Chrome 150 kills popular ad blockers on June 30; open-weight GLM-5.2 reaches frontier-level coding; Apple undermines Hide My Email privacy

Latent Space

GLM-5.2: the Top Frontend Coding Model in the World

Zhipu AI released GLM-5.2, a 744B-parameter MIT-licensed open-weight model that ranks #2 overall in frontend coding benchmarks and #1 in design arena tasks — beating all closed proprietary models except Claude Fable 5. The model features a usable 1M-token context window, two reasoning-effort modes, and technical innovations including IndexShare sparse attention (claiming 2.9× lower per-token FLOPs at long context) and improved multi-token prediction for speculative decoding. The release is being hailed as the strongest open-weight coding and agentic model yet, with same-day ecosystem support across vLLM, SGLang, Ollama, and major cloud providers.

Arseniy Shestakov

Apple Is About to Make Hide My Email Useless

Apple is moving Hide My Email aliases to a distinctive @private.icloud.com subdomain, which will allow websites to trivially identify and block these addresses — defeating the feature's entire purpose. The distinct domain removes plausible deniability and lets services reject Apple's privacy aliases outright, similar to how they already block disposable email providers. The author urges Apple to reconsider, and advises current users to generate additional @icloud.com aliases before the change takes effect.

The Pragmatic Engineer (Gergely Orosz)

Why Is Meta Destroying Its Engineering Organization?

Meta has forced 30–50% of its engineers into AI data labeling roles, implemented pervasive keystroke tracking, and conducted layoffs despite record profits — all while product reliability visibly degrades. The reorganization has demoralized the engineering workforce and is eroding the cultural autonomy that made Meta's engineering teams exceptional. The piece examines how the AI pivot is treating engineers as a cost input for model development rather than a driver of product quality.

Sam Sch (GitHub Gist)

Stop Using JWTs

This widely-circulated essay argues that JWTs are the wrong choice for web session management: they were designed for short-lived delegation tokens, not persistent sessions; they still require a backend data store to support logout and invalidation; they are less efficient than simple session cookies; and the JWT specification itself has well-documented security flaws. The recommendation is straightforward — use HTTP-only cookie-based sessions for browser applications and reserve JWTs only for the cross-service delegation use cases they were designed for.

Dexerto

Stop Killing Games Fails to Secure EU Law Despite 1.3M Signatures

The Stop Killing Games initiative, which collected over 1.3 million European signatures calling for legislation requiring publishers to keep discontinued games playable, was rejected by the European Commission as disproportionate. Rather than binding rules, the Commission will facilitate voluntary industry discussions and a voluntary code of conduct by year-end 2026. The campaign will now pursue amendments through the Digital Fairness Act in the European Parliament.

Martin Fowler

Building Reliable Agentic AI Systems

This case study describes PRINCE, a production agentic AI system built with Bayer for preclinical drug research, which evolved from keyword search into a multi-agent system capable of answering complex questions over decades of PDF reports and drafting regulatory documents. The architecture centers on context engineering (what goes into the LLM) and harness engineering (how agents are orchestrated), using RAG, structured citations, and transparency mechanisms to maintain scientific trustworthiness. Key lessons include prioritizing robustness through structured orchestration and iterating based on real researcher feedback rather than benchmark performance.

Argus Systems

A 27-Year-Old Authentication Bypass in OpenBSD's PPP Stack

A critical vulnerability in OpenBSD's PPP daemon allowed attackers to bypass PAP password authentication for 27 years by sending zero-length credentials, exploiting a comparison that treated missing password fields as a match. The flaw enabled unauthorized PPPoE connections and heap memory disclosure. The vulnerability was discovered in June 2026 and patched within days by adding proper length validation.

Stephen Wolfram

Launching Version 15 of Wolfram Language / Mathematica

Wolfram Language Version 15 ships a built-in AI Assistant available in every notebook, alongside a unified ModelFit function, enhanced time series and categorical data frameworks, and new computational music support. Infrastructure upgrades add support for gigabyte-scale files and improved notebook capabilities. The release deepens the integration between Wolfram's symbolic computation engine and LLM-powered AI, positioning the AI assistant as a first-class citizen rather than an optional plugin.

Google DeepMind

Unlocking UK House-Building with AI-Accelerated Planning

Google DeepMind has partnered with the UK government to build an AI planning prototype using Gemini that aims to halve the time required to process homeowner planning applications by automating data extraction, policy identification, and report drafting. Successful trials in Barnet, Camden, and Dorset are informing a planned nationwide rollout in 2027. The system keeps humans in control of final decisions while targeting the government's goal of building 1.5 million homes by 2029.