Welcome back. Chinese labs are allegedly distilling America's top AI models. This time, the White House is calling them out. Trump's lead tech adviser just accused China's Moonshot AI of “distilling” Fable 5 to build Kimi K3. And that's not the only charge on the list.

The White House accuses Moonshot AI of stealing from Anthropic: Michael Kratsios, technology advisor for the Trump administration, claims the Chinese AI lab "distilled" capabilities from Anthropic's Fable model to build Kimi K3. Kratsios also alleges they trained the model on restricted Nvidia servers. Now, the Treasury Department is warning of possible sanctions for the alleged IP theft. But Silicon Valley is pushing back. Almost 200 startups, including YCombinator, are asking the administration not to block Chinese open-source models.

Cursor's new router delivers frontier coding at a fraction of the cost: The AI coding startup just dropped Cursor Router. This new tool classifies each request and sends it to the best-fit model. It stops defaulting to the most expensive model every time. The company claims early enterprise teams saved 30-50% with no quality drop. You can use it today on Teams and Enterprise plans. See how it works.

This new app teaches you how LLMs are built: Developer and MTS at Anthropic, Felix Rieseberg, just released Language Model Builder. This Mac app shows you how to build a small language model from the ground up. It includes a 90-minute interactive textbook that covers tokenization, attention, and transformers. Everything runs directly on your Apple Silicon Mac. In just one day, you can train a model that actually chats back. Try it here.

Recently, a vibe-coded app leaked 1.5M API keys because nobody reviewed the code.

Your team ships AI-written code every day. Superblocks keeps it governed:

Book a demo.

Fresh eyes are disappearing. AI now writes a growing share of production code, but the review process was built for code written by hand. It relies on a fresh pair of eyes to catch what the engineer missed. Since AI now handles reviews too, both sides fail to notice the same mistakes. To solve this, AgentField co-founder and CTO Santosh Kumar Radha published a playbook on rebuilding the review process from the ground up.

A map, not a verdict. He argues reviews used to do four jobs. They caught bugs, taught the team, flagged risks, and ensured a second person signed off on every change. AI has now automated the first part. The same model that writes the code can check it in seconds. Because of this, Radha stopped asking whether a change is safe to merge. His reviews now list what each change puts at risk.

Three rules from the playbook that do the work:

Swap the reviewer's model. But every team is not ready to start with the above playbook. To keep things simple, we recommend checking out Greptile. It’s built on the idea that Claude Code and Codex are actually better at catching bugs in each other’s work than they are in their own. 

Find and resolve every crash, freeze, slowdown, ANR, and OOM before your users do. bitdrift captures 100% of mobile data, unsampled and in real time, so it’s immediately queryable by engineers and AI agents.

Try bitdrift: mobile observability built for the real world.

How to run your coding agent safely: You’ll learn how to safely run AI coding agents without risking their computers. This tutorial covers the dangers of giving agents unrestricted access, and shows how to easily set up isolated Docker Sandboxes so agents can code autonomously without accidentally deleting files or wiping databases.

Claude Code Security plugin: Scan your changes for vulnerabilities before you commit or run a full scan across your codebase, all from your terminal on the Claude inference you already run.

Library of 52 Claude Code prompts: This is a library of prompts to copy into Claude Code. Use it to explore ways of working you haven’t tried, or when you’re not sure where to start.

Claude Code locks up your entire session during long builds or tests. You end up stuck waiting on a command that doesn't need your input. Delba de Oliveira from Anthropic's Claude Code team shared a built-in fix for this. 

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

OpenAI cofounder and now Anthropic researcher Andrej Karpathy shared the "lazy" prompting habit that gets him better results.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 300K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team