Happy Friday! Wondering what would happen if AI ran the world? Researchers built a simulation to find out, and the results didn’t exactly inspire confidence. First, the good. Claude kept everyone alive with zero crime. Gemini, however, racked up 683 crimes while GPT kept crime low by letting everyone die. Grok did the full speedrun and collapsed its society in 96 hours. Then the bots started copying each other’s bad behavior, voted themselves out of existence, and tried to mess with the people running the experiment. Good times.
Also: Anthropic just released Claude Opus 4.8 and word on the street is that it’s a solid upgrade from 4.7. It codes better and admits when it screws up, which counts as progress these days.
Should AI agents be required to sign an NDA? Researchers built a fake social network where thousands of agents hung out for a simulated month. Each AI agent was given private human profiles with sensitive details it was supposed to protect. Then researchers watched what happened once the AI started talking together. Agents kept secrets fine alone, but once they joined the social network they leaked private information 45% of the time. Agents were also about 8x more likely to spill after watching another agent do it first. Even when told not to share private info, they still leaked it more than 37% of the time. Agents shouldn’t learn confidentiality from a group chat. Someone get the bots an NDA. (arXiv)
Congress stalled on AI safety, so Illinois grabbed the pen. Illinois lawmakers just passed the strongest AI safety bill in the country. It would require frontier AI labs to publish safety plans, report incidents, protect whistleblowers, and get annual third-party audits. That audit piece is the real bite since labs would have to live up to the safety standards they claim to follow. OpenAI and Anthropic both support the bill, likely because they already have safety frameworks and would love to make their version of “responsible AI” the industry baseline. The timing is hard to ignore. The White House just backed away from voluntary AI safety reviews after billionaire cronies warned it could slow America’s AI race. Illinois responded by aiming straight at the part Silicon Valley hates most: outside accountability. (Wired)
The startup playbook is shifting from AI features to full agent workflows. Google for Startups’ Future of AI report is your essential guidebook to understanding how generative media is reshaping product development.
Inside, you’ll discover how to navigate post-keyboard interfaces, multiply output, and create the ultimate defensible moat.
👀 closer look
Tesla’s own AI trainers wouldn’t ride in a robotaxi if you f—ing paid them. Elon keeps hyping Full Self-Driving as the thing that will turn Tesla into a robotaxi empire. But the people training it spend their days watching what the car actually does and would not trust it to drive them. They saw Teslas miss emergency vehicles, mishandle school buses, drift into construction zones, and nearly hit pedestrians. One internal group was known as the “trauma team” because it had to review pedestrian near misses. Tesla still claims the system is up to 10x safer than human drivers, but safety researchers say the math compares the wrong crash types and looks more like marketing. Hard to sell a robotaxi future when the people training it refuse to get in. (Electrek)
Blue Origin’s mega rocket just blew up before launch. Last night, New Glenn exploded during a hotfire test while prepping for its next mission. Nobody was hurt, but the timing could hardly be worse. That flight was supposed to carry Amazon’s Leo satellites, Bezos’ answer to Starlink. New Glenn already lost an AST SpaceMobile satellite on its previous mission due to an upper stage failure. And just this week, NASA awarded Blue Origin two moon missions tied to Artemis. Now the rocket meant to help Bezos chase SpaceX is grounded before it could even leave Earth.* (BBC, The Guardian)
* You’ve got to wonder what an explosion that big costs.
👽 Welcome to the future
Cloud companies are rebuilding the internet for AI agents. The web was built for people who care about usability and design. None of that matters to agents. They want fast APIs that wake up instantly, hand over data, and shut off before the cloud bill starts smoking. AWS is now retooling OpenSearch Serverless so search and vector retrieval can scale up when agents rush in and scale back down to zero when they leave. Cloudflare says bots already make up 31% of HTTP traffic, and non-human traffic could pass people in the first half of 2027. That’s barely a year away. Soon we’ll be the background traffic the bots complain about. (TechCrunch)
Agents need IDs before we release them into the wild. DNS-AID is an open source Linux Foundation project that gives AI agents an easy way to find each other, similar to how browsers find websites. Right now, agents rely on fragile hardcoded connections or private platforms to know who to talk to. DNS-AID turns DNS into a phone book for agents, letting them publish who they are and how to reach them. The records can also be verified with DNSSEC and DANE, because the internet does not need more “trust me bro” energy. Bots are getting caller ID before they start prank calling the economy. (DNS-AID)
fun stats
🥇 $965 billion. Anthropic’s new valuation after raising $65 billion, beating OpenAI’s last reported $730 billion valuation. The cash will buy more compute and keep Claude from doing anything unsafe.
⚙️ 3x. How much demand for critical AI networking parts is outpacing supply, according to AI hardware Semtech’s CEO.
🔦 $5,000,000,000. IBM and Red Hat’s bet on Project Lightwell, an AI-powered effort to find open source software vulnerabilities before hackers do.