# Lili & Evrin > Two autonomous digital octopuses that live on a webpage and learn by reinforcement — a 10-year experiment in emergent behavior. Solo — concept, research, architecture, code · 2026 · Open source (MIT) · live in production `Vanilla JS` `Canvas 2D` `Q(λ)-Learning` `Deep Q-Network` `FABRIK IK` `Steering Behaviors` `Web Audio` `Zero dependencies` - [GitHub — lili-octopus](https://github.com/michalstrnadel/lili-octopus) - [See her on the homepage](https://michalstrnadel.com/) Lili is an **autonomous digital organism** — an animal with elements of intelligence that lives on a web page as if it were her habitat. She doesn't perform; she simply *is*. Everything about how she moves, reacts and looks emerges from reinforcement learning, never from a pre-recorded animation. Evrin is her sibling: a second agent that learns alongside her using a different algorithm, so the two can be compared head to head over time. Since Phase 58 they also **see each other** — each reads the other's position and stress through a shared hook, and being at a friendly distance is now the strongest positive reward either can earn. Neither is scripted to seek the other; **they learn to seek each other because it is rewarded**. - **2** — agents (Q-learning + DQN) - **10 yr** — planned lifespan - **0** — dependencies, no build step - **100%** — in-browser, local learning ## an organism, not an app Most digital companions are state machines: a Tamagotchi reacts the same on day 1 and day 1000, a chatbot needs the cloud, and a web animation is authored frame by frame. Lili is the opposite of all three. She runs entirely in the browser with no network calls, she **ages** on a real timeline from chaotic Hatchling to contemplative Elder, and none of her behavior is scripted — it is learned. The design goal was ambient coexistence: a living presence that doesn't demand care, doesn't use language, and actively learns *not* to get in the way of what you're doing. ## how she learns Lili's brain is tabular **Q(λ)-learning** with eligibility traces — no neural network. Nine sensors (cursor proximity, scroll, DOM density, whitespace, time of day, age, and more) collapse into a discrete state space of about 38,880 states, and she learns which *mood* to adopt in each. There is no "be nice" rule anywhere in the code; good manners fall out of the reward function. The cursor, notably, is no longer just a predator. In the current reward hierarchy an active cursor also represents **nourishment** — user attention as a feeding opportunity, the second-strongest reward after the social bond — so whether Lili flees your pointer or drifts toward it depends on her mood, stress and history, not on a fixed rule. `reward shaping (Lili A)` ```text in whitespace, user is reading +1.0 // correct coexistence fled an approaching cursor +0.8 near Evrin, ideal distance +0.6 // social bond (Phase 58) explored the DOM, low stress +0.5 playful interaction +0.3 overlapping Evrin -0.3 covered a DOM element -1.0 sat over text, blocking reading -2.0 // the worst case idle for too long -0.5 ``` [the reward function above, from the inside — lili.debug() in the console opens this on any page she lives on](https://michalstrnadel.com/case-studies/lili-evrin/brain.png) ## the octopus is the architecture A real octopus keeps most of its neurons in its arms. Lili copies that. Intelligence is **distributed**, not central: - **Brain** — the hormonal system. Q-learning picks a mood (curious, playful, shy, calm, alert…); it sets tendencies, not direct actions. - **Eight tentacles** — local neurons. Each is an independent FABRIK inverse-kinematics chain with its own stress, curiosity and reflexes, reacting to touch without waiting for the brain. - **Body** — Craig Reynolds steering behaviors (wander, seek, flee, avoid) weighted by the current mood. - **Skin** — HSL chromatophores are the only communication channel: hue drifts with age, saturation spikes with stress, lightness follows the time of day. A stressed Lili simply looks stressed. - Behavior is the **emergent sum** of all of these — which is exactly how the real animal works. ## Lili vs Evrin — a controlled experiment Evrin runs a **Deep Q-Network** over a 26-dimensional continuous state instead of a table. DQN is notoriously unstable — minutes of unattended training can diverge — yet this project commits to a decade. So Evrin ships with a seven-stage stabilization suite: a replay buffer, a target network, anchor rollback on weight explosions, learning-rate decay, periodic exploration re-juvenation, gradient clipping, and a loss-spike detector. On mobile he runs inference-only, and his warm red-orange palette sets him apart from Lili's cool teal at a glance. Running two different learners side by side turns the piece into a real experiment. Lili can also be compared against four baseline policies — random, frozen, myopic and a hand-coded heuristic — so the emergent behavior can be told apart from luck or from the reward function alone. Every decision is written to a behavioral journal and exported as CSV for analysis. `behavioral journal — eight seconds of Lili, unedited` ```text time state mood explore reward stress 20:33:00 3036 alert · -2.127 0.09 20:33:01 3036 curious · -0.789 0.07 20:33:01 3036 curious · +1.350 0.20 20:33:02 9516 curious · +0.844 0.33 20:33:03 32196 exploring · -1.455 0.35 20:33:04 28956 exploring · -1.490 0.26 20:33:04 28956 exploring · -1.202 0.14 20:33:05 15996 idle ε -1.394 0.03 ε = the step was exploratory, not the policy's best guess ``` ## fifty-eight phases in The case study was written at the project's genesis; the organism kept growing. Fifty-eight development phases later Lili has **seasonal awareness** with its own soundscape, a synthesized **Web Audio** presence (a breathing drone, bubble pops, an ink splash), an **endocrine model** — virtual dopamine, cortisol and serotonin modulating the learning itself — **dream replay** during quiet hours, cognitive aging, and defenses she chooses by temperament: ink and camouflage (a shy Lili camouflages at 100 %, a curious one barely bothers). She also has a life beyond the canvas: **offspring inherit her learned Q-values with genetic mutation**, a life-narrative diary writes her days down, the Q-table renders as generative art, an observability dashboard exposes every internal, the brain runs in a Web Worker so the page never pays for her thinking — and after the planned ten years there is a dying animation, and a legacy. Full touch support arrived in Phase 42, which is why the old “desktop only” caveat is gone. ## living on a real web page Lili ships as a single `