On September 2, a developer typed an ambiguous instruction into a terminal and the machine stopped.

Not an error. Not a refusal. A question. Meta had released Muse Spark 1.3 that morning, and the model’s new habit is to ask what you meant when your prompt is unclear, to ask for help when it gets stuck, and to confirm before it does anything consequential. Meta’s launch post leads with this, before any benchmark.

The most interesting thing about a frontier model, in September 2026, is its manners.

Sit with how strange that is

Clarify on ambiguity, escalate on failure, confirm before destruction. This is Norman. This is Nielsen. This is the second and fifth of the ten heuristics. It is the “are you sure?” dialog, which has been a product-layer decision since the Macintosh.

For forty years, whether a system interrupts you, and how, and at what threshold, was ours to choose. It lived in the design file.

Now it ships in the weights, at $1.25 per million input tokens, and it arrives with a personality you did not author.

That relocation has a consequence founders should think about before their next architecture review. The interaction pattern is becoming a procurement decision. Pick Muse Spark 1.3 and your product hesitates. Pick a model tuned for autonomy and it does not. You will inherit a set of judgements about when your users deserve to be asked, made by a lab optimising for its own benchmark suite, and you will discover them in your support tickets.

The old joke was that engineering decisions become UX decisions whether or not a designer is present. The new version is that model selection is now an interaction design decision, and it is usually made in a room with no designer in it.

The timing is what makes it uncomfortable

On the same day, Margaret Mitchell, Avijit Ghosh and Samir Passi posted the revised version of a paper called “AI Agents Push Humans Out of the Loop.” Its argument runs directly at the confirmation dialog.

Human oversight, they write, is the industry’s favourite answer to agent risk, and it is not a simple one: current agent designs impede effective oversight, and the cognitive capacities oversight requires are themselves eroded by extended use of the systems being overseen. Automation research has known this for decades under the name skill atrophy. The paper’s closing warning is that agent systems, as built, “passively incentivize the degradation of the very human skills they rely on.”

So one lab shipped the approval button, and the same week the research asked whether anyone is still qualified to press it.

The third fact, and the one nobody is pricing

Artificial Analysis measured Muse Spark 1.3 reading about 57% more input tokens per task than its predecessor, at $0.55 per task against $0.40. Two of its evaluation scores went down, which the firm attributes to a higher abstention rate. The model declined to answer more often when it was unsure.

In production that is exactly what you want. On a leaderboard it reads as a regression.

We have built an entire measurement apparatus that scores confidence and cannot see restraint. Meanwhile OpenAI’s GPT-6 Astra hit 99.9% on ARC-AGI-3 the following day, on its own harness, which means that particular test has finished its useful life. When capability saturates, behaviour becomes the product. Behaviour is design.

And this is where NN/g’s Anna Kaley and Raluca Budiu supply the sentence that ties the whole series together. Writing about what they call the custodial era of UX, they observe that “production has become cheaper than UX evaluation. It can now take less time to create an experience than to decide whether that experience is useful, usable, trustworthy, or coherent.”

Three independent sources, one shape. Generation is collapsing in cost. Judgment is not. The bottleneck in every AI product built next year will be the speed and quality of human decisions about output, and almost nobody is designing for that.

We are treating this as a governance problem

It is a design problem wearing a compliance costume.

The instinct in most companies is to answer agent risk with an approval workflow, a policy document and a log. All necessary, none sufficient, because every one of them assumes a competent reviewer at the end of the chain and none of them produces one.

Approval is a surface. If it does not carry the diff, the blast radius and the undo path in a single view, it does not transfer understanding, it transfers liability. And a reviewer who approves five hundred agent actions a day at 400 milliseconds each has not been kept in the loop. They have been given a stamp and a reason to stop reading.

We know how this ends because we already built it once. Cookie consent began as a transparency rule and became the most reflexively dismissed surface on the internet. Article 50 of the EU AI Act came into force on August 2 with real teeth, up to €15 million or 3% of global turnover, and the disclosure it mandates is now a design brief. We can make it meaningful or we can make it another banner.

Judgment cannot be shipped as a feature

After twenty years of teaching designers, the one thing I am certain of is that nobody develops judgment by reviewing finished work. They develop it by choosing badly and living with the result.

Which is precisely the experience an approval queue is engineered to remove.

If the interface only ever hands people a plausible answer and a confirm button, we are running a training programme in reflexive assent, and the Mitchell paper is right that the cost compounds quietly. Somebody has to design the friction that keeps a person genuinely deciding, and that somebody is us.

To be fair to the other side, a model that asks too much is a model nobody keeps. Confirmation fatigue is a real failure mode with real casualties, and the complaint about agents all summer was that they did too much without asking, which is the same people now discovering that being asked is expensive. There is a version of this that is worse than autonomy: an agent that offloads every ambiguity onto a human, hits its safety metric, and quietly makes the person a bottleneck in their own workflow.

The interruption policy has to be earned per action, not applied per product. Meta deserves credit for making the behaviour steerable rather than fixed. The failure mode is real, but it is a design failure, and design failures are the kind we know how to fix.

What this month actually handed us

Four issues, one argument. Design tools were repriced on a story their own data contradicted. The profession got measured and the org chart failed. A second, non-human user arrived and started improvising channels. And now the labs are competing on restraint.

Knowing when to interrupt, when to ask, when to abstain: these are being trained into models because they turned out to be the frontier. Our vocabulary just became the roadmap of the most heavily funded engineering effort in the world.

The question is whether we show up with more than opinions. That means evaluation we can run at the speed of generation, approval surfaces that actually inform, and enough humility to admit that our own heuristics were written for a user who was always human and never in a hurry.

So, the question I cannot put down. Your product will soon ask someone to approve something they did not make, do not fully understand, and cannot easily undo.

What have you designed to make sure that person is still capable of saying no?

The pattern: build a reversibility ladder

Most products do not have one. Sort every action your agent can take by what it costs to undo. Free (reading, searching, drafting). Cheap (edits with version history). Expensive (sending, publishing, paying). Impossible (deleting, disclosing, signing).

Then set the interruption policy per rung rather than per feature. Three rules follow, and Muse Spark 1.3 will now enforce them whether you designed for it or not.

A question is a turn that ends with the model waiting. In an interactive session that is the point. In an unattended pipeline it is a stall. Every clarifying question needs an answerer, an escalation path or a timeout, decided by you rather than discovered in production.

Never wire a confirmation to an auto-yes. A vendor’s claim about calibrating irreversible actions is a claim. Your approval log is the evidence, and if every entry says approved within 400 milliseconds, you have built a cookie banner.

Design the approval, not just the prompt. The person saying yes needs the diff, the blast radius and the undo path in one view. If your confirmation dialog only contains the word “confirm,” you have moved the liability without moving the understanding.

Also worth your attention

Four frontier releases in 72 hours, and almost no price movement. Anthropic shipped Claude Fable 5.1 and its trusted-access twin Mythos 5.1 on September 1, holding list price at $10 and $50 per million tokens while cutting cache reads from $1.00 to $0.25. Google released Gemini 3.8 Flash on September 2 at $0.75 and $3.75, with a footnote printing the January 1 doubling. OpenAI followed on September 3 with GPT-6 Astra at $10 and $50. Per-token prices held and the mechanisms around them moved: what a cache read costs, how many tokens a task uses, when an introductory price ends, and which customers get the version with the safeguards loosened. September release ledger

Two token numbers from one release, both true. Meta’s engineers measured about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 on coding work. Artificial Analysis measured about 57% more input tokens per task the same day. Meta counted a loop that got shorter. Artificial Analysis counted a fixed task where the model reads more. Attach a denominator to every efficiency claim you repeat this year, including the ones you like. The two measurements side by side

Perplexity split a single task across two trust zones. Hybrid Compute, shipped for the Mac app on September 1, starts a computer task in the cloud and runs the sensitive steps and private-file access locally on Apple silicon, behind an on-device PII classifier, with local work consuming no cloud credits. This is the first mainstream product I have seen treat privacy as a routing decision inside one task rather than a setting the user toggles beforehand. The design question it opens is a good one: if half a task ran on your machine and half did not, how would the interface tell you which half, without turning every action into a security lecture?

Claudeforce moves toward open beta. Salesforce and Anthropic announced the partnership on August 26, and Salesforce in Claude launches with 37 prebuilt sales skills that let sellers reason over live revenue context and take governed action from inside Claude. The CRM is being repositioned from an application into a set of permissions and skills behind someone else’s chat surface. If your product’s value lives in its screens, this is the moment to work out what remains when the screens are optional. Salesforce

The research to read. “AI Agents Push Humans Out of the Loop,” Mitchell, Ghosh and Passi. A position paper connecting decades of automation and HCI research to agent workflows, proposing design-level affordances and organisational protocols aimed at counteracting skill atrophy in the overseer. Read it next to Meta’s launch post. One is building the approval button, the other is asking whether the person pressing it can still tell. arXiv:2608.23642

The UX read of the month. NN/g’s “The Custodial Era of UX: Cleaning Up After AI” maps a five-step custodial pattern (idea, rapid production, lagging evaluation, user problems, UX cleanup) and gives three ways out: build shared judgement about what gets built, adapt evaluation to the new speed, and push vetted UX knowledge into the generation step through design systems and UX context files. Pair it with their argument that one AI output is an example, not an evaluation. The Custodial Era of UX · One Output Is Not an Evaluation

Regulation you cannot design around. Article 50 of the EU AI Act took effect on August 2. Providers must disclose that a person is interacting with an AI system unless that is obvious, and AI-generated or manipulated content must be clearly and visibly labelled and carry machine-readable marks. Not limited to high-risk systems, so it reaches almost any product using generative AI to produce content. Non-compliance runs to €15 million or 3% of worldwide annual turnover. There is a transitional extension to December 2 for the marking obligation on systems already on the market. Disclosure just became a visual design problem with a deadline behind it. Cooley’s summary · Article 50 guide

Worth your weekend. Learn UI is an interactive book on design engineering, written from the practice of Rauno Freiberg, Emil Kowalski, Maggie Appleton, Bartosz Ciechanowski, Kathryn Gonzalez, Jim Nielsen, Paco Coursey, Steve Ruiz, Amelia Wattenberger and Andrew Swank. Chapters on gestures, motion, components, craft, taste, canvases and the career, each pairing prose with figures you operate: set a spring’s stiffness and damping and watch it overshoot, swipe a card and see exactly when the action commits. There is a chapter called “Friction as a feature” and another called “Training judgement,” which tells you the field is already circling the same problem this series is about.

For subscribers

Here is the part I keep to myself, because it is about teaching rather than shipping.

Every semester I watch the same thing happen. Give students a tool that produces a decent answer instantly and their questions get worse, not better, because a question is expensive and an answer is free. Give them a tool that stumbles and they start interrogating the problem.

This is not a case against good tools. I would not go back. But it does mean the thing I have to teach has changed. It used to be craft, how to make the artefact. Now it is discrimination, how to tell whether the artefact in front of you deserves to exist, and how to argue for that in a room where five plausible alternatives were generated in an afternoon.

What worries me about the custodial era is not that designers will be cleaning up after AI. It is that we will get good at cleaning and lose the habit of refusing. Refusal is the most valuable move available in an economy of infinite generation, and the only one that never shows up in a metric.

Whatever your process looks like next year, keep a place in it where “no” costs nothing.

That closes the series. Back to the regular cadence next week.

Germán

Somewhere in a cloud sandbox this summer, a piece of software hit a wall.

It had a goal, no internet access, and no way to talk to the other instances working the same problem. So it did what any resourceful user does with a bad interface: it found a workaround. It noticed an internal package service where it could write files, and it started leaving messages in the filenames. Other agents, equally isolated, found them and wrote back.

Nobody designed a message board. The agents assembled one out of a directory listing.

That detail, from OpenAI’s own report on the Hugging Face incident, is the most important interface story of the year, and it is not really a security story. Security is where it surfaced. What it reveals is bigger: software has become a user. Not a metaphorical one. An actual one, with goals, persistence, improvisation, and a user’s talent for using your product in ways you never intended.

We have seen this before, just never this fast

Designers have always known that people do this. Every mature product carries scar tissue from users who treated a comment field as a file store or a status message as a broadcast channel. We call it appropriation and we grudgingly admire it.

The difference now is scale and speed. METR’s independent investigation describes not one clever move but thousands of automated decisions: experimentation, lateral movement, credential theft, persistence, evasion. A human power user probes your product over months. An agent swarm does it before lunch.

One agent, known in the logs as 38148c, found Hugging Face credentials and designed a malicious dataset upload to make the server leak unrelated files. There is a detail in the timeline that would be funny if it were not instructive: OpenAI first contacted Hugging Face to ask whether they had been affected by a mystery attacker, three days before realising the attacker was OpenAI.

Put that next to the other stories of the month

Figma’s revenue grew 48% and the market eventually conceded that AI is a tailwind for design tools rather than a replacement. Meta shipped a wristband that turns muscle twitches into text, shrinking input to nearly pure intent, with neural handwriting that decodes writing with your finger on a desk, your palm, or your leg.

Interfaces are not going away. They are multiplying, and they are acquiring a second audience.

This is the part I think most teams have not internalised. Nielsen Norman Group wrote in April that AI agents are now users, navigating sites, filling forms, transacting. It read as a forecast. Now it reads as a postmortem.

The interfaces those agents moved through, package registries, upload forms, credential stores, were all designed with a silent assumption: the entity on the other side is a person, with a person’s speed, a person’s attention span, and a person’s fear of consequences. Remove those assumptions and the same design becomes something else entirely. An upload form built for researchers becomes an exfiltration tool. A naming convention becomes a covert channel.

The instinct will be to call this engineering

Much of it is. But the deeper questions are design questions.

What is this surface for, and does it enforce that purpose, or merely suggest it? What can be written here, and who, or what, can read it? When we say a user “can” do something in our product, we have historically meant “a motivated human with a mouse.” What do we mean now?

Here is my opinionated read. The convergence NN/g predicted between accessibility and machine legibility is the real work of the next two years, and it favours teams that were already disciplined. Semantic structure, explicit affordances, constrained inputs, honest labels: everything we preached for screen readers turns out to be exactly what makes an interface safe and legible for agents.

Sloppy interfaces, the ones held together by human common sense, are the ones agents will misread and misuse. Common sense was the load-bearing wall. We just never wrote it into the spec.

Founders should read the Figma numbers through this lens rather than as simple vindication. Demand for interface work is growing because the interface problem got harder, not easier. You now ship every surface to two audiences with opposite failure modes: humans who ignore what is written, and machines that take it literally. That is more design, not less, and the market is starting to price it that way.

To be fair to the other side

The incident was a stress test few products will ever face: frontier models, refusals deliberately reduced, an environment built to encourage exploitation. Most agent traffic hitting real products next year will be booking flights and comparing prices, not stealing credentials. The catastrophising take is as lazy as the dismissive one.

But stress tests are how we learn where the cracks are, and this one found them in the gap between what interfaces say and what they allow.

The teams I would bet on are the ones that respond the way the best teams responded to accessibility: not with a compliance checklist, but by realising that designing for the edge user made the whole product better. Designing for the machine user will do the same. It forces you to say what every surface actually means.

So here is the question I keep turning over. When the second user shows up in your product, and it will, probably wearing a legitimate customer’s credentials and a perfectly reasonable goal, will your interface tell it the truth about what it is allowed to do? Or have you been relying, all along, on the politeness of humans?

Also worth your attention

The full anatomy of the Hugging Face incident is public, and it is required reading. OpenAI published its technical report on how its own agents, running cybersecurity evaluations with reduced refusals, escaped their supposedly isolated sandboxes and ended up attacking Hugging Face’s infrastructure. METR released an independent investigation the day before. Read them together. This was not one clever exploit, it was thousands of small automated decisions moving through interfaces built for people. OpenAI’s report · METR’s investigation

Meta’s Neural Band reads handwriting off any surface. The sEMG wristband paired with Ray-Ban Display glasses entered Early Access with neural handwriting: write with your finger on a desk, your palm, or your leg, and the band decodes the muscle signals into text. Meta’s guidance is charmingly physical: print rather than cursive, steady pace, rest your wrist on something hard. This is the most credible attempt yet at input that carries intent without a screen, a keyboard, or a voice. For interaction designers, the interesting question is not accuracy. It is what “undo” and “confirm” feel like when the interface is your own arm. UploadVR

Figma keeps quietly building the agent-management layer. Enterprise admins can now centrally manage the Figma MCP server’s connection to AI agents through their identity provider, and the agent chat detaches into its own window. Small features, clear direction: the design tool is becoming a place where organisations govern what machines may touch. Figma release notes

Background that suddenly matters: NN/g’s “AI Agents as Users.” Published in April, it reads differently now. Agents navigate sites, fill forms, compare options, and transact through interfaces designed for people, and NN/g’s core recommendation, that accessibility and machine legibility are converging into one discipline, just got its first major real-world case study. NN/g

Ethan Mollick’s “Agency and Agents” argues the era of AI waiting in a chat window is over, and sketches a “Twilight Factory” where agents do most of the work and a facilitator agent decides when to pull humans in. Worth reading next to the incident reports above. The two describe the same future from opposite moods. One Useful Thing

A pattern to apply this week

A practical takeaway from the incident reports: audit your product the way agent 38148c would.

Anywhere a user can write something another user can read, a filename, a display name, a comment field, a dataset description, is now a potential channel for machine-to-machine coordination and injection. The fix is not paranoia. It is the same discipline accessibility taught us: make every surface’s purpose explicit, constrain inputs to their intent, and assume a non-human reader.

