Happy Friday! And please be kind to VCs today. Many are suffering from Not Owning Anthropic. Symptoms include sweating, desperation, and unsolicited proof of funds screenshots. Anthropic stock is so hard to buy that holders are getting ridiculous offers. One investor was offered millions in cash and a VC partnership for his shares. Then a banker dangled his $4.8M Marin County luxury home to get early employees to sell. Anthropic is still private, but the FOMO has already gone public.
Now AI agents are learning by watching people work. Standard Intelligence is training a new breed of AI agents called “computer use models” that learn from videos of people using computers. Most models learn from text, which captures instructions and finished answers. For agents, Standard argues video is better because it shows the messy part: wrong clicks, trial and error, and the moment someone figures it out. They trained their foundation model on 11 million hours of user videos, then tested it by putting a laptop on the dashboard of a Toyota Rav4 and using it to steer around San Francisco. Investors treated that like a signal and threw $75 million at them. (The Information)
Dreame’s AI fridge has a robot arm that puts your groceries away. The company showed off its N1 Refrigerator concept at its NEXT event in San Francisco. The robot fridge will unpack your food, organize shelves, track expiration dates, and suggest recipes using AI, cameras, and sensors. Sounds helpful… until it starts acting less like an appliance and more like airport security. The N1 scans your face and hand to recognize you, then uses health data from your smartwatch and scale to recommend meals. Great, now you need biometric clearance just to grab a snack. (Gizmodo)
There’s a shortcut to ship scalable AI agents. Google Cloud’s Startup technical guide gives you pre-built frameworks to design, build, and deploy intelligent autonomous systems faster.
Inside the guide, you’ll discover:
- Frameworks to accelerate autonomous agent design from day one
- Best practices for prompt engineering workflows
- Frictionless deployment strategies using Google Cloud infrastructure
👀 closer look
Big Tech’s AI profits are starting to look suspiciously circular. Google and Amazon both posted blowout earnings, but a giant chunk came from paper gains on Anthropic. Alphabet’s stake jumped by about $28.7 billion. Amazon booked $16.8 billion from its Anthropic bet. The weird part is both companies also host Anthropic’s models, sell its cloud capacity, and keep pumping its valuation higher. It’s Silicon Valley’s version of recycling, but for valuations. (Fortune)
The US just lost one of its top scientists to China. Charles Lieber was once ranked the world’s top chemist and spent 30 years at Harvard. Then he was convicted for lying about his China ties and filing false tax returns. Now Reuters found him running i-BRAIN, a state backed brain computer interface lab in Shenzhen with dedicated chip-making equipment and access to a facility with 2,000 primate cages. China has made brain computer interfaces a national priority, so naturally it made him an offer he couldn’t refuse. The US convicted him. China handed him a top-tier, well funded brain lab. (Reuters)
state of play
The White House is trying to block Anthropic from expanding Mythos. Anthropic wants to let more companies into Project Glasswing, a controlled rollout for the model. The White House says Mythos is risky because it can find and exploit software bugs, which is fair. Except OpenAI’s GPT-5.5 reportedly benchmarks in the same league, so this is not exactly one rogue lab with a flamethrower. First, the Pentagon labeled Anthropic a supply chain risk, ordering fed contractors not to use it. Now Washington wants to limit who Anthropic can sell to in the private market too. The US still has no real AI regulation, but somehow it wants to control the guest list. (WSJ)
Elon Musk admitted using OpenAI to train Grok under oath. In federal court, OpenAI’s lawyer asked Musk if xAI had used OpenAI’s models through distillation, meaning one AI learns from another. Musk first said all the AI companies are doing it. When pressed, he claimed “it is standard practice to use other AIs to validate your AI.” OpenAI has been trying to stop rivals like DeepSeek from distilling its models, while warning Washington that China could “appropriate and repackage American innovation.” Anthropic also cut off OpenAI and xAI from Claude over similar concerns. Everyone wants model distillation treated like theft right up until they’re the ones doing it. (Verge)
fun stats
⭐ $16.20. Average monthly revenue per active user for Anthropic, versus $2.20 for OpenAI. Fewer users, richer enterprise nerds.
👨✈️ 90%. Fortune 500 companies using AI agents built with Microsoft’s low code Copilot tools. Less than half have security controls for them.
🦾 $2/hour. Operating cost for humanoid robots in 2026, per a new study. Cheap robot labor could bring production back to the US and hand robotics companies a $750 billion payday by 2035.