React 19.3 adds a react-dom browser() API that lets components bail out of server rendering cleanly, suspending on the server instead of relying on mounted flags or typeof window checks. It's Canary-only for now, but it already works across any SSR setup, not just Next.js or Astro's bundler-specific escape hatches.
— via InfoQ, Daniel Curtis
GitLab 19.4 lets Duo CLI agents chase an open-ended goal instead of following fixed prompts, with one model doing the work and another checking it against that goal. New hosted models promise up to four times more calls per credit, but GitLab hasn't published benchmarks to back the comparison.
— via InfoQ, Matt Saunders
Xcode 27.2 ditches the old property-list project file for a JSON format designed to cut merge conflicts and let AI agents read and edit project settings directly. iOS developers calling it overdue say it could have saved "years" of pain, but check your tools first: fastlane, CocoaPods and Tuist need to support it before you convert.
— via InfoQ, Sergio De Simone
GPT and Claude can read Georgian fine. Producing it is where they fall apart, especially with numbers, according to a Kaggle benchmark testing both on one of the world's hardest grammars.
— via dev.to, Giorgi Kobaidze
Letting language models edit their own context, rather than relying on fixed summarization or retrieval rules, pushed accuracy up 11.4% while cutting compute 21.5% on one benchmark. Reinforcement learning drove even bigger gains, but researchers warn the models can still delete information they later need.
— via InfoQ, Sergio De Simone
AWS rebuilds EventBridge's custom event bus from scratch, adding ordered delivery, replay and a publisher/subscriber pricing split meant to kill the multi-bus workarounds teams built to dodge routing costs. AWS calls it the biggest launch since 2019; early reviews call the pricing steep but fair.
— via InfoQ, Renato Losio
Gremlin's new Foresight AI hunts for weaknesses in live services, proposes fixes, then reruns the original test to prove they worked, drawing on a decade of fault-injection data rather than an LLM trained on customer systems. It won't run a test or apply a change without human approval.
— via InfoQ, Matt Saunders
Why do AI changelog generators keep inventing reasons that were never written? One engineer's fix treats "why" as a selection problem, not a generation one, scanning pull requests for rationale markers before ever asking a model.
— via dev.to, nyaomaru
Cloudflare now traces the entire request path, not just Workers, turning blocked rules, rewrites and origin delays into OpenTelemetry spans. A pricing model switch, the second in two months, now bills by data ingested and stored rather than per span.
— via InfoQ, Steef-Jan Wiggers
Google's new Android Bench 2.0 ditches pass/fail grading for AI coding tests and adds week-long tasks that mimic real engineering work. The top model still only clears 32% of them, and porting apps across platforms remains an unsolved problem even for the best performer.
— via InfoQ, Sergio De Simone
A SpaceX training lead who once onboarded nearly every new hire says the jump from top engineer to manager is still one of tech's worst-handled transitions. His fix: borrow nothing wholesale from giants like Amazon, because what works there reliably fails everywhere else.
— via InfoQ, Matt Gjertsen
CircleCI's Machine Runner Orchestrator hits general availability, letting self-hosted CPU, GPU and ARM runners scale from one deployment across GKE, AKS, EKS and KubeVirt. The catch: preview users can't upgrade in place and must uninstall and reinstall under the product's new name.
— via InfoQ, Craig Risi
