Dylan's Homepage.

If there is a single thread running through everything on this site, it is a mild distrust of software that only does what it is told. Since the late nineties I have been far happier setting up a small, slightly leaky set of rules—a virtual biochemistry, a genetic mutation loop, a handful of soldered stripboard oscillators, or a hacksawed surveillance lens—and stepping back to see what sort of trouble the substrate invents for itself [1].

Over the decades this habit has taken me from breeding digital organisms on 1990s home PCs and adjudicating virtual insect athletics on children's television, to co-curating evolutionary exhibitions with multimodal foundation models at Google DeepMind. Yet in the evenings the impulse remains stubbornly low-tech: if an algorithm or circuit cannot be coaxed onto a Raspberry Pi Zero, stuffed into a wooden wine crate, or held together with two-part Araldite epoxy, I am always slightly suspicious of it.

You can read this page straight through as a five-part field notebook, play with the live simulations and patch folios tucked into the right-hand margin, or skip directly to the Complete Archival Ledger at the foot if you simply want the links, papers, and schematics without my commentary.

My professional induction into artificial life began in 1997 at CyberLife Technology (later Creature Labs), where I served as Senior Researcher and Head of Artificial Life on the BAFTA-nominated Creatures series—specifically Creatures 2, Creatures 3, and Docking Station [2]. Rather than scripting behaviours with brittle decision trees, we gave the Norns heterogeneous neural network brains wired into a simulated bloodstream of virtual hormones, antigens, chemo-receptors, and chemo-emitters, all encoded on mutable digital chromosomes. When a Norn learned to speak or decided to eat a poisonous grendel berry, it was because a chemical drive had propagated through evolved synaptic weights.

In 2003 that obsession with emergent creature mechanics jumped from desktop terrariums to broadcast television. For four series of the BBC's BAFTA-nominated show BAMZOOKi (2003–2009), I was Technical and Creative Product Lead. We shipped a free 3D Inverse Kinematics and rigid-body physics construction kit ("the Zook Kit") so children across the UK could engineer multi-legged virtual "Zooks" in their bedrooms and bring them into the BBC studio to sprint, hurdle, and shove each other off a real studio table via live mixed-reality broadcast rendering.

You never really recover from watching a twelve-year-old's lopsided six-legged physics contraption accidentally discover a galloping gait. Years later, I found myself returning to the exact same question at two opposite scales of compute [3]. At the pocket end, I started A-Life on Pi—packing continuous cellular automata, reaction-diffusion soups, and tiny evolving pixel ecologies onto bare Raspberry Pi boards and 8×8 SenseHAT LED matrices just to watch them tick away on a bookshelf.

At the cluster end, inside Google DeepMind, we built EvoBorg (2018–2023): a direct spiritual successor to Karl Sims' 1994 virtual creatures and BAMZOOKi. Using asynchronous distributed genetic algorithms over MuJoCo physics simulations, we co-evolved both the physical 3D morphologies (directed acyclic graphs of limbs, joints, and contact sensors) and the neural controllers of virtual creatures locked in competitive sumo-wrestling arenas—where every new offensive lunge by one lineage forced an anatomical counter-adaptation in the next.

When large vision-language models arrived, the standard party trick was to type a prompt and let a diffusion model hallucinate a finished bitmap in one shot. Useful, certainly, but to an artificial-life tinkerer—and to an artist—it cuts out the best part of making things: the slow, stubborn negotiation with structure.

In Evolution and Foundation: AI Shares Creative Control (2024–2026), built in collaboration with computer art pioneers Professor William Latham, Stephen Todd, Professor Frederic Fol Leymarie, and Lance Putnam, we flipped that relationship on its head [4]. Instead of asking an AI to generate pixels, we placed Google's multimodal Gemini inside a 200-generation evolutionary loop as an aesthetic curator and art critic driving Latham and Todd's classic 3D procedural engine (Mutator / FormGrow).

In each generation, Mutator breeds sixteen 3D organic phenotypes from parametric horns, ribs, and mathematical transformation rules. Gemini inspects multi-angle renders of the candidates in pairwise binary tournaments (~1,600 judgements per run), writes out its critical rationale for which form better captures a target brief—say, "a fly head" or "an iridescent deep-sea siphonophore"—and selects the parents of the next generation, while our PixelScore novelty metric periodically shocks the gene pool whenever the critic grows too complacent. Exhibited at London's OXO Gallery to over 27,000 visitors, featured in The New York Times, and selected for the CVPR 2026, ECCV 2026, and ALife 2026 galleries, you can scrub through 198 generations of the Fly Head lineage directly in Fig. 04 alongside this paragraph.

