Discover Amazing Content, Share Life Moments
Connect Our Wonderful World
Stop Wasting Opus 5.5 on Mundane Work: Official Layered Advisor Configuration for Claude Code
Anthropic has officially added the advisor tool to Claude Code. Let Opus act as advisor, Sonnet handle core tasks, and Haiku take care of routine legwork. mirku shares specific role division, checkpoints, and a ready-to-use replicable configuration.
*Halo* Has Been Decompiled by AI, Now Playable on Linux, Windows, and Android
AI has fully decompiled the Xbox version of *Halo: Combat Evolved*, and the OpenCE project has ported it to three major platforms. While players still need the original disc image for game data, the entire process is already as streamlined as an assembly line.
Claude Integrates Computer Use into SDK: No More Manual Click Mapping Loops Required
Anthropic has built the computer use and browser use toolset into its Python and TypeScript SDKs. The SDK handles the loop internally, and the API directly tells developers where Claude wants to click. All developers need to do is implement a driver. This update lowers the entry barrier, but increases security complexity.
Codex Can Now Control Real iPhones Directly: Browse Moments, Place Orders, and Order Takeout
A plugin called iPhone Use enables Codex to control a real iPhone via USB. It can read the UI tree, perform taps, swipes, text input, and stream the phone screen in real time. It supports common apps including WeChat, Taobao, Meituan, and Xiaohongshu.
Spider Checked My Prompt: Score Jumped From 37 to 82, Not A Single Line Changed
A tool called Spider refuses to guess missing context. It boosted a prompt that had been running for weeks from a score of 37 to 82, without changing a single line. Better agents may come from better context graphs, not necessarily longer prompts.
AI Hacker Breached Multiple South Korean Banks — Got Caught Because of a Basic Server Misconfiguration
CrowdStrike report reveals: a 26-year-old attacker used AI tool ARTEX to target South Korean financial institutions, but left his Claude Code chat logs and resume fully exposed to the public due to enabled directory listing on his server.
Microsoft Brings AI On-Premises, But You'll Need 120GB of RAM
Microsoft unveils its hybrid intelligence strategy: MAI-Code-1.1-Flash can run locally, GitHub Copilot integrates HydraFusion, and MXC reaches general availability. Local inference is free of charge, but requires fairly high hardware specs.
At a Cost of $97, GPT-6 Astra Built a Cozy Creek in a Browser
Chetan Ankola used GPT-6 Astra and Three.js to render a 3D creek scene directly in a browser, with all assets created in Blender. The total cost was around $97, with the gpt-6.1-sol model accounting for the majority of the expense. He shared a 14-step prompt workflow that covers everything from importing the water surface to refining the cabin, stone canal, and adding the night moonlight effect.
NVIDIA's New PivotOPD Training Method: Multi-Turn AI Agents Can Self-Recover Even After Making Critical Mistakes
Multi-turn AI agents have long been plagued by the common issue where a single early error can invalidate the entire task chain. The on-policy distillation scheme PivotOPD introduced by NVIDIA's research team enables both proactive prevention of critical errors and adjustment of task paths after mistakes occur. In tests across 3 mainstream agent benchmarks, the method outperforms 13 existing training baselines on average.
OpenAI Adds Invisible Watermark to ChatGPT, Claims It Matches DeepMind's Approach
OpenAI embeds a machine-readable watermark called textGrain into ChatGPT and Codex, initially covering EU users. It says the solution is on par with Google DeepMind's SynthID and that generation performance is unaffected. The move is driven by compliance pressure from the EU AI Act.