← ALL DISPATCHES

◆ SOURCES BRAIDED

Sources for 2026-04-19

Every tweet, thread, and article cited in today's narration, in reading order.

Tweets

  1. Grok 4.3 research capabilities demonstration

    ns123abc ·

    thought for 13 mins and 17 seconds, created a project folder on its own, produced multiple files, cross-referenced with source verification, 8,365 words

    Extended thinking time is becoming the new frontier - models that can reason for minutes rather than seconds are showing qualitatively different capabilities in research depth.

    Key points

    • 13+ minutes of thinking time on complex research tasks
    • Autonomous project folder creation and file management
    • 8,365 word comprehensive research output
    • Can convert to PowerPoint in one shot
    • Significant capability jump from previous Grok versions

    Open source →x.com/ns123abc/status/2045772621376086142

  2. Qwen3.6-35B-A3B running on M5 Max

    ivanfioravanti — MLX ecosystem developer ·

    mlx_lm server worked flawlessly with Qwen3.6-35B-A3B-8bit, and on M5 Max, the much faster prefill, gives a very pleasant coding experience

    The gap between local and cloud models is rapidly closing - developers can now run Claude-competitive models on their laptops with performance that makes them practical for daily coding.

    Key points

    • Qwen3.6-35B-A3B-8bit running smoothly on M5 Max 128GB
    • Multiple OpenCode instances working concurrently
    • MLX optimizations delivering major speed improvements
    • Local models approaching Claude-level quality
    • 90+ tokens/second on consumer hardware

    Open source →x.com/ivanfioravanti/status/204585332881951…

  3. Robin Hanson on AI wave patterns

    robinhanson — George Mason economist, author of 'The Age of Em' ·

    We have seen waves of interest & activity in AI/automation every ~25-30yrs for ~ century. Each based on new tech w/ real but limited impact. I predict current wave will be like past waves.

    Hanson's contrarian view offers important perspective - while the capabilities are undeniable, the economic transformation may still follow historical patterns of overpromise and correction.

    Key points

    • AI waves repeat every 25-30 years historically
    • Each wave has real but limited impact
    • Current wave comparable to dotcom boom
    • Eventually a wave may be transformative
    • Skeptical that this is the transformative wave

    Open source →x.com/robinhanson/status/2045949985439592538

Articles

  1. Humanoid robot breaks half marathon world record in Beijing

    Al Jazeera ·

    A humanoid robot from Chinese smartphone maker Honor completed the half marathon in 50 minutes and 26 seconds

    This marks a watershed moment for embodied AI - robots aren't just thinking better, they're moving better than humans in endurance tasks that have defined our species for millennia.

    Key points

    • Robot beat human world record of 57:20 by nearly 7 minutes
    • Average speed of 25km/h using autonomous navigation
    • Massive improvement from 2:40:42 in 2025 inaugural race
    • 112 teams participated with 40% using autonomous navigation
    • Fitted with 90-95cm legs mimicking elite human runners

    Open source →www.aljazeera.com/sports/2026/4/19/humanoid…

  2. Vercel April 2026 security incident

    Vercel Security Team — Infrastructure platform powering Next.js and millions of deployments ·

    The incident originated from a small, third-party AI tool whose Google Workspace OAuth app was the subject of a broader compromise

    As AI tools proliferate with OAuth permissions across enterprises, supply chain attacks through compromised integrations become the new attack vector - this won't be the last incident like this.

    Key points

    • Third-party AI tool Context.ai OAuth app compromised
    • OAuth ID: 110671459871-30f1spbu0hptbs60cb4vsmv79i7bbvqj.apps.googleusercontent.com
    • Affected hundreds of users across many organizations
    • Environment variables not marked as sensitive potentially exposed
    • Google has deleted the malicious account

    Open source →vercel.com/kb/bulletin/vercel-april-2026-se…

  3. Changes in the system prompt between Claude Opus 4.6 and 4.7

    Simon Willison — Creator of Django, prolific AI tools developer ·

    Claude keeps its responses focused and concise so as to avoid potentially overwhelming the user with overly-long responses

    Anthropic's transparency about system prompts reveals the subtle engineering required to shape model behavior - each iteration shows how human preferences are encoded into AI systems.

    Key points

    • New Claude in Chrome, Excel, and PowerPoint agents
    • Enhanced child safety with persistent caution after refusals
    • Less pushy behavior when users want to end conversations
    • Tool search mechanism to check for available capabilities
    • Removed restrictions on asterisk emotes and certain words

    Open source →simonwillison.net/2026/Apr/18/opus-system-p…

  4. Making illegal state unrepresentable

    Nicolas Fränkel ·

    As a developer, I want to make illegal states unrepresentable, i.e., users of my API can't create non-existent transitions

    As AI writes more code, type systems that make invalid programs uncompilable become crucial - the compiler becomes the first line of defense against AI hallucinations.

    Key points

    • Static typing prevents illegal state transitions at compile time
    • Phantom types enable type-safe builders without runtime overhead
    • Kotlin, Rust, and Gleam handle this elegantly with extension functions
    • Java lacks the language features for clean implementation
    • Type systems are becoming more sophisticated for API safety

    Open source →blog.frankel.ch/illegal-state-unrepresentab…