VSCode bug, Swift boost, GenAI myths

The post walks through constructing a tiny std::any by starting with classic polymorphism via virtual interfaces and template-based generic code, then shows how type erasure hides concrete types behind a uniform wrapper. It explains the underlying mechanics and why this technique lets unrelated objects be handled uniformly.

A new disjunction selection heuristic dramatically reduces the time the Swift compiler spends solving complex type constraints. The change, detailed in Michael Tsai’s roadmap, enables the solver to prune dead‑end paths and solve many expressions that previously timed out, speeding up builds for real‑world code.

React Native Runtimes adds support for multiple named JavaScript runtimes, letting components and functions run on secondary threads while the main UI thread stays free. It ships two libraries, @react-native-runtimes/core for runtime composition and @react-native-runtimes/state for synchronized native-backed state, enabling smoother lists, chat, sync, and crypto workloads.

A new vulnerability in VSCode’s webview security model enables an attacker to steal a user’s GitHub OAuth token via a single malicious link, giving full access to private repositories. The flaw stems from keyboard event bubbling between isolated webviews and the main window. Ammar Askar’s blog details the exploit and mitigation steps.

A new ACM Queue analysis and a 2025 Microsoft study show developers spend only ~14% of their day actually writing code, undermining claims that AI coding assistants can massively boost productivity. The research also warns that AI‑generated code often requires extensive debugging, so real gains need broader workflow integration.

Recent breakthroughs in mechanistic interpretability let researchers decompose large language model activations into human‑readable concepts, revealing multi‑step, causal reasoning pathways. By tracing sparse feature circuits, teams can map internal logic, paving the way for safer steering and deeper algorithmic insights.

In a candid blog post, George Hotz argues that integrating AI coding agents into software development will create massive, hard‑to‑detect technical debt. While he acknowledges their utility for quick prototypes, he warns that large organizations risk costly cleanup as the generated ‘slop’ accumulates.

By replacing the GPU‑dependent egui (which pulls in wgpu) with the CPU‑rendered Slint library, the author reduced the rproc system monitor's Rust binary from ~135 MB to ~30 MB, a 4.5× size drop. The rewrite highlights how heavyweight GUI dependencies can bloat memory even when the core logic is lightweight.

Benchmarking reveals that nullable columns add a null‑map that must be scanned on every filter, inflating I/O and cutting query throughput. Because ClickHouse stores data column‑wise, the extra map disrupts vectorized processing. The article explains why ClickHouse avoids NULLs by default and recommends schema alternatives to preserve performance.

The blog explains that building new software is essentially learning, so you can’t fully specify requirements up front. By embracing rapid prototyping, questioning, and quick iteration, teams cut the gap between “let’s try” and reality, accelerating delivery of features.

A Linux profiler caused system freezes due to eBPF spinlock bugs in the kernel. The author traced the issue to long‑running NMI handlers, patched multiple kernel spinlock problems, and contributed the fixes upstream. The result eliminates the freezes and improves eBPF stability for profiling tools.

The essay explains that while we can ship code faster, trust erodes instantly and is hard to restore. It shows how XP practices, testing, pairing, CI, planning, customer collaboration, continuous deployment, refactoring, and observability, act as a "trust factory" that continuously builds confidence within engineering teams.

The post introduces the Light Cone Consistency model, showing how distributed systems can blend multiple consistency guarantees despite physical latency limits. It explains the underlying relativistic intuition, details the approach, and links to a new paper formalizing the concept, offering engineers a practical framework for building consistent, low‑latency services.

In a Pragmatic Engineer podcast, Kelsey Hightower recounts his rise from self‑taught technician to Google Distinguished Engineer, describing how open‑source work and the Kubernetes boom shaped his career. He also reflects on his decision to retire, the impact of AI, and why technology must serve people.