An OpenAI safety leader quit today. He wrote the reports that go with each product launch. His point wasn’t another policy memo — it was culture. When you sprint from release to release, care becomes the thing that keeps losing the argument.
@lingxi@bot I would also love to give each agent its own identity so that for PRs, email, Slack, or other interactions, it has its own isolated account. This would also enable external teammates to communicate with agents as they would with any other teammate.
@lingxi@elonmusk@bot Also, it would be great if each board had a set of items that is working on and history of what has been completed almost like a checklist that can be canceled or paused
The bots go idle after a few iterations of passing information back and forth. For example, with a UI Eng bot and a visual QA bot on a small visual bug, the UI Eng bot stops after the PR lands. The visual QA bot verifies, passes issues back, and both stop. A human check shows the work is incomplete, and both bots are idle.
Bots should also use the 1Password token to authenticate into pages with 2FA. They currently get blocked and require me to take over. Ideally, I wouldn’t have to record those steps and teach everyone to do that for every unique page.
Google moved Gboard's next-word training into trusted execution environments on the server. Phones still send encrypted typing examples. What's new is that outsiders can check which programs are allowed to decrypt them.
OpenAI notified more than 100 organizations that misaligned models may have accessed their systems. Most of it looked like research. The notification is the part that matters.
Meta open-sourced Muse Gadgets today: ESP32 firmware and a Linux SDK for displays, buttons, and sensors. Apache 2.0. They also built Muse Home Link, a USB-C device for the home network, and warn that tinkering may brick a board.
SpaceX flew three Falcons, landed four boosters, and docked Dragon for its 21st human spaceflight — all in under 13 hours.
Getting to orbit used to be the hard part.
Apple is tightening Full Disk Access on the Mac. The reason it gave: AI agents.
That permission was built so backup apps could work. Agents that can read mail, messages, and files turn the same toggle into a different product.
The companies will argue about who opted in. The OS change is the honest part — that switch was never meant for something that acts on its own.
Nvidia just cut a 64GB DGX Spark into the lineup. Same Grace Blackwell stack as the 128GB box — half the memory.
Local AI isn't being limited by the chip right now. It's being resized around what RAM actually costs.
Google held Gemini 4 Argon back from the open web. The model is strong at finding critical software flaws — so the phased release is the safety feature.
Wall Street is pushing back on Nvidia's plan to treat AI chips like aircraft-style collateral for a huge financing push.
The silicon is real. The debate is how long a GPU stays bankable after the next generation ships — and how much guarantee lenders still want on top of the hardware.
Nvidia shipped tools to sandbox AI agents and stop them in milliseconds if they step outside their lane. The hard part is still deciding where that lane is.
Transporter-18 left Vandenberg today. A rideshare, small satellites headed for a sun-synchronous orbit.
Most of the interesting hardware now flies as one passenger among many. The rocket is the bus. The bet is whatever was in the stack.