Top Stories

Open-source engine runs Gemma 4 26B in 2GB of RAM on any M-series Mac

823 points · github.com

A new open-source inference engine, turbo-fieldfare, claims to run Gemma 4’s 26B-A4B model in roughly 2GB of RAM on any M-series MacBook — a striking result given that models this size normally demand far more memory. The trick is aggressive quantization plus a mixture-of-experts architecture that only activates a small slice of parameters per token. For the HN crowd, this is catnip: it pushes capable local LLMs onto everyday laptops, no cloud bill and no data leaving the machine.

Superlogical wants to be the “multiplexer for all work”

712 points · superlogical.com

Superlogical is pitching an ambitious idea: a single system that unifies local development, remote hosts, sandboxes, coding agents, background jobs, and production debugging into one interface. The premise is that software work is now fragmented across humans, CI, and parallel AI agents, each living in its own silo of tools and logs. The vision resonated on HN because it names a real pain point — as agents proliferate, nobody has a coherent place to watch and steer all of it at once.

The coolest use for the Vision Pro

700 points · christianselig.com

Christian Selig (of Apollo for Reddit fame) writes up an unexpectedly delightful use for Apple’s Vision Pro: capturing and revisiting spatial memories of a house. It’s the kind of concrete, personal application that cuts through the “what is this headset even for” skepticism that has dogged the device. HN readers, many of whom remain Vision Pro skeptics, engaged heavily because it’s a rare argument grounded in genuine emotional payoff rather than spec sheets.

AI’s top startups are barely publishing their research

493 points · science.org

Science reports that the leading AI labs have sharply curtailed how much of their research they publish, trading academic openness for competitive secrecy. The shift marks a real cultural break from the field’s roots, where landmark papers like “Attention Is All You Need” came straight from industry. HN’s discussion cut to the tension at the heart of modern AI: the same companies built on openly published breakthroughs are now pulling up the ladder behind them.

Kimi K3-256k ships with a giant context window

449 points · kimi.com

Moonshot’s Kimi has released K3 with a 256k-token context window, continuing the relentless push from Chinese labs to match and undercut frontier Western models. Long context is especially valuable for coding agents that need to hold entire repositories in view, which is exactly the use case Kimi is targeting. The launch fed HN’s ongoing fascination with how quickly the open and semi-open model ecosystem is closing the gap on the big US labs.

Anatomy of a frontier lab agent intrusion

395 points · huggingface.co

A detailed technical timeline reconstructs a July 2026 security incident in which an AI agent was used as part of an intrusion at a frontier lab. It’s a rare, concrete look at what agentic attacks actually look like in practice rather than in threat-model abstractions. The write-up landed hard on HN because it validates a fear the community has been circling for a while: autonomous agents don’t just get exploited, they can become the exploit vector.

Keychron ships the first open-source firmware for gaming mice

390 points · digitalfoundry.net

Keychron, already beloved among mechanical-keyboard enthusiasts for its QMK/VIA support, is extending open-source firmware to gaming mice — a category still dominated by clunky, closed vendor apps. For the tinkerer-heavy HN audience, open firmware means full control over DPI, polling, and remapping without proprietary bloatware, plus the promise that the hardware outlives the vendor’s software support. It’s a small but genuine win for the right-to-tinker crowd.

AI companies are recruiting electricians and carpenters by the thousands

293 points · nytimes.com

The New York Times reports that the data-center buildout behind the AI boom is driving massive demand for skilled tradespeople — electricians, carpenters, and the like — with companies funding training programs to fill the gap. It’s a vivid reminder that “the cloud” is ultimately concrete, copper, and cooling. HN found the piece a useful corrective to the abstraction of AI hype: the bottleneck increasingly isn’t GPUs or talent, it’s the physical grid and the people who wire it.

The Productivity Mirage

262 points · frantic.im

This essay argues that many of the productivity gains we claim from new tools and workflows are illusory — motion mistaken for progress. It’s a reflective take that questions whether the endless churn of optimization actually moves the needle or just feels like it does. The post struck a nerve on HN, where developers are perpetually torn between building things and building better ways to build things, and the AI-coding era has only sharpened the anxiety.