Welcome to Jumble, your go-to source for AI news updates. This week, the White House asked AI labs to hand their most powerful models to the government before the rest of us get them. Meanwhile, Microsoft rolled out a fleet of homegrown models to lean less on OpenAI. Let’s dive in ⬇️
In today’s newsletter:
🏛️ Washington asks for a head start on frontier AI
🧱 Microsoft builds its own model fleet
🦾 One open brain for the next wave of robots
🛒 Your shopping app may start faking product shots
🔌 Weekly Challenge: Run a real AI on your own laptop
🔐 Trump's New AI Order Wants an Early Look at Frontier Models
On Tuesday, President Trump signed an executive order asking AI labs to voluntarily give the government early access to their most powerful models, up to 30 days before release. It is a real shift for an administration that had mostly kept its hands off AI.
Why do you think Trump signed this executive order now?
📋 What It Actually Asks
The program is voluntary, with no mandate and no liability shield for joining. Labs that opt in would let federal testers probe frontier models for up to a month before launch.
🧪 Testing for Cyber Teeth
Agencies are told to build benchmarks for a model's cyber capabilities and stand up a clearinghouse for sharing flaws. The trigger was newer models that can sniff out weak spots in critical software.
🫤 The Version That Almost Shipped
An earlier draft gave reviewers 90 days and stricter terms before it was scrapped over competitiveness fears. Reported lobbying calls helped soften it into the lighter order signed this week.
🧠 Microsoft Built Its Own Models to Lean Less on OpenAI
At Build this week, Microsoft introduced seven homegrown models under its MAI brand, led by its first reasoning model. After years of running on OpenAI's tech, it is framing the move as long-term self-sufficiency.
⚙️ Trained From Scratch
MAI-Thinking-1 carries 35 billion active parameters and a 256,000 token context window. Microsoft says it was built on licensed data, with no distillation from OpenAI.
🥋 Going After the Entire AI Industry
In blind tests, raters preferred MAI-Thinking-1 to Claude Sonnet 4.6 and matched Claude Opus 4.6 on a coding benchmark. A smaller coding model reached every paying GitHub Copilot user the same day.
They’ve also revealed their very own OpenClaw app for windows, allowing agents to fully control your laptop.
🧲 The Real Message
Azure still runs OpenAI, so nothing got cut off overnight. But Microsoft now has its own option at every tier, which quietly shifts who holds the leverage.
Weekly Scoop 🍦
🧠 Weekly Challenge: Run a Real AI on Your Own Laptop
Challenge: This week is all about models shrinking down to your own hardware, so let's make it personal. Get a capable AI running entirely on your machine, with no cloud and no account, so your data never leaves the laptop.
Here’s what to do:
📥 Step 1: Install LM Studio Grab LM Studio free for Mac, Windows, or Linux. It is a one-click app, no terminal needed.
🧰 Step 2: Pull a model that fits Download a small open model like Qwen, DeepSeek or Mistral. An 8B to 12B model runs comfortably on roughly 16GB of memory.
✈️ Step 3: Go offline and use it Switch off your wifi, then chat, summarize, or draft. Push it on a real task and note where it keeps up with your usual cloud model.
Would you want the government test-driving frontier models before you can? And how long before a laptop-sized model is all most of us actually need? See you next time! 🚀
Stay informed, stay curious, and stay ahead with Jumble!
Zoe from Jumble
