Happy Friday! Childhood is changing fast. Summer camp used to mean bug spray, friendship bracelets, and at least one kid crying because the canoe tipped over. Now it means AI tutors, omakase lessons, student-built surfboards, aerial combat, and luxury real estate training. That’s the actual pitch from Alpha, the notorious “AI-powered private school” bringing its 2 hour academic model to the Hamptons for $4,500 a week. Baby Burning Man has found its tribe, where rich kids learn teamwork by burning a Trojan horse on the beach.
Scientists just took away the only excuse for not editing embryos. For years, CRISPR gave scientists an easy reason to avoid human embryo editing because it made too many mistakes, scrambling DNA in ways that could be passed down for generations. Two new studies used base editing, a newer technique that swaps a single DNA letter without cutting both strands. It produced far fewer of the major DNA errors that made CRISPR a nonstarter. Researchers say the tech is still nowhere near ready for pregnancy because it can leave behind unwanted edits. But the conversation just changed from “Can we do this?” to “Should we do this?” Fertility companies are already selling embryo selection and openly planning gene editing for future genetic optimization packages. Ethics just became engineering’s hardest problem. (Nature, WP)
The government now decides who gets frontier AI first. OpenAI was preparing to roll out GPT 5.6 until the White House stepped in and took over the guest list. Now the model will go first to a small group of US companies and orgs approved by the Trump administration, with Washington reportedly greenlighting customers individually. This follows the Anthropic fiasco, where the government pulled Fable 5 offline over concerns about its guardrails and cyber risks. The latest executive order requires AI labs to submit frontier models for review 30 days before release, even though the review process is still being invented. Meanwhile, OpenAI is delaying its IPO as investors figure out how to value a company who can be shut down by Washington at any time. If the government picks the first customers, who do you think gets to skip the line? (Verge, NYT)
Wispr Flow works inside every Al tool you use. Dictate detailed prompts 4x faster than typing.
Flow handles camelCase, snake_case, and acronyms. Better input, better output. Used by millions of developers and Al professionals.
👀 closer look
Robot investors may have backed the wrong brain. Robotics startups spent years betting vision language action models could become the AI brain for real world robots. That sounds great until robots actually have to do stuff like plug in cables, grab slippery objects, or handle food without causing a crime scene. Chef Robotics CEO Rajat Bhageria says VLAs are still too slow and unreliable for real tasks. So attention is shifting to world models, which try to predict what happens before robots act and lets them practice in simulations instead of failing one expensive mistake at a time. The problem is world models still hallucinate physics, which is a big deal when your whole job is understanding physics. Real world experience is messy and expensive. You can’t scrape it off the internet. Wild how quickly the robot revolution stalls once gravity gets a vote. (The Information)
China is catching up to Claude and OpenAI faster than anyone expected. Z.ai released GLM 5.2 just days after the US restricted Anthropic’s two most powerful models, and developers love it. They say it’s nearly as powerful as Anthropic’s Fable and Mythos. It already cracked the top 10 global AI leaderboard, where six of the leading models now come from China. What makes GLM 5.2 stand out is its strength in coding and AI agents, areas US frontier models were supposed to dominate. It also costs about one eighth as much as Claude Opus 4.8. DeepSeek is pushing a similar narrative, training its new frontier model V4 entirely on homegrown Huawei chips. The marketing says America is way ahead. The leaderboard says China is already in the room, helping itself to the snacks. (NYTimes)
A startup claims it can make AI use 1,000x less power. Right now, the biggest constraint to scaling AI is energy. Unconventional AI says inference can get dramatically cheaper with oscillator based chips, which use pulsing circuits to run AI with less wasted energy. This isn’t startup vapor. A Nature report says oscillator computing has clear advantages for pattern recognition and machine learning. To prove the idea, Unconventional tested an image model on a simulation of its chip and says the results matched leading diffusion models. That is interesting enough to watch closely, and nowhere close to proving the giant power savings. AI’s power appetite keeps growing, and Unconventional is betting the industry is feeding the wrong hardware. (Nature, TC)
fun stats
💻 25%. How much Apple is raising Mac and iPad prices as memory chip costs soar. RAMageddon is officially upon us.
💅 100 billion. Transistors IBM crammed onto a chip the size of your fingernail. It’s now the world’s smallest sub-1 nanometer processor.
🍇 $4,500. Cost of AI summer tutoring in the Hamptons, compressing school into 2 hours so kids can spend the rest of the day foraging for wild edibles and learning to sell luxury homes.
😡 44%. Americans who resort to yelling “human!” or “person!” to escape customer service AI hell.