Good morning. Remember when every company decided the secret to productivity was “more AI”? Very visionary. Now the bill has arrived, and the same CEOs who wanted employees mainlining tokens are acting like someone left the lights on in the data center. Tokenmaxxing got so dumb some companies turned AI use into a leaderboard. One spent $500 million in a month, another had an employee burn $150k alone, and the most AI-pilled companies are now hitting $7.5k per employee every month. Bosses are suddenly adding token caps and pushing cheaper models, pretending restraint was the plan all along.
Up next: Loopmaxxing. Give an agent a goal to figure out, and it’ll keep asking the model what to try next. It’s like burning tokens on steroids and calling it autonomy.
The first credible signs of AGI in robots are here. Boston Dynamics’ Atlas is finally approaching the kind of autonomy factories actually need. Recent demos show early signs of humanoid general intelligence, meaning Atlas can walk into unfamiliar spaces and handle tasks without being scripted in advance. The secret is training speed. Boston Dynamics simulates millions of Atlas training hours per day, then moves new skills onto a real robot in about an hour. They’ve also simplified Atlas’s hardware, so the simulated bot behaves just like the physical one. That’s how you get from flashy YouTube backflips to smoothly lifting a 100 pound refrigerator, balancing, anticipating weight, and controlling force. AGI was supposed to arrive in a chat window, but it might show up first wearing steel toes. (Interesting Engineering)
Washington gave the world a reason to stop trusting American AI. Anthropic’s sudden shutdown made global companies realize their AI products could vanish overnight if someone in DC gets nervous. China couldn’t have bought a better ad. Its AI models are cheaper, self hosted, and harder for Washington to switch off. DeepSeek’s flagship model costs $0.87/million output tokens – about 60x cheaper than Anthropic’s Fable 5. In early June, four of the five most popular models on OpenRouter were Chinese, and among the top 20, Chinese models processed 2x as many tokens as US models. American AI labs are valued like they’ll sell intelligence to the world, but customers now know US AI comes with a DC panic button. The absurd part is why Washington panicked. Security researcher Katie Moussouris says the supposed “jailbreak” was basically the model being asked to “fix this code.” Washington turned bug fixing into a geopolitical crisis. (Financial Times, The Register)
Zuck’s new AI search engine will make you cringe. Meta is rolling out AI Mode inside Facebook Search, letting people ask questions and get answers pulled from public posts across Facebook, Instagram, and Threads. That means its turning vacation flexes, neighborhood panic, and half-baked posts into official search results. Meta wants the chatter turned into a product, which is either brilliant distribution or the fastest way to make AI sound like the internet after two glasses of boxed wine. Morgan Stanley sees a $10 billion opportunity if Meta keeps a billion users and monetizes just 10% of daily queries. But it also means every brand, scammer, and influencer now has incentive to seed posts for the bot to quote. (TechCrunch)
fun stats
🤝 $3.6 billion. What Salesforce paid for customer AI agent platform Fin as it goes all in on enterprise automation.
🛒 75%. Consumers who are nervous about letting an AI agent complete a purchase and pay on its own, even with spending limits set in advance.
🌍 13+. Countries planning to ban social media for kids under 16, including the UK. Critics say age gates are invasive and could force digital ID checks on everyone.