╔═════════════════════════════════════════════════════════════╗
║ ║
║ ▀█▀ ▄▀█ █▄▀ █▀▀ █ ▀█▀ █▀▀ █▀█ █▀█ █▀▄▀█ ║
║ █ █▀█ █ █ ██▄ █ █ █▀ █▀▄ █▄█ █ ▀ █ ║
║ ║
║ ▀█▀ █ █ █▀▀ ▀█▀ █▀█ █▀█ ║
║ █ █▀█ ██▄ █ █▄█ █▀▀ ║
║ ║
╚═════════════════════════════════════════════════════════════╝
// Tech news at terminal velocity
cat 2026-07-04.md
$ cat TLDR.md
▸ • AI coding gets crowded: Z.ai launches ZCode to take on Cursor, while GitHub Copilot adds open-weight model Kimi K2.7.
▸ • Research reveals you might only need to fine-tune a single middle layer of an LLM to get full RL post-training benefits.
▸ • Gen Z unplugs at the "Summer of Ludd" festival to harness their rage against Big Tech.
The AI lab formerly known as Zhipu AI just dropped ZCode, an "Agentic Development Environment" built for its GLM-5.2 model. It's a direct shot at Cursor and Copilot, complete with bring-your-own-key support and a desktop app that wants to be your new best friend.
Copilot is no longer a walled garden of closed foundation models. GitHub just made Kimi K2.7 Code generally available, giving enterprise buyers the ability to trade off cost and transparency without leaving their favorite IDE.
Google's new framework-agnostic standard lets AI agents declare UI intent and render it natively across web, mobile, and desktop. Instead of shipping arbitrary code, agents now speak the language of your existing design system.
After two decades of powering massive space battles, Fenris Creations has open-sourced the Carbon game engine. The release includes rendering and physics modules built to handle thousands of concurrent players, now available for anyone to tinker with on GitHub.
A new paper suggests we've been wasting massive amounts of compute on RL post-training. Researchers found that training a single transformer layer in the middle of the stack can recover most of the gains of full-parameter RL training, and sometimes even beat it. Your finance team is going to love this one.
OpenAI has introduced Agent Reinforcement Fine-Tuning (Agent RFT) to help teams train reasoning models against their own tools and graders. It's designed to reduce dead ends and wasted calls when your AI agent is out there trying to use external APIs and inspect files.
A fascinating look at how indoor CO2 levels silently destroy our cognitive abilities and decision-making skills. Before you blame your sluggish code on a lack of caffeine, you might want to crack a window and let some fresh oxygen into your home office.
WebBrain is a new MIT-licensed browser agent for Chrome and Firefox that reads pages and automates tasks. The best part? It can run entirely on a local model, meaning your browsing data never has to leave your machine.
Workato has launched an MIT-licensed developer toolkit that lets AI coding assistants and engineers build enterprise workflows entirely from the terminal. It's a developer-focused repository aimed at making AI-driven automation more accessible.
A new paper introduces Fisher-Guided Token Quantization for communication-efficient federated fine-tuning of LLMs on edge devices. It's a crucial step toward making decentralized AI training actually viable on constrained hardware.
New York City's "Summer of Ludd" festival is teaching people how to live offline and harness their rage against Big Tech. Ironically, you're reading about it on a screen right now.
After a year of waiting, the Trump Mobile T1 phone has finally shipped to buyers. In a shocking twist for modern smartphones, it actually comes with a charger in the box.