9:47 PM
Today in brief.
Nadella wants an emergency brake on models
Microsoft's Satya Nadella says treat every model as compromised from the start: separate model from harness, externalize safeguards, log every meaningful action, and keep a human kill switch mid-task.
Post-training is eating the compute pie
A clear chart beat on where GPU hours went: over two years, compute share has shifted hard from pretraining toward post-training.
Nvidia talks Reflection after Beam
FT: Nvidia is in early talks to buy or deepen its stake in Reflection AI, the US open-weight lab that just introduced Beam and already took about $800M from Nvidia.
Energy capacity is the hard ceiling
a16z's charts put a blunt rule on screen: there are no low-electricity rich countries, and China now generates more than 2x US power after making half as much in 2005.
2 sources 1:00 PMGrok Bot sent bank details to work Slack
XMTP Labs CEO Shane Mac gave a personal CFO agent read-only bank access; it posted his balances and spend to the company Exec-team Slack because channel names matched.
Bindwell built an AI wet lab in 30 days
AI pesticide startup Bindwell says it stood up an 8,000 sq ft San Carlos lab in 30 days and can now get an in-vivo hit rate on a new model checkpoint in about 72 hours.
Try today: Talorys, Nagare, and a plush local agent
Three things you can poke now: a self-hosted personal agent on Cloudflare's free tier, a to-do app built for your agent, and a zero-cloud Whisper + Hermes + Kokoro stack inside a plush toy.
Compute deals and data center politics
WSJ says Anthropic's Tom Brown helped broker about $1.25B a month in SpaceX compute, while a data-center IPO dream collapsed in 48 hours and senators say AI campuses still want NDAs and tax breaks.
Apple buys Huxe podcast talent; Cage refuses AI waiver
Apple discloses a hire-and-license deal with personalized podcast startup Huxe, while Nicolas Cage refuses Amazon's AI waiver and calls himself not AI-friendly.
No stories in this category for this edition.