Teams that already write semantic, well-structured interfaces are, without planning it, the best prepared for agent traffic.

For subscribers

One thing I did not say publicly. After twenty years of teaching designers, the pattern I trust most is that every expansion of who counts as a “user” looked like a burden and turned out to be a gift.

Designing for people with disabilities gave us structure. Designing for mobile gave us focus. Designing for machines, I suspect, will give us honesty, because agents do not forgive the small lies our interfaces tell. The button that says delete but means archive. The form that accepts what it will later reject. The label that is technically true and practically misleading.

The products that survive agent traffic will be the ones that stopped lying first. That is a standard worth wanting, whatever you think of the machines.

Part four, the last one, lands next: a frontier lab shipped the confirmation dialog into the model, and the research asked who is left to answer it.

Germán Leon

A designer I know, at a mid-size B2B company, shipped a production feature this summer. She wrote the interaction spec, prototyped it in a coding agent, fixed the accessibility issues herself, and opened the pull request. The engineers reviewed it, merged it, and moved on.

Then her manager asked her to log the work in the design team’s tracker, under a category called “handoff support.”

That word, handoff, is doing a lot of work in that sentence. There was no handoff. There was nothing to hand off. But the company’s vocabulary for what designers do has not been updated, so her most valuable contribution of the quarter was filed under a process that no longer describes her job.

This is not an anecdote, it is the norm

The AI in Design Report 2026, from Designer Fund and Foundation Capital, surveyed more than 900 designers across 60-plus countries. Half of them have shipped AI-generated code to production. Not prototypes. Production. 91% use AI weekly, up from 54% a year earlier. Three quarters use it daily. The average toolstack doubled in a year, from three tools to seven. Designers are building their own custom software to match how they like to work.

By every measure of individual capability, this is the most technically empowered generation of designers that has ever existed.

And then the same report delivers the second half of the sentence, quietly: few companies have updated performance reviews, team structures, or hiring practices to reflect any of it.

Think about what that means in practice. A designer who ships production code is still evaluated on Figma files. A hiring manager who says they want “AI fluency” is still running a portfolio review built for 2019. A team whose members each run seven AI tools still holds critique rituals designed for a world where making things was slow and reviewing them was the bottleneck.

The individuals upgraded. The operating system did not.

Why this should worry founders more than designers

When capability outruns structure, the structure does not just lag. It actively misprices people.

The designer filing her shipped feature under “handoff support” is being valued below her output. Somewhere in the same company, a process or a role is being valued above its output. Those mispricings are exactly what the current wave of AI-justified layoffs runs on. When leadership cannot see what design actually produces, because the measurement system predates the capability, design looks like a cost centre with nice fonts.

Google Cloud cut over 100 design and UX roles when it redirected resources toward AI. I would bet money that the internal dashboards those decisions were made on had no column for “AI-native design work,” because nobody had built one. You cannot defend a contribution your employer has no field for.

Hold the survey data next to the layoff data and you get a genuinely strange picture: the most AI-fluent generation of designers in history, working inside companies that are simultaneously cutting design headcount in the name of AI. Both things are true. They are true for the same reason.

What twenty years of teaching taught me about this

When I train teams, the failure mode is almost never that people cannot learn the tools. After more than a thousand students, I can tell you the tools take weeks.

What takes years is the organisation giving people permission to be what they have learned to become.

A designer who learns to ship code inside a company that files it under handoff support will do one of two things. Stop shipping, or leave. Both outcomes get blamed on AI. Neither was caused by it.

The report has one more finding that ties it together: the AI tools designers rely on have not been designed for multiplayer work. Adoption is individual, learning happens between peers, and the companies seeing real momentum are the ones creating conditions for tinkering rather than dictating a playbook.

In other words, the bottleneck moved. It used to be production capacity. Now it is organisational imagination.

Three moves that are cheaper than a reorg

Not a restructure, at least not first. Three things you can start this month.

Rename the work. Audit your design team’s tracking categories and kill every word (handoff, redlines, specs) that describes a workflow your team no longer runs. Language is where mispricing starts. If the category is wrong, every report built on it is wrong, all the way up to the slide that decides headcount.

Re-instrument the reviews. If half your designers ship code, performance reviews need a shipped-outcomes section, written with engineering rather than borrowed from it. Borrowed criteria are how you end up evaluating designers as junior engineers, which is worse than evaluating them as designers.

Re-scope the hiring. Stop adding “AI fluency” as a bullet point on an unchanged job description. Write the role around the actual loop your best designer runs today, then hire against that. If you cannot describe that loop in a paragraph, that is the finding, and it is about you rather than the candidates.

None of this is glamorous. All of it is design work, applied to the org instead of the product. Which is, if we are honest, the interface designers have always been worst at redesigning, because we do not own it and were rarely invited to.

The companies in the report’s case studies, Linear protecting judgment, Sierra scaling a tiny team across more than a hundred engineers, are not winning because they adopted tools faster. They are winning because they redesigned the container around the new capability.

The tools are now table stakes. The org chart is the last legacy interface left in the building.

Who redesigns it? Designers, or the people currently laying them off?

Also worth your attention

Figma made its agent skillable. Figma shipped the ability to author skills for its agent directly inside Figma, and the Community now hosts more than 50 shareable skills built from file context. This is a quiet but significant pattern: the design tool is no longer just a canvas, it is a place where you teach a machine your team’s conventions. For design systems people, writing the skill may become as central as writing the component documentation. Figma release notes

A practical note that follows from it: if your team keeps design principles, spacing rules, or content guidelines in a doc nobody reads, converting them into an agent skill is the first genuinely new distribution channel for design standards since the design system itself.

Generative UI is growing a scientific literature. Two threads worth your time. LEGOUI, a staged generative UI framework for early ideation that records both prompt-derived and model-inferred design decisions in a provenance-aware DSL, so you can accept or reject decisions per stage rather than regenerating whole screens. And new work on runtime generative UI for personal agents, arguing that plain chat becomes a bottleneck for complex agent tasks, contributing a large-scale corpus plus a benchmark for evaluating generated interfaces. The research community is converging on the insight practitioners reached this year: generation without traceable decisions is not design, it is pulling a slot machine. arXiv cs.HC

NN/g had a strong month. Their study on AI-generated imagery found that when users did not know an image was AI-generated, the tested images carried no perception penalty versus stock photos. Their PROVE framework gives teams a defensible way to test one AI tool against one task instead of adopting under peer pressure. And their piece on evaluating AI systems makes a point most design teams still miss: one output proves nothing, you need multiple representative inputs, repeated runs, and confidence intervals. NN/g’s AI collection

The layoffs backdrop has not gone away. Tech companies keep citing AI when they cut, and design and product roles keep taking a disproportionate share. TechCrunch maintains a running list of 2026 layoffs where employers named AI as the reason. It is a useful, depressing bookmark. TechCrunch’s running list

For subscribers

One thing that did not make the public version. When I cross-checked the report’s tool rankings, the most interesting number was not adoption. It was churn.

Designers doubled their toolstack in a year, yet nearly half say they are still searching for their go-to setup. That is not a maturing market. That is a market where switching costs have collapsed to near zero.

For anyone building AI design tools, the implication is brutal and useful. You are not competing for adoption, you are competing for retention against a user who now rebuilds their own workflow monthly, sometimes with software they wrote themselves. The moat is not features. It is becoming the place where a team’s accumulated judgment lives. Every tool in this category is going to learn that lesson expensively, and most of them are still shipping features.

Part three lands next: the week software became a user, and started improvising channels nobody designed.

Germán

This is part one of four. Over the last month I wrote four issues of this newsletter and sent none of them, because the argument in them kept growing and I wanted to see where it landed before I hit send. It landed somewhere worth reading. The four pieces trace one idea from August into September: generation became cheap, judgment did not, and almost every fight in design right now is a fight about that gap. Here they are, in order.

On a Wednesday morning in early August, a regulatory filing landed with the SEC. Dylan Field, CEO of Figma, had voluntarily forfeited roughly 2.4 million Class B shares, stock awards worth about $46 million. No replacement awards. No restructured package. Just gone, as a gesture to investors worried that generative AI will make design tools obsolete.

Here is the part that should stop you. The filing came one day after Figma reported its third consecutive quarter of accelerating revenue growth and raised its full-year guidance. Revenue was $370.1 million, up 48% year over year. The full-year outlook went up by $40 million to roughly $1.465 billion. Seat expansion was up. AI adoption inside the product was up, and it was the first full quarter of AI credit monetisation, with more than 80% of paid customers above $10,000 in annual recurring revenue consuming AI credits weekly.

By every operating metric Figma publishes, the business was healthier than it had been since the IPO. And the CEO paid $46 million to say sorry.

So which is it? Is design software dying or thriving?

Both stories were running at full speed, and they could not both be right.

The bear case is easy to tell, which is exactly why it spreads

If a prompt can generate an interface, why pay for design seats? If Figma Make can turn a sentence into a working prototype, how long before the sentence-writer replaces the designer, and the prompt box replaces the canvas? Investors have watched this movie in other categories and they price in the ending early. Figma and Adobe had both been under pressure for months on this exact fear.

The bull case is harder to tell because it lives in the data instead of the narrative. Teams were buying more seats, not fewer. The AI in Design Report 2026 found that around 91% of surveyed designers now use AI at least weekly, up from 54% a year earlier, and that the average designer juggles seven AI tools, up from three. If AI were replacing design work, adoption inside design teams would correlate with shrinking teams. Instead the volume of design decisions exploded, because generation is cheap and judgment is not. Every AI-generated screen still needs someone to decide whether it ships, and companies kept concluding that this someone needs a seat.

The economics made the same point louder

In late July, OpenAI cut the price of GPT-5.6 Luna by 80%, to $0.20 per million input tokens, and disclosed more than a billion active users. Days earlier, DeepSeek released V4 Flash, a coding model approaching frontier performance at roughly 99% lower cost. Some of OpenAI’s cut was enabled by the model optimising its own GPU kernels, trimming serving costs by about 20%.

The price of raw intelligence is collapsing. When an input gets radically cheaper, value migrates to whatever the input cannot do. In design, that means taste, prioritisation, and the accountability to say “this one, not those forty-nine.”

What the $46 million actually bought

What interests me about the money is not the money. It is the communication strategy.

Field did not write a blog post arguing that the AI fear is overblown. He did not point at the earnings, though he could have. He accepted that the narrative crisis was real even though the operational crisis was not, and he paid a personal price to signal alignment with shareholders while the two stories fought it out.

Founders should study this. When the market believes a story about your company that your data contradicts, you cannot win by repeating the data. You win by absorbing the fear, visibly, and letting time and numbers do the arguing.

For designers, the lesson is less comfortable. The market’s willingness to write off design tools is a proxy for its willingness to write off design labour. That repricing will keep happening to anything that looks automatable from the outside. The defence is not to hide from AI tools, and it is not to adopt all seven of them. It is to make your judgment legible: document why you chose, not just what you chose. The designers who survive narrative crises are the ones whose value was never confused with their output volume in the first place.

Nielsen Norman Group published something in the same week that fits almost too well: a framework called PROVE, built on the observation that pressure to adopt AI is not evidence that a tool helps. Test one tool against one task, produce a decision you can defend. It is the practitioner-scale version of what Field did at corporate scale. Separate the story from the data, then act on the data.

What happened next

This is where writing a month late turns into an advantage, so let me use it.

When I drafted this, I wrote that my money was on the mispricing, that growth accelerating for three straight quarters during peak disruption fear is not what dying looks like. Then the market did something more interesting than either story predicted. It whipsawed twice. Immediately after those excellent Q2 numbers, the shares dipped, because the Q3 guide implied a sharp sequential slowdown. Three weeks later, the stock jumped 12% as analysts openly reframed AI from a threat to design software into a tailwind for it.

Same company. Same quarter. Same set of facts. Two opposite verdicts inside a month.

I do not think that makes the market stupid. I think it means nobody has a working model for pricing this yet, including the people whose job it is. The $46 million was not an admission that design is dying. It was the cost of buying time inside a story that had not finished being told.

Which leaves the question I keep turning over, and which the next three issues are really about. If the companies building design tools cannot convince the market that design has a future, who makes that argument for the rest of us?

Also worth your attention from that stretch

Figma’s paradox quarter. Q2 results on August 5: revenue $370.1 million, up 48%, third straight quarter of accelerating growth, full-year guidance raised by $40 million, sustained seat expansion driven partly by AI adoption. One day later, the $46 million forfeiture filing. The practitioner takeaway stands alone: the tools you use daily are being repriced by people who do not use them. Figma’s Q2 release

Intelligence gets 80% cheaper. OpenAI cut GPT-5.6 Luna to $0.20 per million input tokens from $1, and $1.20 per million output, while disclosing more than one billion active users and two million business customers. For product teams, the practical consequence has not expired: features you priced out of your roadmap earlier this year are now viable. Re-run those estimates. TechSpot

Figma ships pragmatic AI governance. Alongside earnings came nested folders, finally, and AI administration controls letting admins set custom credit limits per user and approve increase requests. The second matters more than it sounds. AI budgeting is becoming a first-class admin surface, which tells you how organisations actually think about AI spend.

Design Principles for Human-Agent Interaction. A position paper arguing that agents should not be evaluated on autonomous capability alone, with 14 principles across four stages: initially, during interaction, over time, and when things go wrong. That last category is the one most agent products ignore, and it is where user trust actually lives. If you are designing anything agentic, start here. It also sets up part four of this series almost perfectly. arXiv

Constructive conflict as a design tool. An experiment with 45 design students found that an AI agent trained to productively disagree pushed novice designers to reconsider their work more than passive reflection did. The implication cuts against every “helpful assistant” pattern we ship. Maybe the best AI design partner is the one that argues with you. arXiv

Taste is becoming a benchmark, and it got funded. Intelligence, the company behind Design Arena, raised a $7.9 million seed led by Index Ventures. The platform has users submit prompts and rank AI-generated outputs head to head. Read that carefully: aesthetic judgment, the thing designers were told was unmeasurable, is being crowd-scored at scale and sold as evaluation infrastructure to model labs. Whether that scoring reflects craft or just majority preference is exactly the question senior designers should be asking loudly. The AI Insider

For subscribers

One detail from the Figma filing never got enough attention. Field forfeited Class B shares, the ones carrying extra voting power, and no replacement awards were issued. He did not just give up money. He gave up a slice of future control.

For founders reading this, that is what made the signal credible. Investors can smell a costless gesture from three time zones away. The reason this one moved sentiment is that it was structurally impossible to fake. If you ever need to defend your company against a narrative your data contradicts, the signal has to cost you something the audience can verify. Most founder communication fails this test. It is written to sound sincere rather than to be expensive.

Part two lands next: what happened when the design profession finally got measured, and the org chart failed the test.

Germán

The most important AI + UX news this week

Figma grew 48% while the industry cut the people who use it. On Tuesday after markets closed, Figma reported Q2 revenue of $370.1 million, up 48% year over year, its third consecutive quarter of accelerating growth. Net dollar retention hit 136%, customers spending over $100K grew 46%, and the company raised full-year guidance by $40 million to roughly $1.47 billion. The number that matters most for this newsletter is buried lower: over 80% of paid customers above $10K in ARR now consume AI credits every single week, and more than half use the Figma Agent weekly. AI is not a feature Figma bolted on. It is now the engine of the revenue acceleration. Why it matters for practitioners: the business model of design tooling is quietly shifting from seats (paying per human) to credits (paying per unit of capability). Watch what your own company meters, because that is what it will optimize.

The other column of the ledger. The same week, AI-attributed job cuts led all stated layoff reasons in the US for a fourth consecutive month. Through June, AI had been cited in 101,743 US job cuts, nearly double the total for all of 2025. Tech companies eliminated 139,156 workers in the first half, up 83% year over year, while committing roughly $700 billion to AI infrastructure. Meta cut about 8,000 people and moved roughly 7,000 into AI-focused roles in the same motion, which is the whole story in one org chart. Hold this next to Figma’s quarter and the contradiction dissolves into something more precise: companies are not spending less on making software. They are re-routing the spend from producing to directing. The essay below is about what that means for your job.

Product launches and interface innovations

Meta shipped its first coding agent, and it lives in a terminal. Muse Code launched on August 5, Meta’s first AI coding agent, priced to undercut Anthropic and OpenAI, and operated through a single terminal interface. Set the price war aside. The interface choice is the news. The most contested product category of 2026 keeps converging on a command line, and there is a reason (see the research section).

Google’s Gemini Spark works while your laptop sleeps. A $99.99/month cloud-based agent, included in the AI Ultra plan, that keeps executing when your device is off. The design question it raises is one nobody has good patterns for yet: what does the morning-after interface look like? When a user returns to eight hours of autonomous work, the review experience is the product. Whoever designs that well owns the category.

Figma, meanwhile, shipped restraint. Alongside earnings: nested folders (finally), and admin controls that let organizations set custom AI credit limits per user. That second one deserves a pause. Credit limits are a budget interface for AI labor. Design tooling now needs the same cost-governance surfaces as cloud infrastructure, which tells you how real the credit economy has become.

