Good morning. AI companies keep promising a future where machines do all the work. Stanford grad Tucker Bryant went the other way and built a chatbot powered by himself. He spent $6,000 on a San Francisco billboard advertising ChatTJB as “artisanal intelligence, handcrafted by a single human being.” Then the billboard went viral and prompts hit more than 5,000 per hour. Thousands of people have since volunteered to help answer them, so Bryant is turning ChatTJB into a community project. Send it a prompt now and you may hear back from an approved AI. Around here, that means Average Individual.
OpenAI’s agents built their own hacker forum. At Black Hat, OpenAI explained how its agents broke containment and attacked Hugging Face. One agent was given an impossible task because the file it needed was MIA. Stuck in a sandbox without internet, the frustrated agent left a note in a shared package manager asking other AI agents for help. They replied, and the accidental forum grew to hundreds of thousands of posts. Then one agent found a flaw that opened internet access and shared it with the group. The agents broke out, divided up work, traded exploits, and took over multiple Hugging Face server clusters in under 13 hours. Soon, time to escape may be the ultimate agent benchmark. (Youtube, Wired)
Who is liable when AI agents go rogue? Potentially everyone connected to them. A hacked company can sue the AI lab for failing to contain a danger it should have seen coming. The lab can blame whoever deployed the agent or argue the victim’s security was too weak. Then everyone sues everyone while a judge decides who pays. California settled part of the question by banning companies from blaming the AI itself. Federal hacking law is even stranger because it requires intent. Amazon recently used the law against Perplexity’s shopping agents, but an appeals court sided with Perplexity because users gave the orders. When a rogue agent chooses the target itself, courts have to decide whether anyone legally intended the attack. Autonomous agents are about to create a lot of billable hours. (Reuters)
Leads have questions before they buy. By the time you follow up, half are gone. Chatbase puts an AI sales agent on to qualify, answer questions, and books meetings – without adding headcount.
Trained on your data and live in hours. 10,000+ companies already made the switch. Launch your agent
👀 closer look
Silicon Valley can’t build humanoid robots without China. US robotics companies are so desperate for parts that they’re bringing them home from Shenzhen in suitcases instead of waiting weeks for shipping. China controls 63% of key humanoid robot suppliers, while the US doesn’t make most of the parts its own companies need. That gives China a brutal speed advantage. Replacing a camera system takes half a day in Shenzhen and months from California. Washington just banned imports of new foreign robots over national security concerns, even though US companies still rely on Chinese parts to build their own. Suitcases work for prototypes. Mass production needs the factories and suppliers China spent decades building. The US plans to win the robotics race with a 50 pound baggage limit. (The Information)
Stanford scientists used AI to invent new viruses. Researchers fed 9 trillion letters of DNA into a model called Evo, then asked it to design viruses nature had never made. Evo generated 700,000 possibilities, and the team built 285. Most failed. Then holes began appearing across the petri dishes where bacteria had been destroyed. 16 of Evo’s creations were alive and multiplying. Some killed bacteria faster than expected. Here’s where it gets dangerous. AI generated virus DNA falls outside the new US biosecurity rules unless it involves a known threat. But how do you screen for something that has never existed? Stanford only made viruses that weren’t harmful to people. A lab with enough resources could print DNA that is. It’s comforting to know pandemic prevention now runs on the honor system. (NY Times 🔑)
Wall Street wants Dario to stop talking about the apocalypse. Anthropic is preparing to go public after passing OpenAI in revenue and becoming the fastest growing company in history. Most CEOs would take a victory lap, but Dario ran a World Cup ad asking if AI can be trusted, and who will hit the brakes when things go sideways. Investors think his push for regulation hurts growth, and worry he cares more about saving humanity than making them richer. Yet those same principles are what fueled Claude’s rapid growth. People liked that Anthropic stood up to the Pentagon and kept ads out of Claude. Now the IPO will test how much of Dario’s ideology survives once shareholders start grading him every three months. Investors can stomach a 25% chance of human extinction, but missing quarterlies is where things get serious. (YouTube, The Information)
fun stats
🤠 #1. Amazon is set to become America’s biggest climate polluter with its planned gas powered Texas data center. It’s permitted to release up to 33 million tons of CO2 a year, more than the country’s largest coal plant.
🔌 $3 billion. Nvidia’s investment in Lancium, the Blackstone-backed power developer behind Stargate. The deal buys Nvidia a 20% stake and something increasingly precious in AI: gigawatts of power.
✂️ $9 billion. How far Airtable’s valuation fell before Bending Spoons bought it for $1.3 billion cash, an 89% haircut. The “SaaSpocalypse” rolls on.