Wikipedia says traffic is falling due to AI search summaries and social video

Wikipedia's human page views have declined 8% year-over-year, a trend attributed by the Wikimedia Foundation to generative AI search summaries providing direct answers and younger generations seeking information on social video platforms. This shift poses risks to volunteer engagement and donor support, prompting the foundation to urge AI and search companies to encourage direct visits to Wikipedia for proper attribution and to maintain content integrity. Wikipedia is also developing new attribution frameworks and outreach initiatives.

Read more

LLMs can get "brain rot"

This paper introduces the "LLM Brain Rot Hypothesis," demonstrating that continual pre-training on low-quality web text causes persistent cognitive decline. By training LLMs on "junk" Twitter/X data, defined by either high engagement or low semantic quality, researchers observed significant degradation in reasoning, long-context understanding, and safety. The primary failure mode was identified as "thought-skipping," where models truncate reasoning chains, and the induced damage was not fully reversible by subsequent instruction tuning or pre-training on clean data, framing data curation as a critical training-time safety problem.

Read more

AI is making us work more

Generative AI and LLMs are creating a psychological pressure for constant productivity, as their "always-on" nature makes any downtime feel like a missed opportunity. This dynamic is driving the adoption of extreme "996" work cultures in AI startups, where developers feel compelled to match the machine's tireless pace to stay competitive. The author argues this internalizes a "tyranny of can," leading to a self-defeating cycle of burnout that ultimately stifles the innovation these tools were meant to enhance.

Read more

Don't use AI to tell you how to vote in election, says Dutch watchdog

The Dutch data protection authority (AP) has warned against using LLMs for voting advice, finding them to be unreliable and biased. Testing revealed that chatbots consistently suggested the same two political parties, regardless of user input or its alignment with other party platforms. The watchdog concluded this bias is not deliberate but an inherent consequence of how the AI models operate, posing a threat to election integrity.

Read more

AI Weiwei: What I Wish I Had Known About Germany Earlier

The text is an opinion piece by Ai Weiwei, which was commissioned and then rejected by a German magazine. He critiques German society for its deference to authority, lack of individual moral judgment, and underlying authoritarianism. The article also includes reflections on bureaucracy, the nature of art, and cultural observations on German food and humor.

Read more

OptPipe: Memory- and Scheduling-Optimized Pipeline Parallelism for LLM Training

This work addresses inefficient, heuristic-based pipeline parallelism (PP) scheduling for LLM training by reformulating it as a constrained optimization problem. The proposed method generates fine-grained schedules that jointly optimize for memory capacity, activation reuse, and pipeline bubble minimization. This dynamic approach significantly reduces idle pipeline time by up to 50% and improves throughput, enabling the training of larger models under the same memory constraints.

ArXiv

Tensor Logic: The Language of AI

The paper argues that AI progress is hindered by the fragmentation between scalable neural frameworks like PyTorch and symbolic languages like LISP. It proposes "tensor logic," a new language that unifies neural and symbolic AI by treating logical rules and Einstein summation as a single fundamental operation: the tensor equation. This unification enables novel capabilities like sound reasoning directly in embedding space, combining the scalability and learnability of neural networks with the reliability and transparency of symbolic reasoning.

ArXiv

Measuring the Impact of Early-2025 AI on Experienced Developer Productivity

A randomized controlled trial with experienced open-source developers found that allowing the use of frontier AI tools, such as Cursor Pro and Claude 3.5/3.7 Sonnet, surprisingly increased task completion time by 19%. This slowdown contradicts strong predictions of a speedup from both the developers themselves and external experts. The study suggests this counter-intuitive result is a robust finding and not primarily an artifact of the experimental design.

ArXiv

Large Language Models Inference Engines Based on Spiking Neural Networks

This work proposes NeurTransformer, a methodology for converting pre-trained transformers into efficient SNNs for inference to address the quadratic complexity of standard models. The approach replaces the self-attention mechanism with a spike-based version (SSA), converts the feed-forward blocks, and then fine-tunes only the SSA component. Applied to GPT-2, this method showed a minimal performance drop, including a 9.7% perplexity reduction on the small model, while achieving a 64-85% estimated energy reduction for the self-attention block on digital hardware.

ArXiv

Query Decomposition for RAG

This work frames RAG's subquery and document selection as an exploitation-exploration problem to manage the trade-off between retrieval breadth and precision. It employs bandit learning methods to sequentially retrieve documents, dynamically deciding whether to continue exploiting a subquery or explore an alternative. By estimating document relevance using rank information and human judgments, this approach yields a 35% gain in document-level precision and a 15% increase in α-nDCG, improving downstream generation performance.

ArXiv

Show HN: Katakate – Dozens of VMs per node for safe code exec

Katakate is an open-source tool for creating and managing self-hosted, secure VM sandboxes designed for executing untrusted code from AI agents at scale. It provides hardware-level isolation by leveraging a stack of Kubernetes (K3s), Kata Containers, and the Firecracker VMM for fast, lightweight execution. The platform is managed via a CLI, API, and Python SDK, and enforces strong security defaults including network egress policies and minimal Linux capabilities.

Repo

LightlyStudio – an open-source multimodal data curation and labeling tool

LightlyStudio is an open-source, Rust-powered tool for unifying data workflows, including curation, annotation, and management. It provides a powerful Python interface for handling various dataset formats like COCO and YOLO. The tool enables complex, code-based queries and features automated data selection to build optimal training subsets by balancing sample typicality and diversity.

Repo

Show HN: Docker/Model-Runner – Join Us in Revitalizing the DMR Community

Docker Model Runner (DMR) is a tool integrated into Docker Desktop and Engine for managing and running LLMs and other AI models. It provides a docker model CLI and a backend server to pull models from any OCI registry and serve them via a REST API. The system integrates with backends like llama.cpp, exposes Prometheus-compatible metrics for monitoring, and includes experimental Kubernetes support.

Repo

Show HN: LunaRoute – a high-performance local proxy for AI coding assistants

LunaRoute is a high-performance local proxy for AI coding assistants that provides complete visibility into LLM interactions with sub-millisecond overhead. Written in Rust, it offers comprehensive session recording, debugging, and performance monitoring, including token usage and tool call analysis. Key features include zero-config passthrough, automatic PII redaction, API dialect translation, and a web UI for session analysis.

Repo

Show HN: Distributed Storage System to 8x LLM Inference, GPU Training Efficiency

Blackbird is a high-performance, RDMA-based distributed storage system designed for ML training/inference pipelines and feature stores. It leverages UCX for zero-copy performance and features intelligent, policy-driven tiered caching across GPU memory, DRAM, CXL, and NVMe. The system includes a topology-aware placement engine and native support for modern interconnects, positioning it as a high-performance alternative to systems like Redis or Alluxio.

Repo