Thursday — July 24, 2025

Zed editor now allows users to disable AI features, Cerebras launches Qwen3-235B, a frontier AI reasoning model achieving 1.5k tokens per second, and researchers introduce Prompt Injection 2.0, a hybrid AI threat that combines with traditional cybersecurity exploits to evade security controls.

News

You can now disable all AI features in Zed

Zed, a code editor, now allows users to disable its AI features by adding a setting to their settings.json file or with a single switch during onboarding, catering to developers who prefer not to use AI or have concerns about data privacy, environmental impact, or other issues. The editor aims to provide control over the development environment, respecting users' preferences, and will continue to improve its experience for both AI and non-AI users.

Cerebras launches Qwen3-235B, achieving 1.5k tokens per second

Cerebras Systems has launched Qwen3-235B, a frontier AI reasoning model, on its inference cloud platform, delivering production-grade code generation at 30 times the speed and one-tenth the cost of closed-source alternatives. The model achieves unprecedented speeds of 1,500 tokens per second, reducing response times from minutes to seconds, and is available at $0.60 per million input tokens, making it a significant breakthrough in AI model performance and affordability.

Building better AI tools

The tech industry is building AI tools in a way that undermines human capabilities, prioritizing automation over augmentation, and neglecting the importance of human learning and collaboration. By designing AI tools that facilitate human retrieval practice, process reinforcement, and collective knowledge transfer, we can create more effective and sustainable systems that enhance human abilities rather than replacing them.

US AI Action Plan

The US government has introduced an AI Action Plan, a roadmap to achieve global dominance in artificial intelligence, with three pillars: accelerating innovation, building AI infrastructure, and leading in international diplomacy and security. The plan aims to promote American AI systems, computing hardware, and standards worldwide, while preventing adversaries from exploiting US innovation and investment, and ultimately ushering in a new era of economic competitiveness, national security, and human flourishing for the American people.

AI overviews cause massive drop in search clicks

New research from the Pew Research Center found that Google's AI Overviews, which provide summary answers to search queries, cause a significant drop in search clicks, with some searches experiencing a reduction of almost half. The study's findings contradict Google's claims that AI Overviews do not take traffic away from websites, and instead support the concerns of SEO experts who have argued that the feature hurts web traffic.

Research

Prompt Injection 2.0: Hybrid AI Threats – Paper and Open Source Testing Toolkit

Prompt injection attacks, which manipulate AI systems into following unauthorized commands, have evolved into a significant security threat, particularly with the emergence of agentic AI systems that can autonomously perform tasks. This paper analyzes the latest generation of prompt injection attacks, known as Prompt Injection 2.0, which can combine with traditional cybersecurity exploits to evade security controls, and presents architectural solutions to mitigate these hybrid threats.

Frugal Machine Learning for Energy-Efficient, and Resource-Aware AI

Frugal Machine Learning (FML) is a practice that aims to design efficient and cost-effective ML models that minimize the use of computational resources, time, energy, and data. The field of FML involves various strategies, including input, learning process, and model frugality, and utilizes technological enablers such as model compression and energy-efficient hardware to achieve acceptable performance while reducing resource consumption.

CT-ScanGaze: A Dataset and Baselines for 3D Volumetric Scanpath Modeling

The CT-ScanGaze dataset is introduced as the first publicly available eye gaze dataset for Computed Tomography (CT) reading, addressing the lack of resources in this area. A novel 3D scanpath predictor, CT-Searcher, is also presented, which can process CT volumes and generate radiologist-like 3D fixation sequences, and its effectiveness is demonstrated through evaluations on the CT-ScanGaze dataset.

Gemini 2.5 Pro Capable of Winning Gold at IMO 2025 with Prompting

The International Mathematical Olympiad (IMO) poses uniquely challenging problems that Large Language Models (LLMs) typically struggle with, but a recent test using Google's Gemini 2.5 Pro was able to solve 5 out of 6 IMO 2025 problems correctly. This result highlights the potential of LLMs for complex reasoning tasks, but also underscores the need for optimal strategies to fully harness their capabilities.

Dynamic Chunking for End-to-End Hierarchical Sequence Modeling

Researchers have introduced a new technique called H-Net, which enables language models to learn from raw data without relying on pre-processing steps like tokenization, allowing for true end-to-end learning. The H-Net model outperforms traditional token-based language models, particularly in languages and modalities with weaker tokenization heuristics, and demonstrates improved scaling with data and character-level robustness.

Code

Just Open Sourced NeuralAgent: The AI Agent That Lives on Your Desktop

NeuralAgent is an AI personal assistant that automates tasks on your desktop, including typing, clicking, and navigating the browser, using modern large language models. It has a range of features, including desktop automation, background automation, and support for multiple model providers, and can be customized and extended through its open architecture and modular design.

Show HN: Interactive explainable AI as mlflow artifacts

Xaiflow is a library that integrates with MLflow to generate interactive HTML reports for SHAP analysis, providing rich, interactive visualizations that stakeholders can explore and understand. The library offers features such as interactive feature importance charts, SHAP value visualizations, and deep dive analysis, making it easier to share comprehensive model insights with non-technical teams and debug model behavior.

Creating agentic users, not just agentic AI

The goal of this project is to create a human-machine interaction paradigm where users can benefit from AI tools without compromising their agency, by designing interfaces that leave users feeling in control. The prototype, called "BuildTogether", is a chat interface that uses a language model to make minimal edits to the user's work, displaying the changes clearly and allowing users to learn from and maintain control over the edits.

100% AI-built WebDAV client streams 100 GB text, greps in seconds

This project is a modern, high-performance WebDAV browser built with Tauri, React, and TypeScript, designed to handle large text files and featuring efficient streaming, fast in-file search, and a clean, responsive interface. The application is 100% AI-generated, cross-platform, and supports a wide range of file types, including text, code, document, and media files, with features like virtualized rendering, media preview, and internationalization support.

SQLite AI – Local AI Inference, Powered by SQLite

SQLite-AI is an extension for SQLite that enables artificial intelligence capabilities directly within the database, allowing developers to run, fine-tune, and serve AI models using simple SQL queries. It supports various features, including embedded AI inference, streaming I/O, and offline-first functionality, making it ideal for on-device and edge applications, such as building conversational agents and memory-aware assistants.

    Zed editor now allows users to disable AI features, Cerebras launches Qwen3-235B, a frontier AI reasoning model achieving 1.5k tokens per second, and researchers introduce Prompt Injection 2.0, a hybrid AI threat that combines with traditional cybersecurity exploits to evade security controls.