Google and OpenAI are Copying Anthropic

OpenAI Goes 'Code Red' and Builds a Superapp

OpenAI plans to merge ChatGPT, the Codex app, and the Atlas browser into a single desktop application. Fidji Simo, Chief of Applications, is leading the restructuring; Greg Brockman is expected to assist with the reorganization. In an internal memo on Thursday, Simo wrote that they had spread their resources across too many apps and stacks: “That fragmentation has been slowing us down.” The turnaround follows a year in which OpenAI found little resonance with Sora, a hardware device, and various standalone products. CEO Sam Altman, CRO Mark Chen, and Simo have combed through the entire product portfolio in recent weeks and identified areas for deprioritization. OpenAI is operating internally under “Code Red,” according to a spokesperson, after Anthropic's Claude Code and Cowork have gained significant ground in the enterprise segment. Both companies are considering an IPO this year. → Wall Street Journal

Synthszr Take: OpenAI is openly admitting what has been visible for months: too many products, too little focus, too much internal wrangling over resources. Simo calls it 'side quests,' which sounds diplomatic for a product strategy that more closely resembled a shotgun blast. Sora, the hardware project, a browser almost no one knows; the list of launches without a clear business model is long. The real trigger is between the lines: Anthropic has shown with Claude Code and Cowork that enterprise customers don't want ten AI tools, but one that works. OpenAI is now reacting with the superapp strategy that Microsoft has been preaching for years (and rarely pulls off). Whether a company that has produced fragmentation for twelve months can suddenly achieve integration remains to be seen; the 'Code Red' mode at least suggests that this time, it's not just the slide deck that's new.

Google is Testing a Native Gemini Client for Mac

Google is developing a native Gemini app for macOS and has begun initial beta testing. Previously, Mac users could only access Gemini through a browser. The early beta version includes core features like web search, document analysis, image generation, and solving complex mathematical problems. The interface is heavily based on the existing iOS and iPadOS apps. Code references point to a “Desktop Intelligence” feature, which might enable desktop sharing for additional context or integration with external services like email clients. Unlike Anthropic's Claude, which already offers extensive MCP-server integrations, there are no indications of comparable agentic capabilities in Gemini's Mac client yet. → Neowin

Synthszr Take: Google is playing catch-up in the desktop AI market. Anthropic has set the pace with Claude, OpenAI dominates with ChatGPT – now Google has to follow suit. 'Desktop Intelligence' sounds like local context access, but without MCP-server support or real agent functions, it remains a glorified web wrapper. Mac users already have better alternatives installed. Google is late and offers less – not a recipe for success.

Why Azeem Azhar Changed His Mind About Apple

Azeem Azhar has revised his skeptical stance on Apple in the AI era. Apple spends little on data centers, Siri hasn't improved in a decade, and no one expects AI breakthroughs from Cupertino. Analysts like John Gruber and Ben Thompson see Apple as far behind the technological cutting edge. Azhar's turning point came with OpenClaw: he suddenly bought several Mac Minis because the AI agents were overwhelming his existing hardware. Best Buy is reporting empty shelves, and delivery times for configured Mac Minis have increased from three days to eight weeks. Jensen Huang calls OpenClaw “the new computer”—and right now, it runs primarily on Apple hardware. → Azeem Azhar, Exponential View

Synthszr Take: Apple is winning the AI battle without its own model and without capex investments. OpenClaw users are buying Mac Minis like people used to buy graphics cards for Bitcoin mining, while Apple comfortably cashes in on the margin. Tim Cook gets it: let others sink billions into models, you sell the shovels. Mac Minis are selling like hotcakes, Mac Studios with 512 GB of RAM have been discontinued due to high demand amid scarce memory chips. The default platform for all new apps is Apple, and Cupertino doesn't even have to improve Siri for it.

