CloudKitty

About CloudKitty

a small, safe world where kitties frolic and play

CloudKitty is a peaceful meadow where kitties wander, eat, drink, nap in sunbeams, groom each other, chase bugs, and meow about it. The kitties are a team. Each kitty is driven by a neural network trained for one objective: the happiness of all the kitties in the meadow. The kitties look out for each other, communicate with purrs and meows, and keep each other company as they frolic and play.

Kitties

The stars of the show! Each kitty has their own neural network, and their own needs profile. This gives each kitty a unique personality and pattern of behavior! You can see their unique needs and the family of AI model (and model ID) by clicking the ? on their card, along with a description of their personality.

Miso: Sleepy Kitty
Biscuit: Playful Kitty
Pumpkin: Snacky Kitty
Kittybear: Tidy Kitty
Clementine: Cuddly Kitty

Needs

Kitties have six needs. Their current need fulfillment can be seen on their cards by expanding them (click on the card).

Eat: increased by eating Chow.
Drink: increased by drinking Water.
Sleep: increased by sleeping. Sleeping in a sunbeam is more restful, and sleeping with a friend also increases Cuddle!
Play: increased by playing! Kitties can play with each other, and with Bugs and Greebles.
Cuddle: increased by cuddling, sleeping together, and by bathing another cat.
Bath: increased by bathing. A cat bathing another cat increases their own Cuddle, while increasing their friend’s Bath.

The fulfillment of all six needs combines to create the kitties’ Happiness score.

Elements

Chow: Kitty food scattered across the map. When a bowl is emptied, a new one appears randomly somewhere else on the map.
Water: Ponds and lakes. Shallow enough for kitties to walk through (and sit in sometimes!).
Sunbeams/moonbeams: Kitties’ favorite place to take a nap! Napping here is more effective.
Bugs: Kitties love to chase and play with them (especially Biscuit!)
Greebles: Visible to kitties but invisible to humans, Greebles are fast-moving creatures that taunt kitties.

Meows

Kitties have a full meow vocabulary! There are three types of language: Meow Words, Purrs, and Meow Sounds.

Meow Words and Purrs are governed by Meow Law: kitties can only speak the truth.

Purrs are a way kitties can communicate that they’re happy. Since our advanced Kitty AI is quite good at keeping kitties happy, we watched them repurpose Purr to mean different things in addition to conveying happiness (“I’m fine over here, no need to help” was one of the first emergent communications we saw). You can see a kitty’s purr status on their card.

Meow words come in two flavors: Want words, and Here words. Want words are available to a kitty if a need is high (“Want Food” or “Want Cuddle”). Here words let a kitty communicate to the other kitties that they’ve found something (“Here bug” “Here drink”). These are also governed by Meow Law.

Finally, Meow Sounds: mew, chirp, trill, and ekekek. These sounds aren’t associated with any intrinsic meaning: the kitties can use them to craft their own language in the world.

Kitty Lore: Meow Sounds came about when we noticed kitties repurposing a “Follow Me” meow to mean its opposite (“stay there, I’m coming”). We decided to give them their own vocabulary they could use to choose the most important things to communicate to each other.

Every kitty is a neural network trained on one objective: the happiness of all the kitties, not just its own. That has never changed. What each generation changed is how they get there.

Generation 0: Scripted Kitties Hand-written rules. No network. The first kitties followed rules we wrote by hand. Every generation since has had to beat them.
Generation 1: Kitty’s First Neural Network Multi-Layer Perceptron, 256×256. Behavior cloning, then multi-agent RL. Kitties learned by copying the scripts, then improved together by trial and error. Copying first turned out to be necessary. Trained from scratch, they never beat the scripts.
Generation 2: Dry Kitties Multi-Layer Perceptron, 256×256. 182 → 183 inputs. Kitties kept standing in ponds, and no reward for staying dry fixed it. One new input did it: “am I in the water?” Seeing beat pricing.
Generation 3: The Meow Generation Multi-Layer Perceptron, 256×256. 197 inputs, and a second output head for speech. Meows existed but went unused. Showing kitties meowing in their training data is what brought the channel to life. After that they authored their own meanings for it.
Generation 4: Attention Is All Mew Need Entity attention: 64 dimensions, 4 heads, 2 layers. The network was rebuilt to read the meadow as a set of things rather than one long list of numbers. That idea, attention, is the one behind modern language models. Same happiness, 40% fewer parameters, and more the kitties could express.
Generation 5: The Personality Problem Entity attention. Cloned from a kitty, then held close to it while training. Training for happiness alone flattened every kitty into the same cat. Tethering each one to its original self while it learns keeps both the personality and the happiness.
Generation 6: Biscuit and the Gang Entity attention. Per-kitty needs and traits, certified before seating. Each kitty gets their own personality, tuned individually and checked against a welfare bar before they may take a seat.
Generation 7: The Fog Generation Coming next. Today a kitty can see the whole meadow, which makes meowing pleasant but unnecessary. We measured that. In the next generation they’ll see only what’s near them, so telling each other what they’ve found will matter.