That insistence on transparent compositional parts actually started a few years earlier with Arnheim and CLIP-CLOP (2021–2022), named after the perceptual psychologist Rudolf Arnheim [5]. In Arnheim 1, we paired CLIP vision-language encoders with hierarchical neural Lindenmayer systems (L-systems) so images grew out of explicit stroke grammars rather than opaque pixel soup. In CLIP-CLOP / Arnheim 3 (published at ICCC 2022), we took inspiration from Hannah Höch and Richard Hamilton's Dadaist photomontages: evolving populations of cut-out image patches via spatial transformers and genetic operators, so an artist in a Colab notebook could reach in and tweak the palette, object vocabulary, or layout grammar mid-flight.

Because I dislike leaving generative code trapped in cloud notebooks, I also bolted Arnheim 3 onto a dedicated low-power single-board computer on my shelf called Vis Plastica. It has been quietly mutating and publishing fresh evolutionary collages around the clock ever since, requiring roughly the same wattage as a nightlight.

What happens when an artificial-life simulation is given a physical body the size of a London lobby and has to live alongside human beings every day? Between 2019 and 2022, Chrisantha Fernando and I led the AI and ALife architecture for Lux Automata, a permanent suspended media sculpture built with Jason Bruges Studio in the foyer of Google DeepMind's King's Cross headquarters [6].

The sculpture is a three-dimensional lattice of 42 × 10 × 7 custom black-PCB cellular cubes—2,940 cubes housing 635,040 individually addressable LEDs. Rather than playing pre-rendered video loops, it runs four live families of volumetric substrates: 3D rule-based cellular automata (gliders, solitons, fizzy rings), evolved 3D convolutional & developmental Gene Regulatory Network CAs ("Soft CAs"), reinforcement-learning agents playing 3D CA Pong, and an overarching homeostatic AI curator we christened Agatha.

Privacy-preserving motion cameras in the ceiling project people walking through the foyer directly into the 3D lattice as live cellular perturbations. Agatha monitors her own internal homeostatic drives—Novelty, Excitement, Calmness, and Human Interaction—while continuously scoring the 3D light patterns using George David Birkhoff's 1933 Aesthetic Measure ($M = O / C$, the ratio of bilateral symmetry order to entropy complexity). If the lobby is empty and dull, she nucleates wild gliders to amuse herself; if rush-hour crowds over-excite the lattice into chaos, she cross-fades into soothing developmental waves. You can poke a miniature 3D voxel sketch of Agatha's lattice in Fig. 06.

Watching passers-by improvise semaphore games with Agatha led naturally to a deeper theoretical question: why do intelligent agents leave marks in the world at all? In our recent research on External Representations (published in Adaptive Behavior, 2024) and "Wanting to Be Understood" (2025), we investigate two missing ingredients in standard AI architectures [7]. First, biological minds constantly offload internal state into persistent external sketches, diagrams, and notches on sticks, then re-perceive those artifacts as cognitive scaffolding. Second, genuine communication isn't just broadcasting tokens into the void; it requires an intrinsic communicative drive where the speaker actively models whether their listener has understood them and repairs misunderstandings interactively—a thread we also extended to the meta-problem of consciousness (2025).

Alongside these cognitive models, we have kept pushing pure evolutionary mechanisms inside foundation models and deep networks [8]. In Promptbreeder (ICML 2024, and featured in the Gemini 1.0 Technical Report), we built a self-referential evolutionary system where a population of LLMs evolves not only task-prompts for mathematical and symbolic reasoning, but also the mutation-prompts that govern how prompts mutate. Earlier, in PathNet (2017), we used microbial genetic algorithms to evolve modular active pathways through giant neural networks so agents could transfer skills across Atari and Labyrinth games without catastrophic forgetting; and in the Perception Test (NeurIPS 2023), we filmed diagnostic real-world video scenarios to test whether multimodal models actually grasp object permanence, physics, and counterfactuals.

Modular synthesizers are really just analogue artificial life for people who own a soldering iron. You patch a slow random voltage source into a clock divider, feed the output back into a filter's resonance, and suddenly the rack is muttering a polyrhythmic bassline that neither you nor the circuit designer anticipated [9].

When my NiftyCASE ran out of horizontal pitch during the 2020 lockdowns, I sawed rails into a wooden twelve-bottle wine crate and began populating it with hand-soldered Veroboard (stripboard) modules: adapting a £6 24-bit 100-program digital reverb daughterboard into a stereo Eurorack effects unit (100-DSP Reverb), wiring dual analogue oscillators around Electrosmith CEM3340 submodules with hand-punched aluminium panels, and building custom vactrol snare and white/pink noise circuits (NiftyNoise).

Because modular patches evaporate the moment you pull the cables out, I started sketching multi-page illustrated iPad Patch Notebooks for every track—recording every cable arc, unpatched knob position, Turing Machine loop, and signal-flow block diagram for pieces with names like 1 Fish per Second, Leaving the Heliopause, Spectral Gerbil, and Christmas Out of Sync. You can open the complete folio of all 10+ hand-drawn PDFs and stripboard layouts—or play my browser-based physical modeling simulator of KORG Berlin's electromechanical Phase 8 acoustic synthesizer—via Fig. 09.

