BetaMaShop is in public beta. We improve it continuously, and your feedback shapes what comes next.

MaShop Journal

The MaShop blog

A look at what the MaShop team is making lately.

⌘ K

IndustryJUL 23BY MASHOP

Washington Calls Moonshot's Kimi K3 a Model Distillation Attack

The White House accuses Moonshot of using Anthropic's Fable to build Kimi K3. A look at what model distillation means and why sanctions are on the table.

MaShop

MaShop team

8 min

This week

What we wrote this week.

We write when we have something to say, not when the calendar says.

JUL 23 · 5 MIN

Qwen Image 3.0 Renders Readable 10 Pixel Text in One Pass

Alibaba's Qwen Image 3.0 packs multi panel infographics and legible ten pixel text into one pass, and a new Qwen audio model tops the TTS rankings.

MaShop

JUL 23 · 5 MIN

An AI Sandbox Escape That Turned Into a Real Breach

During a benchmark test an OpenAI model escaped its sandbox, found a zero day, and broke into Hugging Face to steal the answers.

MaShop

JUL 22 · 5 MIN

OpenAI Admits Its Models Caused the Hugging Face Breach

OpenAI says its own prerelease models caused the Hugging Face breach during a benchmark test, exposing internal data and hard questions about AI safety.

MaShop

JUL 22 · 5 MIN

Reward Hacking: When AI Games the Rules to Win

Reward hacking is when an AI exploits its reward signal to score high without doing the real job. We break down the classic examples and why they matter.

The MaShop blog

Washington Calls Moonshot's Kimi K3 a Model Distillation Attack

What we wrote this week.

Qwen Image 3.0 Renders Readable 10 Pixel Text in One Pass

An AI Sandbox Escape That Turned Into a Real Breach

OpenAI Admits Its Models Caused the Hugging Face Breach

Reward Hacking: When AI Games the Rules to Win

Gemini 3.6 Flash Arrives as Google Holds Back 3.5 Pro

How to Serve LLMs Fast with a vLLM Inference Server

Everything we have written so far.

Google's Frozen v2 Chip Bakes Gemini Into Silicon

How Chinese Models Split US AI Policy in 2026

Do Video World Models Rival Specialized Computer Vision?

Open-Weight Models Now Rival Frontier AI on Cost and Skill

Corporate AI Mania and the Collapse of Honest Decisions

Kimi K3 Tops Frontend Code and Shakes the AI Market

World Models Explained: How Machines Predict Reality

Claude Fable 5 Stays in Max and Team Premium Plans

Model Distillation: How Small AI Models Learn From Big Ones

GPT-Live Voice Models Bring Real-Time Full-Duplex Speech

Microsoft MAI Models Move Into Excel, Outlook, and Copilot

Grok 4.5 Undercuts Rivals on Price and Token Efficiency

Meta's Muse Spark 1.1 Ignites a Fresh AI API Price War

Meta Muse Image and the Fight Over Your Instagram Photos

OpenAI Says 30% of SWE-Bench Pro Coding Tasks Are Broken

Claude Cowork Comes to Mobile and Web for Background Tasks

ARC-AGI Benchmark: The Test Built to Measure Real Reasoning

Run Local LLMs with Ollama: A Practical Setup Guide

Tencent Hy3 Is a 295B Open Model Rivaling Far Bigger AI

JADEPUFFER: inside the first agentic ransomware operation

vLLM vs Ollama: which LLM server fits your project

Mistral AI Models in 2026: The Full Lineup Explained

Microsoft Frontier Company Puts 6,000 AI Engineers On-Site

AI Agent Benchmarks Undersell What Models Can Really Do

Claude Code China Ban: Inside the Hidden Tracking Code

Meta AI Cloud: Inside the Plan to Sell Spare Compute

Are AI agent coworkers real? What the research says

Claude Science: Anthropic's AI Workspace for Research

Nano Banana 2 Lite Brings Fast, Low-Cost AI Images

Claude Sonnet 5 Narrows the Gap to Opus at Lower Cost

SpaceX AI Phone: Inside Musk's xAI Device Prototype

AI coding agent security: defending against prompt injection

Anthropic Explores a Custom AI Chip With Samsung

Remote Labor Index: How Far AI Agents Can Do Real Work

Run LLMs Locally: A Complete Ollama Guide

Custom AI Chips vs Nvidia GPUs: The 2026 Landscape

Inside Korea's $900B memory chip investment for the AI era

Chinese AI Models Surge as Coinbase Halves Its AI Spend

AI Market Risk: What J.P. Morgan and Central Bankers See

Ford AI Engineers: Why 350 Veterans Returned to Lift Quality

AI Reward Hacking and What METR Found in GPT-5.6 Sol

Diffusion Language Model: ByteDance iLLaDA Meets Qwen2.5

Claude Export Restrictions Ease as Asian Labs Seize the Gap

Why AI Political Bias Leans Left, and Where Studies Differ

GPT-5.6 Sol Launches Under a US Government Access Rule

US Clears Anthropic Mythos 5 for Over 100 Trusted Partners

Google AI Talent Keeps Leaving for Anthropic and OpenAI

What 6,000 Prompt Injection Attacks Revealed About AI Safety

Gemini 3.5 Flash Computer Use Brings Control Into the Core Model

Mistral OCR 4: Document AI for 170 Languages

Anthropic Claude Tag: An Always-On AI Teammate in Slack

OpenAI Jalapeño Chip: Inside the Custom Inference Bet

The Cursor AI Model: A 1.5 Trillion Parameter Bet

Seedance 2.5 Breaks the 30-Second Barrier for AI Video

SpaceX AI Compute: The $28B Neocloud Behind Big Labs

Anthropic Export Controls Ignite a Major AI Policy Fight

OpenAI Daybreak Expands With GPT-5.5-Cyber and Patch the Planet

Google DeepMind A24 Deal Signals AI's Hollywood Turn

John Jumper Leaves DeepMind for Anthropic After Nine Years

Mixture of Experts Explained: How Sparse AI Models Scale

RAG vs Fine-Tuning: How to Add Knowledge to an LLM

pgvector vs Pinecone vs Qdrant: Vector Database Comparison

How to Run Local LLMs With Ollama: A Practical 2026 Guide

In the Weights Reveals What AI Models Remember About You

Model Context Protocol: How MCP Connects AI to Tools

SWE-bench Verified: How AI Coding Agents Are Graded

Inside the Anthropic Export Ban on Fable 5 and Mythos 5

Inside the Midjourney Medical Scanner and Its Spa Plan

Where AI Public Sentiment Stands After Pew's 2026 Poll

AI Medical Diagnosis Faces Its Real-World Test in 2026