💥 AI rebuilt Next.js in one single week

A Cloudflare engineer rebuilt Next.js in a week using AI — $1,100 in tokens, 4x faster builds. Chrome shipped WebMCP so AI agents can interact with websites directly. And Latent Space declared code reviews officially dead. If the ground feels like it's moving under your feet, it's because it is. Here's what's worth your time.

This guide distills 10 AI strategies from industry leaders that are transforming marketing.

Outpace your competitors by mastering AI.

Get Your Free Guide

How We Rebuilt Next.js with AI in One Week (16 min)
One engineer at Cloudflare used AI agents to rebuild 94% of the Next.js API surface on Vite. The result: 4.4x faster builds, 57% smaller bundles, and it cost $1,100 in API tokens. The project is called vinext and it's open source. Whether or not you'd use it, this is a glimpse of what one developer can ship with the right tools.

WebMCP Is Available for Early Preview (2 min)
Chrome 146 Canary now includes WebMCP — a proposed W3C standard that lets websites expose structured tools to AI agents. Instead of screen-scraping, agents get proper function calls. This is a big deal for the future of browser-based AI. Available behind a flag today.

We Deserve a Better Streams API for JavaScript (42 min)
A deep dive into why the Web Streams API was designed for a different era and how an async-iterable-based alternative is 2–120x faster. Long read, but if you work with streaming data in Node or the browser, this is the most thorough analysis you'll find.

MicroGPT (33 min)
Karpathy distilled a complete GPT implementation into ~200 lines of pure Python — no PyTorch, no TensorFlow, zero dependencies. Covers tokenization, autograd, attention, and training from scratch. If you want to truly understand how LLMs work under the hood, this is the single best resource available.

Node.js Memory Cut in Half with Pointer Compression (8 min)
V8's pointer compression reduces heap pointers from 64-bit to 32-bit, cutting Node.js memory usage by ~50% with only 2–4% latency increase. Available now via a Docker image called "node-caged." If you're running Node pods on Kubernetes, you can halve your memory limits.

If AI Writes Code, Should the Session Be Part of the Commit? (3 min)
A new tool called Memento captures AI coding sessions as commit metadata — so you can trace not just what changed, but how and why the AI made those changes. Interesting question for teams adopting AI-assisted workflows. 472 points on HN this week.

When Does MCP Make Sense vs CLI? (6 min)
Thoughtful pushback on the MCP hype. The argument: language models are already great at using CLI tools, so MCP adds unnecessary overhead when well-documented command-line interfaces exist. Worth reading as a counterpoint if you're evaluating MCP for your stack.

I Planted Fake API Keys in Online Code Editors (5 min)
A developer planted honeypot API keys in CodePen, JSFiddle, and other editors, then monitored network traffic. Key finding: CodePen sends your code to servers as you type — every keystroke, before you hit save. Worth knowing if you've ever pasted credentials into an online editor.

Popover API or Dialog API: Which to Choose? (9 min)
These two browser APIs look similar but behave very differently. This CSS-Tricks guide breaks down when to use each — popovers for most lightweight overlay use cases, dialogs specifically for modals requiring focus trapping and accessibility.

Making WebAssembly a First-Class Language on the Web (14 min)
Mozilla explains why Wasm is still a second-class citizen requiring JavaScript glue code, and how the Component Model could let Wasm access Web APIs directly. If you've been waiting for Wasm to feel native in the browser, this is the roadmap.

How to Kill the Code Review (11 min)
Latent Space argues that human-written code died in 2025 and code reviews will follow in 2026. Their alternative: shift human oversight upstream to spec writing and verification instead of line-by-line review. Provocative, but the logic is worth engaging with whether you agree or not.

Software Engineers Should Be a Little Bit Cynical (11 min)
A well-argued case that modest cynicism about organizational dynamics actually helps engineers ship meaningful work. Not about being negative — about thinking clearly through political constraints and competing incentives.

AI Isn't Replacing SREs. It's Deskilling Them. (11 min)
The concern isn't that AI replaces SRE jobs — it's that automation degrades the deep expertise needed when systems fail in ways AI can't handle. Argues for deliberate practice and skill maintenance alongside automation.

Bryan Cantrill on Stack Ranking, Competing with Bezos, Regrets (30 min)
Candid career lessons from the former CTO of Joyent and current CEO of Oxide Computer. Covers navigating industry booms and busts, the politics of performance reviews, and what he'd do differently looking back.

Vercel AI Gateway
Vercel's new AI Gateway provides a unified API for routing requests across multiple LLM providers with built-in caching, rate limiting, and observability. One integration point instead of managing individual provider SDKs.

Apple Xcode 26.3 Unlocks Agentic Coding
Xcode now supports AI agents like Claude and Codex directly inside the IDE via MCP. This is a significant validation of MCP as a cross-platform standard — when Apple adopts a protocol, it tends to stick around.

Docker Model Runner Brings vLLM to macOS (6 min)
Run vLLM locally on Apple Silicon through Docker. Uses Metal acceleration for inference. If you've wanted high-throughput local LLM serving without leaving the Docker ecosystem, this removes the setup friction.

llmfit — Right-Size LLMs to Your System (3 min)
Automatically selects the optimal LLM model based on your available RAM, CPU, and GPU. Useful if you're running local models and tired of guessing which quantization fits your hardware.

Vercel Queues Now in Public Beta
Background job processing built into Vercel's platform. If you've been using external queue services for deferred work in your Next.js apps, this might simplify your stack.

Coaction v1.0 — Multithreading for Web Apps
A library for effortless multithreading in web applications using Web Workers. Designed to move heavy computation off the main thread with minimal API surface.

Cloudflare Just Slop Forked Next.js (7 min)
Fireship breaks down the Cloudflare vinext story — one engineer, one week, AI-written code replacing one of the most popular frameworks. Covers the implications for framework lock-in and the future of open source.

Cursor, Claude Code and Codex All Have a BIG Problem (18 min)
Theo argues that current AI coding tools have fundamental UX problems compared to what developers had before. A candid look at where the tooling falls short despite the raw capability.

Software Engineering Is Dead Now (15 min)
Block (Cash App, Square, Tidal) fired half their engineering staff. Theo unpacks what happened, what it means for the industry, and whether this is a one-off or the start of a trend.

Your Codebase Is NOT Ready for AI (14 min)
Matt Pocock explains why software architecture matters more than ever for AI-assisted development. The key insight: AI agents are like new starters with no memory — your codebase structure is the biggest factor in their output quality.

Why I Chose Electron Over Native (25 min)
Scott from Syntax built a multi-source video recording app with Electron and explains why the tradeoffs were worth it. A practical perspective on when web tech wins over native development.

linear() Is So Good (12 min)
Kevin Powell on the CSS linear() timing function and why it changes how you think about animations. Practical demos of what's possible when you can define custom easing curves as piecewise linear functions.

That's it for this week. Reply if you have something I should include next time.

Until next week,
Niall