VoidZero joins Cloudflare to steward Vite ecosystem; Anthropic AI writes 80% of its own code; Meta glasses hide dormant face recognition

Anthropic

When AI Builds Itself: Anthropic's Progress Toward Recursive Self-Improvement

Anthropic reports that Claude now writes over 80% of production code at the company and is executing research tasks autonomously, marking a significant step toward AI systems that meaningfully improve their own development cycle. The post argues this trend could dramatically accelerate AI progress but also introduces compounding risks around control and alignment that current governance frameworks were not designed to address. Anthropic calls for international coordination mechanisms before recursive self-improvement loops outpace human oversight capacity.

Buchodi

Meta's Stella App Contains a Complete, Dormant Facial Recognition System

Security researcher Buchodi reverse-engineered Meta's Stella companion app for smart glasses and found a fully built facial recognition pipeline—three ML models, a SQLite database for 2048-dimension face embeddings, and UI elements for displaying recognized people—all present but currently disabled for standard users. The discovery reveals significant engineering investment in infrastructure that could be activated at any time via a server-side flag, without a new app release. The finding intensifies concerns about the privacy implications of always-on cameras in wearable form factors.

Electrek

Wind and Solar Generated More Power Than Gas Globally for the First Time

In April 2026, wind and solar together produced 22% of global electricity, surpassing natural gas at 20%—a first in recorded history. Output from renewables has more than doubled over the past five years, driven by rapid expansion across China, the EU, the UK, the US, Australia, and Brazil. The milestone signals that the energy transition has passed an inflection point where clean sources now set the pace of global power generation.

404 Media

Google Employees Internally Share Memes About How Its AI Sucks

While Google CEO Sundar Pichai publicly claimed that 75% of new code at the company is AI-generated, internal communications obtained by 404 Media show engineers privately sharing memes that mock the same AI coding tools they are expected to use. Employees report the tools make their work harder rather than easier, exposing a sharp disconnect between executive messaging on AI productivity and engineers' lived experience. The story echoes broader industry tensions between optimistic AI adoption narratives and the frustration of developers actually working with the tools.

GitHub / Anthropic

Anthropic Open-Sources AI-Powered Vulnerability Discovery and Remediation Harness

Anthropic published a reference implementation for autonomous vulnerability discovery and patching that runs a seven-stage Claude agent pipeline inside sandboxed containers—scanning source code, verifying bugs, and generating fixes without human intervention. The system targets C/C++ memory safety vulnerabilities but is designed to be adapted to other languages and vulnerability classes. The release is paired with a managed production service called Claude Security, positioning AI-driven static analysis as a practical complement to traditional security tooling.

Privacy Guides

South Korea Mandates AI Image Scanning for All Online Forums Starting July 1

A new South Korean law taking effect July 1, 2026 requires operators of online communities to independently acquire enterprise-grade GPUs and run AI models to scan every user-uploaded image and video for illegal content. Critics report the system is flagging swimwear photos, anime artwork, and memes, placing impossible compliance costs on small forum operators while raising serious questions about censorship scope and due process. The regulation is being watched internationally as a test case for government-mandated AI content moderation at the infrastructure level.

GitHub / Alibaba

Alibaba Open-Sources Enterprise-Grade AI Code Review CLI

Alibaba released Open Code Review, an open-source CLI tool that runs LLM agents over Git diffs to produce structured, line-level review comments covering bugs, security issues (NPE, XSS, SQL injection, thread safety), and code quality. The tool supports multiple model providers including OpenAI and Anthropic, integrates into CI/CD pipelines, and was validated at scale across tens of thousands of Alibaba engineers before open-sourcing. It bridges deterministic engineering logic with LLM reasoning to keep review comments precise and actionable rather than generic.

htmx.org

Code is Cheap(er): The Sorcerer's Apprentice Problem for AI-Assisted Development

As AI tools slash the marginal cost of generating code, the htmx author argues that developer value is shifting from writing code to constraining it—exercising judgment about what not to build, evaluating AI output critically, and keeping systems comprehensible to humans. The essay warns that blindly accepting AI-generated code creates a compounding complexity trap where no one understands the system and maintenance becomes unsustainable. The prescription is a subtractive engineering mindset: measure developer success by how little code survives review, not how much gets shipped.

arXiv

Sharing QKV Projections Cuts Transformer KV Cache 96.9% with Minimal Quality Loss

A new ICML 2026 paper asks whether transformer attention mechanisms actually need separate query, key, and value projections—and finds that sharing key-value projections alone achieves 50% KV cache reduction with negligible performance degradation. When combined with existing multi-query and grouped-query attention techniques, the approach reaches 96.9% KV cache reduction, making previously impractical model sizes viable on edge hardware. The findings suggest that standard QKV architecture carries significant redundancy that can be systematically removed without retraining from scratch.