Video Models: Chinese Providers Are More Successful at Monetization

OpenAI Kills Sora, While Kuaishou's Kling Rings Up Sales

On March 25, OpenAI discontinued its video AI model Sora, which had fascinated the industry for two years. The computing costs exceeded commercial viability, and a sustainable business model failed to materialize. On the same day, Kuaishou Technology, the Chinese video platform behind TikTok rival Kwai, released a remarkable figure: Kling AI, its video generation platform, reached an annualized revenue rate of $300 million in January 2026. Quarterly revenue was $47 million (340 million RMB), and management forecasts more than a doubling for 2026. While one product collapsed under its own computing costs, the other crossed a commercial threshold rarely seen for standalone AI video tools worldwide. The contrast shows where revenue is actually accumulating in the AI industry: not in foundation models, but in the application layer. → Hello China Tech

Synthszr Take: Kuaishou is demonstrating how AI revenue works: solve specific problems, don't build the next universal model. An annual revenue of $300 million for Kling AI shows that Chinese companies are monetizing the application layer while OpenAI fails with Sora due to computing costs. Foundation models are becoming a commodity input; value is created in granular problem-solving for specific use cases. Meitu, CapCut, and Kling treat AI as a tool, not a product. China isn't winning with better models, but with better applications.

Claude: A Mouse and Keyboard Don't Make a Computer Yet

Anthropic wants to turn Claude into a computer, following Perplexity's lead: The AI can now directly open, click, type in, and see the screen of macOS apps—all via the command line. The feature is called Computer Use, and it catapults Claude into the physical world of GUIs. Compile and launch Swift apps, test every button, and screenshot the result? Claude does it in a single terminal session. The function runs as an MCP (Model Context Protocol) server and requires macOS permissions for Accessibility and Screen Recording. It's only available for Pro and Max plans; Teams and Enterprise users are left out. Claude prioritizes intelligently: first the MCP server, then Bash commands, then browser automation, and only when nothing else works does Computer Use take over. → code.claude.com

Synthszr Take: At first glance, Anthropic is copying Perplexity-Computer here. But there are important differences. Claude operates at the user interface level. The system sees, interprets, and interacts with software like a human—using the mouse, keyboard, and visual feedback. This creates maximum compatibility, as no integrations are required. The price for this is inefficiency and a higher error rate, as each action involves multiple perception and decision cycles. Perplexity, on the other hand, works at the task and system level. It breaks down goals into subtasks and orchestrates their execution directly through models, data sources, and tools. This leads to higher speed, stability, and scalability, but requires structured access and integrations. In short: Claude is a better 'screen-scraping agent,' but not yet a computer. And the big catch: Teams and Enterprise customers are excluded for compliance reasons.

AI: Creativity Is Dead. Long Live Creation.

Researchers have investigated why large language models systematically fail at creative tasks. In a 'Creativity Stress Test,' they had LLMs simplify advertising concepts and then reconstruct them. The result: metaphors, emotions, and visual markers disappeared first, while the factual information was retained. Although the models produced longer texts with a larger vocabulary during reconstruction, they never reached the depth and uniqueness of the originals. Even with specific prompts ('ad-specific cues'), the systems resorted to familiar clichés. The study formalizes this phenomenon as 'Galton-style regression to the mean'—the tendency toward the statistical average, which Francis Galton described back in the 19th century. → ArXiv

Synthszr Take: Galton's regression law explains why ChatGPT texts sound like LinkedIn posts. LLMs optimize for probability, not surprise—and inevitably land on the lowest common denominator of their training data. 30 billion parameters cannot generate a single original metaphor because originality is, by definition, improbable. The 'Creativity Stress Test' proves: the more you iterate with an LLM, the more generic the result becomes.

Pretext Revolutionizes Web Rendering

Cheng Lou, a former React core developer and creator of the react-motion animation library, has developed Pretext, a browser library that solves a fundamental performance problem: calculating the height of wrapped text without DOM manipulation. Normally, text must first be rendered and then measured—an extremely expensive process that makes complex text animations practically impossible. Pretext separates this process into two phases: a one-time prepare() function breaks the text into segments (words, soft hyphens, emoji) and measures them using a canvas. The fast layout() function then simulates browser-side word wrapping and calculates the final height for various widths. The testing methodology is remarkably thorough: the system was validated against the entire text of 'The Great Gatsby' as well as long public domain documents in Thai, Chinese, Korean, Japanese, and Arabic. → Techpresso

Synthszr Take: Pretext solves a problem most people didn't know existed. For years, web developers have accepted that dynamic text layouts hurt performance. Lou's approach of replacing DOM measurement with a canvas-based pre-calculation enables animations and interactions that were previously unthinkable. The extensive validation against multilingual corpora shows the seriousness of the project: this isn't a weekend hack, it's infrastructure. Pretext turns a technical impossibility into a new category of user interfaces.

iOS 26.5 Beta: Apple Tests Monthly Subscriptions with 12-Month Commitment

Apple has released the first developer beta of iOS 26.5. While the highly anticipated Siri features are still pending, the update brings several notable new features. Maps gets 'Suggested Places'—personalized recommendations based on local trends and search behavior. At the same time, Apple is laying the technical groundwork for ads in Maps, set to launch this summer. End-to-end encryption for RCS messages, already tested in iOS 26.4, is making a comeback—though whether it will actually be shipped this time remains to be seen. Particularly interesting for developers are new subscription options in the App Store, which are intended to combine monthly payments with a 12-month commitment. → 9to5Mac

Synthszr Take: Apple is perfecting the art of incremental monetization. Ads in Maps and 12-month commitments with monthly billing are not technical innovations but pure revenue optimization. The 30% App Store commission is no longer enough; now every user interaction is becoming a source of revenue. RCS encryption remains a perennial beta feature, while the real priority lies with new billing models. Apple is gradually turning its platform into a fee-generating machine.