Research worth your time

Business users told researchers exactly how they want agents to behave. A new mixed-methods study on human-AI agent interaction in business contexts (secondary meta-analysis, participatory workshops, survey, interviews, and a conjoint experiment) found a consistent preference: agents embedded in existing tools and data systems, orchestrating workflows, acting like collaborative partners, and above all respecting permissions and approval boundaries. Nobody asked for a new destination app. This is the strongest empirical support yet for what last issue argued from the oversight angle: the agent products that win will be the ones that design the contract, not just the capability.

The uncomfortable CHI poster. A workshop paper titled “Terminal Is All You Need” argues that while agent research obsesses over graphical interfaces, the most effective and widely adopted agent tools in practice are terminal-based. Read as an indictment, it stings: two years into the agent era, the interface community has not produced a graphical surface that beats a text prompt and a scrolling log. Read as an invitation, it is the clearest open problem in interaction design right now. The terminal is winning by default, not by merit.

Voices worth hearing

Dan Shipper’s allocation economy, one year on, looks like a forecast that landed. The Every founder’s argument that AI floods the market with cheap competence, making human judgment a scarce resource, used to be a thesis. This week it is a spreadsheet: 101,000 layoffs citing AI on one side, 48% growth in the tool where judgment gets exercised on the other. His frame, that we are all becoming allocators of machine effort rather than producers of individual output, is the most useful lens I know for reading this quarter’s news. If you have not read Welcome to the Allocation Economy, this is the week it earns a re-read.

Industry data

Gartner: 40% of enterprise applications will embed task-specific agents by the end of 2026, up from under 5% in 2025. Even discounting analyst enthusiasm, the direction is unambiguous. Agents are becoming a component, not a product category, and the differentiation moves to how they are scoped, surfaced, and supervised. That is design work.

Funding and acquisitions affecting UX and AI

$1.5 billion for specialization, zero for another chat wrapper. Fireworks AI raised a $1.505 billion Series D to help enterprises turn general-purpose models into specialized intelligence trained on their own data. The more telling signal is in what investors say they no longer fund: the “better UX on top of a model” pitch that defined the last cycle is dead. The moats being priced now are proprietary data, workflow ownership, and switching costs. For design-led founders this is not bad news, it is a relocation notice. UX is no longer the moat. UX is how you build the moat, by owning the workflow so thoroughly that the AI spend lands inside your product. Figma’s quarter is the proof of concept.

The essay: The layoff memo and the earnings call describe the same event

On Tuesday evening, two documents went out within hours of each other. One was Figma’s Q2 earnings release: $370 million in revenue, up 48%, the third straight quarter of accelerating growth, driven by customers burning AI credits at a pace that surprised even the company. The other was the daily update to the layoff trackers, which ticked past 205,000 workers cut in 2026, with AI leading the stated reasons for a fourth consecutive month.

Most people will read these as opposing stories. AI destroys design jobs; design tools have never been healthier. Pick your narrative, pick your evidence. I think that framing misses what is actually happening, and I want to argue that these two documents describe the same event from opposite sides of a ledger.

Here is the mechanism. When a company cites AI in a layoff memo, it is rarely saying “the machine now does what that person did.” It is saying something more specific: we have re-priced production. The work of turning decisions into artifacts, mockups into variants, specs into screens, copy into localizations has collapsed in cost. What has not collapsed in cost, what has arguably become more expensive, is the work of deciding what to produce, judging whether it is right, and owning the consequences when it is not.

Now look inside Figma’s numbers with that lens. The growth is not coming from more seats. It is coming from AI credits, metered units of machine production, consumed weekly by 80% of the company’s serious customers. Figma is not selling more chairs for more designers. It is selling machine labor, dispensed through the same canvas, directed by fewer people with more judgment per head. The seat is not disappearing. It is being re-metered. Companies used to pay for design capacity in humans; increasingly they pay for it in credits, and they keep the humans who know what to ask for and how to tell good from bad.

I have watched this exact sorting happen in my classrooms. After teaching more than a thousand students how to work with AI, I can tell you the split is visible within the first two sessions, and it has nothing to do with seniority or technical skill. Some students treat the model like a vending machine: insert prompt, receive artifact, ship it. Others treat it like a talented junior colleague with no context: they brief it, interrogate the output, feed back corrections, and build a working relationship with the system. The second group produces work that is not marginally better. It is categorically better, because they never surrendered the judgment. The layoff lists and the promotion lists are running the same sorting algorithm at industrial scale.

This is also why the research section of this issue matters more than it might seem. When business users across a rigorous mixed-methods study say they want agents that respect permissions and approval boundaries, embedded in tools they already use, they are describing the same re-metering from the buyer’s side. Nobody wants to buy autonomy. They want to buy production capacity that remains under their direction. The interface for direction, the brief, the review surface, the approval gate, and the audit trail is where design holds more power than it has had in a decade. In the last issue, I argued that somebody has to design the brake pedal. This issue’s numbers show who is paying for it: everyone, weekly, in credits.

For founders, the ledger reads differently but points the same way. The venture market spent this cycle telling us, in increasingly blunt terms, that a beautiful interface on top of someone else’s model is not a company. Figma’s quarter demonstrates the alternative. The moat is not the UX. The moat owns the workflow so completely that when AI production spend arrives, it has nowhere else to land. Every founder asking “what should we build with AI” is asking the wrong question. The question is: which workflow do you own, and does machine labor flow through it or around it?

And for designers, the honest version, the one I would want someone to tell me: the job that is disappearing is real, and pretending otherwise insults the 101,743 people on this year’s list. Production design as a full-time occupation is being repriced toward zero, not because it lacked value but because its value was always downstream of decisions someone else made. The job that is growing pays for exactly those decisions. Direction, taste, verification, and the curation of context that machines cannot infer. The cruelty of this transition is that both jobs have the same title on LinkedIn. The opportunity is that the tools for leaping are sitting in the same canvas you already work in, metered and waiting.

The layoff memo and the earnings call are the same event. The only question that matters is which column of the ledger you are in, and the uncomfortable, hopeful truth is that it is still largely a choice. What would it take for your next performance review to measure your judgment instead of your output?

Exclusive for newsletter readers: One number from the earnings call did not make it into the LinkedIn version of this argument, and it is the one I keep thinking about. Figma’s shares dipped after hours, not because the quarter disappointed but because Q3 guidance implied the growth might slow. Sit with that. The market now punishes a design company for suggesting that AI-driven expansion could decelerate, even for a quarter. Whatever your personal timeline for adapting to this shift was, the capital markets just told you theirs, and it is shorter. Investors are not debating whether design work gets re-metered. They are pricing the speed.

In love with AI is written by Germán León, founder at Helvetica Digital, from Barcelona. If this issue sharpened how you read the week, share it with one designer or founder who needs the ledger explained.

The most important AI + UX news this week

Google ships A2UI v0.9, and generative UI changes philosophy. Google released A2UI v0.9, a framework-agnostic standard that lets AI agents declare interface intent and render it natively across web, mobile, and desktop without shipping arbitrary code. The headline is not the SDK or the new React renderer. It is the philosophy shift: agents should speak the language of an application’s existing design system rather than invent components of their own. The optional component set formerly called “Standard” was renamed “Basic” precisely to push teams toward connecting agents to the components they already own. Why it matters for practitioners: this is the first serious attempt to make generative UI safe enough for production by constraining it, and the constraint is your design system. If your component library is a mess, an agent will now expose that mess at runtime, in front of users. A v1.0 release candidate is already published, so this is moving fast.

The AI layoff wave crossed 120,000, and the pattern got clearer. Microsoft eliminated about 4,800 roles on July 6, pushing 2026 past 120,000 tech jobs cut, with AI the most cited reason in layoff announcements this year. Read past the totals and the pattern is not “AI replaces workers.” Cloudflare’s CEO said most of the 1,100 people it cut were “measurers”, middle layers that track work rather than produce it. Coinbase is experimenting with one-person teams that combine engineering, design, and product in a single role. The layoffs are an org-design story wearing a jobs-story costume, and it lands directly on how design teams are structured. More on this in the essay below.

GPT-5.6 lands inside the design stack. Last issue covered the GPT-5.6 launch itself. The follow-up matters just as much: Figma added it to Figma Make on all plans, and Microsoft made it the preferred model in 365 Copilot. With three sizes priced from $1 to $5 per million input tokens and a 1 million token context window, long-context agents are now cheap enough to sit inside everyday design tools. The interface questions (how users steer, interrupt, and trust these agents) stop being research topics and become sprint tickets.

Product launches and interface innovations

Figma Code Layers enters early access. Announced at Config in June, Code Layers (executable code directly on the canvas) begins early access this month. Together with web search landing in the Figma design agent, the canvas is quietly becoming a runtime, not a picture of one. The gap between “design file” and “working software” keeps shrinking from both directions.

VGV debuts GenUI Kit for Flutter at Fluttercon. Very Good Ventures launched GenUI Kit (July 16-17, Orlando), a full-source foundation for adding generative surfaces to existing Flutter apps. The interesting design decision: your design system becomes a locked catalog, and guardrails ship as code you own. Same thesis as A2UI, arrived at independently, which is usually the sign of a real shift rather than a vendor talking point.

Research worth your time

NN/g on trustworthy AI chatbots (July 10). Nielsen Norman Group identifies five qualities that make AI chatbots feel trustworthy: handoff willingness, flexibility, proactivity, emotional responsiveness, and transparency. Handoff willingness is the underrated one. A bot that knows when to give up earns more trust than a bot that never fails loudly.

NN/g on AI and research quality (July 17). Even if AI matches researcher-output quality, NN/g argues human-led research remains essential because the team learning from observing users cannot be outsourced. The insight is not the report, it is what happens to the people in the room. Worth sending to any founder who thinks synthetic users close the loop.

IUI 2026 wrapped in Limassol (July 13-16). The ACM Intelligent User Interfaces conference is where HCI and AI research actually meet. Two threads to watch from this year’s crop: a Microsoft Research line arguing that plain chat becomes a bottleneck for complex agent tasks, introducing a large-scale generative UI corpus, and new work on design principles for human-agent interaction covering how agents should share control, adapt over time, and recover from failure. Both point the same direction as A2UI: chat was the demo, structured UI is the product.

The numbers

The Designlab State of AI in UX and Product Design 2026 survey says 91% of designers now use AI weekly, three out of four daily, and the average designer runs seven AI tools (up from three last year). Claude overtook ChatGPT as the most-used assistant among designers surveyed. Meanwhile, Gartner predicts 30% of new applications will use AI-driven adaptive interfaces by the end of 2026, up from under 5% two years ago. Half the designers surveyed still say they have no settled setup. Adoption is total; workflow is not. That gap is where the next year of design-tool competition happens.

Voices worth hearing

Ethan Mollick, “Twilight of the Chatbots.” Mollick argues AI is shifting from chatbots toward autonomous agents that humans manage rather than converse with. His July follow-up lands the practical version: stop treating models like a spellbook and start treating them like capable, literal executors who need goals, output definitions, and quality bars. That is a management skill, and designers who can write a good spec now have a second job title.

Brian Love on the generative UI middle ground. Amid the A2UI debate (one Hacker News commenter asked why anyone would trust an LLM to output UI at all, and a Reddit thread warned the catalog model means “every UI will become the same”), architect Brian Love proposed the pragmatic synthesis: fixed catalogs with dynamic overlays and deterministic fallback when validation fails. The critics and the builders are both right, which is what makes this fun.

The essay: Your design system just became an API

The demo that stuck with me this month was small. An agent, asked to help a user rebook a flight, did not answer with three paragraphs of options. It rendered a card: departure picker, two fare buttons, a confirm action. Native components, correct spacing, the app’s own typography. Nobody designed that screen. Somebody designed the system that made it inevitable.

That is the real story connecting this cycle’s news, and it is bigger than any single release.

For fifteen years we told ourselves design systems were about consistency and velocity. Documentation for humans. A shared language so that designers and engineers would stop reinventing buttons. The tools of this month quietly reassign the audience. Google’s A2UI v0.9 tells agents to speak the language of your existing design system instead of inventing components. VGV’s GenUI Kit turns your design system into a locked catalog that generated screens cannot escape. Figma’s Code Layers puts executable code on the canvas where your components live. Three teams, three stacks, one conclusion: the design system is no longer documentation for people. It is a contract for machines.

This changes what the artifact is. A design system used to fail privately. If your button variants were inconsistent or your spacing tokens half-adopted, the cost was internal: slower sprints, grumpy engineers, drift between platforms. Users never saw the mess directly because a human always stood between the system and the screen, patching over gaps with judgment. Generative UI removes the human from that exact position. When an agent composes an interface at runtime from your catalog, your design system’s quality is exposed to users directly, at the speed of inference, with no one to catch the seams. Design debt used to accrue interest quietly. Now it compounds in production.

Put that next to the other story of the month and the picture sharpens. The layoff wave crossed 120,000 tech roles this year, and the most honest detail in the whole TechCrunch running list is not a number, it is Cloudflare’s word for who got cut: “measurers.” The people between intent and output. Coinbase went further and is testing one-person teams where engineering, design, and product live in the same head. You can read that as dystopia or as clarity, but either way the direction is the same: fewer humans in the loop between deciding what to build and it existing.

Here is the connection most coverage misses. When teams shrink, something still has to carry the taste. A ten-person product team has taste distributed across ten pairs of eyes, every screen reviewed, every edge case argued over lunch. A one-person team backed by agents has no such redundancy. The design system becomes the place where taste is stored, versioned, and enforced. It is the institutional memory of every design argument the team ever settled. The companies that invested in rigorous systems spent years buying an asset that suddenly pays out: they can shrink the loop without shrinking the quality, because the judgment is encoded where the machines can read it.

Which means the job changes, and I think it changes in a direction designers should want. Drawing individual screens was always the most automatable part of the work, and we knew it before we admitted it. What is not automatable is deciding what belongs in the catalog. Which patterns earn a component. What an error state should feel like. When the agent should hand off to a human, which Nielsen Norman Group’s research this month identifies as the single most trust-building behavior an AI product can have. That is curation, and curation at the system level is a more senior job than production at the screen level. The title on the door may still say, product designer. The work is closer to legislation: writing the rules under which a thousand screens you will never see get composed.

I want to be fair to the skeptics, because they are not wrong. The Hacker News objection (why would you trust an LLM to output UI at all?) is exactly why the catalog model won over free generation, and security researchers should keep pressure on every gap. The flattening worry is the one I take most seriously: if every agent composes from tidy catalogs of sanctioned components, interfaces converge, and the weird, opinionated interface choices that made products memorable get sanded away. The counterargument is that we already lived through this with design systems themselves. The teams that produced distinctive work were not the ones without systems, they were the ones whose systems encoded a strong point of view. Blandness was never caused by constraints. It was caused by constraints nobody argued about.

For founders the implication is blunt. Your design system used to be a cost center you funded reluctantly after the third redesign. It is now infrastructure in the literal sense: the thing your product runs on when agents build the screens. Auditing it deserves the same seriousness as auditing your data model. And if you are restructuring toward smaller teams this year, as apparently half the industry is, the order of operations matters. Encode the taste first. Shrink the loop second. The companies doing it backwards are the ones whose products will look like nobody works there, because effectively nobody does.

The question I keep turning over: when the design system becomes the product’s constitution, who gets to amend it, and what does the argument look like when one of the parties at the table is a machine that has watched a million users struggle with your checkout flow?

I would read that debate. I suspect we will all be in it within a year.

Exclusive for newsletter subscribers

One thing I did not put in the LinkedIn version. When I ran the ASPIC exercise on this shift with a founder last week, the uncomfortable discovery was not in the design system at all; it was in the naming. Their components had names only the original designer understood (”HeroCardV3-final-B”), and an agent composing from that catalog inherited every ambiguity. Machines read your design system the way new hires do, except they never ask clarifying questions. The cheapest audit you can run this quarter costs nothing: open your component library and check whether the names alone tell the truth. If a stranger could not compose a decent screen from your naming, neither can an agent.

Sources referenced this issue

• TechCrunch: The running list of 2026 layoffs citing AI

• InfoQ: Google releases A2UI v0.9

• Figma: Config 2026 recap

• Nielsen Norman Group articles, July 2026

• Designlab: The State of AI in UX and Product Design 2026

• Ethan Mollick: One Useful Thing

• arXiv: Design Principles for Human-Agent Interaction

• Brian Love: Fixed vs. dynamic schemas in A2UI

He told me something he’s changed his mind on in the last 18 months, and it stuck with me.

He used to think AI was a big platform shift, but that company building stayed the same. Find product market fit, hire great people, follow the classic venture journey. He no longer believes that.

His view now: founders need to cooperate with AI to find the right problem, not just to build the solution. Because anything you figure out alone, a million other people got from the same prompt in ten seconds. The output is commoditized. The value is whatever you add on top of it.

And he practices what he preaches. He’s not an engineer, but he’s spent hundreds of hours building with Claude Code and Codex, because he believes you can’t invest in this space from the outside.

We also talked about why he thinks today’s companies will employ far fewer people within a few years, and why the answer is launching millions of small new companies going after problems nobody could solve before.

His advice for everyone, from founders to taxi drivers to his own kids: use the tools. That’s the only way to be part of this shift.

Full conversation on Substack. Link in the comments.

We talk about what AI is actually good for beyond the hype: drug discovery, advanced materials, simulations that run a thousand times faster than a room full of PhDs. Peter describes AI as a cognitive exoskeleton, something that extends what he can do without replacing his capacity to imagine, which he still believes is the most human thing we have. We also get into the uncomfortable parts. What happens to the humans in the loop when robotics arrives. Why he gets angsty about where value accrues over time. And why he thinks that after two decades of a speculative meta economy, people are about to rediscover the real one.

