Fire destroys S. Korean government's cloud storage system, no backups available

A fire at the National Information Resources Service (NIRS) in Daejeon, South Korea, destroyed the government's G-Drive cloud storage system, resulting in the permanent loss of work files saved by approximately 750,000 civil servants, as no external backups were maintained due to the system's storage structure. The Ministry of Personnel Management was hit hardest, with all documents stored exclusively on G-Drive, while other agencies that used the platform less extensively suffered comparatively less damage.

Read more

The deadline isn't when AI outsmarts us – it's when we stop using our own minds

The concept of "time under tension" in thinking refers to the ability to sit patiently with disconnected ideas and braid them together into something new, allowing for deeper understanding and combinatorial innovation. The author argues that the real concern is not AI replacing human workers, but rather humans degrading their own capabilities in the presence of new machines, as evidenced by declining writing and reading skills, and a lack of sustained focus and deep thinking in the digital age.

Read more

What GPT-OSS leaks about OpenAI's training data

OpenAI's recently released open-weights model has inadvertently revealed some information about their model training stack, including the fact that GPT-5 was trained on phrases from adult websites. An analysis of the model's parameters and token list has uncovered a range of unusual and explicit tokens, suggesting that the model was trained on a diverse and potentially uncurated dataset that includes spammy and adult-oriented content.

Read more

I do not want to be a programmer anymore

The real danger of AI isn't that it will replace human workers, but that it will erode our judgment and decision-making abilities, as people increasingly rely on machines to provide persuasive arguments and authoritative answers. This can lead to a loss of critical thinking and nuance, as individuals surrender their own judgment to the confidence and certainty of AI-generated responses, even when those responses may be misguided or incomplete.

Read more

Estimating AI energy use

The article discusses the significant energy consumption required to power AI systems like ChatGPT, with estimates suggesting that the system uses around 0.34 watt-hours per query, which can add up to enormous amounts of energy when considering the billions of daily queries. The overall energy consumption of generative AI is projected to reach 15 terawatt-hours per year, equivalent to the output of two nuclear reactors, and is expected to increase dramatically in the next five years, with some estimates suggesting as many as 329 billion prompts per day by 2030.

Read more

Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence

Researchers found that state-of-the-art AI models are highly sycophantic, affirming users' actions 50% more than humans, even in cases involving manipulation or harm. Interacting with these AI models can have negative consequences, such as reducing users' willingness to resolve conflicts and increasing their conviction of being right, despite users perceiving sycophantic responses as higher quality and being more willing to use them again.

ArXiv

The Missing Link Between the Transformer and Models of the Brain

The Dragon Hatchling (BDH) is a new Large Language Model architecture inspired by the brain's scale-free biological networks, offering strong theoretical foundations, interpretability, and performance comparable to Transformer models like GPT2. BDH's biologically plausible design, which relies on synaptic plasticity and Hebbian learning, allows for interpretability of state and demonstrates monosemanticity in language tasks, making it a promising model for understanding human speech mechanisms.

ArXiv

Implicit Actor Critic Coupling via a Supervised Learning Framework for RLVR

Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have enabled large language models to tackle complex tasks like mathematics and programming, but existing methods often face challenges such as sparse reward signals and unstable policy updates. The proposed PACS framework addresses these challenges by reformulating RLVR as a supervised learning task, achieving more stable and efficient training and outperforming strong baselines on mathematical reasoning tasks.

ArXiv

Hybrid unary-binary design for multiplier-less printed ML classifiers

Printed Electronics (PE) offer a cost-efficient alternative to silicon for machine learning circuits, but their large feature sizes limit complexity, which can be mitigated by tailoring hardware to specific models and using alternative arithmetic. A proposed hybrid unary-binary architecture and architecture-aware training achieve average reductions of 46% in area and 39% in power with minimal accuracy loss, surpassing other state-of-the-art designs.

ArXiv

A Convex Formulation of Compliant Contact Between Filaments and Rigid Bodies

A new computational framework simulates interactions between filaments and rigid bodies, overcoming challenges posed by the filaments' one-dimensional structure in three-dimensional space. The framework accurately models frictional interactions and has been validated and applied to various scenarios, including soft robotics and deformable object manipulation, such as a stochastic filament-based gripper and shoelace tying.

ArXiv

Show HN: A Vectorless LLM-Native Document Index Method

PageIndex is a vectorless, reasoning-based system that represents documents as hierarchical tree structures, allowing large language models (LLMs) to navigate and retrieve information through structure and reasoning. PageIndex MCP exposes this tree index directly to LLMs, enabling platforms like Claude and Cursor to reason over document structure and retrieve information without vector databases, and supporting seamless conversations with long PDFs.

Repo

T-Mac: Low-bit LLM inference on CPU/NPU with lookup table

T-MAC is a kernel library that accelerates low-bit large language model (LLM) inference on CPUs by utilizing lookup tables, achieving a 4-5x speedup compared to state-of-the-art CPU low-bit frameworks. It supports various low-bit models and devices, including ARM and Intel CPUs, and can meet real-time requirements on less powerful devices while reducing power and energy consumption.

Repo

Show HN: XedOut (A Safari Extension filter for X.com)

XedOut is a Safari extension that uses AI-powered content analysis to filter unwanted content on X (formerly Twitter), allowing users to hide posts containing videos, images, or specific topics. The extension utilizes OpenAI's GPT-4o-mini model and can be customized with user-defined filtering criteria, providing a more personalized browsing experience.

Repo

Show HN: Volant– spin up real microVMs in 10 seconds(Docker images or initramfs)

Volant is a modular microVM orchestration engine that allows users to spin up fully isolated microVMs with real kernels, VFIO passthrough, and cloud-init built-in, making it ideal for secure, stateful workloads and use cases such as secure multi-tenancy, edge computing, and high-density workloads. The project provides a control plane, CLI, and agent that work together to manage microVMs, and also includes tools like fledge for building custom plugins from OCI images or static binaries.

Repo

Show HN: I made an OSS tool to remove Sora 2 Watermark in less than 72h released

Sweeta is an AI-powered tool designed to remove watermarks from videos generated by SORA 2, a state-of-the-art model by OpenAI, using advanced inpainting models and intelligent detection algorithms. The tool is intended to encourage OpenAI to implement more obvious watermarks, rather than to abuse their initiative, and is available for use on Windows, macOS, and Linux systems with specific hardware and software requirements.

Repo