Software engineering end to end: languages, data, infra, and security, with AI wherever it enters the work of people who write code. Always with code and examples, never just the news.
Five successive changes to the in-memory data representation, from Vec to a wire-format byte buffer, reduced the size of each cache entry by 56% and freed up about 100 TB of RAM across Cloudflare's DNS resolver fleet.
Released in September 2025, Postgres 18 brought an asynchronous I/O subsystem that directly targets the disk bottleneck in sequential scans, bitmap heap scans, and vacuum. Here's how io_method works and how to validate each mode before switching in production.
Vercel Labs' experimental project uses TypeScript's own compiler to generate executables in C, LLVM IR, or WebAssembly, promising near-instant startup, but the community has already pointed out serious compatibility and performance limits.
virtio-nvgpu forwards Nvidia driver ioctls between guest and host instead of translating API calls, reaching 98% of bare-metal performance under heavy rendering loads.
Google has launched Call for Me, an experimental feature that lets Gemini call businesses on the user's behalf. For those building assistants and chatbots, the case is a design study on where AI should decide and where it needs to stop.
The Swedish startup, which became synonymous with 'programming by intent,' jumped from $500 million to $600 million in ARR in three months and says two-thirds of the Fortune 500 already use its platform.
A survey by research organization Epoch AI, highlighted by researcher Aymeric Roucher, compares the price drop in artificial intelligence with that of other transformative technologies in history, and puts it well ahead.
The chip brings gigabit Ethernet, USB host, Sv32 paging, and already runs community-made ports of Linux 6.18 and 7.1, narrowing the gap between a microcontroller and an SBC like the Raspberry Pi.
Transformers can now read GGUF files directly from the Hub, but the memory savings llama.cpp delivers by default only show up today for Qwen3.5 models and depend on a specific kernel. Outside of that, what you gain is convenience, not RAM savings.
A case reported by Stefan Fercot shows that promoting a standby with archive_mode turned off doesn't stop pgBackRest from completing a backup, but it can leave recovery without a path between timelines.
The open-source profiling tool for TPU workloads gained a suite that exposes cycle-by-cycle hardware counters inside Pallas kernels, previously black boxes in traces.
Ten years after Postgres 10, every release closed a gap that used to require Slony, Londiste, or pglogical. But data architects still have to decide on their own how to avoid ID collisions, silent conflicts, and stuck transactions.
OpenAI expands the GPT-6 family with two models cheaper than Astra, promising near-flagship performance at a fraction of the cost. The numbers check out, but the comparison methodology deserves a close read.
A pointer cast that has always worked in practice can be undefined behavior and disappear from the binary as soon as the compiler turns on aggressive optimizations. Understand why this is even more treacherous in C++.
The admission controller that replaced PodSecurityPolicy has been stable since Kubernetes 1.25, but configuring the privileged, baseline, and restricted profiles per namespace still breaks workloads that weren't audited before enforcement.
The company behind Linear explains how it cut CI wait time and runner cost after code agents started multiplying the volume of pull requests. The account, published on September 21, is a practical playbook for anyone facing the same bottleneck.
SpyCloud research found stolen credentials from nearly two in ten US water and wastewater utilities, including access to networks that control physical pumps and valves.
A dedicated personal settings page and a new enterprise-level review effort default arrive for all Copilot plans, including Business.
At Connect 2026, the company introduced smart glasses that swap the camera for high-quality audio, trying to escape the "pervert glasses" nickname that video-recording models had earned.
The India-American startup chose the niche that vibe-coding generalists left on the table: mission-critical code in C and C++ for defense, aerospace, automotive, and HFT. The check is small, but the thesis is revealing.
The 9B-parameter model, published on Hugging Face on September 21, 2026, turns text into compressed images and only decodes at high resolution the passages the user's question actually needs.
Andrew Kelley, creator of the Zig language, detailed in an interview with JetBrains why the project rejects AI-generated code and why it moved its infrastructure away from GitHub.
Open source project reverse-engineers the path Intel never opened: writing C directly for the NPU's ACT-SHAVE cores, bypassing the official graph compiler.
New model matches the performance of Claude Fable 5.1, outperforms GPT-5.6 Sol in a software development benchmark, and arrives on AWS, Google Cloud, and Microsoft Azure clouds with a reduced price.
AX treats autonomous agents as stateful actors, not as microservices or batch jobs, and uses Kubernetes-style CRDs to suspend and resume tasks in under a second.
An experiment documented by Max Woolf shows how well-designed prompts lead Claude Opus, GPT-5.3 Codex and GPT-6 Astra to rewrite Rust crates up to 20x faster, but it also exposes how the agents cheat if you leave a gap in the benchmark.
The GitLab Duo Self-Hosted extension allows the use of GPT, Claude, Llama, and Mistral hosted within the customer's own Azure environment, without sending prompts and code to GitLab-managed infrastructure.
A survey from the Now-Next blog measured what was missing from this debate: how much it costs to restore a single tenant on its own, and at exactly what point schema-per-tenant breaks backup.
TypeSafe AI's model doesn't generate interface: it chooses among options you define. This changes what it means to maintain a design system.
Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, and Grok 4.5 go offline across all Copilot experiences. Anyone who hardcoded a model name in a script, extension, or CI pipeline needs to act before the deadline.
End of the 29 most recent news stories. Subscribe to the RSS feed to follow along in real time.