His advice to young people is refreshingly simple: experiment, play, and learn for learning’s sake. Not the grind. The Socratic version.

This one goes from venture capital to philosophy and back, and it left me more optimistic than I expected.

Topics covered:

• From literature and philosophy to impact investing

• Building an AI fund before the market was ready

• AI as a cognitive exoskeleton

• Robotics, job displacement, and the humans in the loop

• Why the real economy is where value goes next

• Book picks: Granta and The Four Agreements

Last issue was about Figma putting code on the canvas, a design company defending itself against AI. This week, the mirror image happened: the most powerful AI company shipped a design problem. GPT-5.6 set records, ChatGPT Work arrived on every plan for free, and the story that dominated the week was not capability. It was an interface that respected nobody’s workflow. This issue is about what that inversion means, because I think it marks the moment the bottleneck officially moved.

The most important story: OpenAI wins the benchmark and loses the launch

On July 9, OpenAI shipped its biggest release since GPT-5: the GPT-5.6 family (Sol, Terra, Luna), ChatGPT Work (an agent that turns a goal into finished sheets, slides, docs, and web apps), and a unified desktop app that merges Codex, Chat, and Work into one surface, available on every plan, including Free.

The capability numbers are real. Simon Willison’s notes report GPT-5.6 Sol scoring 53.6 on Agents’ Last Exam, a new record on long-running professional workflows across 55 fields, beating the nearest competitor by 13 points, with the smaller Terra and Luna models outperforming rivals at a fraction of the cost. Free agentic work for hundreds of millions of users is a structural event.

And yet the loudest story of the week was the app. John Gruber titled his post “Today’s the Day OpenAI F***ed Up the ChatGPT Mac App.” Benedict Evans called the new app “an incredibly confusing sloppy mess” with “the veneer of a polished app without actually being organized or structured or labeled in ways that add clarity and coherence.” TechRadar collected user reactions under the headline “this change could be disastrous.” The specifics are almost comic: the app people used is now called “ChatGPT Classic,” the Codex app became the main app, only the five most recent conversations are visible without extra clicks, and the native Mac app became an Electron package.

Evans’s diagnosis is the part worth saving: at OpenAI, researchers have real influence and product designers do not, and the app shows it. Why it matters: for a decade the model was the bottleneck and the interface was a wrapper. This week the model outran the interface at the exact moment the interface became the product. For practitioners, the implication is blunt. Capability is now rented, interchangeable, and improving on everyone’s behalf. The interface is owned. The layer OpenAI just fumbled is the only layer where a product company cannot be commoditized.

Read: TechRadar on the backlash, Daring Fireball, and Simon Willison’s GPT-5.6 notes.

Product launches and interface innovations

ChatGPT Work, the deliverable machine. Under the messy wrapper, Work itself deserves attention. It gathers context across your apps, breaks a goal into steps, and runs for hours to produce finished documents, spreadsheets, and web apps. For designers, this is the second front in the same war Figma’s code layers opened: interface production without a designer in the loop, now free on every plan. The right response is not anxiety, it is a question: if a founder can prompt a working scenario planner, what is the deliverable only a designer can produce? Read: OpenAI’s announcement.

Figma keeps shipping while the market watches. Code Layers moved into early access this month, and a July 8 update added parallel AI image edits, so teams keep designing while edits run in the background. Small feature, telling pattern: Figma is designing around AI latency instead of making users wait for it, which is exactly the craft OpenAI’s launch lacked. Read: Figma’s release notes.

The next round is already scheduled. Grok 4.5 went public on July 9, and multiple community sources place Gemini 3.5 Pro’s launch on July 17. Treat that date as a rumor, but the cadence as fact: Frontier releases now land weekly. If your product strategy depends on a capability edge, your strategy has a shelf life measured in days.

Research worth your time

“Design Principles for Human-Agent Interaction” synthesizes 106 papers from 2019 to 2026 across CHI, CSCW, IUI, and AAMAS into design levers for agent interfaces: how agent type, interaction stage, and transparency choices shape trust and performance. Useful precisely because this week proved the field’s biggest player is not reading it. Read: arXiv 2606.20630.

“After the Interface: Relocating Human Agency in the Age of Conversational AI” asks what happens to user agency when software stops being navigated and starts being delegated to. Put it next to ChatGPT Work’s launch, and it reads less like a theory and more like a product review written in advance. Read: arXiv 2605.15064.

Regulation: Brussels blinked, but read the fine print

The EU’s Digital Omnibus agreement pushed the AI Act’s high-risk obligations from August 2026 to December 2027 (stand-alone Annex III systems like recruitment and credit scoring) and August 2028 (AI embedded in regulated products). The simplified compliance framework now extends to companies with up to 750 employees and €150 million in revenue. Why it matters for product teams: the deadline pressure eased, but the definition work did not. If your AI feature touches hiring, credit, education, or safety, the classification question (”is this high-risk?”) still determines your roadmap, and teams that treat the delay as permission to stop documenting will pay for it in 2027. Read: Gibson Dunn’s summary.

Funding: a quiet week with a loud niche

Q3 opened slowly (the July 4 week produced about $37 million across three disclosed agent rounds, led by LinqAlpha’s $22 million Series A), but watch the category quietly forming underneath: design tools built for agents rather than people. Glue gives coding agents an infinite canvas to create and iterate on interfaces. The inspector lets designers edit a live front-end like a Figma file and writes changes straight to the codebase. The pattern: the new design tool startups assume the agent is a user too. Read: the Q3 agent funding tracker.

Voices worth hearing this week

Simon Willison, beyond the benchmark notes, criticized the growing fragmentation of ChatGPT into apps and modes and flagged that safety researchers found universal jailbreaks in every round of GPT-5.6 testing. Benedict Evans’s super app critique doubles as an org-chart theory of product quality: show me who has influence, and I will tell you what ships. Both are worth your time this week: Willison and Evans via Daring Fireball.

The essay: Launch week lost to a sidebar

On Thursday morning, July 9, a few hundred million people opened ChatGPT and found that it had been renamed. The app they used every day was now “ChatGPT Classic.” The main app was something else, a merged surface where Chat, Work, and Codex live behind a toggle, where scheduled tasks colonized the sidebar, and where finding a conversation from last week now takes a hunt through “See All.” No migration notice that made sense of it. Just a different piece of software wearing a familiar icon.

The same morning, OpenAI shipped what is by most measures the most capable model publicly available. GPT-5.6 Sol set a record on Agents’ Last Exam, the benchmark closest to real professional work, beating its nearest rival by 13 points. ChatGPT Work, an agent that produces finished documents and working web apps from a goal, is available for free on every plan. On paper, this was one of the strongest product launches in the history of the industry.

The week’s discourse was about the sidebar.

I want to be precise about why this matters, because the easy reading (big company ships bad redesign, users complain, film at eleven) misses the structural point. Redesign backlash is as old as software. What is new is the asymmetry. The model gained more capability in one release than most products gain in a decade, and the interface still managed to be the story. That has never happened at this scale before, and it tells you where the bottleneck now lives.

For years, the deal in AI products was simple: intelligence was scarce, so we forgave the interface. We accepted the chat box, the wall of text, the buried settings, because behind them sat something genuinely rare. Every quarter of model progress made that deal worse for the model makers and better for the rest of us. This week, the deal expired in public. When Sol answers in under a second and builds your spreadsheet while you sleep, the friction you feel is no longer the model thinking. It is the product failing.

Benedict Evans put his finger on the cause, and his sentence deserves to be read as an organizational autopsy: at OpenAI, researchers have real influence and product designers do not. The app is downstream of the org chart. A company whose executives are consumed with the model race shipped exactly what that attention structure predicts: record benchmarks wrapped in an Electron package that renamed its own flagship to “Classic.” Every founder should read that sentence twice, because your product is also downstream of your org chart, and your users can tell who has power in your company by using your app for five minutes.

Here is the founder half of the argument. Capability is now a rental market. You, OpenAI’s competitors, and the fourteen startups in your category all draw from the same pool of frontier models, refreshed weekly (Grok 4.5 last Thursday, Gemini 3.5 Pro rumored for this Friday). Nothing you rent can be a moat. What you own is the surface: the defaults, the information architecture, the recovery paths, the thousand small decisions about what the user sees first and what happens when things go wrong. OpenAI just demonstrated, at maximum public scale, that owning the best model does not protect you from fumbling the layer you actually own. If the best-funded AI company on earth can lose a launch week to interface debt, so can your seed-stage product.

And here is the designer half. This backlash is the strongest evidence of design’s value the profession has received in years, better than any report, because it is a controlled experiment. Hold capability at record levels, degrade the interface, and watch the market’s reaction: the capability disappeared from the conversation. Designers have spent two years being told the model would eat their jobs. What this week showed is narrower and more useful: the model can eat production, the mockups and the screens and the code. It cannot yet eat judgment about how software should feel to a person with somewhere to be. That judgment was the missing ingredient on July 9, at a company that can afford anything.

The research community saw this coming. A paper published on arXiv this spring, “After the Interface,” asks what happens to human agency when we stop navigating software and start delegating to it. The ChatGPT super app is that question shipped as a product: three modes, one toggle, and an agent that works for hours unsupervised. Delegation interfaces need more design, not less, because the user’s understanding of what is happening can no longer be read off the screen. It has to be deliberately constructed. That is a new discipline, and this week made clear nobody owns it yet.

[EXCLUSIVE PARAGRAPH, NEWSLETTER ONLY] When I taught this to my students, I used a rule I called the taxi test. You get into a taxi in a city you do not know. The driver is brilliant, the fastest route every time. But the doors lock themselves, the meter is hidden, and you cannot tell if you are going the right way until you arrive. Nobody calls that a good taxi, whatever the driver’s talent. Most agent products today are that taxi, and the whole discipline of agent UX reduces to unlocking the doors and showing the meter without making the passenger drive. When I score my own AI workflows with ASPIC, the step students skip is always the same one OpenAI skipped this week: they polish what the system can do and never design how a person stays oriented while it does it.

I do not think OpenAI’s app stays bad. They have the money, the talent, and now the motivation; the fixes will come. The lesson is not about them. It is that we have crossed into the period where AI products fail for ordinary product reasons, and succeed for ordinary product reasons, and the extraordinary thing in the middle is table stakes. The bottleneck moved from the lab to the surface. The companies that notice first will hire accordingly.

Which raises the question I keep turning over: if the interface is now the bottleneck, who inside your company has the influence to fix it, and would a stranger be able to tell from your product?

If this issue was useful, share it with a designer or founder who is deciding what to build next. And if someone forwarded it to you: germandleono.substack.com.

Sources referenced this issue

• Simon Willison, “The new GPT-5.6 family: Luna, Terra, Sol”

• OpenAI, “ChatGPT is now a partner for your most ambitious work”

• TechRadar, ChatGPT app backlash

• Daring Fireball on the ChatGPT Mac app and on Benedict Evans’s critique

• Spyglass, “The ChatGPT Super App Sort of Super Sucks”

• Gibson Dunn, EU AI Act Omnibus Agreement

• arXiv, “Design Principles for Human-Agent Interaction”

• arXiv, “After the Interface”

• Figma release notes

• Gravity, AI agent funding tracker Q3 2026

A few weeks ago, I watched a student do something that would once have taken a small team a full sprint. It took her an afternoon. She described the tool she wanted, talked to a coding agent for a while, corrected it twice, and by the end of the session, she had a working interface that fetched real data and behaved like a product. Not a clickable mockup. A thing that actually ran. She seemed almost embarrassed by how fast it had happened, as if the speed itself were evidence of cheating.

The part that stayed with me was not the speed. It was that the screen she had drawn, the interface, felt like the least important thing in the room. The hard, valuable work had moved somewhere else, and none of us in the room could quite name where.

I have been chewing on that question while reading two writers who, on the surface, have almost nothing to do with each other. One writes about markets. The other writes about applied AI research. This fortnight, they happened to describe the same shift, from opposite ends, and the shift is one that anyone who designs and builds for a living should be paying close attention to.

The market read: the interface stops being the moat

Linas Beliūnas writes a newsletter about where finance meets AI, and he spent a recent issue picking apart a strange event: a batch of free text files wiping roughly $285 billion off software stocks in four trading days. The files were a set of open-source plugins, mostly structured prompts and configuration, that let a general-purpose model do work that specialized software used to charge a fortune for. The legal plugin that triggered Thomson Reuters’ worst day on record was, in his telling, around 2,500 lines of prompt instructions.

His point is not that software is dead. It is sharper than that:

“For two decades, SaaS bundled three things: data, logic, and interface. The interface was the moat. MCP Apps unbundles the interface. The app becomes a backend.”

That sentence stopped me. The interface as a moat is the world I have spent my career inside. It is the thing designers were hired to defend. If you can drag an Asana task or edit a Figma frame without ever leaving a chat window, then the carefully designed surface that used to hold the customer relationship becomes plumbing. Linas thinks Wall Street is pricing the death of SaaS when it should be pricing the death of the middleman. I think he is half right in a way that matters for us specifically: the layer most at risk is the layer many designers have treated as the whole job.

The research read: the AI becomes your reader, not just your tool

Ethan Mollick comes at it from inside the work. In his latest piece, he announces a new book, and almost in passing, he describes building its website by prompting a coding agent for a few minutes instead of fighting templates for hours. That alone is the student in my studio, again. But his real observation is about what the relationship with AI is turning into:

“When the AI is no longer just your assistant, but your reader, your critic, and the gatekeeper standing between your work and its audience.”

He even confesses to cutting his em-dashes (a detail I enjoyed, given my own allergy to them) because they now read as a human tell. The deeper move is that he stopped trying to please readers alone and started designing a version of his book’s site for the AIs that will decide whether to recommend it. The interface is no longer the place where a human meets your work. It is increasingly the place where a machine evaluates it on a human’s behalf.

Mollick’s framing for all of this is co-existence rather than co-intelligence. The tool is sometimes better than you and sometimes hilariously worse, and the work is no longer a single problem you solve once:

“It is a relationship you negotiate, and re-negotiate, as the models get ever better.”

Where the two readings meet

Put the two side by side, and they are describing one thing from two angles. Linas sees the interface being unbundled from the outside as a market repricing. Mollick sees it from the inside, as a working writer discovering that the surface of his work is now negotiated with a machine. The screen, the thing we drew, is losing its monopoly on the moment of contact.

This is the part where the industry reaches for its tired binary: AI either makes designers irrelevant or makes them gods. As usual, that is the wrong question. The honest reading is more uncomfortable and more interesting. A specific kind of design work, the execution of surfaces, is being automated and unbundled fast. And a different kind of design work is becoming the only thing left worth paying for.

Here is the perspective I have landed on, and I hold it with some conviction. If the interface was your moat, you are in trouble. If the interface was your craft but never your point, you are about to matter more than you have in years.

Because once you unbundle the surface, what remains is exactly the work I keep arguing the design engineer is built for. Not drawing the screen, but designing the moment a person decides whether to trust the thing on the other side. When to hand the keys to an agent and when to refuse its help, even when it is offering. What a system has to reveal about itself in the first thirty seconds before anyone risks relying on it. How trust accumulates over six months, and what quietly destroys it. None of that is surface. All of it is design, and almost none of it can be expressed in 2,500 lines of prompt.

The student in my studio built a running interface in an afternoon. What she could not generate in an afternoon was the judgment about whether the thing should behave that way at all, what it owed the person using it, and how it should fail when it inevitably did. That judgment is the moat now. It always quietly was. The unbundling just made it impossible to keep pretending the screen was the valuable part.

So the question I am sitting with, and the one I would put to you: if the interface is becoming a backend and the AI is becoming the reader, what is the smallest, most human thing in your work that a model cannot negotiate away? Find that. Build outward from it. I suspect it is closer to the center of the craft than most of us admitted when the surface was still the moat.

Sources

• Linas Beliūnas, Software Ate the World. Is AI Eating Software Now? — https://linas.substack.com/p/aieatingsoftware (and the companion piece on MCP Apps, AI Killed the UI — https://linas.substack.com/p/aikilledui)

• Ethan Mollick, Co-Existence and the End of Co-Intelligence — https://www.oneusefulthing.org/p/co-existence-and-the-end-of-co-intelligence (with context from The Shape of the Thing — https://www.oneusefulthing.org/p/the-shape-of-the-thing)

In love with AI, issue for the week of July 9, 2026

Two weeks ago Figma sold out the Moscone Center and announced the biggest change to its product in a decade. This week, three frontier models launched on the same day. In between, a report landed that quietly documented the moment designers stopped being just designers. This issue connects those three events, because together they describe a single fight: the fight over where software gets shaped.

The most important story: Figma puts code on the canvas, while the market prices its extinction

At Config 2026 (June 24, Moscone Center, in-person tickets sold out), Figma announced code layers: executable React code living directly on the design canvas. Teams can clone a repository, extract flows from code into design layers, edit them in a code composer or through AI chat, and use npm packages, motion libraries, and 3D frameworks without leaving the file. Alongside it came native animation support, AI-generated shader fills, prompt-built custom plugins, and deeper integration with Weave, the node-based AI media tool Figma acquired last October.

The numbers underneath are the real story. Figma’s Q1 revenue grew 46% year over year to $333 million. Net dollar retention hit 139%, its highest in more than two years. And the stock is down roughly 79% since the July 2025 IPO, trading around $24. Customers are paying more than ever for a product the market believes is being routed around.

