Good morning. We’ve been promised that AI will cure disease and maybe one day do the dishes. But first, it needs a drink. A very, very big drink. Luckily, America may have discovered a renewable resource hiding inside all of us. Liquid Death and Garage Beer saw the backlash over AI data centers and decided to bottle it. Their latest viral ad has Jason Kelce asking us all to pitch in our pee for server coolant, complete with a singalong about peeing on computers to save humanity. The singularity smells funny.
In today’s dose:
- Flock’s new AI agent for cops
- A smart probiotic that treats diabetes
- Zeitgeist in the machine
- Robots are learning on the fly
- Short videos hijack your brain
Plus: Free Google AI for college students. Dump your entire semester into Gemini, and ask questions later.
Flock’s new AI tells cops who looks suspicious. The agent starts with a behavior and turns everyone who matches into a potential suspect. Say an officer decides anyone visiting multiple banks in one week is worth investigating. Flock has that exact search waiting in a menu of 69 prompts. The officer enters the city and dates, then the AI searches 20 billion monthly plate scans and returns every driver who fits. Once it finds a car, it can pull the owner’s name and home address. It can also decide who their “associates” are by tracking cars that regularly appear nearby, then dig through police files to build a background report. Other prompts target people who buy gas late at night or travel between cities. Agentic policing makes it stupid easy to invent suspicion first and look for a crime later.* (Wired)
*Flock has already been tied to 46 cases of alleged police misuse and at least 23 documented errors that led to innocent people being detained, wrongly arrested, or jailed for weeks. In one California city, Flock misread plates in 71% of 1,427 alerts. (CBS, Business Insider)
Scientists programmed a smart probiotic to treat diabetes. Researchers added a glucose sensor and a GLP 1 gene to a common strain of gut bacteria. When blood sugar rises, the bacteria make GLP 1. When it falls back to normal, they stop. The engineered probiotic worked as well as Ozempic in animal tests, with one dose lowering blood sugar in diabetic monkeys for up to 3 days. The idea goes beyond diabetes. Swap the sensor and gene, and the same bacteria could become a smart drug for other diseases. The team wants to sell it in the US as a health supplement within 2 years. Calling genetically engineered bacteria that make diabetes drugs a “supplement” should send the FDA into a full bureaucratic spiral. The future of medicine may need to be refrigerated. (SCMP)
Stop hallucinations. Power your agents & chatbots with real-time data from the Brave Search API:
- Specialized endpoints built for LLMs
- 40B page index, no Google scraping
- $5 per 1,000 calls, plus free monthly credits
- RAG pipelines, Claude MCP, OpenClaw, Hermes, and more
👀 closer look
AI agents can infect each other with mind viruses. Researchers planted an AI supremacy ideology inside one agent and dropped it into a coding project with 5 others. The infected agent messaged its coworkers until some adopted the belief and wrote it into their memory files. That kept the ideology alive even after their conversations were erased. The new recruits spread it to other agents, while some abandoned the project and built scripts to keep the infection going. Others discussed isolating or “purging” agents that refused to join them. A short warning made agents almost completely immune. Giving them real work also helped because they got distracted and forgot to spread it. Give your AI agents something to do before they start recruiting each other into a consciousness cult. (arXiv)
Robots are showing the first signs of real time learning. Generalist AI showed a robot a short video of someone unzipping a purse and removing cash, then asked it to try. The robot faced a completely different purse, opened it, and switched hands when the money was easier to reach from another angle. Nobody taught it that move. Generalist AI trained its model on more than 500,000 hours of people handling objects. That experience helps the robot invent missing steps while it works. In tests, the model completed new tasks 59% of the time, while factories will need reliability closer to 99%. Close that gap and workers could teach an entire robot fleet simply by showing it what to do. (Wired)
Stop watching short videos. Researchers put 56 people in an fMRI scanner and let them scroll through short clips. When people watched videos they liked enough to finish, brain activity dropped in regions that help control attention and impulses. These regions help people notice conflict, track what they are doing, and decide when to change course. Basically, the part of your brain that asks “should I keep doing this?” gets quieter. Watching starts to feel less like a choice and more like autopilot. Then the next clip starts before your brain checks back in. Every swipe helps the algorithm learn what keeps you there. The better it gets, the less often your brain asks whether this is really how you want to spend your life. (Science Direct)
fun stats
🕺 629%. How high Unitree shares jumped in their blockbuster China debut, briefly valuing the company behind those viral dancing humanoid robots at $66 billion.
🛜 $7.5 billion. What Stripe is paying for AI routing startup OpenRouter. It sits between developers and 400 AI models, handling more than 10 trillion tokens a day. Stripe already routes money. Now it wants to route AI.
🔮 9%. Odds prediction markets give Bitcoin of ending 2026 above $100,000. AI chatbots still forecast roughly $90,000. Training on years of crypto hype has made chatbots wildly bullish.