid-1995, Cyberlife signs a publishing contract with Warner Interactive: a one-million-pound advance against a forecast of 200,000 copies. This is money with which you can properly hire a team.

And on July 19, 1995, Creatures, as the chronicle puts it, finally receives a full design description - as with most software projects, this required first building most of the program.

This date matters as a turning point: Grand had been working from intuitive sketches for 2.5 years, and only now, in the fifth year after the first notes, a formal specification appeared.

Beat Six: Developing Life, 1995-1996

Phase one: team and biology, second half of 1995 to March 1996.

During this period, Grand turns from lone programmer into the lead of a growing team. Dave Cliff deserves special mention: probably the most academically significant external consultant on the project. Cliff was an expert in artificial life at the University of Sussex, and later worked at MIT.

Grand asked him to evaluate the project. Cliff saw, under the cute creatures, one of the most complex artificial-life environments then in existence, and was impressed enough to join as a consultant.

In the credits he appears under "Special Thanks To", but his role was broader: he gave the project scientific cover and later became a co-author of academic publications. Cliff is the person who helped the project transition from "a game about virtual pets" to "a commercial product that can produce a paper at the International Conference on Autonomous Agents."

Phase two: first life, March-April 1996.

This is the most beautiful moment in the entire development chronicle. On March 21, 1996, at 10:50 AM, the first captive-bred norn was born: meaning a norn born from two parent norns, not created manually from a preset by developers. His name was Cain; his parents were Ron and Eve. The reference is fairly obvious.

On April 22, 1996, Creatures gained a disease mechanism caused by bacteria. This was a late but critically important feature. It made the world genuinely dangerous and raised the stakes of the player's emotional attachment.

Phase three: the final push and a change of leadership, summer-autumn 1996.

By mid-1996, the amount of work no longer fit into one Grand. He later described Toby Simpson as the producer of the last phase of C1, after Grand stopped being the only developer, and the person who then led C2.

So Toby Simpson is the second key person in the history of C1 after Grand. Before Creatures, Simpson had worked on two Diggers games at Millennium. At Cyberlife he became Creative Director and Executive Producer.

This change of leadership matters for understanding the product: the first half of development is Grand as lone visionary; the second half is Simpson as product manager, turning a scientific project into a commercial game.

Grand himself admitted that he began "losing touch with the product." But that is exactly why Creatures reached release as something that could be put into a box and sold in a shop, not merely as a research demo.

By November 1996, Cyberlife has 10 people on the core team, plus contributors in graphics, QA, and other areas. The credits show about 20-25 names in total.

Release: November 11, 1996.

Creatures is released on November 11, 1996 after, in Simpson's phrase, "4 years of development and more than 20 years of prior research." The "20 years" is a stretch, but if counted from the first Microcosm in 1979, the number is realistic.

The initial reaction exceeds forecasts. Douglas Adams, the same author whose "small furry creatures from Alpha Centauri" gave the project its working title, described Creatures as "more exciting than discovering life on Mars."

One month after release, on December 18, 1996, Cyberlife released the "Norn 6-pack": Buffy, Dion, Jarvis, Melvin, Sharla, and Teesha, the first downloadable set of additional norns. This launched the modding scene.

Beat Seven: How Norns Work

This section contains many technical details. Most of the information comes from Alan Zucconi's 2020 article The AI of Creatures; the link is in the bibliography.

Technically, Creatures is unusual even by today's standards. It consists of three tightly connected subsystems, coordinated through a fourth. Let us unpack the layers.

Architecture: three layers plus glue.

Each norn is not one program, but three separate simulations running in parallel and exchanging data:

  • Genome: a description of how this particular norn is built.
  • Biochemistry: its "body": chemicals, reactions, organs.
  • Brain: a neural network that makes decisions and learns.

The genome connects them. It encodes brain parameters, biochemical parameters, and appearance. When two norns mate, what is inherited is not "behavior", but the construction of the whole machine. Each child receives a slightly different brain and a slightly different metabolism, and behavior emerges from those differences by itself.

Genome: 16 gene types.

