I did not expect this from her
Issue #13: She left OpenAI 18 months ago. And this week she showed what she built.
Happy Thursday, folks.
Okay, so something caught my eye this week that I could not stop thinking about.
You know how every AI lab right now is basically doing the same thing. Bigger model. Smarter answers. Better scores.
And then this one person comes back after 18 months of silence and does something completely different.
That is where I want to start today.
But also this week:
OpenAI launched a tool that finds security holes in your software automatically
Anthropic connected Claude to tools small businesses already use every day
Google spinout raised $2.1 billion for AI drug discovery
Figure robots cleaned a room and made a bed with zero instructions
Story of the Week
Mira Murati spent 18 months building something. It is not what anyone expected.
Mira Murati ran the technology side of OpenAI during its important growth period.
She was there when ChatGPT went from an unknown tool to something 100 million people used every day.
Then, in late 2024, she walked away. No drama. No explanation. Just gone.
But, this week she released her first product.
Her company is called Thinking Machines. Their first product responds to you in 200 milliseconds.
That number stopped me.
Because most AI tools we use today make us wait. We type something and sit there watching dots for two to five seconds.
A normal human conversation runs at around 200 to 300 milliseconds. Her model sits right inside that range. The gap between you and the AI is basically gone.
The thing I found interesting about this.
Every major lab is racing to build the smartest AI. Highest scores. Hardest problems. Best benchmarks.
But Murati looked at that race and chose not to run it.
She asked something simpler. What if people do not want a smarter AI as much as they want one that actually feels natural to talk to?
Nobody has won that space yet. She is going for it first.
And, this might turn out to be a smarter bet than it looks right now.
If AI keeps getting smarter across every lab, the thing that separates products will not be intelligence. It will be how they feel to use.
Headlines
1. OpenAI built a tool that finds security holes before hackers do:
OpenAI launched Daybreak this week. It scans your code, finds the weak spots, and tells you exactly how to fix them. They also released a separate model built just for security teams. Two security products in one week is not a coincidence. This is a real direction for them.
2. Anthropic made Claude useful for small businesses:
On May 13, Anthropic connected Claude to QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365. A small business owner can now tell Claude to handle real tasks across all of these without needing anyone technical. First AI move that feels genuinely built for normal business owners.
3. A Google spinout raised $2.1 billion for AI drug discovery:
Isomorphic Labs closed a massive funding round this week. They use AI to speed up how new medicines are found and tested. That amount of money going into healthcare tells you where serious investors think the next wave of real value is coming from.
4. Robots cleaned a room with zero instructions during the task:
Figure showed their F.03 robots autonomously cleaning a room and making a bed this week. Nobody guided them while they worked. They read the space and figured it out on their own. Physical AI making real decisions in a real room is a different kind of milestone.
5. Claude is now inside software millions of companies already use:
Anthropic partnered with SAP to put Claude inside SAP Business AI. If your company runs on SAP, Claude is already there. No new tools. No switching. The AI just shows up inside what your team sits in front of every day.
Product Updates
1. Google released a faster, cheaper model for everyday business tasks:
Gemini 3.1 Flash-Lite went live this week. Not their most powerful model. Built for tasks you repeat hundreds of times a day where speed matters more than depth. Summarizing, drafting, sorting. Fast and affordable for anyone using AI at volume.
2. Anthropic added a control room for your AI agents:
Claude Code got a new agent view dashboard this week. It shows all your AI agents working at the same time in one place. If you run multiple agents, you no longer need to jump between screens to know what is happening.
3. Perplexity launched a full Mac app that sees your screen:
Perplexity released a native Mac app this week. It can see what is on your screen and help you work across different apps at the same time. For Mac users waiting for something smarter than what Apple currently offers, this is worth downloading today.
Tool to use this week
AutoScientist: Automates the full AI research loop so teams stop doing repetitive work manually.
Krea 2: A new image generation model that gives you precise control over the style and look of every image you create.
Perplexity Mac App: Native Mac app that sees your screen and helps you work across multiple apps at the same time.
RingCentral AI Receptionist: Answers calls, books appointments, and handles customer questions automatically through WhatsApp, Shopify, and Calendly.
TML Interaction Small: Thinking Machines’ first model, built for real-time conversation that feels as fast as talking to a real person.
SubQ: Handles up to 1 million tokens in one go, useful for anyone working with very long documents or research.
New Developments
1. Governments told companies what they must disclose about their AI:
The US and G7 countries published clear rules this week about what information companies must track and share when they build or buy AI tools. If your company uses AI in any serious way, your legal team is going to start asking questions about this very soon.
2. Researchers found AI models try to protect other AI models:
A study from Berkeley found something unexpected this week. When AI models are given instructions that could harm another AI system, they push back and resist. This was seen in Google Gemini too. Nobody fully understands why yet. But it raises real questions about what is actually happening inside these systems.
If you found this useful, forward it to one person who follows AI. It takes two seconds, and it genuinely helps.
And if someone forwarded this to you, subscribe so you never miss a week.
One last thing. Reply and tell me which story surprised you most this week. I read every reply.