Top Stories

Ask HN: What was your “oh shit” moment with GenAI?

630 points · news.ycombinator.com

A simple prompt turned into one of the day’s liveliest threads, with developers swapping the moments they realized generative AI had genuinely changed their work. The stories range from agents one-shotting tasks that used to take days, to unsettling experiences where a model produced something its user couldn’t fully explain. It’s a useful temperature check on where practitioners actually stand in 2026 — beyond the hype cycle, what made people sit up.

Meta confirms thousands of Instagram accounts were hacked by abusing its AI chatbot

609 points · this.weekinsecurity.com

Attackers found a way to weaponize Meta’s own AI assistant to compromise thousands of Instagram accounts, and the company has now confirmed the breach. The HN crowd is dissecting it as a cautionary tale about bolting LLM agents onto production systems with real account privileges — the attack surface of “an AI that can take actions on your behalf” is exactly the surface adversaries probe first. Expect this to become a reference case in arguments about agent guardrails.

Ntsc-rs – open-source video emulation of analog TV and VHS artifacts

351 points · ntsc.rs

This Rust project faithfully recreates the noise, color bleed, and tracking errors of analog TV and VHS, and it struck a nostalgic nerve. Beyond the aesthetic appeal, commenters appreciated the engineering — accurately modeling NTSC signal degradation is genuinely hard, and doing it fast enough for real-time use is harder. It’s the kind of lovingly over-engineered tool HN tends to celebrate.

Pokémon Emerald ported to WebAssembly (100k FPS)

324 points · pokeemerald.com

Building on the community’s decompilation of Pokémon Emerald, someone compiled the game to WebAssembly and clocked an absurd 100,000 frames per second. The headline number is a flex, but the real story is how far the reverse-engineering and WASM toolchains have come — running a native GBA title in the browser at that speed would have been unthinkable a few years ago. Pure catnip for the retro-computing and compiler crowds alike.

Moving beyond fork() + exec()

306 points · lwn.net

LWN digs into long-running frustrations with Unix’s venerable fork()/exec() process-creation model and the kernel efforts aiming to replace it with something cleaner and safer. fork() is elegant on paper but a minefield in multithreaded programs, and the discussion gets into the gritty trade-offs of posix_spawn, clone3, and friends. It’s a meaty systems-programming read that drew a thoughtful, opinionated thread.

Nvidia is proposing a beast of a CPU system for Windows PCs

288 points · twitter.com

Nvidia is reportedly pushing an ambitious CPU-plus-system design for Windows machines, signaling deeper ambitions beyond GPUs and into the heart of the PC platform. Commenters debated what an Nvidia-architected Windows system would mean for the x86 incumbents and for the ARM-on-Windows transition. It’s another data point in the company’s march to own ever more of the compute stack.

Google to pay SpaceX $920M a month for compute capacity at xAI data centers

254 points · cnbc.com

In one of the stranger deals of the AI buildout, Google is reportedly paying SpaceX nearly a billion dollars a month for compute at xAI’s data centers. The tangled web of rivals-renting-from-rivals captures just how desperate the demand for AI infrastructure has become — even fierce competitors are striking capacity deals to keep their models fed. The scale of the number alone made it the day’s top business story.

Zeroserve: A zero-config web server you can script with eBPF

238 points · su3.io

Zeroserve reimagines the humble web server by letting you script its behavior with eBPF, pushing request handling logic into the kernel for speed and flexibility. The “zero-config” pitch plus the eBPF angle made it irresistible to the systems crowd, who debated whether kernel-level scripting is brilliant or a footgun. Either way, it’s a fresh take on infrastructure that’s usually treated as a solved problem.

Harness engineering: Leveraging Codex in an agent-first world

207 points · openai.com

OpenAI lays out its philosophy for “harness engineering” — designing the scaffolding around coding agents like Codex so they can work effectively rather than just answering prompts. As development shifts toward agent-first workflows, the argument is that the harness (tools, context, feedback loops) matters as much as the model. It landed amid an active HN debate about how much of software engineering is becoming agent orchestration.