Norns have 16 different gene types divided into four basic categories: brain genes, which define lobes, neural dynamics, and dendrite properties; biochemical genes, which define receptors, emitters, reactions, half-lives, and initial concentrations; organism genes, which define appearance; and, in C2+, organ genes. Reproduction is sexual, with crossover. Exchange points and the number of exchanged genes are random. Sometimes crossover errors occur, leading to loss or duplication of individual genes. This is not a bug, but a key feature: those "errors" provide evolutionary material. Mutations can make a norn immortal, multi-lobed, or infertile; they can also produce behaviors the developers never anticipated.

Biochemistry: chemistry as the lower layer of emotion.

A norn's biochemistry is the set of all its chemical reactions. The idea is that when the brain is connected to chemical monitoring, it can make decisions such as "maybe I should eat, because glycogen is low."

A concrete reaction is represented as a gene describing a formula. For example, a real gene from a standard norn says: 1 unit of glucose + 2 units of hexokinase -> 4 units of CO2 + 1 unit of hunger; half-life = 24. In other words, when a norn spends glucose, its biochemistry automatically generates a "hunger chemical." Then two types of "sensors" become involved:

  • Emitters generate chemicals under certain conditions: a pain emitter triggers on impact, a cold emitter at low temperature, a stress emitter when drives are excessive.
  • Receptors watch chemical levels and alter brain behavior or other functions. One receptor effect is tracking the aging chemical and switching the norn between life stages.

"Hunger", "pain", "fear", "arousal", and "tiredness" are not flags in code, but concentrations of chemicals that rise and fall according to biochemical equations. Diseases, poisoning, medicines, and drugs are all just other chemicals in the same system.

Brain: 952 neurons, 9 lobes.

First-generation norns are controlled by a four-layer neural network of 952 neurons and about 5000 connections, organized into 9 functionally different lobes. The terminology is borrowed from real neuroanatomy. Each neuron is a place to store a numeric value from 0 to 255. Most neurons lose their stored value over time; some faster, others slower.

Lobes in Creatures 1:

Lobe Size Function
Drive 16, 13 active pain, hunger, fear, boredom, sex drive, and other drives
Stimulus Source 40 vision: one neuron per object class
Noun 40 fires when the player enters an object name
Verb varies fires when the player enters a verb
General Sense 32, about 20 active special events: got slapped, hit a wall, my child is in front of me
Attention 40 Stimulus Source + Noun, focus selection
Perception container copy of Drive + Verb + General Sense + Attention
Concept largest "situations": learned combinations of percepts
Decision 16, 11 active what action to perform

Vision through categories, not pixels.

The most characteristic technical trick is how vision works. In the Stimulus Source lobe there are 40 neurons, each representing one of 26 object classes in the game: carrot, fruit, norn, grendel, incubator, small toy, big toy, machine, elevator. If a norn sees a toy, the "toy" neuron fires. If the toy also makes noise, the neuron fires more strongly. There is no image recognition; the norn operates on abstract object categories. Incidentally, this is surprisingly similar to how animal vision is thought to work today: there are separate neurons for faces, motion, edible things, and so on.

Despite its name, the Perception lobe is not an input lobe. It is a container into which values from Drive, Verb, General Sense, and Attention are copied. This was necessary because of an engine limitation: a lobe could connect to at most two others. To give Concept access to all four sources, they were collected into one proxy lobe. An architectural workaround became a principle.

Decision: Winner-Takes-All.

In the Decision, Attention, Stimulus Source, and Noun lobes, a Winner-Takes-All policy is used: at every moment, only the neuron with the highest value is considered active, and it determines the action or focus. Every tick, the norn literally chooses "the strongest thought" and acts on it. This explains why its behavior looks so abrupt and switch-like: it really is that way, without smoothing.

Behavior is assembled as a Cartesian product of two decisions:

  • Decision chooses the action: 11 actions such as push, pull, stop, come, run, get, drop, think/say, sleep, left, right.
  • Attention chooses the object to focus on.

If Decision is "push" and Attention is "food", the norn eats the nearest food. Yes, in Creatures, "eat" is a special case of "push": the norn physically pushes the food into its mouth.

Concept lobe: a library of situations.

Concept is the largest and most complex lobe. Its neurons correspond to "situations" a norn can find itself in. Each Concept neuron receives inputs from 1 to 3 Perception neurons. For example, one Concept neuron may represent the situation "I am hungry AND food is in front of me AND it is close." In a well-trained brain, this neuron should strongly push Decision toward the action push, because push food = eat food.

