Every frame from this class
25 captures, in order. Click any one to see it full size.
Never incremental
The innovation ladder
The four levers of systems thinking
Multi-order effects
The homework, mapped
The class in brief
Katrin's slides carry a confidentiality footer, so this page teaches her frameworks in the corpus's own words, full credit to her, no imagery. She placed the same object, a taxi, on three altitudes of design, named four mechanisms for how systems actually change, and pushed the room to chase an idea's effects past the obvious first move. After this page you can place any brief on the classic-design-to-computational-design ladder, name which of the four systems-thinking levers you are actually pulling, and map an idea's effects out to the third order before you trust the story its first order tells you.
The night at a glance
Why this matters · 6:44 PM
Where you think you will end up is a straight line. Where you actually end up is a curve that breaks upward without warning.
Katrin opened on the gap between predictable, linear progress and the exponential curve AI actually produces, the "vision gap" between comfortable incremental thinking and a result that outperforms it once you cross into growth. Her framing for the moment we're in: "We are probably at the moment where the Model T was rolling off, and we all were able to drive something that yet had to find stoplights and seat belts and all of the things." Early, useful, and not yet safe by default.
"We do not have yet the systemic understanding of how AI is going to change how we work, how the organizations work that we work in, how the systems, the government, the structures that we live in will be evolving."
Katrin, on why doom dominates the conversation
The framework · 6:47 PM
A color easy enough to spot from a block away. The taxi's yellow, and nothing past the object itself.
Not the cab, the inconvenience of hailing one. This is Uber's whole redesign, and it never touches the vehicle.
What should a city center look like if you balance livability, economic growth, transportation, and climate all at once, for the next 25 years. More variables than one human brain can hold at the same time, and a different class of design problem entirely.
Worked live against a taxi roof sign, the room called it every type at once, product and marketing, and Katrin agreed it "straddles over different concepts." Her cleanest example of a company operating at the top rung: Tesla, not because the car is clever, but because the EV, the charging network, a separate profitable battery-lifecycle business, and the driving data harvested from the fleet toward autonomy are four businesses solving each other's problems at once.
The craft · 6:57 PM
Modern songs run shorter intros than older ones because music is now designed against listening-behavior data. "We are optimizing for listening behavior even in the design of music today," at the cost of creative diversity.
Her example: the EU AI Act, throttling the inflow of what AI is allowed to do to individuals and institutions once a feedback loop's negative side starts winning.
Donella Meadows's ranking of intervention points. The counterintuitive finding: patterns and trends are a weak lever. The strong levers are mindset, culture, and above all trust. Her proof case: Airbnb cut discriminatory outcomes by redesigning the booking flow that shapes trust, not by only reweighting the recommendation algorithm.
Bezos's napkin sketch: lower prices drive traffic, more sellers, a lower cost structure, more selection, lower prices again, with Prime bundled on top. Every student in the room was a Prime member; nobody could name what any single piece of it actually costs.
The exercise · 7:17 PM
Ask a tool what happens in the long run as your chosen system (Spotify, TikTok, Instagram, Maps) keeps optimizing itself. Then do the harder part alone or with a partner: find the loop the AI's answer left out. Like Bezos's napkin sketch, the first pass usually gives you only the virtuous flywheel. What is the balancing loop it skipped: burnout, regulation, filter bubbles, saturation.
The closing question Katrin put to the room: the AI is itself a self-optimizing system. Would it flag its own limits?
Asked an AI about Anthropic to check for bias, and got an answer that was "surprisingly not biased at all, very much complimentary." What it never raised on its own: where humans fit in a future of very capable AI. Katrin's read: "it's not flagging its own limits. It's going on everything can be better and ignoring the reality of the human."
Ran smartphone autocorrect through the exercise and came back with four evaluation dimensions: accuracy, intent, identity, agency. The systems insight underneath: accepting a wrong autocorrection because you're in a hurry reads to the system as endorsement, so the system reinforces the mistake.
Asked what happens if Spotify keeps optimizing itself, got an answer built around a better listening experience. Re-asked with the real goal stated instead ("maximize profit for Spotify") and got a completely different answer. Models default to the flattering objective unless you name the real one.
The judgment · 7:49 PM
The first-order effect is the obvious one. It is rarely the one that costs the most.
