Good morning. Robotaxis have entered the “I’m telling mom” stage. Two California teens were joy riding in a Waymo when the car decided to call it quits. After watching them drink and shoot Orbeez out the window, it pulled into a parking lot, locked the doors and called the police. Officers arrived ready for a high-risk stop, only to find toy guns instead of real ones. Waymo says its interior cameras and microphones help keep riders safe. Very comforting, unless the car decides you’re the emergency.
Meta thinks the problem with perv glasses is branding. In a world where everyone thinks they’re an influencer, Meta’s AI glasses have become the obvious creep tool used to record people without consent. Meta’s fix? A tiny glowing recording light. And a new software update that disables the camera if the LED gets blocked or tampered with. At the same time, Meta is testing always-on glasses that listen and snap photos every few seconds. The pitch is that AI can help people search what they saw and heard that day. The reality is a searchable record of your life with a questionable privacy policy. That’s the kind of training data every surveillance and physical AI company dreams about. (TechCrunch)
The latest version of ChatGPT knows when to shut up. OpenAI released GPT-Live, a next gen voice model built to make chatting with GPT feel more like talking to a person. It can laugh, breathe before speaking, follow messy conversations, and sit through pauses without jumping in every 5 secs. It also drops in the occasional mhmm to remind you the machine is listening. OpenAI’s trying to spin this as “one step closer to accessible AGI,” but it’s really just a less annoying chatbot. It’s far from perfect, but if you like talking to AI while walking, driving, or knitting (like the demo video), this version starts to feel useful. It’s clearly aimed at all the lonely people. Enable your microphone, and ChatGPT can quietly hang out in the background. You might even forget it’s there. (OpenAI, X)
Everyone wants to put AI in front of customers. Very few can make it accurate enough to trust.
Cube is the agentic analytics platform built on a semantic layer, so AI agents answer from trusted metrics instead of guessing. Brex built an AI financial analyst for 35,000+ customers on Cube and lifted answer relevance from the high 50s to nearly 90%.
👀 closer look
Cloudflare wants to make it easier for AI to use your data. Last year it helped websites block AI crawlers unless publishers gave permission or got paid. Now Cloudflare is testing a live view pilot that feeds OpenAI real time signals on what pages changed and which ones are worth crawling. This will help ChatGPT discover new web content sooner. Cloudflare sits in front of over 20% of the web, giving it a massive view of what is changing online. It also puts Cloudflare on both sides of the crawler fight. It sells publishers control over who gets access, then helps OpenAI reach the right pages faster. Either way, Cloudflare wins. (TNW)
Self-improving AI is becoming a DIY project. AI labs talk about recursive AI like it should be locked in a high security airlocked vault with zero access to the internet, watched by researchers with badges. Now the same loop is running in a home lab. WIRED’s Will Knight used Claude and Karpathy’s AutoResearch to train a smaller AI model, test what came out, change the recipe, and run it again. The first version babbled nonsense. A few rounds later, it started making more sense. We’re not talking about building GPT6 level brains in your bedroom or anything. But it’s entirely possible to build software that gets better at building itself. Somewhere, a frontier lab’s safety team just reached for the stronger coffee. (Wired)
What if AI paid your electric bill? Sunrun wants to turn your solar panels and batteries into paid AI infrastructure. The plan is to install small compute nodes in homes and pay homeowners for hosting them. Those nodes would run AI inference for enterprise customers using power from the same solar and battery systems already connected to the house. AI companies need power faster than utilities can build it, and homeowners have unused capacity that could be plugged into a distributed AI network. Instead of waiting years to build another data center, Sunrun is asking whether homes can collectively become one. It’s about time AI starts subsidizing solar. (Barrons)
fun stats
🪫 137x. How much more electricity AI agents use per query vs. chatbots. They also take 154x longer to respond, leaving GPUs idle more than half the time while the agent thinks really hard about clicking a button.
🔦 $25 billion. Amazon’s ‘surprise’ bond sale, flashing a warning sign about how expensive the AI boom is becoming. Hyperscalers have borrowed $194 billion so far this year.
🟡 $2. Cost per million input tokens for SpaceXAI’s Grok 4.5, about 2.5x cheaper than Claude Opus 4.8.