Concept is, roughly speaking, a learned library of associations: world state -> useful behavior. The genome sets constraints on which combinations can form. For example, one Concept neuron cannot use two different drives at once, excluding impossible situations such as "I am hot AND cold at the same time."

Concept -> Decision connections: dual dendrites.

Each Decision neuron receives 256 inputs from Concept, but not as a homogeneous set. 128 connections contribute positively, supporting that action in that situation. 128 contribute negatively, arguing against it. Technically these are separated into dendrite classes D0 and D1. The State-Variable Rule of the Decision lobe sums D0 inputs, subtracts D1 inputs, and adds the neuron's current state. In other words, the norn brain does not simply "activate." It explicitly models a balance of pros and cons in each situation. For neural networks in 1996, this is a rather elegant construction.

Learning: reinforcement, atrophy, migration.

Connections between Concept and Decision are not fixed. The brain physically rewires itself throughout life. Grand designed a system of three mechanisms:

  • Reinforcement. When a norn performs an action that produced reward, meaning the chemical Reward is emitted, currently active connections are strengthened.
  • Atrophy. When an action leads to punishment, meaning Punish or worsening drive, currently active connections are weakened.
  • Migration. A connection weakened too much physically detaches from a neuron and reconnects to another. This lets the network rebuild topology, not only weights.

This triad solves a fundamental engineering problem. If all lobes were fully connected, about a million connections would be needed. Grand wrote in an academic paper that the total number of cells required to represent all possible sensory permutations up to four inputs would be unrealistically large. Out of that million potential connections, Creatures stores only 5000 - but not arbitrary 5000. It stores the ones that reinforcement and migration selected as useful. In essence, this is a sparse representation learned online, an idea that became mainstream in machine learning fifteen or twenty years later.

Instincts: what reinforces behavior when the player is absent.

For a newborn norn to survive at all, instincts are built in. A first-generation norn has 19 of them: special genes that inject Reward or Punish when certain combinations of neural activity and action occur.

C1 instinct distribution:

  • 11 teach the norn to obey verbal commands from the player: push, pull, come, stop, run, get, drop, and so on.
  • 2 support courtship and mating.
  • The rest cover eating when hungry, resting when tired, avoiding crowds, and pushing/pulling/wandering when bored.

Without instincts, a norn would not know what "correct" means at all. In Creatures 2, the list was expanded from 19 instincts to 44.

Real learning during sleep.

Instincts do not operate in real time. They are processed when the norn sleeps: during sleep, instincts are replayed over recent experience, reinforcing "correct" connections and weakening "wrong" ones. So a sleeping norn in Creatures is not an idle state with animation; it is a learning stage.

Grand effectively implemented an analogue of memory consolidation during sleep ten years before neuroscience seriously popularized the concept. If the player does not let a norn sleep, learning from instincts simply does not happen.

State-Variable Rules: evolving neuron logic.

Each neuron processes inputs not by a fixed formula, but by a genetically specified function: a State-Variable Rule, or SV-Rule. It is a small embedded DSL. The Concept lobe uses the rule anded 0:, meaning the neuron activates only if all inputs are active, a logical AND. Decision uses state:PLUS:type 0:MINUS:type 1:, meaning sum D0 inputs, subtract D1 inputs, and add the current state.

In other words, the logic by which a neuron works is itself written in the genome and can evolve. Mutate the SV-Rule of a particular lobe, and the offspring's brain literally "thinks" differently. This is a level of meta-evolution rarely attempted even in modern evolutionary neural networks.

World: COB and CAOS.

The things with which norns interact - carrots, balls, incubators, teleporters, learning computers - are implemented through a separate subsystem. COB, or Creatures Object, is an object in a .cob file containing images, a description, and CAOS code.

CAOS, Creatures Agent Object Script, is an embedded scripting language: register-based, low-level, without local variables. The basic unit is a command reacting to an event, such as timing, collision, or click. The language is deliberately low-level and hard to read: it uses magic numbers instead of named constants.

The command:

stim writ norn 10 255 0 0 66 255 0 0 0 0 0 0
Enter fullscreen mode Exit fullscreen mode

means "inject 255 units of substance ID 66 into the norn." That substance is progesterone. The ID list can be looked up in special tables, but the language does not become more readable from that.

