There is a quiet inversion buried inside the phrase “generative architecture,” and most of the industry noise around it — the rendered blobs, the vendor demos — hides how old and how disciplined the idea actually is. Strip the marketing away and the concept is almost austere: instead of drawing a solution, you describe the problem — rules, goals, constraints — and let a system grow the form. The architect stops being the author of one building and becomes the author of the machine that discovers thousands.

←TODAY: In 2026 a structural engineer can run generative grid predesign directly on the BIM model and validate what the algorithm proposes. →3012: By the Zurich-3012 horizon the constraint file is the design document; the geometry is disposable output. Fulcrum: The value was never the shape — it was always the honesty of the constraints you were willing to write down.

What it is: Generative design, in the definition Wikipedia now settles on, is an iterative process in which software generates outputs that fulfil a set of constraints iteratively adjusted by a designer — human, test program, or artificial intelligence — who refines the feasible region with each pass. That is deliberately broad, and it should be: the same Wikipedia entry illustrates it with something as domestic as Samba, a piece of furniture the Brazilian designer Guto Requena produced this way. It covers rules-driven parametric CAD, evolutionary solvers, topology optimization, and grammar-based systems alike. What unites them is the reversal of the question. The old CAD paradigm asks, “Does this shape meet the requirements?” Generative design asks, “Which shape best meets the requirements?” The difference sounds academic until you watch a population of candidates reorganise itself against a single number you tightened, and realise the fitness function is doing the arguing you used to do by hand.

Why it works: The mechanism rests on three legs, none of them mysterious. First, a way to describe form as a function of parameters rather than fixed coordinates — so a building becomes a point in a solution space, not a drawing. Second, a fitness or objective function that scores each candidate against what you actually care about: material mass, clear-span limits, cost, daylight, code compliance. Third, a search method — genetic, gradient, brute-force — that explores that space faster than a person can. The ScienceDirect structural-design reviews cited in PAZ’s own concept panel frame the ambition as a ladder of autonomy from L0 (fully manual, engineer in complete control) upward, where higher rungs learn from prior drawings and codified mechanical knowledge to optimise for performance, cost, and regulatory compliance simultaneously. That last word is the whole point: humans optimise sequentially and badly; a solver holds all the constraints at once. The economics matter too — computing has become, in relative terms, effectively infinitely fast and vanishingly cheap, which is exactly why heuristic brute-force search stopped being a luxury and became a default.

Origins: This is less a rupture than a homecoming. Architecture was once a branch of mathematics — Isidoros and Anthemius computed the Hagia Sophia before they raised it, and Islamic tiling grammars anticipated formal classification by centuries. The modern strand crystallised in the second half of the twentieth century: George Stiny and James Gips gave us shape grammars in 1971, a formal system for generating designs by rule-substitution; John Frazer wrote An Evolutionary Architecture in 1995, arguing for buildings as evolving organisms rather than fixed objects. Springer’s survey of the field summarises the arc as a path “from mathematical optimization to grammatical customization.” The Swiss anchor is not decorative: Dr Kristina Shea, now a professor at ETH Zurich, wrote in 2005 of “integrated performance-driven generative design systems” aimed at “spatially novel yet efficient and buildable designs” — a decade before the term migrated into mainstream mechanical CAD. The lineage on site is equally concrete: Gaudí’s hanging-chain catenary models at the Sagrada Família (Barcelona, 1882–) were a hand-computed ancestor of form-finding, gravity doing the maths while he read it upside down; Herzog & de Meuron with Arup generated the Beijing “Bird’s Nest” lattice (2008) as one continuous system; Zaha Hadid Architects rationalised the Guangzhou Opera House skin (2010) and computed the seams out of the Heydar Aliyev Center (Baku, 2012) and the Morpheus Hotel diagrid (Macau, 2018).

In practice — Atelier: At a Swiss studio, the honest place to start is not the sculptural envelope but the boring floor plate. Take a 30-metre plate, parameterise the bay spacing of the column grid, set a fitness function that minimises structural material, and hold every clear span under an 8-metre code limit — then tighten that limit by half a metre and watch the “optimal” grid reorganise without you redrawing a single line. Rhino and Grasshopper are the house ground for this, dropping into GhPython or RhinoCommon once the logic outgrows the visual nodes, and wiring the loop into Galapagos when you want the search to run itself. The McNeel forum matters as much as the software: the knowledge there is shared sideways, horizontally, not licensed downward, which is exactly the posture a public-good discipline should want. The Monday move: pick one recurring, constrained decision your office makes by hand — grid spacing, riser layout, panel subdivision — and this Monday write it as a fitness function with one hard constraint and one thing to minimise, before you touch geometry.

Hack: Score a column grid the way a solver would, then feel the constraint bite. The intention: find the coarsest bay count that still keeps every clear span under a code limit, and read off what it costs. The three lines below scan the feasible grids on one axis of a 30-metre plate, weight columns against a rising beam penalty, and return the cheapest legal option — tighten MAX_SPAN and the answer jumps, which is the entire lesson about who the constraint file really serves. Render it in GhPython or a plain interpreter:

PLATE, MAX_SPAN = 30.0, 8.0
legal = [n for n in range(2, 13) if PLATE / n <= MAX_SPAN]
best = min(legal, key=lambda n: (n + 1) + 0.04 * (PLATE / n))
print("bays=%d  span=%.2fm  cols=%d" % (best, PLATE / best, best + 1))

Move: The human role migrates from author to curator — you set the objectives, draw the constraint fence, and choose among the candidates the machine explored. The judgement is still yours; the enumeration is not. And that is where the discipline hardens rather than dissolves. A solver optimises what it can measure and quietly ignores what it cannot, so the architect’s real job becomes writing the constraints that encode the things worth protecting — daylight, dignity, the street, the data-residency clause in the procurement contract that specified the software in the first place. From the vantage of the late 2070s, the offices that stayed sovereign were the ones that read their own constraint files rather than accepting a vendor’s defaults. Generative architecture will hand you thousands of buildings. Choose the constraints before you choose among them — open your next brief and write the one number you refuse to let the cheapest grid override.

Sources & Further Reading

• Primary: Wikipedia — Generative design