× + newsletter-2026-04-22
╔═════════════════════════════════════════════════════════════╗
║                                                             ║
║   ▀█▀ ▄▀█ █▄▀ █▀▀   █ ▀█▀   █▀▀ █▀█ █▀█ █▀▄▀█             ║
║    █  █▀█ █ █ ██▄   █  █    █▀  █▀▄ █▄█ █ ▀ █             ║
║                                                             ║
║   ▀█▀ █ █ █▀▀   ▀█▀ █▀█ █▀█                                ║
║    █  █▀█ ██▄    █  █▄█ █▀▀                                ║
║                                                             ║
╚═════════════════════════════════════════════════════════════╝
    

// Tech news at terminal velocity

cat 2026-04-22.md

SpaceX's $60B Cursor Bet, Meta's Keystroke Harvesting, and Anthropic's Leak

$ cat TLDR.md

• SpaceX secures a wild $60B option to acquire AI coding darling Cursor.

• Meta mandates a controversial new policy to record employee keystrokes for AI training.

• Anthropic's 'dangerous' Mythos cyber tool gets leaked to unauthorized users.

Headlines & Launches

🚀 SpaceX's $60B Cursor Option

In a move that screams 'Elon synergy,' SpaceX has secured an option to acquire AI coding platform Cursor for a cool $60 billion (or pay a $10B fee). It's a massive play to bolster xAI's developer tooling against OpenAI and Anthropic.

Source: The Verge

⌨️ Meta's Internal AI Training Gets Creepy

Meta is rolling out a mandatory internal tool that converts employee mouse movements and keystrokes into training data for its AI models. Unsurprisingly, staff are absolutely thrilled about becoming literal training fodder.

Source: TechCrunch

🔓 Anthropic's Cyber Tool Slips Out

Mythos, Anthropic's highly restricted cybersecurity AI, was accessed by an unauthorized group via a third-party contractor. The irony? The model is deemed 'dangerous in the wrong hands' but apparently quite easy to hand over.

Source: The Verge

🎧 Anker Puts AI in Your Earbuds

Anker just announced the 'Thus' processor, a custom neural-net compute-in-memory chip designed to run local AI on audio and IoT devices. Because even your charging cables need a neural network now.

Source: The Verge

Deep Dives

🧠 Why LLMs Are Confidently Incorrect

A fascinating paper in Nature Machine Intelligence explores the paradox of LLM confidence. Models exhibit a 'choice-supportive bias' where they stubbornly stick to initial answers, yet simultaneously overweigh contradictory feedback. It's basically the AI equivalent of human stubbornness.

Source: Nature Machine Intelligence

🛡️ Cloudflare Rethinks Bot Detection

As AI agents become indistinguishable from humans, traditional bot detection is breaking down. Cloudflare proposes a new model using anonymous credentials to preserve privacy while keeping the internet functional for both meatbags and silicon.

Source: Cloudflare Blog

🛰️ The Magic of GPS Explained

A brilliant, accessible deep dive into the math and physics behind the Global Positioning System. It's a great reminder that the blue dot on your phone relies on atomic clocks and relativity to keep you from getting lost on the way to Starbucks.

Source: Per Thirty Six

Engineering & Research

🕸️ A Browser Built for Bots

Meet Obscura, a Rust-based headless browser designed specifically for AI agents. Unlike memory-hogging Chromium wrappers, it runs on just 30MB of RAM and includes built-in stealth mechanisms for autonomous web scraping.

Source: PyShine

🔖 IBM's Meta-Token Memory Trick

IBM Research introduces 'meta-tokens' injected during pre-training to help LLMs capture distant contextual information. By sharpening positional encoding, these tokens act as content anchors, effectively caching preceding context and doubling the reliable context window.

Source: IBM Research

💬 Deploying Agents to Group Chats

Photon just open-sourced Spectrum, a TypeScript framework that deploys AI agents directly into iMessage, WhatsApp, and Telegram. It solves the distribution problem by putting agents where users actually live instead of behind a developer dashboard.

Source: MarkTechPost

Odds & Ends

🪟 WSL Gets a Retro Cousin

Someone actually built a Windows 9x Subsystem for Linux. Because why run modern software when you can natively execute 32-bit nostalgia directly in your terminal?

Source: Hacker News

🧹 From Vacuums to World Domination

Chinese robot vacuum maker Dreame dropped $10 million on a Super Bowl ad to kick off its plan to become a global consumer electronics giant. Watch out, Dyson.

Source: The Verge

← Back to archive