The architectural idea was modern already in C1: the world is extended through agents. Any object can be added or replaced without modifying the engine. That is why the Creatures modding scene turned out so durable: users wrote new objects in CAOS and plugged them into a running game. This did not apply to norns themselves: their brain and body were hardcoded into the engine; only the environment was open.

What emerges from all this?

This architecture produces things impossible in an ordinary behavior-scripted game:

  • Norns genuinely learn, and they learn differently. Two norns from the same litter can grow into different individuals, because reinforcement is written into their personal dendrites and Concept lobes form differently.
  • Evolution actually works. After several generations, a population in the world begins to drift: it may tolerate hunger better, respond better to calls, or spend energy differently. This is adaptation, not a script.
  • Bugs become biology. In final C1, the "death from old age" gene did not work, so norns died only from infections. The game effectively became an epidemiological simulator, and Grand was fine with that outcome. Mutations can also lead to positive feedback loops in the brain: a norn may get stuck walking into a wall if, at some moment, doing so accidentally triggered pleasure chemistry.
  • Every norn is literally unique. Because of genetic recombination, personal reinforcement history, and stochastic connection migration, two identical norns do not exist in nature. That non-sameness is exactly what generates the player's emotional attachment.

Final Beat: The Player Community, Shelters, and Torture Levels for Norns

If the project looks unusual even from our time, for an ordinary gamer in the 1990s it was a shock. The game sold 500-600 thousand copies, which can be considered an unambiguous success for the time. Many players considered norns not only a form of digital life, but beings with feelings.

It is worth noting that while this was a fairly convincing simulation of a living organism, the norn brain was orders of magnitude simpler than that of insects and closer to a roundworm. C. elegans has around 300 neurons; norns have about 1000. Although the network architecture was radically different, the developers were primarily inspired by mammalian brains.

In any case, even then there were people who started torturing norns for fun or to provoke other users, drawing intense anger from the community. Some even received threatening letters.

But the enemies of norns were not only internet hooligans; their own genetics could also be a threat. Like us, they could have genetic defects, which could lead to genetic diseases.

Even the authors of the game talked about this

Even the authors of the game talked about this.

For example, one user described a curious case where a norn was born completely blind, deaf, and paralyzed.

For sick norns, and for norns who suffered from previous owners, people created full shelters where they tried to nurse them back to health.

Conclusion

The main thing that distinguishes Creatures from everything before it, and almost everything after it, is this: it was the first commercial game where the behavior of creatures was not described by the developer, but emerged from a simulation of mechanisms. And not just any mechanisms, but biologically inspired ones that solved two tasks at once: a scientific task, creating a plausible artificial creature, and an engineering task, making it run on 1996 hardware.

Grand did not ignore hardware limitations; they acted as a forcing function pushing him toward biologically plausible solutions. Out of a million possible connections, only 5000 are stored because there is no room for more. But those 5000 learn to become the most important ones, just as in real living beings, whose brains are also sparse.

Instincts teach the brain during sleep because while the norn is awake, the processor is busy with many other actions. But real animals also consolidate memory during sleep. Vision works through categorical neurons because there is no other practical way - but animal visual cortex, broadly speaking, is not that far from this.

Grand was not merely saving CPU cycles. Out of necessity, he invented a reduced version of what nervous systems really are. That is the bottom-up approach he had been walking toward since 1986, through three employers, the failure of Rome, Maxis cancelling the project, and its near-death in 1994.

Given everything described above, Creatures, released on November 11, 1996, looks like an innovative AI project even by today's standards.

Bibliography

This is a large source list. The most important technical source on the norn AI is Alan Zucconi's article:

Steve Grand:

A. K. Dewdney and The Planiverse:

Logotron, Millennium Interactive, and company history:

Core War:

Alak:

The Adventures of Robin Hood:

Microcosm / Gulliver:

Rome: Pathway to Power / Rome AD92:

Maxis, Jeff Braun, Will Wright, and The Sims:

Birth of Creatures, 1992-1993 timeline:

Little Computer People:

Douglas Adams and "small furry creatures":

MediEvil:

Development of Creatures:

Visual style of Creatures:

Norn torture:

Shelters, nurseries, and places for norns:

Top comments (0)