I think nobody wants AI in Firefox, Mozilla

Mozilla is developing "Window AI," a native, opt-in AI assistant for Firefox that will function as a third browsing mode. The initiative has received overwhelmingly negative feedback from users on Mozilla's forums. This strategy positions Firefox as a middle-ground competitor against other AI-enabled browsers, attempting to appeal to users who want occasional AI integration without full commitment.

Read more

Oracle hit hard in Wall Street's tech sell-off over its AI bet

The provided text is a paywall page for a Financial Times article titled "Oracle hit hard in Wall Street’s tech sell-off over its huge AI bet." The text does not contain the article's content, only navigation links and subscription information.

Read more

Nvidia is gearing up to sell servers instead of just GPUs and components

According to a J.P. Morgan report, Nvidia plans to vertically integrate its supply chain by selling fully assembled L10 compute trays, starting with its upcoming Vera Rubin platform. These pre-built modules will include the Vera CPU, Rubin GPUs, and integrated cooling, shifting partners like ODMs from system design to rack-level integration. This strategy aims to shorten deployment times and capture more hardware value, consolidating profits within Nvidia.

Read more

Microsoft executive closes replies after Windows 11 "Agentic OS" backlash

Microsoft's announcement that Windows is evolving into an "agentic OS" has faced strong backlash from its technical user base. Critics argue the push for deep AI integration is unwanted, citing a preference for bug fixes and a less bloated OS over new LLM-based features. The negative reaction prompted the Windows chief to lock replies on his social media post about the initiative.

Read more

Show HN: European tech news in 6 languages

FMC secured €100M for new memory chips designed for energy-efficient AI data centers. Other notable funding rounds include €6.4M for Lative's AI-driven sales planning tool and €3.7M for Dialog's AI shopping agent trained on brand data. An EU report also highlighted soaring VC investment in AI-first GovTech startups, indicating growing adoption in the public sector.

Read more

Quantifying Long-Range Information for Long-Context LLM Pretraining Data

LongFilter is a data curation framework designed to improve the efficiency of long-context LLM pretraining. It identifies valuable training samples by measuring the information gain from extended context, contrasting model predictions in long-context versus short-context settings. Experiments extending LLaMA-3-8B to a 64K context length using LongFilter-selected data showed substantial improvements on long-context benchmarks like HELMET and LongBench.

ArXiv

Questioning Representational Optimism in Deep Learning

This paper challenges the assumption that better performance from scaling implies better internal representations. By comparing SGD-trained and evolutionarily-searched networks on a simple image generation task, researchers found that SGD induces a disorganized "fractured entangled representation" (FER), while evolved networks develop a more structured "unified factored representation" (UFR), despite both producing the same output. The authors hypothesize that FER may degrade core model capacities like generalization and continual learning, making its mitigation critical for future representation learning.

ArXiv

EnvTrace: Simulation-Based Semantic Evaluation of LLM Code

The paper introduces EnvTrace, a simulation-based method for evaluating LLM-generated instrument control code by assessing the semantic equivalence of execution traces within a digital twin. An evaluation of over 30 LLMs using this trace alignment technique showed that top-tier models can approach human-level performance in this domain. This work positions the symbiosis of LLMs and digital twins as a key step towards autonomous embodied AI.

ArXiv

Chinese co's roadmap for aneutronic fusion

ENN is pursuing proton-boron (p-B) fusion using a spherical torus (ST) design, which new data suggests is feasible with a hot ion mode and high wall reflection, potentially yielding a fusion gain of Q>10. Their roadmap includes the next-generation EHL-2 device, planned for completion by 2026. To support this, the existing EXL-50 device was upgraded to EXL-50U, which successfully achieved its first plasma in January 2024.

ArXiv

Autoregressive or Diffusion Language Models, Why Choose?

TiDAR is a sequence-level hybrid architecture that combines the parallel drafting of diffusion models with the high quality of AR models. It performs parallel token drafting and autoregressive sampling within a single forward pass using structured attention masks, all within a standalone model. This design is the first to close the quality gap with AR models while delivering 4.7x to 5.9x more tokens per second, outperforming speculative decoding and other diffusion variants in both throughput and quality.

ArXiv

RAG Chunk: CLI tool to parse, chunk, and evaluate Markdown documents for RAG

rag-chunk is a CLI tool for preparing and evaluating Markdown documents for RAG pipelines. It allows users to compare different chunking strategies, including fixed-size, sliding-window, and paragraph-based. The tool's key feature is a recall-based evaluation mechanism that scores chunking effectiveness against a user-provided test file of questions and relevant phrases, helping to optimize the retrieval step.

Repo

Show HN: OpEx, an agentic LLM toolkit for Elixir

OpEx is an agentic LLM toolkit for Elixir featuring an OpenAI-compatible client with automatic retry logic. It provides a flexible, hooks-based chat loop for custom tool execution and message handling. The library's core feature is its full support for the Model Context Protocol (MCP), including session management with health checks for both local (stdio) and remote (HTTP) MCP servers.

Repo

Show HN: Agent Playbook – An open-source Storybook-like playground for AI agents

Agent Playbook is a development tool for AI agents, inspired by Storybook for UI components. It provides an interactive web playground to build, test, and debug agents built with the pydantic-ai library in an isolated environment. The tool auto-discovers agents and scenarios, offering real-time visibility into their thinking process and tool executions to accelerate iteration cycles.

Repo

LangDiff: Progressive UI from LLM

LangDiff is a Python library for streaming structured LLM outputs to frontends. It provides intelligent, token-by-token partial parsing that emits granular, type-safe events as data structures are built. The library also automatically generates JSON Patch diffs by tracking mutations on application state objects. This decouples the frontend UI from the LLM's output schema, enabling efficient state synchronization and allowing backend and frontend to evolve independently.

Repo

Show HN: ThinkReview open source browser Copilot GitLab and ADO PRs(Ollama)

ThinkReview is an open-source browser extension that provides AI-powered code reviews for GitLab and Azure DevOps. It supports two modes: a cloud service using Gemini, or a fully private local mode via Ollama. The extension is AGPL-3.0, while its cloud backend service is proprietary.

Repo