*Future Perspectives | Reading time 9 min*
Every Pixel Is a Hallucination
What if nothing on your screen was real? No buttons. No links. No text rendered from HTML. Every single pixel drawn fresh by an AI model, in real time, just for you.
That’s Flipbook. A prototype that dropped last week and split the tech world into two camps: people who think it’s the future of computing, and people who think it’s an expensive parlor trick. Both groups are paying attention, which tells you something.
Flipbook calls itself an “infinite visual browser.” In practice, you type a search term and get back a richly illustrated page that looks like a spread from a visual encyclopedia. Click anywhere on that image. The system figures out what you tapped, interprets your intent, and generates an entirely new page. No hyperlinks involved. No DOM. No CSS. Just pixels, composed on the fly by a generative model.
The team behind it reads like a recruiting brochure for frontier AI. Zain Shah (former OpenAI researcher, YC S13 alum), Eddie Jiao (ex-Humane, ex-Slack), and Drew O’Carr (ex-Apple). When Shah posted the demo on X in late April, the conversation went vertical overnight.
Fast Company compared Flipbook to Apple’s 1987 HyperCard, the tool that first let ordinary people organize information through cards and links. The comparison holds. HyperCard reimagined how non-programmers could structure knowledge. Flipbook reimagines how anyone can explore it.
How It Actually Works
The open-source community reverse-engineered Flipbook within days. A project called OpenFlipbook used Playwright to document the full technical stack, and the architecture is surprisingly legible.
Static mode (the default experience): You submit a query. The backend calls an image generation model and streams back a progressive JPEG via Server-Sent Events. One request takes roughly 19 seconds and delivers about 7.7MB of data. The image sharpens incrementally until the final frame locks in. Generated pages get stored on R2-compatible object storage with permanent URLs.
Click navigation: When you click anywhere on the image, the system sends your click coordinates plus the current image to a vision-language model (VLM). The VLM identifies what you clicked, converts it into a new query, and triggers a fresh page generation. No predefined links. No sitemap. The “navigation” is inferred from visual context every single time.
Video mode (optional, GPU-heavy): Toggle this on and Flipbook connects via WebSocket to Modal Labs’ serverless GPU cluster, running Lightricks’ LTX-2 video diffusion model. It streams 1080p at 24fps, every frame generated by the model. The transport layer uses a custom binary protocol (reverse engineers dubbed it “LTXF”) built on H.264-encoded fMP4 segments played through the browser’s MediaSource API.
The internal codename was “Sketchapedia.” That name captures the ambition better than “Flipbook” does.
One clarification that matters: Shah’s viral claim that “every pixel streams from the model in real time” describes video mode, not the default. Video mode is off by default because it’s computationally brutal. The everyday experience is static image generation.
The Question That Makes This Interesting
The technical stack is cool. But Flipbook’s real provocation is a question: if AI gets good enough, do we still need web pages?
For thirty years, the fundamental contract between humans and digital information hasn’t changed. A server returns a structured document. A browser parses and renders it. Users interact through predefined elements: links, buttons, form fields. Mosaic in 1993 and Chrome in 2026 run on the same underlying logic.
After ChatGPT launched, the conventional wisdom said “the chat box is the new interface.” Flipbook’s founders pushed back with a line that stuck: selling chat boxes and fixed layouts as AI’s ultimate interface is like bolting an engine onto a horse carriage and calling it a car.
Flipbook takes a different path. Instead of layering AI on top of existing interfaces, the AI becomes the interface. You don’t learn interaction rules. You see something interesting, you click it. That’s the entire UX. For humans, this is the most natural interaction pattern there is. More natural than typing queries. More natural than drilling through menus. More natural than conversing with a chatbot.
This connects to a broader shift: interfaces moving from structured to generative. Traditional interfaces are pre-designed, with fixed layouts and fixed interaction paths. Generative interfaces are created on demand. Every user sees something different. Every interaction produces new content.
Where Generative UI Fits in the Broader Stack
Flipbook didn’t emerge in a vacuum. Several companies are chasing variations of the same thesis.
| Approach | Example Products | Core Mechanic | Limitation |
|---|---|---|---|
| Conversational AI | ChatGPT, Claude, Gemini | Text in, text out | Linear, low information density |
| AI-augmented search | Perplexity, Google AI Overviews | Query → synthesized answer + sources | Still text-dominant, fixed layout |
| Generative UI components | Vercel v0, Anthropic Artifacts | Prompt → functional code/UI | Produces traditional web artifacts |
| Full visual generation | Flipbook, OpenFlipbook | Query → explorable image | No text fidelity, no accessibility |
Flipbook sits at the most radical end of this spectrum. It doesn’t generate code that renders into a page. It generates the page directly as pixels. That’s a philosophical difference, not just a technical one. Code-generation tools like Vercel’s v0 still produce HTML and CSS. They’re faster carriage builders. Flipbook is trying to skip the carriage entirely.
The question for B2B SaaS builders: which layer of this stack will your product live on five years from now? If generative interfaces mature, the concept of “a page” as the atomic unit of web experience starts to dissolve. Product surfaces become fluid, personalized, contextual.
The Reality Check
Vision laid out. Now, the constraints.
