The US government just became your AI gatekeeper
Issue #12: Five things happened this week that most people scrolled past
Happy Thursday, folks.
I was scrolling through AI news for this week and something caught me off guard.
The US government quietly signed deals with Google, Microsoft, and xAI to see their AI models before anyone else does. No big announcement. I almost missed it.
And I could not stop thinking about it all week.
A lot more happened too. Let me walk you through everything.
TLDR
US government now previews frontier AI models before public release
GPT-5.5 Instant is your new ChatGPT default as of May 5
Cloudflare and Stripe gave AI agents the ability to make purchases autonomously
Harvard study says AI outperforms doctors in real diagnostic scenarios
DeepSeek closing its first-ever funding round at $50 billion valuation
Story of the Week
The US government just put itself between you and the next AI model
Three months ago the Trump administration was saying leave AI alone. Let the labs move fast. Minimal interference.
That lasted just 90 days.
And this week Google DeepMind, Microsoft, and xAI all signed agreements with the government. The deal is straightforward. Before any new frontier model goes public, the government sees it first.
OpenAI and Anthropic already had this arrangement quietly in place.
Now every major lab is inside that room.
What got me is how it happened. No press conference. No executive order.
Just agreements signed while everyone was focused on the GPT-5.5 launch.
Right now these are previews. The government looks at the model and gives feedback, and the lab decides what to do next. No approval required. No hard veto.
But here is what I keep thinking about.
Social media companies had the same kind of voluntary arrangements once. Friendly conversations with senators. Informal check-ins. Nothing binding.
Then came the subpoenas.
I am not saying that is where this goes. Maybe it stays a light touch forever.
But the door is open now. And in my experience, doors like that do not close.
Headlines of the Week
OpenAI changed what you see when you open ChatGPT
GPT-5.5 Instant is now the default model for every ChatGPT user as of May 5. It is faster, more concise, and makes fewer mistakes in law, medicine, and finance. Memory got better too. If ChatGPT feels different to you right now, this is why. Paid users can still access older models for now.
xAI shipped Grok 4.3 with voice cloning built in
Grok 4.3 launched with a 1 million token context window, stronger math and reasoning, and aggressive pricing. xAI also dropped a fast voice cloning tool at the same time. Online comparisons with GPT-5.5 and Claude have been everywhere this week. The gap between top models is genuinely closing.
Cloudflare and Stripe gave AI agents the ability to spend money
This one flew under the radar. The two companies released an open protocol that lets AI agents create accounts, buy domains, start subscriptions, and handle payments on their own. You set the spend limit. The agent does the rest. This is one of the clearest signs yet of where agentic AI is actually heading.
Harvard researchers say AI is now beating doctors at some diagnoses
A study published in Science found that AI diagnostic support can outperform doctors in certain real-world diagnostic scenarios. This one is from Harvard and landed in a serious journal. Not hype. Worth paying attention to.
DeepSeek is about to close a round at $50 billion
A few weeks ago the number being floated was $20 billion. Now it is $50 billion. DeepSeek has not even closed its first VC round yet. That jump tells you how much appetite still exists for frontier AI outside the US.
The Musk vs OpenAI trial actually started
The courtroom fight between Elon Musk and Sam Altman formally kicked off this week. The core question is whether OpenAI abandoned its nonprofit mission when it went commercial. Whatever the outcome, it will shape how AI companies structure themselves for years.
Product Updates
Mistral built a system for running AI reliably in production
Mistral released Workflows this week. It is an orchestration engine that handles multi-step AI processes, tracks failures, and makes everything auditable. If you are building anything serious with AI at work and tired of things breaking quietly in the background, this is worth a look.
NotebookLM now makes audio shows and they went viral
Google’s NotebookLM added audio digests this week. The output blurs the line between human-made and AI-made audio content. It went viral fast. I tried one. Genuinely hard to tell the difference.
Must Try Tools
AI.cc: One API, one bill, and you get GPT-5.5, Claude, Gemini, DeepSeek, and Llama all in one place.
Dataiku Kiji Privacy Proxy: It sits between your data and any AI model and makes sure your private info never leaves your system.
Brightfin Spend Clearly AI: If your company’s AI and tech costs are getting out of hand, this tool helps you see exactly where the money is going.
OpenBind v1: A free dataset and AI model built specifically to speed up drug discovery research, worth knowing if you work in biotech or pharma.
Pure Signal MCP Server: A new MCP server that makes your AI agents smarter and more useful inside enterprise workflows.
Mistral Workflows: Mistral built a system that runs your multi-step AI tasks reliably in production without things quietly breaking in the background.
New Developments
The Academy banned AI from the Oscars
The Academy of Motion Picture Arts and Sciences voted to ban AI-generated performances and AI-written scripts from Oscar categories for the 98th ceremony. Creative industries are drawing lines. This one is official.
Two AI startups shut down quietly this week
NeuroPixel.AI and Covrzy both announced shutdowns in early May. The funding environment is still active, but it is not forgiving everyone. Worth remembering while the big launches grab all the attention.
The EU AI Act enforcement deadline is getting pushed back
EU negotiators voted to delay the hardest enforcement deadlines by over a year. High-risk AI systems may not face full compliance until late 2027 or 2028. The rules stay the same. Only the timing shifts. Use the extra time to actually prepare.
If you found this useful, forward it to one person who follows AI. It takes two seconds, and it genuinely helps.
Want this every Thursday? Subscribe below.
Follow me on LinkedIn for daily thoughts between issues. Link below.