No April Fool's Joke: Anthropic Leaks Complete Source Code

Anthropic leaks source code for Chinese competitors

Anthropic accidentally made the complete source code of Claude Code public. A 59.8 MB JavaScript source map file was published in version 2.1.88 of the npm package @anthropic-ai/claude-code. Chaofan Shou, an intern at Solayer Labs, discovered the leak yesterday and shared it on X with a direct download link. Within a few hours, thousands of developers were analyzing the roughly 512,000 lines of TypeScript code on GitHub. The company confirmed the incident as a “packaging error due to human failure,” not a security breach. Anthropic → venturebeat.com

Synthszr Take: Anthropic didn't just leak code; it provided structural insights into the next evolution of AI systems—especially regarding memory. The code shows that memory is no longer just a server-side issue (HBM on GPUs) but has been massively extended to the client side. Idle processes of Claude Code consume ~15 GB of RAM, while active sessions use 90–130 GB. With planned always-on agents (DAEMON, PROACTIVE, CRON), the baseline is shifting: developer workstations will soon require double-digit gigabytes permanently just for running AI agents. This implies either significantly more expensive hardware or a further shift to the cloud. Additionally, a previously underestimated cost factor is revealed: memory management itself. Features like 'Auto Dream' regularly run background processes that consolidate sessions (e.g., 900+ sessions within a few minutes). A significant portion of inference power is therefore not used for productive user interactions but for maintaining its own memory. Thus, memory creates recursive overhead—more memory functionality requires more computing power, which in turn requires more memory. The dynamics on the model side are also clear: context windows with up to 1 million tokens increase the memory requirement per session manifold. Combined with multi-agent setups (multiple parallel instances per user) and persistent sessions, this creates not a linear but an exponential demand curve for memory. Strategically, this means: the real bottleneck of the next AI wave is not primarily compute—but memory, distributed across servers and clients.

Hyper, Hyper: OpenAI Raises Another $12 Billion

OpenAI has expanded its latest funding round by an additional $12 billion, bringing the total to $122 billion. The valuation remains at $730 billion, as announced in February. SoftBank is leading the round along with Andreessen Horowitz, D.E. Shaw Ventures, and the investment fund MGX from the United Arab Emirates. Nvidia and SoftBank each contributed $30 billion, while Amazon, as a new investor, invested $50 billion. ChatGPT is now used by over 900 million people weekly, and OpenAI generates over $2 billion in monthly revenue. Despite an annual revenue of $13 billion in 2025, the company remains unprofitable and plans to spend $115 billion over the next four years. An IPO could happen as early as the end of this year. → The New York Times

Synthszr Take: A $730 billion valuation on $2 billion in monthly revenue gives a price-to-sales ratio of 30. Amazon is pumping $50 billion into a company that plans to burn through $115 billion in four years. ChatGPT has 900 million weekly users, but OpenAI generates just $2.22 in revenue per user per month. SoftBank and Nvidia are each throwing $30 billion into a furnace that burns money faster than any other private company in history. The planned IPO at the end of the year will show whether Wall Street is willing to finance a company that spends nearly nine dollars for every dollar it earns. OpenAI is building the most expensive machine in the world, and no one knows if it will ever be profitable.

Apple is Late to the Party but Calls the Tune

Apple is supposedly losing the AI race. Google has long shipped agent-based features on Android, OpenAI has 900 million weekly users, and Anthropic's agents are writing compilers. And Apple? It's still delivering notification summaries that hallucinate headlines. But while all the AI labs compete for the best model, Apple is quietly building the platform layer that every model must pass through to do anything meaningful on a smartphone. Not a chatbot, but an agent-based runtime environment directly in the operating system—Apple is positioning itself as the bottleneck between users, AI agents, and every app on the device. Mark Gurman's latest leak about the Siri overhaul ahead of WWDC on June 8th shows: Apple is hardwiring the MCP (Model Context Protocol) directly into iOS. This changes the calculus for every developer and every AI company. → Nate from Nate's Substack

Synthszr Take: Apple is doing exactly what Apple always does: arriving late and then changing the rules of the game. 1.5 billion iPhones in the pockets of the consumers with the most spending power is not a bad starting point for a platform strategy. MCP directly in the OS means every AI app must pass through Apple's bottleneck, whether it's from OpenAI, Google, or the next startup. Google faces an unsolvable dilemma: the Gemini deal looks like a win-win, but the asymmetry clearly favors the one with model control (Spoiler: that's Apple). OpenAI is pouring billions into hardware bets with Jony Ive to build the distribution that Apple already owns. Apple has never lost a platform game.

Amazon to Launch Rival to Musk's Starlink in 2028

Amazon and Delta Air Lines have signed a multi-year agreement that will bring Amazon's satellite technology, Leo, to hundreds of aircraft. Starting in 2028, installation will begin on 500 Delta planes with the goal of offering free high-speed Wi-Fi from gate to gate for all passengers. Delta CEO Ed Bastian emphasizes his airline's global focus and sees the partnership as the best and most cost-effective solution for connectivity. The agreement builds on the existing partnership between AWS and Delta. Amazon Leo promises low latency and high speeds thanks to advanced satellite technology. The Wi-Fi will be available for free to all Delta SkyMiles members. → Tech Brew

Synthszr Take: Amazon is using its satellite infrastructure as a Trojan horse in the aviation market. 500 aircraft is just the beginning; whoever controls the connectivity layer collects data on millions of passengers and their devices. Delta gets free premium Wi-Fi, and Amazon gets a captive-audience test market for future services. Airlines are becoming flying AWS data centers where Amazon not only provides the internet but will eventually offer entertainment, shopping, and cloud services.

