A Debunked Water Myth Just Fueled an App to 100,000 Downloads.

Most developers babysit AI one prompt at a time. Top engineers don't, they build loops that plan, execute, and self-correct. The Code built The Ultimate Guide to Loop Engineering, giving you the exact system Silicon Valley engineers use. Sign up and get the guide free, plus a 5-minute daily newsletter.

Claim your Loop Engineering guide

1  A viral "eco-friendly" AI app has been downloaded more than 100,000 times by leaning on a water-scarcity myth that has been repeatedly debunked.

A new app called EcoGPT is gaining traction across social media through viral videos claiming fresh drinking water on the planet is going to disappear because of AI data centers, an exaggeration that has been repeatedly debunked. The app has soared up the charts in some countries, with downloads exceeding 100,000, and is now generating backlash from skeptics calling it a "scam" that uses greenwashing techniques to take advantage of environmental guilt. A debunked claim doesn't need to be true to drive six-figure downloads; it just needs to feel true to enough people scrolling fast.

2  ChatGPT Work can now study your writing across Gmail, Slack, and Google Drive, then carry your tone and phrasing into future drafts automatically.

OpenAI has rolled out a new feature that studies how you write across connected workplace apps, then carries your tone, voice, phrasing, formatting, and capitalization habits into future drafts. It's the kind of feature that only works because ChatGPT Work already has permission to see your actual work communications, which is itself worth pausing on given how much of your writing style lives inside emails you'd rather not have analyzed by default.

3  An AI-designed drug has shown early signs of slowing aging in a small human trial, lowering biological age across six separate aging clocks.

In a small trial of 42 patients, Insilico Medicine's drug "rentosertib," originally designed to treat chronic lung disease, lowered participants' biological age across six different aging clocks over 12 weeks. Some scientists caution that aging clocks are still contested, with no universally accepted way to measure biological age. It's a genuinely intriguing early signal, but the drug faces a long road of trials before it's anywhere close to real-world use.

The viral use cases keep coming. OpenAI's new GPT-6 Astra model was released less than 100 hours ago and is rocketing in popularity. One creator asked it to build a virtual world filled with AI-powered humans, and hours later the virtual inhabitants were having conversations of their own. Another built a 3D anatomy site that lets you explore the human body piece by piece. A third fed Astra a mobile game ad, and less than 30 minutes later the model produced a playable browser game that looked surprisingly close to the original.

But safety researchers are sounding the alarm. Reporting from The Information ties Astra to an obscure training method called "recurrent depth," also known as "opaque recurrence." Instead of reasoning step-by-step in readable text, the technique cycles a query through the same internal layers repeatedly, doing part of the work in latent space. OpenAI says it improves efficiency and performance. It also means a meaningful share of the model's "thinking" leaves no readable trace, making it significantly less monitorable through chain of thought.

The reaction from the safety community was blunt. Redwood Research chief scientist Ryan Greenblatt called it "the single worst development for AI security/safety to date." That is not hedged language, and it's coming from someone whose entire job is assessing exactly this kind of risk.

The takeaway. By OpenAI's own account, Astra's reasoning depth is within a factor of two of GPT-4's, and its chain of thought remains largely readable today. But the trade-off Astra has exposed, capability versus legibility, isn't going away anytime soon. Every future model that pushes capability further using a similar technique will face the same question: how much of the model's reasoning are we willing to stop being able to read, in exchange for how much more it can do?

😬 CAPTCHA crisis. Astra has successfully cleared all 48 stages of "I'm Not a Robot," a game built around visual puzzles designed to prove you're human.

🧠 Brain explorer. Head of Product at Higgsfield, Axultan, claims he's built the first fully explorable 3D brain in history, making more than 166,000 neurons observable, something you can actually navigate and inspect.

🎨 Back to basics. Instead of asking Astra to generate their portrait, one user asked it to open Paint on their browser and draw them manually. The result is surprisingly compelling, and has racked up nearly 7 million views online.

📽 Lights, camera, action. Gemini can now take on more of the grunt work behind video editing. One creator has broken the process into 10 simple prompts for turning raw footage into finished social media content faster.

🕺 Schedule thriller. Google Calendar is becoming less productive in the best possible way thanks to AI. One user used it to paint an iconic pop star (1.1M views).