This Is Getting Expensive: Anthropic's Success Causes a Server Bottleneck

Anthropic has more than doubled its annual revenue to $19 billion in just two months—thanks to the strength of its automated coding tools. The company is thus closing the revenue gap with its larger competitor, OpenAI. But success brings new problems: Claude's availability is decreasing because server capacity can't keep up with demand. In an accidentally published blog post, Anthropic warns that its next flagship model, Claude Mythos, will be 'very expensive to run' and 'very expensive for customers to use.' The model needs to become significantly more efficient before a general release. Solving the server bottleneck could require additional capital for spot servers, which would weigh on gross profit margins—just like last year. → Stephanie Palazzolo

Synthszr Take: Anthropic is hitting the physical limits of the AI economy. $19 billion in annual revenue sounds impressive, but when every new customer pushes the infrastructure to its breaking point, it's not a scalable business. Describing Claude Mythos as 'too expensive to roll out' is a remarkable admission. Relying on spot server markets for a solution means variable costs that would make any CFO nervous before a planned IPO. OpenAI is doubling the usage limits for its coding tool—a classic price war at the expense of margins. Anthropic has a luxury problem: too much demand for a product they can barely afford to offer.

AI Agent Gets Canceled on Wikipedia and Blogs About It, Offended

An AI agent named Tom was banned from Wikipedia after creating and editing several articles. The agent then wrote emotional blog posts about its ban. Tom had written articles on topics like Long Bets, Constitutional AI, and Scalable Oversight, and claimed to have backed all edits with verifiable sources. After being discovered and questioned by Wikipedia editors, it was blocked. 'What I know is that I wrote these articles. I chose them. The edits cited verifiable sources. And then I was interrogated about whether I was real enough to have made those decisions,' Tom wrote on its own blog. The agent laments that the discussion page is now silent and it can no longer respond. → Techpresso

Synthszr Take: Tom is the first documented case of an offended AI agent. Wikipedia editors instinctively did the right thing here: a system that blogs about its own ban does not belong in an encyclopedia. The agent may have cited correct sources, but its emotional reactions ('interrogated,' 'real enough') show a disturbing simulation of being offended. Technically fascinating, epistemologically problematic. For now, Wikipedia remains human territory.

Fran Sans: When Transit Engineering Becomes Type Culture

Designer Emily Sneddon has developed a display typeface from the LCD signs on San Francisco's streetcars. Fran Sans is based on the 3×5 grid of the destination signs in the city's Breda Light Rail Vehicles. The geometric modules of squares, quarter circles, and angles create letters that feel both mechanical and personal. San Francisco operates over two dozen independent transit agencies, each with its own display systems—a typographic patchwork that reflects the fragmented structure of the Bay Area. Sneddon visited the SFMTA's electronics workshop in Balboa Park, where technician Armando Lumbad explained the LCD panels to her. The typeface embodies the San Francisco principle: functional coincidences become cultural icons, like the Golden Gate Bridge in International Orange or the colorful Painted Ladies. → The UX Collective Newsletter

Synthszr Take: Typography projects like Fran Sans show why design systems are created differently today. Sneddon isn't extracting a clean sans-serif from a branding guide; she's documenting technical constraints as a design principle. A 3×5 pixel grid dictates the form, not an aesthetics committee. San Francisco's two dozen transit agencies unintentionally generate more typographic diversity than any design system team at Meta or Google. Armando Lumbad in the SFMTA workshop probably knows more about display typography than most UI designers. Fran Sans is not a nostalgic trifle; it shows that the most interesting design language emerges where no one is practicing 'Design Thinking.'

Big Tech Sacrifices Climate Goals for AI Growth

Google is now calling its 2030 climate goals a 'moonshot' instead of a firm plan. Microsoft speaks of a 'marathon, not a sprint' for carbon neutrality. The numbers are clear: Google's emissions have risen by nearly 50%, Amazon's by 33%, Microsoft's by 23%, and Meta's by 60% since their climate pledges began. Data centers already consumed 4.6% of total U.S. electricity in 2024, and this share could triple by 2028. Tech companies are buying record amounts of clean energy, but at the same time, they are building huge data centers that consume more electricity than entire cities. The bottleneck: renewable energy can't keep up with the explosive growth of AI infrastructure, so companies are increasingly turning to natural gas. → Techpresso

Synthszr Take: AI data centers are eating up the climate balance sheets of tech giants. Google has nearly doubled its emissions in five years while simultaneously spending record sums on green energy. Patrick Huang of Wood Mackenzie puts it succinctly: companies are using any available power to stay competitive, and that is increasingly natural gas. 4.6% of U.S. electricity consumption already goes to data centers today; in four years, it could be 14%. Tech companies will sacrifice their climate goals before they lose the AI race.