Apple's new Macs are local AI powerhouses

Welcome back. Devs love running AI workflows on their Macs, and Apple is taking note (remember the OpenClaw Mac Mini frenzy?). The iPhone maker just released new machines that will help devs host heavier workloads locally. But that’s not the only update for local AI enthusiasts: OpenAI is officially taking on Nvidia with their first-ever inference chip. 

Also: See why Shopify’s CEO wants to ban Claude Code, creating stunning art images using JavaScript, and how to run local models with Mac Ultra Studio.

Apple's new Macs are local AI powerhouses: The iPhone maker just dropped two new silicon chips. The M6 debuts in a Mac mini as the first 2nm processor from Apple, and the M5 Ultra powers a refreshed Mac Studio with up to 512 GB of unified memory, so you can run big models on your desk without touching the cloud. Apple even tucked Neural Accelerators into every GPU core for a serious AI speed boost. They're claiming it handles LLM prompts up to 9.8x faster than the M1 Ultra. You can check it out and preorder here.

Stripe blocked $300M in fraud aimed at an AI coding startup: The team behind coding harness OpenCode says Stripe just blocked about $300M in fraudulent charges. That is actually more than the startup even makes, according to co-founder Dax Raad. The target was OpenCode Go, a $10 a month flat plan. Scammers are using stolen cards to open thousands of accounts and resell tokens. If your product sells subsidized flat rate access to models, fraud defense is now a core engineering responsibility. Raad says these charges could have sunk the company overnight. 

OpenAI's debut chip claims a lead over Nvidia's Blackwell: The ChatGPT maker just dropped the first benchmarks for Jalapeño, the inference chip it built with Broadcom. The company says it beats Nvidia's GB200 and GB300 systems on speed and efficiency. It also runs models like GPT-OSS and DeepSeek R1 more effectively than most. For devs, that could mean faster ChatGPT responses and snappier agents. You can also expect cheaper tokens and steadier access as the chip rolls out through 2027. Just a heads-up, it’s built strictly for inference, so it won’t be used for training.

Your analytics in another database? That's pipelines, sync lag, and data drift—hidden costs in every query.

Tiger Data is Postgres, purpose-built for time-series and analytics at scale. Hypertables handle billions of rows. 95% compression cuts storage costs. Continuous aggregates power real-time dashboards without the infrastructure headache.

Everything you chose Postgres for: reliability, SQL you know, ecosystem you trust—plus time-series power at scale.

No sharding. No trade-offs. No separate database.

Built by the TimescaleDB team. One database. Full power. Your analytics stay fast, your pipelines disappear, your data stays true.

Get $1000 credit to start.

Slop wins by default. Your coding agent now writes more in an afternoon than you can read in a week, and that's where slop creeps in. You skim PRs, approve requests, and let code quality drift toward whatever passes a glance. Tech founder David Zhang’s viral essay on "software factories" points to the fix: reading every line isn't feasible anymore, so you test what the code does instead. Think of the codebase as a black box. You stop worrying about the internal mechanics and focus solely on making sure the right input always delivers the right output. 

Here’s how that works:

Where to start. You don't need a total overhaul to get going. Just pick one messy section, define its inputs and outputs, and have your agent generate a ranked decision ledger. Read those decisions instead of the diff and see what jumps out. 

P.S. Zhang has open-sourced the skills he uses to run the full loop, and they drop into Claude Code, Cursor, Codex, or any tool that supports skills.

Gitar automatically reviews pull requests, providing AI-powered feedback on security, bugs, performance, edge cases, and code quality. It analyzes your CI pipelines and fixes CI failures on your PRs. Not just comments: real fixes, until your pipeline passes.

Try it yourself.

How devs should use Grok Bot: This tutorial guides developers through setting up Grok Bot for AI agent automation. You'll learn to configure permissions, connect plugins like Notion, Gmail, and Cursor, and build multi-bot chains that automate web scraping, prompt preparation, and cross-platform data retrieval.

Headlong: Self-guided agents that keep thinking. This open-source microharness gives agents a continuous inner loop, so they don’t stop when the prompt ends. They can set priorities, kick off their own work, and ping you with updates as they go. It's alpha software that runs shell commands around the clock, roughly $1–2/hour in tokens

Build Your Own AI Coding Agent Harness (by Vercel): A basic tool loop breaks down fast on real coding work. This hands-on guide shows you how to build a working harness from scratch, covering tool design, safety gates, sandboxes, context pruning, subagents, human approvals, planning, verification, and extensibility.

Claude Code checks its own code, so it misses its own mistakes. OpenAI released an official Codex plugin that reviews Claude's changes without leaving the terminal.

Codex only reads here. To hand it a bug to fix, run /codex:rescue followed by a plain-English task.

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

Andrew Ng shared the 6 skills you need to build and deploy AI apps, including the one thing that separates top-tier AI engineers from the rest. 

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