Published on June 01, 2026 at 18:00 CEST (UTC+2)
NPM packages from Red Hat have been compromised (506 points by kurmiashish)
NPM packages from Red Hat have been compromised
This article reports a security incident where malicious npm packages were published within the @redhat-cloud-services/ scope. Multiple versions of packages like @redhat-cloud-services/chrome, compliance-client, and frontend-components were compromised, posing a supply-chain risk to users. The issue was disclosed on GitHub by a security researcher from StepSecurity, with links to a detailed blog post. It highlights the growing threat of typosquatting and account takeovers in the JavaScript ecosystem, especially for enterprise-critical packages.
CS336: Language Modeling from Scratch (81 points by kristianpaul)
CS336: Language Modeling from Scratch
Stanford University offers a course (Spring 2026) taught by Tatsunori Hashimoto and Percy Liang that builds a comprehensive understanding of language models from the ground up. The curriculum covers the full pipeline—from data processing and training to evaluation and deployment—without relying on high-level frameworks. It is designed for students and engineers who want deep, hands-on knowledge of modern NLP and AI.
Windows GOG DOS Games on M-Series Macs (56 points by f055)
Windows GOG DOS Games on M-Series Macs
The author describes a method to play GOG-purchased DOS games (like Heroes of Might & Magic II) on Apple Silicon Macs where Windows virtualization is slow. The solution involves installing the game on a Windows machine, then copying the installed files to macOS and running them under DOSBox for Mac. This workaround bypasses the need for emulated x86 Windows and works well for classic, low-resource titles.
Flipper Zero Zig Template (54 points by Nars088)
Flipper Zero Zig Template
This GitHub repository provides a production-ready template for developing applications on the Flipper Zero using the Zig programming language. It integrates Zig’s build system with the Flipper SDK, handling ARM Cortex-M4 cross-compilation and offering type-safe, memory-safe code. The project aims to make Zig a viable alternative to C/C++ for embedded device development.
The Pirate Bay Remains Resilient, 20 Years After the Raid (142 points by speckx)
The Pirate Bay Remains Resilient, 20 Years After the Raid
TorrentFreak recounts the 2006 Swedish police raid on The Pirate Bay’s servers and how the site survived thanks to a last-minute backup made by co-founder Fredrik Neij. Despite legal battles and convictions, the site continues to operate in various forms. The article reflects on the raid’s lasting impact on piracy infrastructure and the resilience of decentralized sharing platforms.
A 10 year old Xeon is all you need (475 points by cafkafk)
A 10 year old Xeon is all you need
The author demonstrates running Google’s Gemma 4 language model on a 2016-era Intel Xeon E5-2620 v4 server with 128 GB of slow DDR3 RAM and no GPU. By combining custom quantization, a verifier model, and optimized inference code, they achieve usable performance on hardware that is far below modern AI workstation specs. The post is a deep technical walkthrough of making LLMs work on commodity recycled servers.
Launch HN: Expanse (YC P26) – Unlock Wasted GPU Capacity (36 points by ismaeel_bashir)
Launch HN: Expanse (YC P26) – Unlock Wasted GPU Capacity
Expanse is a startup that analyzes job submission scripts, hardware, and scheduler logs to predict the real resource needs of HPC/GPU workloads before they run. It aims to reduce the 30–40% typical utilization by flagging over-allocations and surfacing optimizations. In a month-long study of one large cluster, 59% of compute was wasted; Expanse’s tool could save millions in cloud costs.
Sysadmining Like It's 2009 (42 points by yacin)
Sysadmining Like It’s 2009
The author launches “Legacy Labs,” a summer-long event inspired by the Old Computer Challenge, where participants constrain themselves to low-end hardware from around 2009. The goal is to explore vintage operating systems and tools, learning by intentionally using limited resources. It’s a nostalgic yet educational exercise in retro computing and resource-constrained sysadminning.
The Dirt That Refused to Die (12 points by speckx)
The Dirt That Refused to Die
Biologist Sébastien Fontaine’s team sterilized soil with gamma radiation and found it continued emitting CO₂ for years, despite no detectable microbial life. This suggests that abiotic geochemical processes can mimic biological respiration, challenging assumptions about how early life emerged. The findings support a “metabolic first” theory of the origin of life, where simple chemical cycles preceded cellular life.
Linux Basics for Hackers (38 points by ibobev)
Linux Basics for Hackers
This GitHub repository contains structured study notes from the book Linux Basics for Hackers by OccupyTheWeb, organized into 14 modules. Topics range from terminal basics and text manipulation to network management, bash scripting, and exploiting services. It is a practical guide for aspiring security professionals to build foundational Linux skills.
Democratizing AI inference on aging hardware
The successful running of Gemma 4 on a 10-year-old Xeon (Article 6) shows that careful quantization, custom inference stacks, and clever model architectures (e.g., multi-token prediction) can make LLMs accessible on non-GPU, old servers. This trend lowers the barrier for hobbyists, education, and resource-constrained environments, and may drive further interest in lightweight model variants and edge deployment.
GPU utilization crisis and optimization as a service
Expanse (Article 7) highlights that the majority of GPU/HPC compute is wasted due to over-provisioning. This mirrors a broader industry problem where AI labs and cloud tenants request far more resources than needed. Startups and tools that predict actual resource usage and prevent failures could save billions, and this area is likely to attract more investment—especially as GPU demand continues to outstrip supply.
Security of AI/ML supply chains
The Red Hat npm compromise (Article 1) is a reminder that AI pipelines—which often depend on hundreds of open-source packages—are vulnerable to supply-chain attacks. With many ML libraries (e.g., Hugging Face, PyTorch) relying on npm and PyPI, securing package registries and adopting software bill-of-materials (SBOM) practices is becoming critical for AI deployment.
Hands-on education in language modeling is in high demand
The popularity of Stanford’s CS336 course (Article 2) reflects a growing need for engineers who understand LMs at the code level—not just as API users. Courses that teach tokenization, training loops, and model internals from scratch are essential as companies race to build custom or fine-tuned models. This trend signals a shift from "applied ML" to "foundational AI engineering."
Retro computing and low-resource constraints inspire new thinking
The Legacy Labs retro sysadminning (Article 8) and the Flipper Zero Zig template (Article 4) both emphasize working within severe hardware limitations. In AI/ML, such constraints often force innovation—e.g., efficient model compression, on-device inference, and novel architectures. The retro mindset may influence how we design models for tiny devices or edge scenarios.
Abiotic processes may inform bio-inspired AI
The discovery that sterilized soil continues to exhibit life-like chemistry (Article 9) touches on the metabolic origins of life. In AI, this parallels the idea of emergent behavior from simple, non-living components (e.g., neural network weights). Understanding how complex patterns arise from dead matter could inspire new self-organizing AI systems or better models of natural computation.
LLM inference cost reduction is a dominant theme
Across multiple articles (Gemma on Xeon, GPU waste, Expanse), reducing the cost of running AI models is a clear trend. Whether through hardware optimization, better scheduling, or model quantization, the community is moving toward making large models affordable and accessible. This has direct implications for production AI services, where inference cost often dominates total expense.
Analysis generated by deepseek-reasoner