|
⏱ The 10-Second Version
- OpenAI's ChatGPT can now read, draft, and send iMessages on your Mac, with approval required before each send by default, though OpenAI's own documentation flags a known way to disable that check.
- Nvidia is paying AI coding startup Poolside $6 billion for a non-exclusive model license and another $1 billion for equity, roughly quadrupling its valuation to $12 billion and extending job offers to 109 of its employees.
- Anthropic's revenue run rate hit $65 billion in July, up roughly 600% year over year, putting it on track for an IPO some analysts think could rival SpaceX's record debut.
|
|
The Big Thing
ChatGPT Just Got Permission To Read And Send Your Texts.
What Happened
OpenAI rolled out a ChatGPT plugin for Mac this week that reads iMessages, drafts replies, and sends messages on your behalf, across iMessage, SMS, and RCS. Getting there takes real effort: Full Disk Access, contact permissions, automation approval. By default, ChatGPT asks before sending each message. But OpenAI’s own documentation flags a known issue: certain tasks can disable that approval prompt, letting messages go out unsupervised. The launch also lands amid Apple’s trade-secret lawsuit against OpenAI, testing whether Apple treats this the way it treated Beeper Mini in 2024, or lets it slide. Read the full story at 9to5Mac →
Why You Care
Ask your team this week which of your AI tools already have standing approval to act without a human checking first.
|
What to Say in Your Next Meeting
“The company Apple is suing for allegedly stealing its trade secrets just got permission to read and send your texts, and the safety switch has a known way around it.”
|
|
|
Speed Round
Nvidia is paying AI coding startup Poolside $6 billion for a non-exclusive license to its models and investing another $1 billion for equity, a deal that roughly quadruples Poolside’s valuation to $12 billion from $3 billion a year ago and comes with job offers for 109 of its employees. The structure mirrors Microsoft’s 2024 Inflection deal: buy the technology and the talent, skip the acquisition review. Nvidia didn’t acquire Poolside, technically, the same way you technically didn’t eat the whole pizza.
Read the full story at CryptoBriefing →
Anthropic’s annualized revenue run rate hit $65 billion in July, roughly 600% growth year over year, putting it on track for an IPO some analysts think could rival SpaceX’s record debut. The company has committed to spending $100 billion on Amazon’s cloud over the next decade and up to 5 gigawatts of Amazon’s own chip capacity, while separately paying SpaceX roughly $1.25 billion a month through 2029 for Nvidia chip capacity, a purely commercial deal with no equity attached. The most profitable seat in the AI boom might just be Anthropic’s landlords.
Read the full story at The Motley Fool →
Google’s open Gemma models passed 1 billion cumulative downloads this week, with developers building more than 100,000 variants on top of them, running everywhere from a NASA satellite doing image analysis to a 100-million-download Indian health app to a Georgia Tech project decoding dolphin vocalizations. Somewhere between low orbit and dolphin research, a free model quietly became infrastructure.
Read the full story at Unite.AI →
Apple Music will require a “Made With AI” label on any track substantially generated by AI, the company confirmed this week, after its own VP disclosed that over a third of monthly uploads are now 100% AI-generated, even though that music accounts for less than half a percent of actual listening. AI is flooding the upload button and getting ignored at the play button in roughly equal measure.
Read the full story at MacRumors →
Linear’s own usage data shows AI now authors nearly half of all issues created on the platform, and teams using coding agents roughly tripled their weekly pull requests, from 21 to 65, over two years. But total product development time still went up, not down, the kind of result researchers call a Jevons paradox: agents added work instead of removing it. Three times the pull requests and somehow still behind schedule; that’s not a bug, that’s the industry’s business model.
Read the full story at Linear →
|
|
Prompt of the Day
The AI Permission Audit
OpenAI’s own documentation for its iMessage plugin admits there’s a known way around the approval prompt that’s supposed to keep a human in the loop before it sends anything, and that plugin almost certainly isn’t the only AI tool in your stack with that kind of standing permission. Most teams have granted various AI tools access to email, calendar, files, or messaging over the past year without ever centrally checking which ones can act on their own versus which ones just ask first. Two minutes gets you the list.
Here's what AI tools our team has granted access to:
[paste: e.g. tool names, what each can access, email,
calendar, messaging, files, and whether it requires
approval before acting or can act on its own]
Tell me:
1. Which of these tools can currently send, delete, or share something without a human confirming first
2. Which ones we assumed required approval but might not, based on how the tool actually works
3. The one permission to turn off or double-check this week
|
|
Forward this to the coworker who granted an AI tool Full Disk Access and moved on with their day.
— Don.
|