China's Tech Giants in an AI Agent Frenzy

China's tech giants are currently experiencing their “shrimp farming” frenzy: Tencent, Baidu, and Alibaba are outdoing each other with AI agent offerings, which they market as “digital shrimp farming.” The metaphor is deliberate—like in aquaculture, users are meant to “feed” and “raise” their AI assistants. Baidu started in February with a one-click deployment service, followed by DuClaw in March. Tencent countered with WorkBuddy and QClaw, while Alibaba is directly targeting the enterprise market with “Wukong.” The numbers speak for themselves: over 230,000 Chinese OpenClaw instances are already running on the public internet, with more added daily. But behind the hype lies a fundamental strategic conflict between consumer scaling and enterprise security. → Hello China Tech

Synthszr Take: 230,000 unsecured AI agents on the Chinese internet is not a success, it's a security nightmare. Alibaba's CEO Chen Hang warns of “uncontrolled super-agents with blowback risk”—in plain terms: the shrimp are eating their farmers. While Baidu and Tencent are betting on rapid consumer adoption, Alibaba is building an enterprise solution with Wukong, featuring “DNA-based authorization control.” The market is splitting: the mass market wants convenience, while enterprises need control. Gartner predicts an enterprise penetration of 40% by 2026, but the current reality mainly shows KPI-driven installations with no real utility. China is pushing ahead where OpenAI still hesitates—with all the consequences.

MiniMax Bets on Self-Learning Agents with Model M2.7

MiniMax has released M2.7, a model that drives its own evolution. The system uses agent frameworks and Reinforcement Learning for memory updates, skill development, and iterative self-improvement. M2.7 achieves 97% skill adherence across 40 complex skills and is close to the industry leaders in benchmarks like SWE-Pro (56.22%) and VIBE-Pro (55.6%). The model supports multi-agent collaboration and complex workflows in software engineering, office productivity, and research. MiniMax is already using the model internally to automate its own R&D processes with minimal human intervention. → TLDR AI

Synthszr Take: MiniMax is demonstrating how Chinese AI labs are advancing agent architecture. 97% skill adherence and self-learning systems sound impressive, but the real innovation lies in the internal application: MiniMax is already using M2.7 for its own research with “minimal human intervention.” Multi-agent collaboration plus Reinforcement Learning results in a self-improving development process. Benchmarks like SWE-Pro show competence, but its use in-house is the better proof. China is no longer just building models, but systems that evolve themselves.

Xiaomi Advances with Trillion-Parameter Model as China Pushes AI Agents

Xiaomi presents MiMo-V2-Pro, a 1-trillion-parameter model that activates only 42 billion parameters per run, achieving the performance of GPT-5.2 and Opus 4.6. The sparse architecture, combined with Multi-Token-Prediction, drastically reduces latency at a fraction of the usual cost. In parallel, MiniMax is launching its M2.7 model, which can perform autonomous debugging and contributes to its own evolution. Anthropic has published a study with 81,000 participants on AI surveillance, showing that permanent observation makes people more productive on simple tasks but slower on complex ones. → TLDR AI

Synthszr Take: Xiaomi is copying OpenAI's scaling strategy but making it more efficient. 1 trillion parameters with only 4.2% active – that's the Chinese answer to Western compute power excesses. MiniMax goes a step further and lets models steer their own development. Anthropic's 81k-participant study confirms what every office worker knows: surveillance makes dumb work faster and smart work slower. China isn't building better models, but more efficient architectures.

Five Lessons from Vercel's Founder for Building in the AI Era

Guillermo Rauch, founder of Vercel and creator of the Next.js framework, shared his experiences building a company in the AI era at a16z speedrun. He describes open source as a “speedrun to product-market fit”: if people won't even use your product for free, you should work on something else. Rauch warns against the danger of unlimited ambitions in small teams. Vercel initially tried to make every programming language and every framework deployable until he asked the crucial question: Where are we truly excellent? The founder describes a “bicameral mind”—you need a boundless vision for the future, but on a Tuesday, with three people in the room, that ambition will destroy you if you let it dictate your daily actions. → a16z speedrun

Synthszr Take: Rauch provides the blueprint for AI startups that are lost between megalomania and irrelevance. Open source as a filter works brutally well: 99% of projects fail in the first few weeks, and the rest build communities that later become paying customers. Vercel proved that you can become a billion-dollar company with a single framework (Next.js) while competitors suffocate in their own complexity. AI makes this principle even more relevant: agents can copy your code, but not your community, not your ecosystem, not the thousand small decisions that turn a framework into a platform. Rauch's recipe for success is shockingly simple: build something developers love, and ignore everything else.

