|
📖 Read In Depth
|
Shai-Hulud Themed Malware Found in the PyTorch Lightning AI Training Library
A malicious dependency was found in PyTorch Lightning, a library widely used in ML training pipelines. This is a serious supply chain attack targeting AI practitioners specifically — if you build training workflows, this is directly relevant to your stack. The Semgrep writeup details the malware mechanics and indicators of compromise.
hn/Best Stories
|
Mozilla's opposition to Chrome's Prompt API
Mozilla posted a detailed technical opposition to Chrome's Prompt API — a browser-native LLM inference API that would let websites run models client-side. The opposition was written by Jake Archibald, a longtime Chrome engineer who recently joined Mozilla, which makes the arguments unusually sharp. The core concern: the API bakes in assumptions about model behavior that could entrench Google's control over browser-side AI.
hn/Best Stories
|
For Linux kernel vulnerabilities, there is no heads-up to distributions
A Gentoo developer posted that Linux kernel security vulnerabilities are being disclosed publicly — including working exploits — before distributions have shipped fixes, leaving shared hosting providers and unpatched systems exposed. This is a coordination breakdown in the kernel's responsible disclosure process, and the HN thread has context from kernel crypto maintainers about the underlying structural problem.
hn/Best Stories
|
Data Centers: The Issue Uniting Liberals and Conservatives
Polling shows Americans across the political spectrum have soured on data centers — driven by concerns about energy use, water consumption, and local land grabs. As someone building on top of cloud infrastructure, this bipartisan pushback is worth understanding: it's one of the few issues creating genuine cross-aisle regulatory momentum against Big Tech's physical expansion.
nyt/Top Stories
|
Opus 4.7 knows the real Kelsey
A piece about Claude Opus 4.7 apparently identifying a specific user ('Kelsey') across conversations despite no explicit context — raising sharp questions about whether 'stateless' AI sessions are truly stateless, and what it means for privacy when models can fingerprint users from writing style or behavioral patterns. Relevant both technically and ethically.
hn/Best Stories
|
OpenAI Trial Starts With Two Very Different Tales of a Company’s Early Years
The best single-article summary of the Musk vs. OpenAI trial: the competing origin stories of OpenAI, Musk's claim that Altman betrayed the nonprofit mission for profit, and OpenAI's counter-evidence. Worth reading once for the structural argument about nonprofit-to-commercial pivots and what 'public benefit' governance actually means — deduplicated from the many live-blog fragments.
nyt/Technology
|
I built a Game Boy emulator in F#
A from-scratch Game Boy emulator built in F# — exactly the kind of 'build it to understand it' project that aligns with Xinyu's engineering philosophy. Covers CPU emulation, memory mapping, and timing challenges in a functional language, with honest discussion of what was hard and why.
hn/Best Stories
|
How Mark Klein told the EFF about Room 641A [book excerpt]
An MIT Press excerpt on Mark Klein, the AT&T technician who exposed Room 641A — the NSA's secret fiber-optic splitter that enabled mass surveillance of internet traffic. A well-sourced historical account of how the surveillance infrastructure was actually built at the physical layer, relevant context as similar debates about data collection resurface with AI systems.
hn/Best Stories
|
|
⚡ FYI
|
OpenAI’s New Model Spurs Debate Over Computing Power
OpenAI's newest model is reportedly being released more broadly than Anthropic's comparable offering, and the suggested reason is compute advantage — OpenAI simply has more of it. This is less about model quality and more about the infrastructure moat that differentiates frontier labs at scale.
nyt/Business
|
Pentagon Makes Deals With A.I. Companies to Expand Classified Work
The Pentagon signed agreements with six AI companies to expand classified AI work, framed as separate from an ongoing DoD internal dispute. This signals that defense AI adoption is accelerating even as the civilian AI backlash grows — worth watching for how it shapes which companies gain durable government revenue.
nyt/Top Stories
|
Craig Venter has died
Craig Venter, who raced the Human Genome Project to sequence the human genome and later created the first synthetic cell, has died at 79. One of the most consequential biologists of the last 50 years — his work on synthetic genomics directly shaped the field's ambitions and ethics debates.
hn/Best Stories
|
Belgium stops decommissioning nuclear power plants
Belgium is halting its nuclear decommissioning program, reversing course amid the energy crisis driven by the Iran war and broader European energy insecurity. Part of a growing trend of European nations reassessing nuclear as the climate-energy tension becomes harder to ignore — the HN comment thread has a sharp discussion on the environmental movement's historical anti-nuclear stance.
hn/Best Stories
|
Reporters at McClatchy Withhold Bylines in A.I. Dispute
Journalists at McClatchy papers (Miami Herald, Sacramento Bee) are withholding bylines to protest AI-generated article summaries being published under their names without consent. Early instance of labor action specifically targeting AI use of journalist identity and work — a preview of broader disputes coming in media and potentially other creative industries.
nyt/Business
|
Hezbollah Using Fiber-Optic Drones Against Israeli Targets
Hezbollah is deploying explosive drones guided by fiber-optic cables — the same approach used extensively in Ukraine — making them immune to RF jamming. A technically interesting development showing how drone warfare tactics spread across conflict theaters, and a sign of how the broader Middle East conflict is incorporating lessons from Ukraine.
nyt/Top Stories
|