The Last Mile Has Many Losers

Meituan has won the Chinese delivery war, but at what cost? The market leader now controls 70% of the food delivery market in China, while competitor Ele.me has shrunk to 28%. The brutal price wars of recent years have secured market dominance for Meituan, but have also exposed the structural limits of its business model. Despite $75 billion in revenue in 2023, margins in the core business remain thin: just a 6.4% operating margin in food delivery. The real insight, however, lies deeper: even with 7 million drivers and artificial intelligence for route optimization, the physical reality of the last mile cannot be optimized away. Meituan is now trying to diversify with new business areas like hotel reservations and local services, while Alibaba's Ele.me continues to lose market share. → Hello China Tech

Synthszr Take: Meituan proves that market dominance in the platform business does not automatically lead to profitability. 7 million drivers, 70% market share, $75 billion in revenue—and still only a 6.4% operating margin in its core business. China's delivery war has brutally exposed the physical limits of scaling: every order needs a person on a motorcycle, no matter how much AI is behind it. Alibaba's Ele.me is continuously losing market share (from 35% to 28%), while Meituan tries to offset thin delivery margins with hotel reservations and local services. Amazon understood this years ago and invented AWS; Meituan is still looking for its second pillar. The platform economy hits its limits when the product cannot be digitized.

Device Code Phishing Becomes a Nightmare for Microsoft

Sekoia's Threat Detection & Research Team discovered a new Phishing-as-a-Service kit called EvilTokens in March 2026, which has been circulating since mid-February 2026. The kit abuses Microsoft's device code authentication—originally intended for smart TVs and IoT devices—to compromise Microsoft 365 accounts. Victims are tricked into entering a code on the legitimate Microsoft login page, giving attackers access to their accounts. In addition to the phishing kit, EvilTokens also offers advanced BEC (Business Email Compromise) features: access weaponization, email harvesting, reconnaissance capabilities, an integrated webmail interface, and AI-powered automation. The service runs via fully-featured Telegram bots and is expected to soon be expanded with Gmail and Okta phishing pages. → TLDR IT

Synthszr Take: Microsoft built an OAuth feature for TVs, and now criminals are using it for mass phishing. Device code authentication was intended for devices without keyboards—a smart TV shows a code, you enter it on your laptop. EvilTokens turns this into a business model: a phishing page shows a code, the victim types it into Microsoft's site, and the attacker gets the token. Telegram bot, AI automation, and BEC tools included—for a few hundred dollars a month. OAuth 2.0 Device Authorization Grant is becoming a security nightmare because Microsoft can't disable the feature without crippling millions of IoT devices.

Marriott Wants to Industrialize Intimacy

With 270 million loyalty program members and nearly 10,000 hotels worldwide, Marriott International is attempting what the hotel industry considers impossible: personal service on an industrial scale. At the Qualtrics X4 conference, Chief Customer Officer Peggy Roe presented the vision of winning over each of the 1.4 billion travelers worldwide as a lifelong customer. The company opens a new hotel every 12 hours globally and now operates over 22,000 restaurants and bars. The strategy relies on real-time reactions: through the Marriott app, on-site staff will receive prompts on how to address individual guest needs. The goal is to create personalized experiences from standardized processes without sacrificing operational efficiency. → TLDR Marketing

Synthszr Take: Marriott is attempting the impossible: making 270 million customers feel like regulars in 10,000 hotels. Peggy Roe talks about a 'lifelong partnership,' but what she means is algorithmic behavioral control of hotel staff via an app. A new hotel every 12 hours, 22,000 restaurants worldwide—at this speed, 'hospitality' inevitably becomes a digital rulebook. Marriott isn't optimizing for real relationships, but for the illusion of them: the concierge only knows your name because the app tells them. Scalable intimacy is a contradiction, but it's this very contradiction that makes Marriott the perfect machine for the Experience Economy.

Xi's Innovation Paradox: Obey Absolutely, but Think Creatively

The Chinese Communist Party has announced new rules for local officials. They are to 'resolutely uphold the authority and centralized, unified leadership of the Central Committee with Comrade Xi Jinping at its core' while also 'taking local conditions into account and enhancing their initiative and creativity at work.' These two mandates from last Friday reveal the central contradiction in Xi Jinping's leadership style: he demands total loyalty and obedience, but also sincerity and innovation. Chinese officials are told to avoid rigid thinking, but at the same time must devote themselves to studying Xi Jinping Thought. They are urged to abandon outdated rules, yet must follow every directive from above. → FP's James Palmer

Synthszr Take: Xi Jinping is banking on a leadership style that Western management theory considers impossible: central doctrine plus local improvisation. The model is reminiscent of franchise systems, where brand standards apply, but the local manager decides how to implement them. China's economic miracle of the past 40 years was based on this very tension: Beijing set the direction, while provincial governors experimented with special economic zones, subsidy models, and regulations. Whether this can still work under tightened ideological control depends on how much leeway 'taking local conditions into account' means in practice. The rules sound contradictory, but China's bureaucracy has historically proven it can withstand productive ambiguity. Xi is betting that loyalty and creativity are not a zero-sum game.