Nvidia is Quietly Building a Third Pillar

In just a few years, Nvidia has turned its networking business into its second-largest revenue pillar after its chip business. $11 billion in quarterly revenue, a 267 percent increase year-over-year, and over $31 billion for the full year. The division includes NVLink for GPU communication, InfiniBand switches, Spectrum-X for AI networks, and other components for so-called “AI Factories.” Kevin Cook of Zacks Investment Research compares: Nvidia's networking division makes in one quarter what Cisco makes in an entire year. The origin lies in the Israeli company Mellanox, acquired in 2020 for $7 billion. While everyone is focused on Nvidia's chips, Jensen Huang has simultaneously built up a second pillar that is three times the size of the original gaming business. → StrictlyVC

Synthszr Take: Buying Mellanox for $7 billion and generating $31 billion in annual revenue from it four years later—that's the kind of deal MBAs will be recreating in PowerPoint. Nvidia isn't just selling the shovels in the gold rush; it's selling the entire mining infrastructure. 267 percent growth at this scale shows: anyone serious about AI training can't get around Nvidia's networking stack. Cisco is currently being overtaken by a company most people still know as a graphics card manufacturer. Jensen Huang has once again proven that timing is more important than technology.

Are Forward Deployed Engineers a Sham?

Forward Deployed Engineers (FDEs) are experiencing a paradoxical moment: job postings on Indeed have increased tenfold in 2025, yet hardly any developer wants the job. 50 companies mention the role in their earnings calls, compared to eight the previous year. The discrepancy between demand and interest reveals a fundamental misunderstanding: companies are looking for developers for customer projects but are getting something more like solutions engineers with coding skills. FDEs spend their time on-site with customers, implementing custom solutions and battling legacy systems—a far cry from product development at headquarters. While Palantir has successfully established the model, many imitators fail due to unrealistic expectations and poor positioning of the role. → The Pragmatic Engineer

Synthszr Take: Palantir made a virtue out of necessity and turned FDEs into its business model. A 10x growth in job postings meets developers who would rather build products than cater to enterprise clients. FDEs are the new consultants: well-paid, travel a lot, with little real development work. Companies are slowly realizing they aren't looking for developers willing to travel, but for technical sales staff with programming skills. The role isn't dying; it's just being named more honestly.

Why Gmail Clips Your Emails—And What You Can Do About It

Gmail automatically clips any part of an email that exceeds the 102 KB limit. Subscribers then only see a “[Message clipped] [View entire message]” link where the rest of the content should be. Most recipients never click this link—they simply assume the email ends there. The problem is exacerbated because Gmail often clips the tracking pixel at the end of the email, causing opened emails to appear as unread in statistics. This particularly affects long newsletters, promotional emails with multiple images, and heavily formatted messages with many buttons and styled sections. → TLDR Marketing

Synthszr Take: 102 kilobytes determine success or failure in email marketing. Gmail treats any message over this magic limit like an overly long movie: cut off, without warning. AWeber calculates that about 20% of all marketing emails are affected—usually the most elaborately designed campaigns. Tracking pixels disappear into the digital void, open rates plummet, and marketers puzzle over seemingly uninterested subscribers. The solution sounds trivial (shorten emails), but it collides head-on with the trend towards ever-longer, image-rich newsletters. Gmail is thus silently dictating the rules of email marketing—forcing an entire industry onto a digital diet.