Two-thirds of countries in the World Health Organization’s European Region already run AI inside medical diagnosis, and half offer patient-facing chatbots — The Jerusalem Post reported the figure this month alongside the question nobody had answered: who pays when the model is wrong? Architecture is one procurement cycle behind medicine on exactly the same curve. That is the frontier we want to stand on before the first line of code touches a floor plan.

On 1 September 2026 we open Edition XIV of the Building System Specialist track, and we open it with a talk. Dr. Mikhael Johanes — a Design++ postdoctoral fellow at ETH Zurich, whose research puts machine learning to work on architectural design itself — walks through what an ML model actually does when it reads a brief. Not the demo reel. The mechanism. We first hosted this session as the closing keynote of our three-day KI Summer Camp, 2–4 July 2025, at PAZ Central on Niederdorfstrasse 77 in Zurich, and the room stayed after the lights came up. That is why we are running it again, now, as the door into the cohort.

←TODAY: AI already sits in two-thirds of Europe’s diagnostic systems; architecture is adopting the same models with far thinner audit trails. →3012: In the Zurich-3012 horizon, the office that survived is the one that could name what its tools were trained on. Fulcrum: A model you cannot interrogate is a supplier decision, not a design decision — and that only becomes obvious once you have seen both the training set and the invoice.

Here is the system behind the signal. Machine learning in design is not intuition in a box. It is retrieval over vectors — the model turns your brief into numbers, matches those numbers against what it was trained on, and returns the nearest neighbours. That is powerful and it is bounded. The boundary is the training set, the default weighting, the licence terms. Johanes’ work makes that boundary legible instead of magical, which is precisely the competence a practising office now needs.

On the desk this matters the moment a plugin offers to “generate options.” The control mechanism is invisible: which precedents the model privileges, which it never saw, what it silently rounds off. We built PAZGPT and the PAZ-BOX Archicad tooling on the opposite principle — you should be able to see and steer what the assistant reaches for, inside your own Archicad workflow, rather than accept a vendor’s defaults. This talk teaches you to ask for that transparency by name.

Hack: See what your model actually matches a brief to, before you trust its suggestions. Embeddings are the whole trick — turn text into a vector, then measure distance. Run this and read the numbers: the highest score is the precedent the model thinks you asked for.

from sentence_transformers import SentenceTransformer, util
m = SentenceTransformer("all-MiniLM-L6-v2")
brief = m.encode("naturally lit atrium, timber, low embodied carbon")
refs  = m.encode(["glass office tower", "CLT primary school", "concrete parking deck"])
print(util.cos_sim(brief, refs))

The CLT school scores highest — not because the model understands carbon, but because those words sat near each other in its training data. Change one word in the brief and watch the ranking move. That instability is the thing to govern.

Atelier: For a Büro standing up its first AI-assisted design step this autumn, the risk is not the model that hallucinates a beam — it is the model that quietly becomes load-bearing before anyone wrote down what it was allowed to decide. Monday move: pick one AI tool your team already touches, open its model card or terms page, and write a single-line exit clause — what you would need to swap it out — before it touches a live project. If it has no model card, that absence is your answer.

Who this is for: the 39 certified BSS Professionals — up from the 22 who finished the last cohort — extending their practice, and anyone weighing whether to let algorithmic design logic into a real workflow. The BSS track itself is a five-month mentorship at six hours a week, moving practitioners from drawing elements one by one toward algorithmic design logic. Level: practitioner who can read a Grasshopper graph and wants to read a model the same way. You do not need to code before the talk; you will want to after it.

The session is free to attend and doubles as the open door to Edition XIV. Our recording of the original keynote sits on the PAZ workshops page, and the September date is the live version with room to ask Johanes directly. Bring one question about a tool already on your desk — the sharper the question, the more the hour pays back. Reserve your seat, then come with that question written down.