Multi-order effects are the chain reactions that follow an initial change, second, third, and further orders that often outweigh the first. Worked live: the cell phone, built for easier communication, produced isolation as conversation moved to text. Airbnb, more short-term rentals, depleting long-term housing. The car, connecting distant places and widening how far people could commute, while driving climate impact and reshaping whole cities around parking. Asked for a positive chain instead, the room mapped GLP-1 drugs: a diabetes treatment, an accidental weight-loss effect, a population-level drop in heart disease, a stock downturn for soda and fast food, and snack companies reformulating toward a more health-conscious market.
Methods and prompts
Guess what a self-optimizing system's own AI-generated growth story leaves out before you check. The gap between your guess and the real answer is the actual lesson.
Working prompt
Before I ask you anything, my prediction for what a self-optimizing system like [system] would leave out of its own growth story: [my answer]. I answer first, then you check me. Now tell me what happens long term as [system] keeps optimizing itself, and tell me if I called the missing loop correctly.
You will know it worked whenit tells you what actually happens long term as the system keeps optimizing, and confirms whether the loop you predicted was missing was the real one.
Models default to the objective that sounds good. Restate the actual goal and watch the answer change.
Working prompt
I asked you about [system/decision] and got an answer built around [the flattering goal]. The real goal is [the actual goal, stated plainly]. Answer again with that as the stated objective, and tell me what changes.
You will know it worked whenthe second answer visibly changes once the real goal replaces the flattering one, not just a restatement with the same conclusion.
One real change, mapped out to the third through fifth order, refusing the first pass every time.
Working prompt
Act as a multi-order thinker. Here is a real change: [the first-order innovation]. Map its 2nd and 3rd order effects, both good and bad. Then push further: what does each of those produce at the 4th and 5th order? Do not stop at the first pass, keep going until the chain gets genuinely surprising.
You will know it worked whenthe chain actually reaches a fourth and fifth order effect that feels genuinely surprising, not a list that stops at the obvious second order.
Predict the human-layer gap before you compare notes. The gap itself is the design opportunity, not a flaw to smooth over.
Working prompt
Here is the multi-order map you gave me: [paste]. Before you respond, my guess at what a designer, a community member, or an ethicist would add that you missed: [my answer]. I answer first, then you check me. Now tell me what you actually left out, and mark every addition as a human one.
You will know it worked whenit marks each addition it names as a specifically human one, and tells you plainly whether your own predicted gap was the real one.
Classify a brief across all three altitudes before committing to one. The computational-design tier is usually the one nobody tries first.
Working prompt
Here is a brief: [paste]. Answer it three ways: as classic design (the feature or quality fix), as design thinking (the human need or experience fix), and as computational design (the systems-level fix that only makes sense once you factor in data and constraints a person can't hold in their head at once). Show me all three.
You will know it worked whenthe three answers are genuinely different in kind, not the same fix restated as a feature, an experience, and a system.
The close · 8:34 PM
Horizon 1 is improvements to what already exists: variants, extensions, cost reduction, on technology and markets you already serve. Horizon 2 is adjacent growth: a market you don't yet serve, or technology you haven't deployed, her example a gas vehicle moving to electric. Horizon 3 is genuinely new markets, products, and technology, her example a connected EV heading toward autonomous driving. Her own addition, a Horizon 4: AI turned on whole systemic problems at population scale, climate, cancer, the kind of question no single business unit owns.
The homework runs the same discipline on a real problem: pick an industry or process actually going through AI-driven disruption, state its first-order innovation plainly, then use AI to map the effects out to the third, fourth, and fifth order, iterating past the first pass every time. Last step, done by hand: mark what a designer, a community member, or an ethicist would add that the map missed. That gap is the design opportunity, not a flaw in the map. Due the following Tuesday or Wednesday, solo or in pairs, on the highest-power model available.
Try this prompt
Quiz me on the four systems-thinking mechanisms and the innovation ladder's three tiers. Then give me a real industry change and make me map its multi-order effects, my prediction of what's missing first, before you check me.
You will know it worked whenit quizzes you on the four systems-thinking mechanisms and the ladder's three tiers first, then makes you map a real change's multi-order effects, your prediction first.
The shelf
25 captures, in order. Click any one to see it full size.