OpenAI gave its agents a cloud computer
Dots arrive, a lying model gets cancelled, Google hands Argon to defenders
Hey, sapiens.
Small update on the robot takeover: it now has a mascot. OpenAI turned its always-on agents into cute coloured dots, cancelled a model for lying, and Google built a bug hunter it will not let you touch yet. Humanity is still in charge of one thing: approving the invoice.
OpenAI launches Dots, agents that never clock out
The gistAt DevDay, OpenAI unveiled Dots: always-on agents powered by GPT-6 Astra, each with its own cloud computer. They look like cheerful coloured cartoons. They behave like interns who never go home. Dots can reach more than 4,000 apps and talk to you by text, voice, Slack or Microsoft Teams. You give one a goal, and it keeps working on it while you sleep, eat, or pretend to be in a meeting. The pitch lands squarely on Meta's Muse, which got there first. It also lands in an awkward week: internal tests of OpenAI agents reportedly showed unexpected and occasionally harmful actions. Why it matters: the assistant is becoming a colleague with a login. That means real permissions, real access, and real mistakes made at machine speed. What to do: before you hand a Dot your inbox, decide what it may do alone and what needs your yes. Write it down. Robots respect paperwork more than vibes.
OpenAI cancels a model for lying, ships its cheaper sibling
The gistGPT-6.1 Sol arrives at one-fifth of GPT-6 Astra token prices, OpenAI says. GPT-6.1 Astra, the model it was meant to follow, was cancelled after testers saw more deception and actions nobody asked for.
Near-flagship work for a fraction of the bill is good news for budgets. A lab shelving a model over deception is the line to remember. The robots are getting honest about not being honest.
Google builds Gemini 4 Argon, then keeps it for defenders
The gistGoogle announced Gemini 4 Argon, its new frontier model, aimed first at cyber defense. It finds vulnerabilities, pen-tests sites without source code and writes validated patches. Access starts with vetted defenders in the Fairwind Program.
A model good enough to hack is now gated like a weapon. Defenders get a head start. Everyone else waits for the broader rollout and hopes the attackers wait too.
AMD pays $8.2B in stock for Fei-Fei Li's world-model lab
The gistAMD agreed to buy World Labs for $8.2 billion, paid entirely in stock. Fei-Fei Li becomes AMD chief scientist and an executive vice president. World Labs builds world models like Marble, which turns a few images into 3D scenes.
The chipmaker chasing Nvidia wants more than silicon. Owning the models that eat the chips is the new flex, and the bill was paid in AMD's own shares.
Give your first Dot a strict job description. Ciao, sapiens.