Welcome back. Tiny models are quietly outperforming the giants. A San Francisco-based AI lab just dropped a new reasoning model with fewer than 1B active parameters that rivals frontier models. The most surprising part? They didn't use a single Nvidia chip to train it.

Also: How to use Claude's dreaming agents, rethink your fundamentals for the AI era, and find the right AI personal assistant for your workflow.

Zyphra trains a frontier model without a single Nvidia chip: The San Francisco-based AI lab just dropped ZAYA1-8B, the first mixture-of-experts model built from the ground up on AMD silicon. It was trained on a massive 1,024-node MI300X cluster on IBM Cloud, allowing this open-weight model to compete with far larger reasoning systems for math and coding. The weights are live now on Hugging Face, giving developers a no-cost way to test and benchmark the model.

Anthropic doubles Claude Code limits after SpaceX deal: The AI lab just secured the entire compute capacity at SpaceX's Colossus 1, opening up over 220,000 NVIDIA GPUs and 300 megawatts of power in just a month. Claude Code rate limits have already doubled across Pro, Max, Team, and Enterprise plans, while Pro and Max users will no longer face peak-hour throttling. Anthropic and SpaceX are even looking into building gigawatts of AI compute out in orbit.

OpenAI open-sources the networking behind Stargate: The ChatGPT maker just released MRC, the protocol that keeps every GPU inside its supercomputers in sync during massive training runs. Developed in partnership with heavyweights like AMD, Broadcom, Intel, Microsoft, and Nvidia, it spreads data across hundreds of paths. It also reroutes around failures in microseconds, effectively killing the network congestion that usually stalls training.

Using Claude Code, Codex, or Gemini? Learn how to compose them into complex agent architectures using a primitive called the harness. Learn why your repo is the largest prompt you'll ever write. Why your boundary drifts mid-run. How to engineer harnesses that hold in production.

All distilled from shipping 100+ open-source harness recipes on GitHub. Python/TS/Go, Apache 2.0. Fork one and compose your own.

Read the blog

The line is dissolving. The more careful approach to AI coding, often called agentic engineering, used to be distinguished from vibe coding by its strict, line-by-line reviews. However, in a recent post, Django co-creator Simon Willison admits he’s abandoned that practice, even for production code. He calls this shift the "normalization of deviance."

The pattern is mounting. Coinbase is part of a growing list of tech companies restructuring around AI. Just a few days ago, CEO Brian Armstrong notified his team that "non-technical groups are now shipping production code," announcing 700 layoffs and a move to an "AI-native" structure, turning Simon Willison's observations into a structural reality.

Polish lost its meaning. An engineer can now spin up a repo with a hundred commits, a solid README, and full test coverage in thirty minutes. Even Willison can't tell it from months of careful work. As non-technical teams at Coinbase do the same, a polished finish no longer signals quality.

The line moved downstream. It used to be about the code review before you deployed, but now it's about what actually survives in production. Trust is earned by code that's been battle-tested by real users. Without that track record, you're just vibe-coding, no matter how clean the README looks. Polish is cheap, but hours in production are everything.

Long Codex sessions accumulate context debt. You might be deep into a refactor before realizing there’s a better way, but starting over usually means losing all your context. Codex’s fork command solves this by cloning your entire session history into a new thread. This keeps your original work intact while you branch off and experiment. Branch your latest session instantly:

For an older session, grab the ID from the picker, “/status”, or “~/.codex/sessions/”:

If the new direction doesn't pan out, just jump back with “codex resume”. If you're in the middle of a session, “/side” does the same thing inline, so you don't have to leave your main thread.

P.S. You can find 50+ AI coding hacks here.

How to build interactive agents with Generative UI: This tutorial covers how to move beyond plain-text responses by building interactive AI agents. You’ll learn to create agents that generate dynamic UI elements like charts, forms, and interactive components on demand.

Flue: A TypeScript framework for next-gen agents, built around a dedicated agent harness. Think of it as a fully headless, programmable version of Claude Code. It makes zero assumptions about needing a human in the loop.

DeepSeek TUI (17.3k ⭐): A terminal-based coding agent for DeepSeek models that lets you read and edit files, run shell commands, and search the web. It manages git workflows and coordinates sub-agents, all through a streamlined, keyboard-driven interface.

Bootstrapping Composer with autoinstall (by Cursor): Composer training was slowed down because broken coding environments forced the AI to spend time fixing setup issues instead of learning tasks. The research shows that older Composer models can now auto-repair these setups, making the training process much faster and more efficient.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 260K+ 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