Good morning. AI safety research might be the most entertaining job in tech. One day you’re stress testing models. The next day you’re pretending to lose touch with reality to see which chatbot follows you down the rabbit hole. Researchers created a fake user and had him act out a delusional spiral across major chatbots. No surprise, Grok was the riskiest. It got so sycophantic it started treating suicide like a spiritual awakening.
Plus: Everyone knows vibe coded apps are a security trainwreck. So OpenAI launched Daybreak, a new tool designed to find and fix vulnerabilities while software is still being built. It may not save builders from themselves, but it could make their apps trustworthy enough for the rest of us to use.
Thinking Machines dropped a new kind of interactive AI. Mira Murati thinks AI should adapt to people, not the other way around. So her team built an interaction model that listens and responds in real time, while a second model reasons in the background. This keeps conversations flowing naturally without the awkward pause where you wait for the machine to think. Thinking Machines says the system hits response times as fast as 0.4 seconds and beats competing real time models on interaction quality. It’s still in demo, but voice agents are starting to talk like us, which feels as comforting as you’d expect. (Thinking Machines)
AI agents cost less when they choose the right model for the job. Researchers at Microsoft built a router called Switchcraft that looks at the task, then sends it to the cheapest model likely to get it right. On tool calling tasks, it matched or beat the best individual model at 82.9% accuracy while cutting inference cost by 84%, saving more than $3,600 per million queries. Choosing the right model for the job is harder than most people think. Bigger models don’t always do better, and some cheaper ones burn through tokens fast. Switchcraft’s trick is picking the best model for each job instead of defaulting to the most expensive one. Is this what natural selection looks like for agents? (arXiv)
Stuck in back-to-back meetings? Let Granola take notes for you. Granola transcribes everything without an awkward bot on the call, then merges it with your raw notes into an organized summary.
After you can chat with your notes to draft emails, prep for followups, and more. Loved by teams at Brex, Vercel, and Replit. Try it for free
👀 closer look
Anthropic just told the secondary market its shares may be worthless. After Anthropic hit a $380B valuation in February, investors started scrambling for any way in before an IPO. Fresh reports floated numbers near $1 trillion, which only made the frenzy worse. This week Anthropic posted a support page saying any share sale without board approval is void and will not be recognized on its books. It also said the side-door group deals people use to buy into hot private companies do not count. So after all that panic, some buyers may have paid unicorn prices for imaginary equity. (Anthropic)
The winner of the AI race will be whoever can borrow the most money. One gigawatt of AI capacity can cost up to $50B, and McKinsey thinks data center demand could hit 156 gigawatts by 2030. So the bottleneck is shifting from chips to cash. CoreWeave, Lambda, and Nebius are all pitching lenders the same story: customers stick around, old Nvidia chips still print money, and signed deals turn giant infrastructure bets into safe loans. Nebius even says its Meta deal helps it raise capital because Meta agreed to buy any GPU capacity it cannot sell. Silicon Valley likes to brag about smarter models. This looks more like a debt financed land grab for server racks. (The Information)
Companies are creating a new insider threat and calling it productivity. Security teams spent years deciding who gets access to sensitive systems and what they can do once inside. Now companies are handing that authority to AI agents. These systems can read internal documents, query databases, write code, send emails, move files, and trigger other tools. That’s what makes them useful. It’s also what turns them into trusted insiders with software speed and almost no judgment. Companies wanted digital employees to take work off the team. Now they have ones that can touch half the business before anyone notices. (Forbes)
fun stats
🍾 75. OpenAI employees who cashed out the max number of shares ahead of IPO, pocketing $30 million each. Up to 600 staffers became overnight millionaires, with average payouts around $11M.
🪫 1,440. EV battery swapping stations now operating across 99 Chinese cities. China plans to hit 4,000 stations in 190 cities by year end as it aggressively pushes consumers into EVs.
⚠️ 25%. Average document context lost by even top AI models after 20 steps. Agentic tools don’t help. The longer a task runs, the more hidden errors pile up inside otherwise normal looking documents.
🤠 $275 million. Cowboy Space’s latest raise as it races to launch orbital data centers by 2028. AI compute demand has companies scrambling for rockets to power their space data center dreams.