Latency kills the magic. Nineteen seconds to generate a static page is an eternity in web UX. Users expect sub-second responses. Even Google’s AI Overviews, which people already complain are slow, render in two to three seconds. At 19 seconds per click, Flipbook is a demo, not a product.
Text accuracy is broken. Image models still can’t spell reliably. “Speeds” becomes “speds.” “Handlebars” becomes “HANDLEEBRS.” For a product whose entire value proposition is conveying information, garbled text is disqualifying. You cannot ship an encyclopedia that misspells its own entries.
Accessibility is nonexistent. The entire page is a flat image. Screen readers see nothing. Text can’t be selected, copied, or translated. For visually impaired users, this product does not exist. In 2026, building an information product with zero accessibility is not a quirky limitation. It’s a serious design failure that would block any enterprise deployment.
Information provenance is absent. AI-generated “encyclopedia pages” look authoritative but may contain hallucinated facts. No source attribution. No citation links. Users cannot verify accuracy. This is worse than traditional search engine misinformation because at least Google tells you which website a claim came from.
Interaction is shallow. Only one interaction is supported: click to explore. No text input on the generated page. No drag. No zoom. No precise manipulation. Shah himself acknowledges Flipbook’s current feature set is narrow, designed around “visual interpretation” and nothing else.
Open Source Already Caught Up
Within a week of launch, the open-source community shipped OpenFlipbook (also called Endless Canvas), a full-featured clone. It didn’t just replicate the core functionality. It improved on it in several ways:
It swapped the proprietary LLM for Qwen 2.5 VL 72B, an open-source vision-language model accessed through OpenRouter. It supports bring-your-own API keys, removing dependency on centralized infrastructure. It added transparent status indicators so users can see the VLM analyzing their click, planning the page layout, and rendering, instead of staring at a frozen screen for 15 seconds. And it ships as a single Docker Compose file. One command to deploy.
Two takeaways from this. First, the core concept has real pull. People wanted to build on it immediately. Second, the technical moat is shallow. The real differentiation will come from model quality, generation speed, and UX polish, not from architecture secrets.
For SaaS builders watching this space, that’s a familiar pattern. When the implementation barrier is low but the quality bar is high, competitive advantage migrates toward data, speed, and user experience. Think of how image generation commoditized quickly at the model level but Midjourney won on aesthetics and workflow.
What This Means for Product Teams
In the short term, Flipbook changes nothing about how you ship software. It’s a prototype with a waitlist. Compute costs make it unscalable. Information reliability makes it unusable for anything consequential.
But the direction deserves attention from anyone building information-heavy products.
The search-click-read funnel is compressing. Future information interfaces will likely shift from linear query flows to visual exploration patterns. Users won’t type a search, scan ten blue links, click through, read, go back, click another. They’ll enter a topic and explore it spatially. Products that can offer this kind of fluid discovery will have an engagement advantage.
Content format is diversifying beyond text. When AI can turn any topic into an interactive visual space, static articles lose relative value. This doesn’t mean text dies. It means text becomes one layer in a richer information surface. SaaS products that present data, documentation, or knowledge bases as text-only will feel increasingly dated.
Front-end development may fragment. Today, building interfaces means writing HTML, CSS, and JavaScript (or frameworks that compile to them). If generative interfaces become viable for certain use cases, a parallel track emerges where “building” an interface means training or fine-tuning a model. The skills are completely different. Product teams that start experimenting with generative UI primitives now will have a head start when the compute economics catch up.
Accessibility will be the bottleneck for adoption. Any generative visual interface needs to solve accessibility from day one, not as a retrofit. Enterprise buyers in North America won’t touch products that fail WCAG compliance. Whoever cracks the problem of making AI-generated visual interfaces accessible will own a critical enabling layer.
The Timeline Question
How far away is this from production readiness? The limiting factors are compute cost and model capability, both of which are improving on predictable curves.
GPU inference costs have dropped roughly 10x every two years since 2020. If that trend holds, the $2-per-session cost of Flipbook’s visual generation becomes $0.20 by 2028 and $0.02 by 2030. That’s commercially viable for many use cases.
Text rendering in image models is improving with every generation. GPT-Image, DALL-E 4, and Ideogram 3 all show dramatically better typography than models from even a year ago. Give it another 18 months and the “can’t spell” problem likely shrinks from disqualifying to annoying.
Latency is harder to project. Getting from 19 seconds to under 2 seconds requires both model architecture changes and hardware improvements. Speculative decoding, distillation, and specialized inference chips (Groq, Cerebras) are all attacking this from different angles. Two to three years feels realistic for static generation. Real-time video generation at acceptable quality is further out.
The Takeaway
Flipbook is not a product you should worry about competing with today. It’s a concept that should change how you think about interfaces over the next three to five years. The web page as we know it, a structured document rendered from code, has been the dominant paradigm for three decades. Flipbook is the clearest signal yet that this paradigm has an expiration date. Not because the technology is ready now, but because the trajectory is legible and the destination is compelling enough that serious people are building toward it. The teams who start thinking about generative, visual, fluid interfaces now will be better positioned when the economics flip. And based on current trends, that flip is coming faster than the 19-second page load might suggest.
Stay updated with our latest AI insights