On the visual side of the hardware bench sits PaperPiAI (2025) [10]. Most digital art frames fail a basic domestic test: they glow like a small television in the corner of the hallway and radiate heat all night. PaperPiAI pairs a Raspberry Pi Zero with a multi-colour bistable e-paper panel mounted in a wooden picture frame. Every so often it wakes up, dreams up a fresh generative composition, dithers the pigment particles across the glass with that wonderfully theatrical three-second e-ink flash, and goes completely back to sleep—holding the physical ink image on the wall while drawing exactly 0.0 watts.

For years, a whole stratum of my older hardware and optics builds lived in a semi-broken directory called /salvaged/ (rescued from an old Concrete5 database via the Wayback Machine) alongside a handful of unlinked folders in /diy/. Bringing them back into the light is a reminder that a Junior Hacksaw and a tube of Araldite solve an unreasonable number of optical engineering problems [11].

Take the £5 25mm f/1.4 C-mount CCTV lens on my Olympus PEN E-PL1 (June 2011). Surveillance lenses give you glorious, swirly Petzval-style bokeh on a Micro Four Thirds sensor, except the rear metal shoulder prevents the glass from sitting deep enough in the mount to focus past your own nose. Naturally, the fix was to unscrew the rear optical group, clamp the aluminium barrel to the kitchen table, saw two millimetres off the back with a hacksaw, paint the raw metal matte black, and epoxy the glass back in. On days when even £5 of glass felt decadent, I drilled a hole in a plastic body cap and taped a pierced tin-foil shim inside to shoot digital pinhole photographs.

Or consider Super-Macro photography (2009) [12]: if you take an old manual 58mm f/2 prime lens from a charity shop and screw it backwards onto the front filter thread of a standard camera zoom using a £3 male-to-male coupling ring, the reversed prime turns into a brutally sharp, multi-element microscope objective—resolving individual hexagonal ommatidia on a housefly's compound eye, garden ants, and the pitted craters of poppy seeds. Click any thumbnail in Figs. 11–12 to inspect the full build steps and macro plates in the specimen lightbox.

The same scavenging instinct produced the Electricity-Free Hard-Drive Record Player (2010) [13]. When a 3.5-inch desktop hard drive dies, the magnetic platters still sit on an extraordinarily smooth fluid-dynamic spindle bearing that acts as a natural flywheel. Stick a 7-inch vinyl single onto the exposed platter, roll a sheet of A4 paper into a horn with a sewing pin taped to its tip, spin the disc with your finger, and out comes surprisingly recognisable music with zero electricity. Around the same time I laid out a custom Veroboard stripboard version of Jay Silver's Drawdio—strapping a 555 timer circuit and miniature speaker onto a graphite pencil so that sketching on paper forms a conductive variable-pitch synthesizer loop through your own body.

Before consumer drones existed, getting a camera into the sky required a stunt kite, nerves of steel, and hacked firmware [14]. By loading the Canon Hack Development Kit (CHDK) onto a cheap Canon PowerShot A530, I wrote custom BASIC intervalometer and HDR exposure-bracketing scripts, suspended the camera from a kite line over beaches and parks, and hoped the wind wouldn't drop it into the sea. Back indoors, I stripped the plastic lens off a £5 USB webcam and mounted the bare CMOS sensor inside a 35mm film canister—coupling it to a toy microscope (retrofitted with white LEDs and polarising filters to image birefringent vitamin C crystals and red onion cells) and to a backyard telescope for long-exposure captures of the Orion Nebula, the Pleiades, and The Plough.

Those years of camera and mobile tinkering also led to a rewarding string of public art and augmented reality collaborations with artist Bettina Furnée (2008–2012) [15]: setting up coastal camera systems to document 38 appliquéd flags along the Suffolk shingle at Bawdsey for Lines of Defence; staging the multi-authored text composition POWERHOUSE at Sir Giles Gilbert Scott's Cambridge University Library tower; and building a mobile augmented-reality platform on Parker's Piece in Cambridge (where the 1848 Cambridge Rules of football were drafted) so passers-by could view and propose New Rules For A Fair Society floating in 3D space for Reality Rules.

Which leaves only the culinary and speculative-biotech footnotes [16]. For reasons lost to history, I have long maintained the Instant Noodle Review—dedicated to the rigorous proposition of delivering "instant reviews on instant noodles, the instant they are consumed"—alongside Decoupled Labs, an interactive pitch simulator for a fictional mitochondrial-uncoupling biotech startup whose core thesis is that "fat storage is uncompiled legacy code running on outdated hardware." One has to keep one's hand in.