Good morning. Love is getting weird. A new study finds that 1 in 7 Gen Z adults in relationships has a side thing with an AI chatbot. Most keep at least part of it secret from their real life partner. And 69% don’t want them to know how deep their AI romance goes. Getting caught with a secret AI soulmate is a dumb way to lose a real one.
Self improving AI agents keep forgetting how to do their jobs. Researchers tested long running agents designed to continuously learn new tasks and discovered a troubling pattern. Every time an agent picks up a new skill, it loses parts of what it already learned. Researchers call it “capability erosion.” Practically, it means agents theoretically get smarter but steadily become worse at tasks they once mastered. In one test, agents trained on complex workflows scored just 42% on simpler jobs they previously aced. Continuous evolution is starting to look a lot like early onset AI dementia. (arXiv)
Microsoft showed off its next wave of AI tools at Build 2026. Only a few stood out from the usual Microsoft fog. First came MAI, a new homegrown AI model family that Microsoft claims matches Claude Opus 4.6 in coding and Nano Banana 2 in images. A bold claim for a fresh model. Then came Scout, a personal assistant that’s basically Microsoft’s answer to OpenClaw for Office 365. It’s designed to be a habit forming way for non-techies to automate their office work. Arguably the most useful announcement was Agent Control Specification, an open source standard that lets companies set rules for what agents can do, when approvals are required, and what’s logged. Time will tell if any of this beats Copilot. (Microsoft, 404, TechCrunch)
Treat AI projects like onboarding new employees. Metadata is your AI’s “employee handbook,” making it faster, more reliable, and scalable. In this guide from You.com, you’ll discover:
- The key to high-quality results
- How to avoid costly mistakes
- Gold-standard templates
Don’t let your AI initiatives stall. Get the guide
👀 closer look
Trump’s new AI executive order looks more like access control than oversight. The White House says advanced AI models pose national security risks, so Trump signed an order with classified benchmarks, a cybersecurity clearinghouse, and voluntary 30 day reviews for select advanced models. A tougher version got scrapped after Big Tech execs complained it could slow American AI. The final watered-down order lets AI companies keep their latest models hidden. The real move is the “trusted partners” clause, letting the government and AI firms jointly control early access. Policy experts warn Washington might pick winners. AI safety is still fuzzy, but the White House just took control of who gets in. (AP)
Researchers use multiple AI agents for advanced cyber attacks. Their system, CAESAR, splits a cyberattack across several agents. One studies the target, another picks strategies, others run the tools, and a validator checks what worked before the agents try again. CAESAR works as a coordinated system, with agents sharing successes and discarding failures to rapidly improve performance. Tested against 25 Capture the Flag challenges, CAESAR outperformed single agent setups across every major security category. Security teams now have a clear warning, and hackers just got step-by-step instructions. (arXiv)
⏰ UP NEXT
AI companies are strip mining the creative economy. New York Times publisher A.G. Sulzberger called out AI firms for taking creative work without permission or payment. He labeled it a “brazen theft,” warning they’re repackaging stolen news and other content as their own. The fear is that readers will ask AI questions, get answers, and never visit the original creators. That matters to every independent publication like us, built entirely on earned trust. Sulzberger says the damage goes beyond journalism and puts the entire creative economy at risk. Creative professions employ more than 50 million people and generate roughly $12 trillion in economic value each year. Without creative people, AI just paints the world beige. (MarketWatch)
Sixteen mathematicians sound the alarm on AI. They say AI is undermining math, the one field built entirely on proving what is true. The trouble is now AI writes proofs that look correct while hiding errors only human experts can catch. Journals face a nightmare since flawed proofs are easy to generate and brutally slow to verify. Backed by the International Mathematical Union, the mathematicians released the Leiden Declaration calling for strict rules around AI-generated math. They also are warning AI companies against announcing math breakthroughs in press releases before peer review. Math used to be the discipline where showing your work was not optional. Now AI spits out questionable breakthroughs faster than anyone can check them. (Sci American)
fun stats
⚖️ 75%. How often law professors picked AI answers over answers from fellow law professors in a blind Stanford study. They also called the AI answers misleading less often. Awkward faculty meeting incoming.
👕 1,000. Photos of you Google wants before unlocking its new AI wardrobe feature, so you can mix and match outfits. Apparently your closet needs a training dataset. Or maybe Google does.
🟡 100 billion. Tokens OpenAI’s top user burns through each month, according to Sam Altman. So roughly $88,000 to $1.75 million worth/month. And that’s still not the biggest account. Altman claims one external customer uses even more.
🫸 $1,500. Strict monthly token spending limit Uber imposed per AI coding tool for every employee after blowing its annual budget in 4 months. Tokenmaxxing is officially dead.