Welcome back. It's a great time to build something new. Over the weekend, Anthropic gave Claude a million-token memory and doubled your usage across every plan and every Claude product. The promotion runs for two weeks, and the extra usage won't even count against your weekly limits.

Also: How a 30-year veteran ships 5x faster with his agentic setup, how to reverse-engineer Anthropic's locked Architect exam, and how to master spec-driven development.

Claude just doubled its usage limits for the next two weeks: The AI lab just rolled out a promotion that doubles usage limits during off-peak hours. It applies automatically on weekdays outside 5-11 AM PT and all day on weekends, valid through March 27 across all plans and every Claude product. Best of all, this promotional usage won't count toward weekly limits. The company also made its 1M token context window generally available for Opus 4.6 and Sonnet 4.6.

Karpathy builds open-source tool to visualize AI's impact on jobs: The OpenAI co-founder recently open-sourced a weekend project that visualizes 342 U.S. occupations across 143 million jobs using Bureau of Labor Statistics data. Its standout feature is a prompt-driven pipeline where an LLM scores every profession by any criteria you set, from AI exposure to offshoring risk. Karpathy later took it down after the scores were widely misread as job-loss predictions.

A tiny transformer can now run compiled programs on its own: Researchers at AI startup Percepta recently showed that a small transformer can internally execute arbitrary C programs step by step, without needing an external interpreter. The real breakthrough is a 2D attention mechanism paired with a custom HullKVCache, which replaces standard linear-scan decoding with logarithmic-time lookups. This allows for millions of execution steps per second on a standard CPU. In one demo, the system solved the world's hardest Sudoku with perfect accuracy. Read more here.

Need for speed. Former Dropbox CTO Aditya Agarwal recently noted that he produced more code in five days using AI than he had in the previous five years. While teams everywhere are seeing similar gains, Dax Raad, founder of OpenCode, and Karri Saarinen, CEO of Linear, are raising a critical question: what happens when the tools outpace the judgment needed to keep them in check?

The bar quietly dropped. Raad sent his team a memo that's going viral on X. He argues that when a feature can be prompted into existence in twenty minutes, questioning whether it should exist disappears. Even worse, code that would've been refactored six months ago is now just being left alone. And as the mess compounds, no one is close enough to the code to feel truly responsible for it.

Output isn't judgment. Saarinen took it further, arguing that if your success metric is the percentage of agent-written code, you've missed the point. Agent-built features still add complexity for users and maintenance burden for engineers who may not fully grasp the logic. Ultimately, users care whether the product works, not how the code was generated.

The gap is where the risk lives. A feature that takes an entire sprint to build naturally creates checkpoints like reviews, pushback, and friction. In contrast, a twenty-minute feature lacks those safeguards. The real issue is the widening gap between development speed and the time required for honest evaluation, and we still haven't found a way to bridge it.

Claude Code often generates excessive code, including unrequested abstractions that make reviews difficult. To fix this, Anthropic released code-simplifier plugin. It refactors your recent changes to keep them clean and concise without breaking functionality.

Install it from the terminal:

Restart Claude Code, then at the end of any session:

The agent reviews your CLAUDE.md file, follows your specific coding standards, and focuses exclusively on recently modified files.

If your team has specific conventions, define them in CLAUDE.md and the simplifier will follow them automatically.

How a senior engineer actually uses OpenCode: This tutorial teaches developers how to master the AI coding agent OpenCode. You will learn to configure specialized sub-agents for testing and security, securely manage API keys, install custom skills, and seamlessly integrate the tool into workflows using Neovim, GitHub, and web interfaces.

OpenSquirrel: This repo is a native Rust desktop app that lets you run multiple AI coding agents within a single tiled interface. A primary agent can coordinate tasks by delegating work to sub-agents, targeting both local and remote machines via SSH.

Attention Residuals (by Moonshot AI): Standard language models process data from previous layers uniformly, which often causes earlier information to get lost as networks grow deeper. The new "Attention Residuals" approach solves this by allowing layers to selectively prioritize specific past outputs, effectively enhancing the model's overall performance.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 200K+ engineers and 100K+ 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