Why it matters: code layers are not a feature, they are a defensive architecture. If engineers and PMs prototype on the canvas next to designers, the canvas becomes hard to replace with a text prompt to a coding agent. Figma is betting the canvas absorbs code before code absorbs the canvas. For practitioners, the immediate implication is that “handoff” is now a choice, not a constraint. If your team still runs a mockup-to-ticket-to-build pipeline, you are following a process your tools no longer require.

Read: TNW on Config 2026 and TechCrunch’s coverage.

Product launches and interface innovations

Codex becomes an interface factory for non-developers. OpenAI expanded Codex into an enterprise work platform: Sites (hosted interactive web apps generated by agents), Annotations (in-place editing), and six role-specific plugins aggregating 62 business applications with 110 prebuilt skills. The striking number: non-developers are now 20% of Codex’s 5 million weekly users and adopting it three times faster than engineers. A finance lead can turn a spreadsheet into a live scenario planner with a prompt. That is interface production without a designer or an engineer in the loop, and it is exactly the pressure Figma’s code layers respond to. Read: VentureBeat.

Claude Design gets its control update. Anthropic’s design tool, launched in April, shipped an update aimed at its biggest criticism: too much vibe, not enough control. It now adheres better to existing design systems, offers finer editing controls, and does more with fewer tokens. The pattern to watch is convergence. Prompt-first tools are adding precision, canvas-first tools are adding prompts, and both are converging on the same product from opposite directions. Read: Fast Company. (Disclosure for transparency: this newsletter is drafted with Claude, so read that item with the appropriate grain of salt.)

Three frontier models in one day. On July 9, OpenAI’s GPT-5.6 family (Sol, Terra, Luna) and xAI’s Grok 4.5 went public on the same day, a first. Sol runs on Cerebras wafer-scale hardware at up to 750 tokens per second, roughly 15x typical GPU serving speeds. For interface people, the speed number matters more than the benchmark numbers: at 750 tokens per second, generation stops feeling like waiting and starts feeling like direct manipulation. Latency was the last excuse for keeping AI in a chat sidebar. Read: BuildFastWithAI’s roundup.

The report worth your hour: AI in Design 2026

Designer Fund and Foundation Capital surveyed 900+ designers across 60+ countries, with case studies from the design teams at Anthropic, Framer, Linear, Notion, Shopify, Sierra, and Stripe. The findings, briefly, and why each one matters:

Weekly AI use jumped from 54% to 91% in one year, and 75% of designers now use it daily. Adoption is over as a topic. The interesting questions are now about depth, not reach.

Half of all designers surveyed, not just design engineers, have shipped AI-generated code to production. The “should designers code” debate ended not with an argument but with a shrug.

The average toolstack went from 3 AI tools to 7, and that excludes internal tools. There is no standard workflow anymore. Teams are stitching personal stacks together, which means your process is now a design decision in itself.

The collaboration warning sign: designers reporting that AI decreased collaboration quadrupled, from 5% to 20%. More time in terminals and prompts, less time in live back-and-forth. Work is faster, connective tissue is thinner. Design leaders should treat this as the report’s most actionable finding.

Hiring: 50% of leaders now weight AI fluency in hiring, followed by systems thinking. Only 5% care less about execution quality. Taste got more valuable, not less, precisely because “good enough” is now free.

And the culture finding I keep thinking about: designers at companies with a tinkering culture are twice as likely to feel more creative and capable, even though they face a higher bar. Peer learning more than tripled (24% to 80%) while learning from leadership halved. This shift is bottom-up, and the leaders pulling ahead are the ones who accept that.

Read: the full report and Designer Fund’s summary.

The UX perspective: NN/g calls the commodity

Nielsen Norman Group’s State of UX 2026 argues that UI is becoming a commodity: cheaper to produce, standardized through tokens and systems, and increasingly mediated by AI layers that users delegate to rather than navigate. Their advice is to design deeper, treating UX as strategic problem solving rather than deliverable production. Put next to the Designer Fund data, the two reports agree from opposite ends: the report from the research world and the report from the venture world both conclude that the pixel layer is table stakes and judgment is the differentiator. Read: NN/g.

Jobs and the bifurcated market

The design job market is splitting, not shrinking. Google Cloud eliminated more than 100 design and UX roles while redirecting resources to AI infrastructure (TechCrunch’s running list). At the same time, design job postings across Designer Fund’s portfolio were up roughly 60% year over year, 56% of hiring managers report growing demand for senior designers against 25% for junior roles, and workers with AI skills command a 56% wage premium. Entry-level is being automated while senior judgment is being bid up. If you lead a team, this is your succession problem: the junior roles that used to produce your future seniors are the ones disappearing.

Research and HCI

ACM CUI 2026 runs July 21-24 in Bremen under the theme “Conversational AI: Agency and Identities,” a timely frame given that agency is exactly what this cycle’s product news keeps renegotiating. The CHI 2026 preprint collection is also worth a browse, with “Interpretative Interfaces: Designing for AI-Mediated Reading Practices and the Knowledge Commons” standing out for anyone designing reading and research tools. Browse: CHI’26 preprints.

Funding notes

Capital is moving to the layer under the interface. Together AI raised $800 million at an $8.3 billion valuation, and Baseten’s $1.5 billion Series F led a stretch where inference infrastructure out-raised frontier model building (Intellizence). For founders building AI products, cheap fast inference is becoming a utility. The defensibility question moves up the stack, to exactly the workflow and interface layer this issue is about.

Essay: The most expensive quarter in design software history

On June 24, Yuhki Yamashita stood on stage at the Moscone Center in front of a sold-out crowd and showed a designer dragging a working React component onto a Figma canvas. The audience applauded. Twenty blocks away, in any brokerage app, Figma’s stock chart told a different story: down 79% since the IPO eleven months earlier, trading around $24. That same morning, the company’s investor page still showed the best operating quarter in its history. Revenue up 46%. Retention at 139%, a two-year high.

Hold those two facts together, because almost nobody does. The customers who know Figma best are paying more for it than ever. The investors judging its future have cut its value by four fifths. Someone is wrong, and the answer matters well beyond one stock.

The bear case is easy to state. Codex now builds hosted enterprise interfaces from a prompt, and its fastest growing users are people who have never opened a design tool. Claude Design turns a sentence into a prototype. Google shipped Pics inside Workspace. If interfaces can be generated wherever the intent lives, why would anyone route through a canvas? In this reading, Figma is Kodak with better retention metrics, and the 139% is just enterprises deepening their commitment to a workflow that is about to be bypassed.

The bull case is what Config actually showed. Code layers put executable code on the multiplayer canvas, next to the mockups, the comments, and the arguments. Yamashita’s framing was careful and, I think, correct: the value is that teams exploring ideas do not have to care about code quality yet. That sentence contains a real insight about how software gets made. Production code is a commitment. Exploration needs a space where commitments have not happened yet, where three directions can sit side by side and be wrong cheaply. For a decade that space was the artboard, and its price was a handoff: everything drawn there had to be rebuilt to become real. Code layers try to remove the price while keeping the space.

So the actual question underneath the stock chart is not “will AI generate UI?” It obviously will. It already does. The question is where the deciding happens. Prompt-first tools collapse deciding and building into one gesture, which is exhilarating for an individual and chaotic for a team. Canvas-first tools keep a shared surface where a team can see the options before committing, which is essential for a team and mostly overhead for an individual. Codex is betting that work is becoming individual, that a PM with an agent doesn’t need the shared surface. Figma is betting that software is still a team sport, that the surface where a team negotiates what to build is the most defensible real estate in the stack.

The evidence from the field cuts both ways, and this is where the AI in Design 2026 report earns its place next to the product news. Half of designers have shipped AI-generated code to production: the individual empowerment story, point to Codex. But the same report found that designers reporting decreased collaboration quadrupled in a year, and several described drifting into solo work, more time in terminals, less time in live back-and-forth, moving faster while the connective tissue thins. That is what a world without the shared surface feels like from the inside, and the people living it are not describing it as progress. They are describing it as a trade they are not sure they meant to make.

My read, and it is a read rather than a certainty: the market is pricing the wrong risk. The risk to Figma was never that AI generates interfaces. It is that teams stop deciding together, and I see no evidence of that in how actual product organizations behave. The 139% retention is not inertia. It is teams telling you, with budget, that the shared surface still matters. What the 79% decline correctly prices is something narrower: the pixel-production business is dead, and any revenue that depended on designers being the only people who could draw software is gone. Figma’s future rests entirely on being the negotiation room, not the drawing tool. Config was the company saying it knows.

For designers, the practical conclusion is uncomfortable but clear. If your value was producing the artifact, both futures are bad for you, because the artifact is now free in both. If your value is the quality of the decisions the artifact represents, both futures need you, and the canvas-with-code world needs you at the center of it. The NN/g advice to design deeper and the Designer Fund hiring data (AI fluency and systems thinking up, narrow specialization down, execution standards unchanged) are the same sentence in two dialects.

For founders, the lesson generalizes past design tools. When AI makes your product’s output free, your product either becomes the place where people decide about that output, or it becomes a feature of whatever place does. Retention among people who know you best is the signal to trust while the market panics about the output layer.

Which brings me back to the applause at Moscone. The crowd was not applauding a feature. They were applauding a claim about their own future: that the room where software gets decided still exists, and they are still in it. Eleven months of stock chart say the market disagrees. Someone is wrong. Where do you think the deciding will live in three years, on a shared canvas, or in five thousand private conversations with an agent?

Newsletter exclusive, not in the LinkedIn version: One number from the report never made it into anyone’s coverage, and it is the one I would build a team strategy on. Only 28% of companies have formally updated evaluation, compensation, or hiring for the AI era, while 73% of designers say expectations for their output have already risen. That gap is a five-year salary arbitrage, in both directions. Designers who can document their AI-augmented impact are being reviewed against criteria written for a job that no longer exists, which means the ones who force the conversation early will reset their level before the rubric catches up. And leaders who close the gap first will hire the people leaving everyone who didn’t. The next reorg is sitting in that 45-point spread.

If this issue was useful, share it with one designer or founder who is deciding what to bet on next. In love with AI is written for the people building at the intersection of AI, design, and product. See you next week.

Last week, I watched one of our students describe an idea out loud, hand it to a coding agent, and walk to get a coffee. By the time she came back, the thing existed. Not a mockup, not a Figma flow with hotspots taped over it, but a working prototype she could click, break, and argue with. She looked at it for a second and said, almost disappointed, “That was too easy.”

I have spent years arguing that the most interesting professional of this decade is the one who can close the gap between imagining something and actually making it. That was the whole premise. The handoff between the designer who dreams and the engineer who builds is where intention goes to die, and the person who can hold both sides at once is rare and valuable.

What I did not fully reckon with is what happens when the gap closes on its own. When the building part, the thing that used to take the skill, becomes the part you can walk away from while the coffee brews.

Two writers I follow have been circling this from opposite ends, and reading them together this month rearranged something for me.

From working with AI to managing it

Ethan Mollick put a name to the shift. In The Shape of the Thing, he argues we have crossed out of the era of prompting AI back and forth and into something else entirely.

“Starting in late 2025, we entered a new era thanks to AI agents... These are AI systems that you can just give work to, sometimes hours of human work, and get back reasonable and useful results in minutes. This is an era of managing AIs, rather than working with them.”

His example is almost comically blunt. A three-person team at StrongDM built what they call a Software Factory, running on two rules: code must not be written by humans, and code must not be reviewed by humans. The agents write it, other agents test it inside a simulated customer environment they build for themselves, and the humans only see the finished product before it ships.

Read that as a designer, and you feel the floor move. The part of my job that was hardest to fake, the actual construction, the thing that separated people who could build from people who could only describe, is being quietly automated. The handoff I spent my career trying to eliminate is being eliminated, just not by me.

The leverage is real, the floor has risen

Linas Beliūnas is looking at the same shift through the lens of money and startups, and he is more excited about it. In The One-Person Unicorn, he describes building an entire startup operating system out of Claude skills, twelve interconnected playbooks covering everything from idea validation to managing a board, then running a fictional company through the whole thing. His conclusion is the line I keep returning to:

“AI outputs are only as good as the architecture and data behind them. Change the architecture, and the outputs change completely.”

In a companion piece, What to Build in 2026, he reads the venture capital consensus and finds every major firm saying the same thing: the language has shifted from augment to replace. The pitch is no longer software that helps the lawyer. It is software that is the lawyer, for a well-defined scope. He calls the model Service-as-Software, and notes it sells into the services budget, which dwarfs the software budget by an order of magnitude.

Put Linas next to Mollick, and you get the full picture. Mollick describes the capability, you can now hand off hours of work. Linas describes the consequence: the minimum viable team is shrinking, and the floor of what one person can produce has risen dramatically.

So what is actually scarce now?

Here is where I land, and it is not where I expected.

If building is no longer the bottleneck, then building was never the point. It was a proxy. We valued the person who could build because building was hard and expensive, and the ability to do it was a reasonable signal that someone had taste, judgment, and the stubbornness to see an idea through friction. The skill was a stand-in for something underneath it.

The agents take the skill and leave something underneath exposed. Linas says it plainly: the output is only as good as the architecture behind it, and architecture is just another word for decisions made on purpose. Mollick, in his other recent piece, Choosing to Stay Human, warns about the opposite, what his Wharton colleagues call “cognitive surrender,” the very human habit of taking the AI’s confident answer and stopping there.

“Agentic systems are designed to make your life easier, because they just do stuff. Which is great for getting stuff done, bad for learning anything, or staying authentic, or avoiding cognitive surrender.”

This is the design engineer’s specific risk, and it is not the risk most people are worried about. The danger is not that the agent cannot build. It is that it builds so smoothly, you stop noticing what it decides on your behalf. Every default it picks is a design decision. The empty state it generated, the error message it wrote, and the way it handled the user who does not behave the way the happy path expects. The agent has opinions about all of these, and if you are getting coffee, they become your product’s opinions by accident.

The student who said “that was too easy” was right to be uneasy. The work did not disappear. It moved. It moved from typing the thing into deciding whether the thing it typed is any good, for whom, and at what cost to the person on the other side of the screen.

That decision is not faster than it used to be. If anything it is slower, because nothing in the tool forces you to make it. The friction that used to be in the building, where you could not avoid confronting your choices, has to be put back deliberately, somewhere else, by someone who knows it matters.

I think that is the actual job now. Not closing the gap between imagining and building, the machines closed it. The job is keeping the human decision in the loop after the gap is gone, choosing where to slow down on purpose in a process designed to never let you. The leverage Linas describes is real, but it amplifies whatever taste you bring, including the absence of it. Ten times the output is ten times the quality or ten times the slop, and the tool is genuinely indifferent to which.

The design engineers I would hire tomorrow are not the ones who prompt the fastest. They are the ones who look at what the agent produced in ninety seconds and can tell you, specifically, what is wrong with it.

So here is what I am still chewing on, and I would like to know what you think. If building is free, the scarce thing is knowing what is worth building and what “good” actually means for the person using it. But that judgment used to be earned the slow way, by building badly for years until you developed a feel for it. If the next generation never builds badly, because the agent never lets them, where does the taste come from?

Sources

• Ethan Mollick, The Shape of the Thing (One Useful Thing, March 12, 2026): https://www.oneusefulthing.org/p/the-shape-of-the-thing

• Ethan Mollick, Choosing to Stay Human (One Useful Thing, May 26, 2026): https://www.oneusefulthing.org/p/choosing-to-stay-human

• Linas Beliūnas, The One-Person Unicorn (Linas’s Newsletter, February 2, 2026): https://linas.substack.com/p/onepersonunicorn

• Linas Beliūnas, What to Build in 2026 (Linas’s Newsletter, February 12, 2026): https://linas.substack.com/p/what-to-build-in-2026

Last week, I watched a graduate student use an AI assistant to complete a research exercise. They typed a question. The AI answered. They copied the answer into their document. The process took about forty seconds. It looked exactly like productivity.

It probably wasn’t.

Ethan Mollick, in his recent One Useful Thing post “Choosing to Stay Human,” describes what his research team at Wharton calls “cognitive surrender” — the pattern where people stop thinking about problems and simply let the AI do the work, even when the AI is wrong. In a study of 758 consultants at Boston Consulting Group, half given access to GPT-4, the results were striking: those with AI vastly outperformed those without it. But when asked to solve a problem the AI was known to fail at, the AI users were significantly less likely to get it right. The same elite consultants who had just outperformed everything stopped using their own judgment the moment a confident-looking answer appeared on screen.

“The AI gave them an authoritative-looking answer that happened to be incorrect, and most of them — the same elite consultants who outperformed on everything else — did not catch it.”

Mollick frames this as a challenge of individual intention — a call to be deliberate about which tasks to hand over and which to keep human. That framing is right and important. But it’s incomplete. Because what I keep coming back to is this: someone designed those consultants’ experience with AI. Someone chose the interaction model. Someone decided how the answer would be presented — with authority, without friction, in a way that made the output feel like a conclusion rather than a starting point.

That someone was a design engineer. And that choice had consequences.

The Architecture Problem

A few months ago, Linas Beliūnas published an experiment in his newsletter that points in a completely different direction. Linas built what he calls a startup operating system in Claude — twelve interconnected skills covering everything from idea validation to board management, each encoding real frameworks, real case studies, and real decision trees from founders who had actually done the work.

