╔═════════════════════════════════════════════════════════════╗
║ ║
║ ▀█▀ ▄▀█ █▄▀ █▀▀ █ ▀█▀ █▀▀ █▀█ █▀█ █▀▄▀█ ║
║ █ █▀█ █ █ ██▄ █ █ █▀ █▀▄ █▄█ █ ▀ █ ║
║ ║
║ ▀█▀ █ █ █▀▀ ▀█▀ █▀█ █▀█ ║
║ █ █▀█ ██▄ █ █▄█ █▀▀ ║
║ ║
╚═════════════════════════════════════════════════════════════╝
// Tech news at terminal velocity
cat 2026-05-17.md
$ cat TLDR.md
▸ • Vercel Labs drops "Zero," a systems programming language built specifically for AI agents to read and write native code.
▸ • OpenAI is reportedly considering legal action against Apple over a lackluster ChatGPT-Siri integration that failed to drive subscriptions.
▸ • Google DeepMind open-sources Gemma 4, while NVIDIA drops a world model that generates 720p video on a single GPU.
Vercel Labs just released Zero, a systems programming language designed explicitly for AI agents rather than humans. It compiles to native executables, ditches mandatory garbage collection, and provides version-matched agent guidance directly through the CLI. Because why teach AI to write Python when you can just build them their own language?
DeepMind has unleashed Gemma 4, featuring a 26B Mixture of Experts variant and a 31B Dense model built on Gemini 3 tech. In a surprising shift, they're releasing it under an Apache 2.0 license, giving developers full control to deploy across cloud environments without the usual corporate strings attached.
NVIDIA just dropped SANA-WM, a 2.6B-parameter open-source world model that can generate 60-second, 720p videos without requiring a massive server farm. It spits out a minute of video in just 34 seconds on a single RTX 5090, proving that you don't need a data center to make high-res AI video anymore.
The honeymoon is officially over. OpenAI is reportedly exploring legal action against Apple because the highly-touted ChatGPT integration into Siri failed to bring in the expected wave of new subscribers. Nothing says "strategic partnership" quite like hiring an independent law firm to draft a breach-of-contract notice.
Everyone is rushing to adopt the Model Context Protocol (MCP), but most teams are doing it wrong. This deep dive explores the anti-patterns of production agentic systems, explaining why wrapping three tools in an MCP server for a weekend POC usually leads to skyrocketing token usage and unpredictable agents in production.
The scientific research repository ArXiv is finally cracking down on the careless use of LLMs. If authors are caught letting AI do all the heavy lifting on their scientific papers, they'll be slapped with a one-year ban. It's a necessary step to keep the platform from drowning in hallucinated citations and synthetic slop.
The popular Python agent framework just dropped v1.97.0, bringing a new MCPToolset that uses fastmcp-slim under the hood. It also splits out Google Cloud providers and adds online evaluators for failed calls, making it easier to debug when your agent inevitably goes off the rails.
Researchers are testing CAR T cell therapy—traditionally used for cancer—as a way to completely reset the immune system in patients with autoimmune diseases. It's a fascinating crossover of medical engineering that could fundamentally change how we treat chronic conditions.
After getting roasted online for a demo of the Xperia 1 XIII's AI Camera Assistant, Sony is doing damage control. They're desperately trying to clarify that the feature doesn't actually edit your photos, it just makes "suggestions" based on lighting and depth. Sure, Sony, whatever you say.
Deep Care has released a $350 offline desk gadget that sits on your desk and judges your posture. It's pricey, but it runs completely offline and might actually be the thing that finally gets you to stop hunching over your keyboard like a goblin.