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.
The deciding never left, it just got harder to see. When the tool writes the code, the leverage moves to whoever framed the problem and knows which version to kill. I’ve watched designers hand execution to AI and panic, then realize the taste and the no were always the job. The canvas was never the point.