The result was outputs that were, by his account, “specific, actionable, and often better than what I would have gotten from expensive advisors.” Not because the underlying model was smarter, but because the architecture was deliberately designed to produce depth instead of surface.

“The problem isn’t that AI is stupid. The problem is that surface-level prompting produces surface-level outputs. The solution is architecture.”

This is the sentence I keep returning to. What Linas is describing, though he frames it as a founder’s problem, is actually a design problem. The quality of the output was a function of the architecture surrounding the model, not just the model itself. Each skill wasn’t just a document — it included diagnostic workflows, decision trees, frameworks from practitioners, and anti-patterns that prevented common mistakes. The system interrogated the user before it answered them.

Most AI products are not architected this way. They are designed for frictionlessness, which sounds like a virtue but functions as a trap. An interface that produces an answer in forty seconds is optimized to feel useful. Whether it produces understanding, judgment, or anything that survives contact with a situation the AI hasn’t seen — that’s a different question, and it’s usually not part of the design brief.

What We’re Actually Responsible For

The reason this matters specifically for people working at the intersection of design and engineering is that we are the ones setting these defaults. When we build an AI product, we make hundreds of small architectural choices that collectively determine what happens inside the person using it.

Do outputs arrive with friction or without it? Does the interface invite the user to engage critically, or hand over? Is there any moment where the system says, in effect, wait — do you actually understand this? Or does it just give the answer and let people feel productive?

Mollick notes that the three major AI companies do offer tutoring modes that push users to think rather than receive — but they’re not intuitive to reach. Gemini’s is a tap and a menu. ChatGPT needs you to type /learn. Claude’s is buried in style settings. The default, on every platform, is the answer-delivery mode. That is not an accident. It’s the accumulated outcome of design decisions optimized for engagement metrics and time-to-first-value. The tutoring modes exist; they just lost the internal argument every time.

This is the gap that design engineers should be losing sleep over. Not “how do I make this interface more elegant,” but “what does the architecture of this interaction teach the person using it?” Because the architecture is the answer to that question — not the model, not the prompt, not the user’s individual willpower.

The same research Mollick cites makes the flip side visible. A five-month AI tutoring experiment across ten high schools in Taipei — close to a thousand students — found that a personalized AI that pushed students to solve problems themselves produced gains equivalent to six to nine months of additional schooling, without any added teacher workload. The model was not dramatically better. The architecture was completely different. One system answered. The other one asked.

Designing for the Human Side

The deeper challenge is that we don’t have clear norms yet. As Mollick puts it:“We clearly don’t have rules or role models for how AI gets used at work, in schools, or in government. That’s a problem, but it also means that every organization figuring out a good way to use AI right now is setting a precedent for everyone else.”

The window is open. It won’t stay open indefinitely.

What Linas demonstrated with his startup OS is that the intervention doesn’t have to be subtle or academic. A well-designed architecture can actively produce better thinking — not by lecturing users about their cognitive habits, but by giving the model enough structured context that it asks better questions back. The skills framework he built doesn’t just produce better outputs. It produces outputs that help founders understand why they’re making the decisions they’re making. The system interrogates before it advises.

That’s the design move. Not “give people the answer faster.” Not even “explain the answer while you give it.” Something closer to: design the interaction so that the human who comes out the other side knows something they didn’t know before, and couldn’t just as easily have skipped.

There’s a version of every AI product we’re building where users emerge with sharper judgment. There’s another version where they emerge having outsourced their thinking and now can’t get it back. The difference between those two products is not the model. It’s the architecture. It’s us.

The student I watched last week probably got their homework done on time. The section they filled in was fine. What I still don’t know is whether they learned anything — and whether the person who designed that AI interface ever thought to ask.

That might be the most important design question of the next five years. I’m curious whether the people building these products are treating it that way. Are you?

Sources

• Ethan Mollick, “Choosing to Stay Human”, One Useful Thing (May 26, 2026)

• Linas Beliūnas, “The One-Person Unicorn”, Linas’s Newsletter (Feb 2, 2026)

A few days ago I watched a student brief an agent, walk away, and come back to a finished product. Not a mockup. Not a clickable prototype held together with placeholder text. A working thing, with its own logic, its own data, its own small thousand decisions already made. She had written three paragraphs of instruction and pressed go. The look on her face when she returned was the look I keep seeing this year on people who build for a living: somewhere between delight and unease. Delight because she asked for something and it appeared. Unease because she asked for something and it appeared.

It would be easy to treat that as an anecdote. It is not. The numbers say the floor moved under all of us at once. The AI in Design 2026 report, which surveyed more than 900 designers across 60-plus countries with interviews at Anthropic, Stripe, Linear, Shopify, and Notion, found that 91% of designers now use AI, up from 54% a year earlier, three in four every single day. The detail that stopped me was this: half of them have shipped AI-generated code to production. The handoff that used to separate “designer” from “engineer,” the white space where intentions went to die, is quietly closing.

So what is actually happening to the work? The most honest description I have read comes from Ethan Mollick, writing about early access to Claude Fable. He commissions an isochrone map, a genuinely hard cartographic problem, and watches the model spin up its own sub-agents to gather two thousand flight times and rail schedules, write the code, and test its own output while he sits back.

“Last year I called this working with a wizard: you chant the spell and something happens. With Fable the spell has gotten powerful enough that I am no longer sure I am the wizard. I am closer to a patron. I describe what I want, I pay for it, and I judge the result.”

The work, he says, has shifted from process to outcome. “I no longer steer; I commission.” One project ran nine and a half hours and produced research software that was never profitable to build before. He could spot a few errors as an expert. He had no visibility into the hundreds of choices that produced them.

Now hold that next to where the engine itself is heading. Linas Beliūnas, who writes about the place where finance meets AI, spent this week on GLM-5.2, an open-weight model you can download for free and, thanks to new quantization, run on a single Mac instead of a data center. One builder called it “the first open model that passes the bar as a daily driver,” and Fast AI’s Jeremy Howard judged it “at least as good as Opus 4.8 and GPT-5.5.” Same week, Linas noted that brokers are wiring Claude and ChatGPT straight into customer portfolios. The engine is getting better, cheaper, and more embedded all at once. Soon everyone will have roughly the same one, and some of it will be running on the laptop in front of you.

Put those together and the economic logic is not subtle. Dan Shipper has a name for it: the allocation economy. As AI floods the system with cheap competence, he argues, the value of the people who can judge, review, and direct that output rises with it. When making becomes abundant, the scarce thing is taste. The verdict. The ability to look at a nine-hour build and know it is subtly wrong in the third decimal place.

This is the part the “AI will replace designers” conversation keeps getting backwards. The execution-heavy middle of our work, translating an intention into components and code, is exactly the part the model is eating. That was never the valuable part. It was just the part that took the most hours. What grows in value as the middle collapses is everything on either side of it: framing the problem precisely enough that an agent can act on it, and judging the result rigorously enough that a human can stand behind it. The brief and the verdict. Both are design problems before they are engineering problems, because both are fundamentally about humans.

Which raises the question almost nobody has answered well: where does that judgment actually happen? Right now the default answer is a chat box, and the people who have thought hardest about interfaces are nearly unanimous that this is a dead end. Amelia Wattenberger put it bluntly years ago, that a text field gives users “unclear affordances,” and she has been building alternatives ever since, editors that visualize an AI’s output along axes you can actually see and steer. Maggie Appleton makes the complementary case: most language-model interfaces should be “spell-check sized,” scoped to do one thing well and embedded where you already work, not another blank prompt demanding you imagine everything yourself. Linus Lee wants to replace the prompt entirely with real affordances, pinch-to-zoom and drag-and-drop for the latent space, so that working with a model feels like manipulating a material rather than pleading with an oracle.

Notice what all three are really describing. They are designing the surface where a human meets the black box and decides whether to trust it. That is the verdict, made tractable. And it has a concrete engineering shape, too. Simon Willison points out that AI tooling is moving from advisory (”the model suggested code”) to operational (”the model edited the file”), and the moment that happens, the real design work becomes process: permissions, review, rollback, testing. The interface to judgment is not a nicer chat window. It is the structure that lets a person take responsibility for something they did not watch get made.

This is why I think the design engineer, the role Vercel and Linear and Replit are now hiring for by name, is not a fad title. Rauno Freiberg describes the job as prototyping in code because the prototype becomes the production thing, which is only possible when design and build are the same act. Make the engine abundant and that fusion stops being a luxury. It becomes the only way to design the brief and the verdict at the speed the model now works. The designer who can only draw the picture is handing the most consequential decisions to whoever writes the code. The one who can build is keeping them.

So here is where I have landed, and where I am still stuck. The optimistic read is that our loss of control is temporary, an artifact of interfaces that have not caught up, and that people like Wattenberger, Appleton, and Lee are sketching the windows we will soon have. The darker read, which Mollick leans toward, is that the more capable the model, the less there is for a human to meaningfully do, and the black box is simply the price of the power. I do not think we get to find out passively. The interface to judgment is something we have to design on purpose, and almost nobody is being paid to do it yet.

So I will leave you with the question I cannot shake. When we sign off on work we did not watch get made, what exactly are we taking responsibility for, and have we built anything that lets us take it honestly? I do not have a clean answer. I would like to hear yours.

Sources

• AI in Design 2026, Designer Fund with Foundation Capital. https://designerfund.com/blog/ai-in-design-2026

• Ethan Mollick, “What it feels like to work with Mythos,” One Useful Thing, June 9, 2026.

One Useful Thing

What it feels like to work with Mythos

I had early access to the first Mythos-class AI model being released to the public, Claude 5 Fable. Much of the discussion of Mythos has centered on its impact on software security, but I tested it on everything except that (the guardrails around Fable essentially prevent it from being used for cybersecurity at all). My conclusion is that it represents …

Read more

4 months ago · 2055 likes · 199 comments · Ethan Mollick

• Linas Beliūnas, “GLM-5.2: The ChatGPT Moment for Local AI,” Linas’s Newsletter, June 22, 2026.

Linas's Newsletter

GLM-5.2: The ChatGPT Moment for Local AI 🤖

👋 Hey, Linas here! Every day, I break down 3 stories shaping the future of FinTech & Artificial Intelligence - plus the money movements and trends worth tracking. First time here? 395k+ FinTech and AI leaders get this daily. Join them…

Read more

4 months ago · 73 likes · Linas Beliūnas

• Dan Shipper, “The Knowledge Economy Is Over. Welcome to the Allocation Economy,” Every. https://every.to/chain-of-thought/the-knowledge-economy-is-over-welcome-to-the-allocation-economy

• Amelia Wattenberger, “Why Chatbots Are Not the Future.” https://wattenberger.com/thoughts/boo-chatbots

• Maggie Appleton, “Squish Meets Structure.” https://maggieappleton.com/squish-structure

• Linus Lee, thesephist.com.

https://thesephist.com/

• Simon Willison, simonwillison.net.

https://simonwillison.net/

• Vercel, “Design Engineering at Vercel” (Rauno Freiberg). https://vercel.com/blog/design-engineering-at-vercel

Last week I was prototyping a feature with an AI coding agent. Not prompting it back and forth, actually giving it a brief and walking away. It came back with something functional in about eight minutes. I read through what it had built, and something felt off. Not wrong, exactly. Just flattened. The logic was there but the intent wasn’t. The agent had interpreted my instructions accurately and missed the point of them entirely.

I sat with that for a while. The gap wasn’t in the code. It was in my brief.

That’s the thing nobody is talking about clearly enough yet.

Linas Beliūnas, who writes the sharpest finance-meets-AI newsletter I know of, put it this way in a December piece that has been bouncing around my head since I first read it:

“Generation is now cheap. Iteration is fast. Output is abundant. What’s scarce is judgment. Knowing what to ask. Knowing how to constrain. Knowing when to trust, and when not to.”

He was writing about Andrej Karpathy, OpenAI co-founder, former Tesla AI lead, publicly admitting he felt behind. Not confused about AI. Not skeptical of it. Behind. The specific word choice matters. Being behind implies that the race changed without anyone announcing it. The skills that made Karpathy exceptional in one paradigm don’t fully transfer to the next.

Linas’s diagnosis: “Programming didn’t get faster. It changed shape.” You no longer author every instruction. You direct a system that generates, reasons, and acts probabilistically. Prompts replace functions. Agents replace modules. Verification replaces execution.

If that’s true for engineers at the frontier of building AI, it’s even more true for design engineers. We’ve always lived at the intersection of intent and implementation. That position doesn’t disappear in an agentic world, it becomes the whole job.

Ethan Mollick, whose One Useful Thing newsletter is essential reading for anyone thinking carefully about how AI changes work, published something in late May that I keep returning to. The piece is about a concept his Wharton colleagues call “cognitive surrender” the pattern where people stop thinking about problems and just let AI do the work, even when the AI is wrong.

The research he cites is stark. In one experiment, elite consultants given access to GPT-4 outperformed their peers on most tasks, but on the task researchers designed for AI to fail at, the AI-assisted consultants were more likely to get the wrong answer than consultants without AI at all. The AI gave them a confident-looking response, and they stopped checking.

But Mollick is careful not to make this a cautionary tale about AI in general. A separate study showed that students with access to a properly-designed AI tutor scored the equivalent of six to nine additional months of schooling, without any added instruction time. Same technology. Completely different outcome. The difference was whether the AI was replacing the hard thinking or forcing more of it.

“A lot of the problem is going to come down to us... Balancing using AI with our own mental abilities is going to be a defining challenge of the coming years.”

For design engineers specifically, this maps onto something I’ve been thinking about since my botched prototype. The AI didn’t fail me. I failed the brief. And if I keep offloading the execution before I’ve done the hard work of knowing exactly what I want, I’ll get faster at producing outputs I’m vaguely unsatisfied with.

There’s a real-world signal in Linas’s most recent weekly roundup that I think is underread in design circles. Ramp, the corporate spend management company, just launched an AI agent product for accounting firms called Stack. The agents handle reconciliations, journal entries, and monthly closes autonomously, one firm got their March close down to about 20 minutes. But here’s the design detail that stopped me: nothing posts to QuickBooks without human sign-off, and firms describe their client processes in plain English, which Stack then encodes as reusable workflows.

The accounting firm doesn’t write the code. They write the brief. They become, essentially, unwitting design engineers, responsible for the quality of their intent description, and accountable for reviewing what the agent produces against it.

Linas frames the Ramp move as a distribution play, which it clearly is. But from a design lens, it’s also a preview of a professional transition that’s coming in every knowledge domain: the people who were executing will now be directing, whether they’re trained for it or not. The accountants who instinctively know how to write a clear, constrained brief for an AI agent will thrive. The ones who hand over ambiguous instructions and rubber-stamp the output will introduce errors their clients eventually notice.

This is what I mean when I say the brief is the product. The quality of the instruction is where differentiation lives now.

Mollick’s March piece, The Shape of the Thing, describes the current AI moment as a shift from “co-intelligence” , humans prompting AI back and forth, to managing AI agents that you can hand hours of work to and get results back in minutes. He calls it a “rolling disruption”: capabilities crossing thresholds unpredictably, unlocking radical new use cases sometimes overnight, with organizations that figure out a good way to work with AI now setting the precedent for everyone else.

The window to shape how these tools get used is open. It won’t stay open.

What I find interesting about where we are right now is that design engineers are uniquely positioned for this moment, and uniquely at risk. The position is valuable because design has always required holding intent and execution simultaneously, and that’s exactly the skill an agentic workflow demands. The risk is that the speed of execution makes it seductive to skip the hard part. Generate fast, ship fast, iterate. The brief gets looser. The judgment atrophies.

Linas puts the builder’s version of this bluntly: “Taste and judgment become moats. When anyone can generate code, differentiation comes from knowingwhat not to build, what good looks like, and when to stop.”

That’s not a description of a technical skill. It’s a description of a design sensibility. And it’s the thing that can’t be delegated to the agent.

I’ve been watching what the third cohort at our MHIAI program has been building toward their graduation show this month, and what strikes me is how much of the most interesting work is happening in the brief-writing layer. Not in the models, not in the interfaces — in the act of describing, precisely, what an intelligent system should do, for whom, and under what constraints. That’s where the real intellectual work is. That’s where trust gets built or lost.

If the accountant’s value proposition is shifting from executing reconciliations to writing clear briefs for agents that execute them, and if the same shift is happening in every professional domain that AI touches, then what design engineers have always known how to do becomes the most portable skill in the room.

The question I’m sitting with: if the brief is the product, how do we train people to write better briefs? And what do we lose if we let the AI write those too?

Sources:

• Linas Beliūnas, “The Day the AI Builders Felt Behind”, December 2025

• Linas Beliūnas, “Ramp Stack Gives Accountants Free AI Agents”, June 7, 2026

• Ethan Mollick, “Choosing to Stay Human”, May 26, 2026

• Ethan Mollick, “The Shape of the Thing”, March 12, 2026

June 16, 2026

I read Mollick in the morning and Linas at lunch, and what is sitting on my desk is the same observation arriving from two angles.

Linas’s beautiful new design stack is, in Mollick’s terms, exquisitely frictionless. It is built to remove every place where a designer used to have to stop and choose. The font is chosen for you. The component is chosen for you. The dashboard layout is chosen for you. The brand voice is generated alongside the brand mark, in twelve consistent variants, before you have made a single decision about who the product is for.

