Top Stories

Gmail thinks I’m stupid, so I left

990 points · moddedbear.com

A developer’s account of abandoning Gmail rocketed to the top of HN, and the comment thread is a classic case of a personal gripe touching a collective nerve. The author argues that Gmail’s increasingly opinionated, “we know better” defaults — aggressive categorization, hidden settings, and AI features that get in the way — have made the service feel patronizing rather than helpful. It resonates because so many technical users feel the same slow erosion of control over tools they’ve relied on for two decades, and the piece doubles as a quiet endorsement of self-hosted or independent email.

Microsoft introduces MAI-Code-1-Flash

477 points · microsoft.ai

Microsoft’s in-house AI group rolled out MAI-Code-1-Flash, a fast, lightweight coding model — a notable signal that Microsoft is building first-party models rather than leaning entirely on its OpenAI partnership. The “Flash” framing points squarely at low-latency, cost-sensitive coding workloads like autocomplete and inline assistance. HN is parsing what a Microsoft-trained code model means for the competitive landscape against Anthropic, OpenAI, and the open-weight crowd, and whether it’ll show up inside GitHub Copilot and VS Code.

1-Click GitHub Token Stealing via a VSCode Bug

447 points · blog.ammaraskar.com

A security researcher details how a single click could exfiltrate a user’s GitHub token through a flaw in VS Code’s handling of certain URIs. It’s the kind of supply-chain-adjacent vulnerability that makes developers nervous: the editor is trusted infrastructure, and a stolen token can mean access to private repos and CI pipelines. The writeup walks through the exploit chain clearly, which is why HN’s security-minded readers are dissecting both the bug and the broader pattern of editors handling untrusted links.

CT scans of BYD car parts

396 points · lumafield.com

Lumafield put components from a BYD electric vehicle through industrial CT scanning, revealing the internal engineering choices behind one of China’s dominant EV makers. The scans expose how BYD achieves its aggressive cost structure — integration, packaging, and manufacturing decisions you can’t see from the outside. HN loves teardowns, and this one carries an extra geopolitical and industrial subtext about how Chinese automakers are engineering their way to a price advantage Western rivals are struggling to match.

Use your Nvidia GPU’s VRAM as swap space on Linux

334 points · github.com

A clever hack exposes unused GPU VRAM as a Linux block device so it can serve as fast swap space. It’s the sort of resourceful systems trick HN adores — repurposing hardware you already own to claw back performance, especially appealing to people with beefy GPUs and memory-hungry workloads. The discussion digs into the real-world tradeoffs: latency over the PCIe bus, reliability, and whether this is genuinely useful or just a delightfully cursed experiment.

AI outperforms law professors in Stanford Law study

268 points · law.stanford.edu

A Stanford Law study found AI models outperforming law professors on certain legal tasks, reigniting the debate about how much of knowledge work is automatable. The HN thread is appropriately skeptical — what exactly was measured, how the tasks were scoped, and whether “outperform” means accuracy, speed, or both. Still, the headline lands because it comes from a credible academic source and feeds directly into ongoing anxiety and excitement about AI’s encroachment on high-status professional work.

My thoughts after using Clojure for about a month

227 points · acdw.net

A developer shares honest reflections after a month with Clojure, covering the REPL-driven workflow, immutable data structures, and the mental shift the language demands. These “I tried language X” posts reliably draw out HN’s passionate Lisp and functional-programming contingent, and this one is no exception. The appeal is the candor — what clicked, what frustrated, and whether the productivity gains people rave about actually materialize for a newcomer.

How we index images for RAG

144 points · kapa.ai

Kapa.ai walks through the practical engineering of making images searchable inside retrieval-augmented generation pipelines — a genuinely tricky problem as RAG systems move beyond plain text. The post covers embedding strategies, captioning, and how to surface the right visual context to an LLM. It’s catnip for HN’s growing cohort of practitioners actually shipping RAG in production, who are hungry for concrete implementation details rather than another high-level overview.

Pluto.jl 1.0 release – reactive notebook for Julia

141 points · discourse.julialang.org

Pluto.jl, the reactive notebook environment for Julia, hit its 1.0 milestone. Unlike Jupyter, Pluto’s cells automatically re-run when their dependencies change, eliminating the stale-state problem that plagues traditional notebooks. The 1.0 stamp signals maturity and a stable API, which matters for educators and scientific computing users who’ve been building courses and reproducible workflows around it. HN’s data-science readers are weighing the reactive model’s pros and cons against the Jupyter status quo.