A small studio above a bakery, in a town nobody intended to visit.
We work from a converted stone building in Midsomer Norton, Somerset, twelve miles south of Bath. There are four partners, six commissions a year, two GPU rigs in the cellar and a cat called Vellum. The studio is deliberately small because the work we want to do is impossible at scale.
The Island House, second floor.
The Island House sits in the middle of Midsomer Norton's high street, on a strip of land that used to be a millrace and is now a car park. The building dates to 1796 and was, in order: a wheelwright's, a draper's, a Methodist reading room, an estate agent, and finally, since February 2025, the home of Design Labyrinth Limited.
We occupy the second floor. The ground floor is a bakery whose ventilation reaches us by about 11 a.m. and decides the second half of the studio's day. The first floor is an accountant. We chose the building partly because it has eight windows, all of them north-facing, which is the light an architect's drafting room wants. We did not choose it for the price, which was honestly a stretch.
The studio has four rooms inside it: the survey room, with the tracing paper and the big tables; the writing room, with the books; the operator room, with the two workstations that the models run on; and the kitchen, which is where the actual conversations happen. The cellar is a fifth room and it is full of GPUs and dehumidifiers; we will get to it.
Four practitioners. One cat.
We are deliberately not a hierarchy. The studio is owned in four equal parts by the four founding partners; new commissions are signed by whichever partner has the bandwidth and the right instincts for the job, and the others critique. Vellum the cat is non-voting but vetoes by sitting on documents she does not like, which has happened twice.
The architecture partner
Trained at the Bartlett and practised for nine years in a London residential studio before training a model that put her out of that job. Leads commissions where the deliverable is a building or a building system.
The industrial design partner
Twelve years of furniture and homeware, the last four spent quietly building constraint-first pipelines inside a Danish design house. Leads commissions where the model has to know about manufacture.
The identity partner
A type designer and identity practitioner; drew three commercial typefaces before discovering that drawing typefaces is, technically, a form of generative modelling. Leads identity engagements.
The engineering partner
The one with the PhD. Spent six years at a London research lab building small models for scientific applications; now does the same on the design side. Leads no commissions; reviews all of them.
We are not naming individuals on this sheet because the partner introductions are something we prefer to do over coffee, and because two of us have previous employers who would rather we kept our heads down for another quarter or two. If you are starting a serious commission, you will know all our names by the end of the survey room.
Pencils, ink, code, weights.
The studio's tool wall is a deliberate spread across centuries. We use Staedtler 2H pencils, Pentel Brush Pens, Rotring isograph nibs (the 0.18 mm and the 0.35 mm are our standards), Fabriano 5 paper for surveys, Bovill tracing paper for drafting, and three different scales of architects' rule. None of this is performance; it is the part of the work where the model is not allowed in.
For the computational side we run a small, opinionated stack. We train in PyTorch on two workstations with consumer GPUs — nothing in our cellar costs more than a used estate car. We use diffusion models for image work and small encoder-decoder transformers for layout and typography; we have a strong preference for models we can read end-to-end on a long afternoon. We do not use third-party hosted inference for client work. The weights belong to the client and they do not leave the building until the handover.
We write a lot of small validator programs in Python and Rust — the constraint layer that says "this output is allowed; this one is not". These are usually the most valuable artefact of an engagement, and the most boring to look at.
Two workstations. One dehumidifier.
The training cellar is the only room in the studio that an electrician has been involved with. Two workstations, each with a single high-end consumer GPU, live in a small acoustically-treated alcove behind what was once a coal chute. We do all training on these machines; we have so far never needed cloud compute for a client engagement, which we mention partly as a point of pride and partly because the alternative would be wrong on principle.
The reason small models on local hardware is enough is that we are not building general intelligence. We are building specialised tools that need to know one studio's style very well. A 500-million-parameter model trained on a well-prepared 4,000-drawing archive will, on average, beat a 30-billion-parameter model with the same archive shoved into its context window. We have run the comparison several times; the lopsidedness is consistent.
The dehumidifier is because the building is from 1796.
The books that argue with each other.
The studio library is two walls of the writing room. It is unsystematic on purpose; the partners shelve their own books and the disagreements are visible. Some shelves we re-read often: Christopher Alexander's Notes on the Synthesis of Form; Vilém Flusser's writings on technical images; Donald Schön on the reflective practitioner; the proceedings of the early Architecture Machine Group; Yona Friedman's Pro Domo; Ellen Lupton's Thinking with Type; and a complete set of 9H magazine that we inherited.
For the AI side we keep a smaller, more sceptical shelf: the early Wolfram NKS, Marvin Minsky's The Society of Mind, and the assorted technical papers that we have actually re-read. We avoid the airport-bookshop side of the AI literature; not because it is wrong but because it is the same book.
The borrowing rule
Books leave the library only with a written note in the loans book, in pencil. Vellum is not allowed to borrow but does anyway.
What the day actually looks like.
The studio opens at 9 a.m. and the first hour is silent. We draw. Nobody writes email; the workstations remain off. At 10 a.m. somebody puts the kettle on, and the day's first conversation about the work in progress happens. The model in the cellar may have run overnight; we look at outputs together over coffee and decide which ones get the red pencil.
From 11 to 1, training runs are kicked off and we work on whatever room of whichever commission we are in. The bakery vents at this point are at maximum. Lunch is together at the kitchen table at 1; we do not work at lunch. From 2 until 5 the operator room is busy — this is the slow, careful sitting-with-the-model time. At 5 the workstations go off and we draw again until 6. The cat comes in at 6:30. We close.
Fridays are reserved for the writing of constraints, the maintenance of the library, the answering of long emails, and the slow argument about the next revision of the manifesto. We do not take client meetings on Fridays. The pattern is the same most weeks; the slowness is the point.
You are welcome, on Wednesdays.
Clients and prospective clients are welcome to visit the studio on Wednesdays. The train from London Paddington to Bath is ninety minutes; we will collect you from Bath Spa or, if you would rather walk a bit of the Somerset Coal Canal, from Midford. The visit takes a half-day; we make the lunch.
We mention this because a great deal of the work of choosing a studio is the work of standing inside it for an hour. The website is honest as a website can be, but the building is more honest.