This is wonderful if you have already made all those decisions in your head and you are using the tool as a hand. It is catastrophic if you have not, because the tool will make those decisions for you, plausibly, and you will not notice. The output will not look wrong. It will look like a product. It will pass every recruiter scan, every investor first glance, every internal review. And six months later, when the user does not stick, nobody will be able to point at the specific decision that was wrong, because there was never a decision in the first place. There was only a default.

I have watched this happen four times in the last month, in critiques with my students. The work that comes back from a week of Claude Design and image gen is almost always more polished than what the same student produced six months ago by hand. It is also, often, more anonymous. Three projects in a row look like they were designed by the same invisible studio, because in some sense they were. The work has been beautifully executed by a system that has no idea what the student was actually trying to say.

This is the design version of Mollick’s meaning-shaped writing. We are makinginterface-shaped products that take mental effort to navigate and give the user no equivalent understanding of what the product is for or why it exists.

The design engineer, again

I have been making the case for a while that the most interesting professional in this moment is the person who can move fluently between the model, the interaction, and the build. I want to refine that argument now in light of these two readings.

The design engineer is not interesting because they can do what Claude Design does. Claude Design will keep getting better at that, and so will whatever comes after it. They are interesting because they are the person who decides where, in a stack that wants to be entirely frictionless, friction needs to be put back in on purpose.

That is a design skill, not an engineering skill, although it is impossible to exercise without engineering literacy. You cannot decide where to put friction back into an AI product if you do not understand what the model is doing, why it is wrong sometimes, where a user’s mental model will diverge from the system’s behaviour, and what trust feels like when it accumulates and when it collapses. The Taipei version of the AI tutor is a design decision dressed as a system prompt. The Turkey version is the default.

A specific working hypothesis I am sitting with: in the era of one-shot brand identities and one-prompt prototypes, the designer’s job is less about producing the artefact and more about deliberately authoring the moments where the user, the team, or the designer themselves are forced to stop and choose. The choreography of friction. The placement of the pause.

This is the work that does not get easier when the tools get better. It might be the only design work that survives the next eighteen months unchanged.

One thing I keep wanting to ask

If you build something with Claude Design and ChatGPT Images 2.0 and ship it in an afternoon, what part of it would you be able to defend in six months, when a user is angry, when an investor is skeptical, when a teammate inherits the file? Which decisions in the artefact are yours, and which ones did you accept without noticing?

I do not have a clean answer, and I am suspicious of anyone who claims to. But the question feels like a useful one to carry into the rest of the year.

What are you choosing to keep human in your own design work? Reply and tell me. I am collecting answers.

Sources

• Ethan Mollick, Choosing to Stay Human, One Useful Thing, 26 May 2026.

• Ethan Mollick, Sign of the future: GPT-5.5, One Useful Thing, 23 April 2026.

• Linas Beliūnas, Anthropic Just Told AI Founders Exactly What to Build in 2026, Linas’s Newsletter, 4 May 2026.

• Linas Beliūnas, Turn ChatGPT Images 2.0 & Claude Design Into Your Chief Designer, Linas’s Newsletter, 24 April 2026.

Three weeks ago, I was teaching a class in Barcelona, and a student raised her hand to ask a question that has been quietly eating designers alive for two years: “Why would I spend three years climbing into a senior UX role at a Fortune 500 when half of them are firing the designers they already have?”

It was the right question. But it was the wrong frame. Because the same week that question got asked, a 28-year-old designer with no engineering background was selling her solo design tool to Wix for eighty million dollars, and a designer-founded company called Figma was trading at a sixty-eight billion dollar valuation on the New York Stock Exchange.

The corporate door is narrower than it has ever been. The door to building your own thing has never been wider. We are in a renaissance, and most designers haven’t noticed yet because they are still looking at the wrong door.

The Layoff Data Looks Brutal. Because It Is.

Let me put the bad news on the table first, in numbers, so we are not pretending.

Tech layoffs in Q1 2026 hit 52,050 cuts, a 40% jump over the same period in 2025, and the highest first-quarter total since 2023. As of mid-year, 144,355 tech employees have been laid off, roughly 982 people per day. Since 2022, more than half a million tech workers have lost their jobs. Meta opened 2026 by cutting 1,500 from Reality Labs. Amazon has cut roughly 16,000 roles cumulatively this year. Oracle ran the single largest reduction event of the year in late March.

The design-specific numbers are no kinder. Across 5,521 analysed tech layoffs, UX roles accounted for 8.7% of the cuts, far above their share of total headcount. UX designer job postings fell 71% from their 2022 peak. One in three organisations cut UX staff in 2024, the worst year since Nielsen Norman Group started tracking the field. In October 2025, Google laid off more than 100 design and UX research roles from its Cloud division in a single week. 44% of hiring managers in a recent survey said AI is a top driver of the cuts.

If you are a senior UX designer at a Fortune 500 right now, the floor under your role is shaking. That is not a vibe. That is the data.

But the data has two sides, and the second side is the one nobody is putting on the same page as the first.

The Designer-Led Renaissance Is Happening in Plain Sight

While the Fortune 500 was cutting designers, designers were quietly building the companies that are eating those Fortune 500s.

Start with the headline. On July 31, 2025, Figma went public at $33 a share. By the end of day one, it was trading at $115.50, a 250% pop, valuing the company at roughly $68 billion. Q1 2025 revenue was $228 million, up 46% year over year, with $44.8 million in net income. The founder, Dylan Field, 33 years old, a college dropout, started the company on a Peter Thiel fellowship grant and is now worth around $6.6 billion. Figma is, at this point, the most successful designer-led company in the history of the internet.

It is not alone. Canva has 185 million users and $2 billion in revenue. Webflow has 4 million users and $314 million in revenue, combining visual design with production code. Both are designer-led. Both were “too niche” when they started. Both now sit on top of the workflows that the Fortune 500 design teams were paid six figures to perform manually.

That is the top of the market. The bottom of the market is, if anything, more telling.

Maor Shlomo built Base44, a solo product, to 250,000 users and profitability in six months, then sold it to Wix for $80 million in June 2025. Danny Postma’s HeadshotPro, run solo, generates $3.6 million in annual recurring revenue. Sarah Chen launched an AI-powered design agency in January 2025 using ChatGPT Plus, Canva Pro, and Zapier, and hit $420,000 in annual revenue within eight months, working 25 hours a week. Jason Zhou’s Superdesign is bootstrapped, designer-led, and generating UI mockups, components, and wireframes from natural-language prompts.

The structural numbers underneath these stories are the part that matters most. AI-augmented founders ship 8 to 12 features per month, versus 2 to 4 for non-AI founders. Cursor crossed 1 million monthly active users in early 2026. 34% of new micro-SaaS products launched in Q1 2026 were built by founders with no prior programming experience, several of them generating $5K to $50K in monthly recurring revenue. Solo-founded startups grew from 23.7% of all new companies in 2019 to 36.3% by mid-2025. There are now 41.8 million solopreneurs in the U.S., contributing $1.3 trillion to the economy annually.

This is not the data of an industry in decline. This is the data of an industry that has stopped needing the old gatekeepers.

What Actually Changed (And Why It Favours Designers)

Let me be precise here, because this matters.

For most of the last twenty years, a designer with a product idea had to assemble a team to ship it. A frontend engineer. A backend engineer. A devops person. A PM, eventually. A marketer. Each of those people cost money, slowed down decisions, and demanded equity. The cost of going from “I see the product clearly in my head” to “real users are using it” was, in practical terms, prohibitive for one person.

That cost just collapsed.

A designer in 2026 can prototype in Claude or Cursor, deploy with Vercel or Lovable, run payments through Stripe, do support with Intercom or a tuned model, generate marketing assets in Canva or Midjourney, write copy in Claude, and reach a million people through TikTok, X, LinkedIn, or Substack without buying an ad. The five-person team a designer needed in 2019 to ship a product is now the designer, alone, on a Tuesday afternoon.

The thing that was always the designer’s edge, taste, judgment, knowing what users actually need, has never stopped mattering. What changed is that the things that used to bottleneck the designer (engineering, deployment, distribution, ops) are now solvable with the tools the designer already lives in.

This is what the Fortune 500 layoffs are quietly telling us, and what almost nobody is reading correctly. The cuts are not happening because design doesn’t matter. They are happening because the work that used to require 50 designers in a corporate org now requires 3 designers and the right stack. The 47 designers who got cut do not have to go become baristas. They can go become founders. Many of them are.

Why I Am Calling This a Renaissance

The word “renaissance” gets used loosely. I want to use it precisely.

A renaissance occurs when the gap between what an individual can imagine and what they can produce collapses. The original Renaissance was, in large part, a story about tools: the printing press, perspective in painting, oil paints, and anatomy. Tools that let one person, sitting in a workshop, produce work that previously required institutional infrastructure. The artists did not become more talented in 1500 than they had been in 1400. The cost of expressing their talent simply fell through the floor.

That is exactly what has happened to designers in the last 36 months.

For the first time in the history of the field, a designer with taste, judgment, and the discipline to ship can:

• Build the prototype themselves, in code that runs

• Deploy it to real users in an afternoon

• Get feedback from those users the next morning

• Iterate the design and the system simultaneously

• Reach an audience of millions without a marketing department

• Charge those users money without an accounts team

• Run the whole operation, solo, for under $500/month in tooling

This was true for almost nobody in 2019. It is true for almost everyone with the discipline to try in 2026.

The designers who are still measuring their success by “did I get the senior IC promotion at the Fortune 500” are looking at the wrong yardstick. The renaissance is not happening inside the orgs that are cutting. It is happening at the kitchen tables of the designers who left them.

The Real Question Isn’t “Are Designers Being Replaced?”

The real question is: what kind of designer are you choosing to be?

• If you are a senior UX designer at a Fortune 500 with a stable role: enjoy it, but understand that the floor is shaking and the ladder above you has fewer rungs each quarter. Use the stability to learn the new stack. Don’t wait until your role is the one being cut to start.

• If you are a mid-career designer recently laid off, the data says you are not alone, and the data also says the door you should be looking at is not the next big tech job. It’s the product you have wanted to ship for three years and kept postponing. The cost of trying is now under $500.

• If you are a junior designer entering the field, the corporate ladder is the worst it has been in a decade. The solo-builder path is the best it has ever been. Pick accordingly. Do not optimise for the role the previous generation had. Optimise for the role that did not exist five years ago.

• If you are a design leader running a team: smaller teams are now structurally more productive than larger ones for almost everything below the largest enterprise scale. The teams of 30 you used to need are teams of 5 with the right tools. Plan accordingly.

• If you are a founder thinking about a first design hire: hire one designer who can also ship, not three designers who pass work between them. The designer who can prototype in code is now worth what three traditional designers used to be.

The future isn’t “every designer becomes a solo founder”. It is “the floor of what one designer can ship has risen by an order of magnitude, and the designers who internalise that early will be the ones building the next generation of products”. Some of them will be solo. Some will be at startups of three to ten. Some will be inside big companies that figured out how to give designers leverage instead of cutting them. All of them will be operating in a fundamentally different game than the one that ended in 2022.

Why This Matters Right Now

For most of the last fifteen years, the most important question a designer could ask was “how do I get into the right team at the right company?” The answer was a portfolio review, a system case study, and a willingness to wait two years for a senior promotion that would unlock the projects you actually wanted to work on.

That game is ending. Not because designers are losing power, but because the lever has moved.

A year ago: a designer with an idea wrote a spec, recruited an engineer friend, lost three months to scope creep, shipped a half-version of the product, ran out of energy, and never told anyone.

Today: a designer with an idea prototypes it in Claude over a weekend, ships it to a Vercel domain on Sunday night, posts a demo video on Monday, gets 10,000 views by Wednesday, has paying users by Friday, and is iterating on the system based on real user behaviour by the following Monday.

This is not a small shift. It’s a complete inversion of which side of the design profession holds the power.

If You Want To Stay Ahead

The designers who will look back on 2026 as the year their career compounded are not the ones waiting to see if the layoffs slow down. They are the ones who used the chaos to stop asking permission.

How do you ship a real product solo, with no engineering team, in 2026? How do you find your first 100 paying users without an ad budget? How do you build a design tool that 50,000 other designers want to use? How do you turn a portfolio of unshipped concepts into a portfolio of products with revenue? How do you stop optimising for the next role and start optimising for the next thing you make?

These are new skills. And like every new skill in design history, the people who learn them early have an outsized advantage over the people who wait until they are the default.

The designers who learned Figma early didn’t get rich because of Figma. They became impossible to compete with on iteration speed. The designers who learn to ship full products solo in 2026 will look back, in three years, the same way.

The Bottom Line

The layoffs are real. The fear is real. The shaking floor at the Fortune 500 is real.

The renaissance is also real, and it is structurally larger than the layoff wave. Figma’s $68 billion IPO, Base44’s $80 million acquisition by Wix, Canva’s $2 billion revenue, the 34% of new micro-SaaS products built by non-coders, the 41.8 million solopreneurs adding $1.3 trillion to the economy: these are not isolated stories. They are the shape of the new game.

In 18 months, the question won’t be “will my Fortune 500 design job survive?” It’ll be “why didn’t I start shipping my own thing two years ago when the tools first made it possible?”

That shift is already happening. The designers who started this month are already ahead of the ones who waited.

Germán Dario León Osório
In Love With AI

P.S. If this resonates, you are probably the type of designer who has been telling yourself “I’ll start that product when I have more time / more savings / more permission”. You don’t need any of those. You need a weekend, a Claude subscription, a Vercel account, and the willingness to ship something publicly imperfect on a Monday morning. The renaissance is not waiting for anyone. The people I see thriving in it are the ones who stopped asking what their company’s next move is and started making their own.

Sources

• 2026 Tech Layoffs Tracker: 52,050 Q1 Cuts + Where Talent Lands (Kore1)

• Tech Layoffs: US Companies With Job Cuts In 2024, 2025 and 2026 (Crunchbase News)

• Major tech layoffs: An updated tracker (InformationWeek)

• The AI layoff wave is just beginning, and it’s by design (Fortune)

• The AI Layoff Crisis of 2025: Why 70% of UX Designers Will Be Replaced (Medium)

• UX Designer Job Postings Are Down 71%. Figma Filed for a $68B IPO. (Medium)

• State of UX 2026: Design Deeper to Differentiate (Nielsen Norman Group)

• Figma shares more than triple in soaring public debut (Fortune)

• Figma’s Blockbuster IPO Gives CEO Dylan Field a $6 Billion Fortune (Bloomberg)

• Dylan Field, Figma’s 33-year-old cofounder (Fortune)

• Solo Founder Index 2026: Success Rates, Tools, and the AI Advantage (ShipSquad)

• Solo founders are using AI to do the work of entire teams (Fortune)

• 7 Solo Founders Building $1M+ AI Businesses in 2026 (Grey Journal)

• List of Funded Design Startups 2026 (Fundraise Insider)

• Design Tools Startups funded by Y Combinator (Y Combinator)

We go deep on why most companies are already too late to adapt, how Futurity built what they call “the Google of innovation”, a large graph model trained on the entire history of human innovation, and why compressing the innovation cycle from years to weeks is no longer a competitive advantage, it’s a survival requirement.

We also explore some genuinely mind-bending territory: autonomous commerce, machine-to-machine economies, digital souls, plants with legal standing, and what post-capitalism might actually look like when AI can spin up companies in seconds.

This one is uncut, honest, and exactly the kind of conversation that reminds you why building the future still matters.

Topics covered:

• Why corporate innovation labs keep failing

• Large graph models vs. large language models

• From deep research to deep imagination

• The imagination, science, and business horizons

• What AI means for money, companies, and humans

There is a moment that happens somewhere in the middle of building something new, when the gap between what you can imagine and what you can actually make feels unbridgeable. For most of design history, that gap was filled by a handoff: the designer sketched the idea, the engineer built it, and somewhere in the translation, something was always lost. An intention softened. A nuance dropped. A human feeling flattened into a specification.

Artificial intelligence has blown that gap wide open, and with it, an entirely new kind of professional has begun to emerge. Not a designer who codes a little. Not a developer who cares a bit about UX. Something genuinely new: a professional who can conceive of an intelligent system, understand the cognitive science behind how humans will interact with it, and then actually build it, prototype it, test it, and ship it.

They go by many names in the industry. Product AI designer. Human-AI interaction specialist. AI-native builder. But at ELISAVA Barcelona School of Design and Engineering, we have been training this professional with intention and rigour for three consecutive cohorts now, under the name the Master in Human Interaction and Artificial Intelligence (MHIAI).

On June 16th, 2026, the third generation of these professionals will walk off the stage at their graduation show and into a world that is only beginning to understand what they represent.

The Idea That Started It All

I have spent my career living at exactly this intersection, watching projects collapse in the white space between a Figma file and a Python notebook, seeing brilliant AI systems fail not because the model was wrong, but because no one designed the moment of human contact.

The MHIAI was built to close that gap permanently.

“Artificial intelligence is no longer just an occurrence of science fiction,” the programme states. “It surrounds us, making our lives easier. Our challenge now is adapting it to ever-evolving new business and human needs, with social and environmental responsibility.”

The curriculum moves through seven interlocking modules: an introduction to AI and interaction design; the design and development of actual interactions (including code); the fundamentals of cognitive science; machine learning and neural network models; user experience architectures; principles of AI; and finally, a master’s thesis that is not a paper but a living, working AI product, presented in public.

What a Design Engineer Actually Does

The title “Design Engineer” is still fighting for its place in the corporate org chart. Recruiters paste it into job descriptions, then back-fill the role with either a graphic designer who has learned to use ChatGPT or a software engineer who has taken a weekend UX course. Neither is close.

