× + newsletter-2026-04-29
╔═════════════════════════════════════════════════════════════╗
║                                                             ║
║   ▀█▀ ▄▀█ █▄▀ █▀▀   █ ▀█▀   █▀▀ █▀█ █▀█ █▀▄▀█             ║
║    █  █▀█ █ █ ██▄   █  █    █▀  █▀▄ █▄█ █ ▀ █             ║
║                                                             ║
║   ▀█▀ █ █ █▀▀   ▀█▀ █▀█ █▀█                                ║
║    █  █▀█ ██▄    █  █▄█ █▀▀                                ║
║                                                             ║
╚═════════════════════════════════════════════════════════════╝
    

// Tech news at terminal velocity

cat 2026-04-29.md

The Agentic Era is Here (And It's Driving a Cadillac)

$ cat TLDR.md

• NVIDIA drops a multimodal model that actually remembers what it saw and heard

• GM is stuffing Google's Gemini into 4 million cars

• Sentry and WordPress launch AI agents that might actually fix your bugs instead of creating them

Headlines & Launches

🧠 NVIDIA's "Everything Everywhere All at Once" Model

AI agents usually play a game of telephone between vision, speech, and language models, losing context along the way. NVIDIA just dropped Nemotron 3 Nano Omni, an open multimodal model that unifies all three into a single system. It promises up to 9x more efficiency and tops six leaderboards, giving developers a production-ready path for agents that don't have the memory of a goldfish.

Source: NVIDIA Blog

🚙 Your Next Uber Might Talk Back

General Motors is rolling out Google's Gemini AI assistant to roughly four million vehicles across the US. If you own a 2022 or newer Cadillac, Chevy, Buick, or GMC with Google built-in, your dashboard is about to get a lot chattier. Let's just hope it's better at navigating traffic than it is at generating historically accurate images.

Source: The Verge

🛑 China Hits the Brakes on Robotaxis

After a series of chaotic incidents involving Baidu's Apollo Go fleet, China has reportedly suspended new licenses for autonomous vehicles. It turns out that unleashing dozens of driverless cars into the wild without a flawless backup plan leads to gridlock and headaches. The robot uprising has been temporarily delayed by traffic congestion.

Source: The Verge

📦 Box Jumps on the Agentic Bandwagon

Box just launched Box Automate, a workflow solution designed to dynamically route tasks between humans, Box Agents, and enterprise systems. It's essentially an attempt to untangle the messy web of corporate document management using AI. Because if there's one thing enterprise software needs, it's more autonomous agents reading your PDFs.

Source: Business Wire

Deep Dives

🐛 GitHub's Six-Hour Fire Drill

Wiz Research used AI models to uncover a terrifying remote code execution vulnerability in GitHub's internal git infrastructure that could have exposed millions of repositories. GitHub managed to validate, fix, and investigate the critical flaw in under six hours. It's a fascinating look at how modern security teams respond when the absolute worst-case scenario lands on their desk.

Source: The GitHub Blog

📉 Why Your AI Agent Suddenly Gets Stupid

A new paper explores the inherent instability of training multi-turn LLM agents using reinforcement learning. Researchers found that even when models appear to have diverse reasoning, they often rely on fixed templates that just look diverse but fail to actually respond to different inputs. It's the AI equivalent of nodding and saying "that's crazy" when you aren't listening.

Source: OpenReview

🗜️ Squeezing the Fat Out of Chain-of-Thought

Reasoning models are great at self-reflection, but their excessively long outputs create prohibitively massive key-value (KV) caches during inference. This paper introduces a redundancy-aware compression method specifically for reasoning models, preventing the catastrophic failures that happen when standard compression techniques try to truncate chain-of-thought processes.

Source: OpenReview

Engineering & Research

🕵️ Sentry's AI Will Now Read Your Stack Traces

Sentry just launched Seer Agent in beta, allowing you to literally ask questions about your app's errors and get answers based on your entire Sentry history. You can ping it in Slack or hit Cmd + / in the dashboard to figure out why production is on fire without having to manually parse through thousands of generic LLM failure logs.

Source: Sentry Blog

⌨️ A Senior WP Dev in Your Terminal

Automattic has released Studio Code, an agentic CLI tool that acts as a WordPress expert right in your terminal. It can read your codebase, edit files, spin up local sites, and actually understands WordPress best practices. It's currently in beta, so it might still occasionally suggest a plugin you don't need, but it's a massive step up for WP development.

Source: WordPress Blog

Profiling AI Agents on Arm

Arm just dropped Performix, a free performance analysis toolkit specifically designed for modern agentic development workflows. It provides deep insights to help developers optimize AI applications running on Arm-based cloud platforms, with early backing from Microsoft, MongoDB, and Redis.

Source: Arm Newsroom

Odds & Ends

🐧 Your PS5 is Now a Linux Box

Developer Andy Nguyen has officially published the GitHub steps to soft-mod your PlayStation 5 into a fully functional Linux PC. It won't let you pirate PS5 games, but it will let you run Ubuntu and play PC games on Sony's hardware, which is arguably much funnier.

Source: The Verge

🌍 The Internet is Held Together by Duct Tape

Cloudflare's Q1 2026 report on internet disruptions is a wild ride through global chaos. From nationwide shutdowns in Uganda to unprecedented drone strikes taking out cloud infrastructure, it's a stark reminder that the "cloud" is actually just a bunch of very vulnerable computers on the ground.

Source: The Cloudflare Blog

← Back to archive