Meta’s new model runs on two Macs

Welcome back. Meta is coming for China’s lead in open-weight coding models, with a little help from Apple. Meanwhile, Anthropic and Cognition want to help devs ship faster and spend less time putting out fires.

Also: A former NASA engineer spills on life at SpaceX, SF founders and billionaires predict AI’s next act, and Elon Musk teases a Tesla Roadster demo unlike any you’ve seen before.

Meta’s new open model now runs on two Macs: The social media giant just dropped Muse Spark Open, calling it the most capable US open-weight model yet. Built for multi-step reasoning and complex software engineering, it packs 42B active parameters and a 1M token context window. You can run it locally on two M5 Ultra Mac Studios or serve it through vLLM, SGLang, and llama.cpp.

Cognition wants AI to take your on-call shift: The AI coding startup just shipped Devin Oncall, its push toward self-driving software. The agent picks up production pages and bug reports before you do, then digs through logs and code. It either escalates with the severity, likely root cause, and a suggested next step or opens a tested PR. It remembers past incidents too, helping it skip roughly 60% of repeat investigations. It reportedly resolved half of AngelList’s incidents on its own.

Anthropic makes AI code review 3x faster: Coding agents are churning out code faster than humans can review it, and the AI lab wants to close the gap. It just upgraded Code Review to Claude Team and Enterprise plans, cutting most reviews to 6–10 minutes from about 20. Eligible enterprise orgs can claim up to $50K in credits. Meanwhile, the Claude Code desktop app now splits parallel tasks into nested sessions, with a sidebar showing which ones need your input.

Deepgram Flux TTS reads the whole conversation rather than isolated text strings, keeping agent speech natural across multi-turn calls, while offering time to first audio as low as 80ms and native interruption handling out of the box.

Test custom scripts, adjust speed & expressivity parameters, and sample 36 voices in real time directly in the Flux TTS Playground.

Test it on your stack.

Git can’t keep up. When teams hand more coding tasks to AI agents, the bottleneck inevitably shifts to the repository. Anthropic saw its merge volume jump from roughly 2,500 merges per day in May to more than 30,000 in September. Traditional Git simply could not handle that relentless scale. In a recent blog post, the company detailed how a small team of seven engineers built an entirely new replacement in just five months. 

GitHub sees it too. While Anthropic's numbers are extreme, GitHub is feeling the exact same pressure. In September, developers and agents generated 7.38 billion commits on GitHub, more than five times the volume from a year earlier. Pushes surged 4.9x, and actions runs more than quadrupled. To handle the load, GitHub is actively rebuilding its Git infrastructure while keeping the site fully operational.

Writes are the hard part. GitHub explains that serving the same code to multiple agents is fairly straightforward because caches handle most of that workload. However, saving their changes creates a much bigger challenge. The system must store each update and make it visible before the next agent can build on top of it. Because agents commit after nearly every step, these merges quickly pile up on the main branch. 

Two very different fixes.

The repo is next. Anthropic designed Scion specifically for massive monorepos, so the team is not telling everyone to abandon Git just yet. If your developers run AI agents in tight execution loops, you should closely monitor push latencies and merge queues. Your agents might already be ready for their next assignment while your repository is still struggling to catch up. 

Try it: GitHub’s merge queue docs cover tuning build concurrency and merge limits.

How to build your own open-source Grok Bot in a day: This tutorial walks through how this dev built Rakazo (a GrokBot alternative) from the first plan to a working agent. You’ll learn how he used Claude to build it, screenshots and emulators to verify that tasks actually worked, and bots to review and merge PRs as the project scaled.

Humanizer: A local 12B model that rewrites AI-written drafts so they sound more human without changing the facts. It works on emails, essays, reports, and posts in English or Chinese, runs offline, and is trained to preserve every number, date, name, unit, and quote.

Build a scheduled repository reviewer with Claude (by Anthropic): This recipe shows how to set up a read-only agent that reviews your repo on a schedule and remembers what it found last time. Each run picks up from the previous review, flags new problems, marks fixed ones as resolved, and stays bounded with read-only tools, turn limits, and a spending cap.

Agents ship sloppy work unless you supervise every step. SpaceXAI engineer Lauren Tan shared the 10-second setup for pstack, her skill stack for rigorous autonomous engineering.

The mode stays active for the session. Run /poteto-help when you're unsure which skill fits.

A senior founder says this thread will change your mind if you believe “humans will find a way” through AI disruption. He called it heart-breaking.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.

We put together a few guides on coding agents, agentic engineering, and leadership frameworks to help you level up in your career. Browse all our guides.

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

Until next time — The Code team