AI News Digest, July 29, 2025

Generative AI

Claude Introduces Weekly Usage Limits, $200 Plan Costs Users Thousands

Anthropic announced weekly usage limits for Claude Pro and Max users starting late August, affecting less than 5% of subscribers.
Some users ran Claude Code 24/7, with extreme cases seeing a $200 plan incur tens of thousands in costs.
Users report a lack of transparency in usage data, unable to track consumed tokens or remaining quotas, prompting many to seek alternative products.

Microsoft Edge Browser Transforms into an AI Agent

Edge introduces “Copilot Mode,” enabling cross-tab contextual awareness to analyze all open pages simultaneously.
A streamlined interface with a unified input box auto-detects user intent, supporting voice control and thematic journey features.
Currently free in all Copilot markets, this feature may later be bundled with Copilot subscriptions, potentially ending Edge’s free software status.

MIRIX: Open-Source Multimodal, Multi-Agent AI Memory System

Researchers from UC San Diego and NYU launched and open-sourced MIRIX, the world’s first multimodal, multi-agent AI memory system, with a desktop app.
MIRIX divides memory into six modules—core, contextual, semantic, procedural, resource, and knowledge vault—managed by a meta-memory controller and six sub-modules.
In ScreenshotVQA tests, MIRIX outperforms traditional RAG by 35% in accuracy with 99.9% less storage; it achieves a record-breaking 85.4% on the LOCOMO long-conversation task.

World’s Most Accurate Solar Storm Prediction: First Chain-Based AI Space Weather Model

China’s National Satellite Meteorological Center, Nanchang University, and Huawei released “Fengyu,” the world’s first chain-based AI space weather forecasting model.
The model pioneers a chain-training structure with three components: solar wind (“Xufeng”), Earth’s magnetic field (“Tianci”), and ionosphere (“Dianqiong”).
Fengyu achieves ~10% error in global electron density predictions, excelling in major geomagnetic storm events, with 11 Chinese national invention patents filed.

Shanghai AI Lab Open-Sources Intern-S1, a Multimodal Scientific Model

Shanghai AI Lab released and open-sourced Intern-S1, the top globally open-sourced multimodal model, surpassing closed-source Grok-4 in scientific capabilities.
Features a “cross-modal scientific parsing engine” for precise interpretation of chemical formulas, protein structures, seismic signals, and more.
The team’s unified-specialized data synthesis method delivers strong general reasoning and top-tier specialized capabilities, significantly reducing reinforcement learning costs.

a16z Partner: No Technical Moat, Future Lies in Infrastructure and Vertical Focus

a16z’s Martin Casado predicts AI model competition will mirror cloud computing’s oligopoly, forming a new brand-driven landscape.
The application layer lacks a technical moat; rational business strategies involve “sacrificing profits for distribution,” with value emerging from model infrastructure and vertical specialization.
AI doesn’t turn average developers into super engineers but makes “10x engineers 2x better” by eliminating platform complexities, refocusing programming on creative essence.