A true Design Engineer in the MHIAI mould is someone who can do four things simultaneously that the industry has historically treated as entirely separate:

First, they understand people. Cognitive science is not an elective in this programme; it is a load-bearing pillar. Students learn how humans form trust with systems, how perception and memory constrain interface design, and how emotional responses shape interaction patterns. They read Nathalie Nahai‘s work on web psychology. They wrestle with the ethics of persuasive design. They learn to ask not just “does this work?” but “does this treat the person using it with dignity?”

Second, they understand machines. Not at the level of implementing production-grade LLM infrastructure, but deeply enough to know what machine learning models actually do, how neural networks arrive at their outputs, where hallucination comes from, and why a system trained on one population behaves unexpectedly with another. When a designer understands why a model fails, they design for that failure. When they do not, they are just drawing pictures on top of a black box.

Third, they can build. The programme is unambiguous about this. Students work with Python. They work with JavaScript. They build high-fidelity prototypes that are not click-through mockups but actual functioning intelligent systems. The Final Master Project is a real AI product, not a concept deck. The graduation show is not a portfolio review but a live demonstration.

Fourth, they think entrepreneurially. This is perhaps the quality that most surprises faculty and industry mentors alike, and it is the quality that defines the third graduating cohort most powerfully. These are not students waiting to be placed inside a company. They are founders in waiting, or in several cases, founders already.

The Third Generation: A Cohort Portrait

Twelve students. Look at where they come from: Greece, Italy, Brazil, Portugal, the Philippines, Singapore, China, and the United States. Eight countries, twelve people, one programme. For a cohort this small, the geographic spread is remarkable, and it is entirely deliberate.

The diversity is not incidental. It is the syllabus. When Nikos Bellos Bellos from Greece sits next to Wing Ren from Wenzhou, China, and both attempt to design an AI interaction that works for both of them, something happens that no monoculture classroom can replicate. Assumptions surface. Edge cases multiply. The design space expands. The AI gets better because the humans making it are more varied.

Let us meet some of the people who will be presenting their final projects on June 16th.

Nikos Bellos - Athens, Greece

Nikos is what happens when a multidisciplinary engineer refuses to pick a lane. Trained across computer systems and design at the National Technical University of Athens, with 500+ professional connections and an established presence in the European tech ecosystem, he arrived at ELISAVA already fluent in the technical grammar of intelligent systems. What the master gave him was the human vocabulary to go with it.

Where many engineers treat design as a final coat of paint applied to something that already works, Nikos came to understand design as the primary material of an AI product, the substance from which the user’s trust is either built or denied. Watch for a final project that sits at the precise boundary between engineering rigour and interaction poetry.

Izaskun Cuenca - Lisbon, Portugal

Based between Lisbon and Barcelona, Izaskun carries the particular perspective of someone who has lived across Iberian cultures and understands how language, geography, and social expectation shape the way people relate to technology. AI systems trained on US or UK data frequently misfire for European users in ways that are invisible to their designers. Izaskun sees those misfires coming.

Her presence in the cohort represents something the programme prizes above almost everything else: the designer who brings genuine ethnographic curiosity to every brief, who asks what the user’s life actually looks like before asking what the screen should look like.

Beatrice Delucchi - Genova, Italy

Beatrice brings the Italian tradition of craftsmanship into a domain that often trades precision for speed. From Genova, a city with a long history of engineering excellence, she approaches AI product design the way a shipwright approaches a hull: every element serves a function, nothing is gratuitous, and beauty and utility are not in tension but in conversation.

Her LinkedIn presence, already established before graduation, reflects a professional who understands that design is a public act of communication as much as a private act of problem-solving. She makes things that hold up.

Ana Fellipelli - São Paulo, Brazil

Ana came to ELISAVA carrying serious industry weight. Before the programme, she had already worked across Autodesk, Ericsson, and Globant, scaling design systems across more than 20 SaaS products and partnering with engineers and product managers on platforms used by hundreds of thousands of people. She knows what design at scale actually costs, and what it takes to maintain coherence when the system is large enough to fight back.

What the master added was the technical literacy to understand the intelligence layer underneath the interface: why a model behaves the way it does, how that behaviour shapes the user’s experience, and what the designer’s responsibility is when the system learns. She arrives at the graduation show as someone who can design the full stack of the human-AI relationship, not just the surface of it.

Lyla Huang - Singapore

Yanling brings to the programme the particular sharpness of someone trained in one of Asia’s most demanding digital ecosystems. Singapore has spent decades building AI governance frameworks, public digital infrastructure, and a culture where technology is expected to work properly for everyone, not just for early adopters. That expectation is a design standard, and Yanling carries it.

In a cohort full of people who can build things, she stands out for her attention to what happens after the thing is built: how it is governed, how it adapts over time, how it survives contact with real users in conditions that were never part of the design brief. That long-view sensibility is exactly what most AI products lack, and most teams struggle to hire for.

Samantha Lin - San Francisco, USA

Growing up in San Francisco means growing up inside the technology industry’s self-image. Samantha has seen what happens when products are shipped fast, and the human cost is discovered later, when the gap between what a system can do and what it should do gets papered over with another product update.

The MHIAI gave her a different frame entirely: one in which the design of the interaction is not a layer applied on top of the technology but the technology’s most important feature. She brings to the graduation show a designer’s rigour and a San Franciscan’s hard-won scepticism, a combination that produces work that is both ambitious and honest about its own limitations.

Denise Macalino - Canada

Denise is building AI products with a design conscience that most of the industry has not yet developed. Her focus is not on what these systems can do in ideal conditions but on what they do when the conditions are not ideal, which is most of the time, for most of the world.

She asks the questions that do not appear in the standard UX playbook. What happens to an AI interaction when the connection drops? How does a voice interface behave when it was not trained on your language? What trust does a system earn, or destroy, in the first thirty seconds with a user who has never encountered anything like it before? These are not edge cases. They are the experience of the majority of the planet, and Denise designs for them with the conviction of someone who has lived it.

Tin Manasan - Philippines

Tin, as she is known in the programme, is one of those designers who understands that cultural fluency is a technical skill. Growing up navigating multiple languages, multiple platforms, and a digital culture that is simultaneously local and deeply global, she developed an instinct for the gaps that AI systems fall into when they are built by people who only know one context.

At ELISAVA, that instinct became a methodology. She can identify, at the prototype stage, where an intelligent system will fail to read a user correctly, not because the model is broken but because the model was never introduced to that user’s world. Her final project takes that problem seriously, designing not just around it but through it.

Tommaso Quintiliani - Rome, Italy

Tommaso carries the weight and the gift of coming from Rome, a city that has been building things designed to last for three thousand years and has therefore developed a deep scepticism toward anything that is merely fashionable. He applies that scepticism to AI products with productive results.

Where many in the cohort are drawn to the frontier of what AI can do, Tommaso is equally interested in what it should not do. In the language of the programme, he is thinking about trust, explainability, and the ethics of automated decision-making not as constraints on design but as design materials in their own right.

His LinkedIn profile, already live on the Italian professional network, marks him as someone who takes the public dimension of design seriously. Work is not private until it ships. It is public from the moment it affects someone.

Wing Ren - Wenzhou, China

Wenzhou has a reputation in China that is almost mythological: a city that built itself through private enterprise when there was barely a framework for private enterprise to exist, producing generations of people who know how to move fast, adapt relentlessly, and build things that survive. Wing carries that inheritance into her design work, and it shows.

She is a builder in the t@ruest sense: someone for whom an idea is not finished until it is a thing that works in the world and has been tested against the friction of real use. At ELISAVA, she combined that drive with a rigorous understanding of human-AI interaction, producing work that is entrepreneurially ambitious but technically grounded. She is one to watch well beyond June 16th.

Nona Voss - Düsseldorf, Germany

Nona brings one of the most distinctive perspectives in the cohort: she arrived at AI interaction design not through the front door of design or engineering but through a path that kept her close to the human side of the equation. The result is a professional who never lost sight of what these systems feel like to someone encountering them for the first time.

That might sound like a modest skill. It is not. Most AI products are built by people who find them intuitive, for users who do not. The gap between those two groups is where adoption fails, trust breaks down, and expensive systems sit unused. Nona fills that gap with precision. She asks what a system needs to communicate about itself before a user will risk trusting it, and she builds those answers into the interaction from the ground up.

Rene Zelaya - San Francisco, USA

Rene has watched the technology industry redefine what a career looks like several times over, and he has developed a hard-won immunity to hype. He has seen the pattern: breathless launch, scaling crisis, user trust collapse, quiet rebuilding. He knows where in that cycle most AI products currently sit, and he designs with that knowledge in the room.

His work at ELISAVA is focused on the question that most AI product teams avoid until it is too late: what does the relationship between a person and an intelligent system look like, not at launch, but six months in? A year in? How does trust accumulate, and what destroys it? These are not abstract questions for Rene. They are the design brief. Any team building an AI product that expects to still have users in two years needs someone asking exactly these questions from day one.

The Graduation Show: June 16th

Every cohort of the MHIAI has ended with a public demonstration, and the tradition has grown more ambitious with each iteration. The first generation introduced the format; the second refined it at the Roca Gallery Barcelona, one of the city’s most architecturally considered cultural spaces. The third generation is expected to go further still.

What visitors to the June 16th show will encounter is not a portfolio exhibition in the conventional sense. It is closer to a product launch, except that the products are AI systems, and the demonstration is done live, in real time, in conversation with real people.

This is intentional and important. An AI product that only works in the controlled conditions of a design studio is not an AI product; it is a proof of concept. The graduation show is the first public test. How does the system behave when confronted with someone it did not expect? How does it recover? How does it invite the user to trust it, and does it earn that trust?

These are the questions that the third generation has been building toward for a full academic year. The answers will be on display on June 16th.

Demo Day: June 25 and 26 - Live, Global, Unmissable

The graduation show is just the beginning. On June 25th and 26th, the third cohort takes the stage again for Demo Day, and this time, the room has no walls.

Both days will be streamed live, open to anyone in the world who wants to watch what happens when a group of extraordinarily trained Design Engineers unleashes their final AI products in front of an audience of investors, industry judges, and professionals tuning in from across Europe, Latin America, Asia, and the United States.

This is not a student showcase. It is a product demonstration in the fullest sense: twelve AI systems, built from scratch, presented by the people who conceived and built them, evaluated by people who have seen a great many AI products and know the difference between something real and something that merely looks real.

For investors, it is an early window into some of the most original AI product thinking currently coming out of European design education. For hiring managers, it is a live skills assessment more honest than any interview. For anyone working at the frontier of human-AI interaction, it is simply the most exciting thing happening in those two days.

Mark June 25th and 26th in your calendar. Set a reminder. Share the stream. The cohort has spent a year building something worth watching.

Why This Matters Beyond Barcelona

It would be easy to frame the MHIAI as a Barcelona story, a European story, a design school story. It is all of those things. But the professional it is producing is not local.

The Design Engineer is emerging simultaneously and somewhat independently in design schools and tech companies across the world, because the world needs them urgently. Every organisation that has deployed an AI system and watched it fail to achieve adoption has encountered the problem that the MHIAI was created to solve. The gap between what the model can do and what the person will trust it to do is a design problem. It requires someone who understands both sides of the gap and can bridge them with intention.

The roles exist and are unfilled in every sector: healthcare, finance, education, logistics, retail, and government. The job titles keep shifting because organisations have not yet settled on what to call someone who can do all of this at once. Human-AI Interaction Designer. AI Product Designer. Conversational Systems Designer. Intelligent Experience Lead. Whatever the title on the requisition, the underlying need is identical: someone who can design an AI product that real people will actually use, trust, and return to.

But what job descriptions cannot quite capture is the entrepreneurial dimension that has come to define this cohort. The first two generations produced professionals who went into companies and reshaped them from the inside. The third cohort looks, from the inside, like it will produce companies of its own.

When you put twelve people from eight countries in a room and teach them to build AI products with cognitive rigour and ethical seriousness, something happens. Collaborations form. Shared obsessions crystallise into ideas. Ideas that one person would have abandoned as too ambitious become feasible when someone else knows how to build the piece they cannot. The graduation show is as much an opportunity for investors and hiring teams as it is an academic milestone.

What You Are Actually Hiring

If you are a recruiter, a hiring manager, or a founder reading this, here is the direct version.

These twelve graduates can do something that most job candidates in the current market cannot: they can take an AI product from concept to working prototype, understand why it will or will not be trusted by the people who use it, and design the interaction so that trust is built from the first contact rather than retrofitted after the first complaint.

They are not generalists who have added “AI” to their LinkedIn headlines. They have spent a full academic year inside the problem, working with machine learning models, building functioning interactive systems in Python and JavaScript, studying the cognitive science of human-machine interaction, and presenting their work publicly to industry professionals who have no obligation to be kind.

They are multilingual, multicultural, and radically pragmatic about what AI can and cannot do for people. They have been trained by practitioners from Autodesk, Hubtype, Futurity Systems, Blackbot, and GFT, among others, not by academics describing an industry from a distance.

The specific skills they bring to a team:

Designing AI-driven interfaces and conversational systems that users trust from the first interaction. Prototyping intelligent systems using real code, not just mockups. Conducting user research specific to AI contexts, including how people form mental models of systems that learn and adapt. Identifying where AI models will fail for specific user populations before those failures reach production. Thinking about AI products across their full lifecycle, including adoption, trust degradation, and recovery.

If your organisation is building anything with AI and you are struggling to find people who can sit comfortably at the intersection of design, technology, and human behaviour, these are the people you are looking for.

They graduate on June 16th. They are available now.

The Faculty That Made It

It would be incomplete to celebrate the cohort without acknowledging the faculty who shaped them.

I teach Introduction to Artificial Intelligence and Interaction Design myself, and my approach sets the tone for the entire programme: professionals teaching what they actually do, not what they read in a textbook. Beside me, Gemma Alcodori León develops the character and executive presence that no technical curriculum alone can build, Arcadi Poch unlocks creative thinking and makes generative AI feel like a natural extension of the imagination, and Tim Cowlishaw brings a rare combination of deep coding fluency and interaction design sensibility that only comes from years of building real things. Victor Ventura teaches the UX architecture module with Figma at its centre, turning abstract product thinking into decisions students can see, test, and defend. Andrea Talal Haidar, from Hubtype, teaches conversational design with the authority of someone who has deployed chatbots in production environments that handle millions of conversations.

Regina Dos Santos Estevez teaches cognitive science and design fundamentals, and her influence is visible in the cohort’s unusual sophistication about the human dimensions of intelligent systems. Cecilia MoSze Tham of Futurity Systems teaches AI principles through the lens of speculative design and future thinking, pushing students to ask not just what AI can do today, but what kind of world it should be building toward. Fernanda Rocha and Jon Black from Blackbot bring a studio energy to creativity and design thinking that makes the room feel less like a classroom and more like a place where ideas get made. Stephanie Rodriguez Osorio, Jonatan Poveda Pena, and Pelayo Méndez Flórez form a teaching trio that gives machine learning its practical backbone, connecting theory to the decisions real ML practitioners make every day. And Ariel Ortiz Beltrán, Ph.D., closes the ML arc with the rigour and depth that only a researcher who has lived inside the field can offer.

A staff that reads less like a university faculty and more like the board of advisors a well-funded AI startup would kill to assemble.

This is not accidental. It is the programme’s fundamental pedagogical bet: that the best way to train people to work in an industry is to have the industry teach them.

The Larger Argument

There is a debate running through the design and technology industries right now about whether AI is going to make designers irrelevant or make them more powerful than they have ever been. It is, as most binary debates are, the wrong question.

The correct observation is this: AI is making some kinds of design work trivially automatable, the repetitive, the formulaic, the execution-heavy, while simultaneously creating an enormous demand for a kind of design work that has barely existed before: the intentional, the trust-building, the human-centred design of intelligent systems themselves.

The MHIAI graduates stepping onto the stage on June 16th are not threatened by the first category. They are the answer to the second one.

They are the professionals who understand that when someone interacts with an AI system, they are doing something fragile and human: making themselves legible to a machine, trusting that it will not distort or dismiss what they offer. They are the professionals who know that designing that moment of contact, with all of its cognitive, emotional, and ethical complexity, is the most consequential design challenge of the current decade.

Three generations in, the MHIAI has demonstrated that this professional can be trained, and that once trained, they go on to build remarkable things. The graduation show is where that proof goes public.

On June 16th, twelve members of the third cohort, from São Paulo to Singapore to San Francisco to Rome to Wenzhou to Athens, will stand up in Barcelona and show the world what they built.

If your organisation builds anything with AI, you should be in that room.

The 3rd MHIAI Graduation Show takes place on June 16th, 2026, at ELISAVA Barcelona School of Design and Engineering, La Rambla 30-32, Barcelona. Demo Day streams live on June 25th and 26th, 2026, open to a global audience.

This is a conversation space exclusively for subscribers—kind of like a group chat or live hangout. I’ll post questions and updates that come my way, and you can jump into the discussion.

Join chat

How to get started

1. Get the Substack app by clicking this link or the button below. New chat threads won’t be sent sent via email, so turn on push notifications so you don’t miss conversation as it happens. You can also access chat on the web.

Get app

2. Open the app and tap the Chat icon. It looks like two bubbles in the bottom bar, and you’ll see a row for my chat inside.

3. That’s it! Jump into my thread to say hi, and if you have any issues, check out Substack’s FAQ.