Disable AI in Firefox

Mozilla has integrated several new AI/ML features into Firefox, enabled by default. These include a sidebar chat, on-page text actions, a page summarizer, and smart tab grouping. All features can be disabled globally by setting browser.ml.enable to false in about:config, or toggled individually using more specific flags.

Read more

Google AI falsely named an innocent journalist as a notorious child murderer

Google's AI Mode falsely identified a graphic designer as a confessed murderer, apparently conflating the designer working on the story with the article's actual subject. This incident highlights the significant risks of LLM misattribution and hallucination in production search systems. The error occurred in a sensitive context where the true subject's name was legally protected.

Read more

The AI-collapse pre-mortem

The author predicts an imminent AI bubble burst but argues that this financial correction should not discredit the underlying technology. While current LLMs are often unreliable and lack clear business models, their capabilities are a stunning technical achievement. Beyond the LLM hype, AI has already demonstrated superhuman performance in specialized domains like transcription with Whisper and protein folding with AlphaFold. The author concludes that even if the bubble collapses, these profound technological advancements will persist and continue to evolve.

Read more

Deepagent: A powerful desktop AI assistant

DeepAgent is an AI platform that builds a wide range of solutions from prompts. Its capabilities include generating full-stack web applications with database and payment integration, creating RAG-powered chatbots from documents, and developing features directly in code repositories by raising PRs. The platform also automates complex tasks like data analysis, financial modeling, and generating reports or AI videos.

Read more

Filmmaker Guillermo del Toro says 'I'd rather die' than use generative AI

Filmmaker Guillermo del Toro expresses a strong artistic and personal opposition to generative AI, stating he would "rather die" than use it. He compares the arrogance of Victor Frankenstein to that of "tech bros," criticizing the creation of powerful technologies without considering their consequences. Del Toro argues that AI, operating on "ones and zeros," lacks the "alchemy of emotion, spirituality and feeling" that comes from lived human experience and is essential for art, concluding his concern is less with artificial intelligence and more with "natural stupidity."

Read more

Fast-DLLM: Training-Free Acceleration of Diffusion LLM

This paper addresses the slow inference speed of Diffusion LLMs by introducing a block-wise approximate KV Cache for cache reuse and a confidence-aware parallel decoding strategy. The confidence-based approach mitigates quality degradation from broken token dependencies by only decoding tokens above a certain threshold. These techniques result in up to a 27.6x throughput improvement on models like LLaDA and Dream, making their performance competitive with autoregressive models.

ArXiv

AI use in American newspapers is widespread, uneven, and rarely disclosed

An audit of 186K US newspaper articles using an AI detector found that ~9% are partially or fully AI-generated, with usage concentrated in smaller local outlets and specific topics. Opinion pieces in major publications were 6.4 times more likely to contain AI-generated content than news articles from the same outlets. Despite this prevalence, a manual audit of 100 flagged articles found that disclosures of AI use are extremely rare, highlighting a significant transparency gap.

ArXiv

A Framework to Support Technical Assessment in AI Regulatory Sandboxes

The Sandbox Configurator is a proposed open-source, modular framework to address fragmented assessment methods within the EU AI Act's Regulatory Sandboxes (AIRS). It enables users to generate custom testing environments with integrated dashboards by selecting domain-relevant tests from a shared library. Through a plug-in architecture, the framework aims to standardize AI evaluation, streamline compliance for providers, and support Competent Authorities (CAs) in building a scalable infrastructure for AI governance.

ArXiv

ChunkLLM: A Lightweight Pluggable Framework for Accelerating LLMs Inference

ChunkLLM is a lightweight, pluggable framework designed to mitigate the quadratic complexity of self-attention in LLMs. It introduces two trainable components, a QK Adapter for feature compression and a Chunk Adapter for detecting semantic boundaries, while keeping the backbone model frozen. During inference, chunk selection is triggered only at detected boundaries, enabling up to a 4.48x speedup on 120K long texts by retaining only 48.58% of the KV cache while maintaining 98.64% of the original performance.

ArXiv

Text or pixels? On the token efficiency of visual text inputs in multimodal LLMs

A novel input compression technique for multimodal LLMs involves rendering long text inputs as a single image. This method of using visual text representations dramatically reduces the number of decoder tokens required for processing. Experiments on long-context retrieval and summarization benchmarks show this approach can nearly halve token usage without degrading task performance.

ArXiv

Show HN: LLM Rescuer – Fixing the billion dollar mistake in Ruby

LLM Rescuer is an experimental Ruby gem that intercepts NoMethodError on nil objects at runtime. It monkey-patches NilClass to send the surrounding code context to an LLM, which then generates a plausible return value to prevent a crash. This allows the program to continue with a "hallucinated" value, trading predictable failures for unpredictable behavior and API costs.

Repo

Edge – Vector Database in Zig Using USearch and RocksDB

Antarys Edge is a lightweight, open-source vector database for small to medium-sized projects, built in Zig. It uses Usearch for HNSW indexing and RocksDB for storage, with client APIs available for Python and Node.js. While search performance is comparable to competitors like Qdrant, indexing is currently slower, positioning it for experimentation or as a foundation for custom vector DBs.

Repo

Show HN: Check what is hogging your disk zpace

Zpace is a fast, Python-based CLI tool for analyzing disk usage, tailored for developers. It identifies large files by category and pinpoints space-intensive directories common in development, such as node_modules, .venv, and .git. The tool features efficient scanning and correct handling of sparse files, making it effective for inspecting resources like Docker images.

Repo

Show HN: AI Agent for Microsoft Grap Red Team Framework

The Microsoft Graph AI Red Team Framework is an autonomous security tool that uses a multi-agent system for penetration testing the Microsoft Graph API. It leverages three specialized agents: a Scout for reconnaissance, a GPT-4 powered Strategist to generate adaptive attack plans based on discovered permissions, and an Operative to execute them. The framework automates the entire attack lifecycle, orchestrating complex attack chains like privilege escalation and persistence.

Repo

Show HN: Understanding LLM fundamentals without frameworks

This resource teaches foundational LLM application development using only Python and direct API calls, bypassing complex frameworks. It demystifies common patterns like RAG, Agents, and Memory, explaining them as simple programming constructs such as sequential API calls and list management. The course provides code examples for both OpenAI and Anthropic models, covering core topics from tool calling and streaming to multi-step prompt chaining.

Repo