Good morning!

We’re here with another day of big AI headlines. Let’s dive in👇

👀 Big Moves, Big Money

💰 Mira Murati’s Thinking Machines Hits $12B Valuation

Former OpenAI CTO Mira Murati just raised $2B for her stealth-mode AI startup, Thinking Machines Lab, now valued at $12B with no product yet. The company promises a multimodal model with a strong open-source core, built to collaborate naturally with users via conversation and vision. First product drops in “a couple of months.”

🇺🇸 Trump Announces $92B AI & Energy Push in Pittsburgh

At Carnegie Mellon, President Trump announced a $92B investment package across AI and energy, declaring America’s “destiny” to be the AI superpower. It includes billions for data centers, electric grid upgrades, and modernizing hydro plants in Pennsylvania, part of a broader vision to outpace global rivals through tech infrastructure.

🧠 Mistral Drops Voxtral, Its Open-Source Voice Model

Mistral’s Voxtral suite just hit the open-source scene. Designed for speech understanding, it includes both a high-power 24B model for large deployments and a compact 3B variant for edge devices. With transcription plus native Q&A built in, it’s another nudge toward fully open voice-first interfaces.

⚙️ Where AI Meets the Real World

🌆 San Francisco’s AI Boom Comes With Growing Tension

Mayor London Breed is giving AI assistants to the people, or at least to San Francisco’s public workforce. In a first-of-its-kind move, the city is rolling out Microsoft Copilot to 30,000 municipal employees. The goal: modernize government operations, automate repetitive tasks, and make the city the AI capital of the world.

🛡️ Pentagon Picks 4 Tech Firms to Scale Military AI

The Department of Defense is partnering with four commercial AI companies to accelerate battlefield tech development. The program focuses on AI agents for real-time decision-making, logistics, and threat detection, signaling a major pivot to adopting Silicon Valley speed and innovation across U.S. military operations.

🧪 LLMs Still Struggle with Long Contexts

New experiments show that as input tokens grow, model performance drops sharply, even on simple tasks. Despite having longer context windows, today’s top models fail to maintain consistent output, suggesting context length remains a fundamental bottleneck in LLM reliability.

🛠️ Tools of the Day

• Voxtral – Mistral’s open-source voice model suite

• Prezi AI – Create smart, visual-first presentations

• Mozart AI – Cursor for music production

• Prompter – Generate AI prompts based on your context & projects

💬 Prompt of the Day

⚡ Quick Hits

• Anthropic launches Claude for financial services

• Google AI agent “Big Sleep” spots security flaw preemptively

• Princeton boosts cybersecurity AI attack rate with $36 compute

• Google invests $25B in data centers and hydropower in PA

• xAI introduces Grok-powered interactive avatars

📌 TLDR

Mira Murati’s new lab is now worth $12B before launch, Trump just pledged $92B to AI and energy in Pittsburgh, and Mistral’s Voxtral goes open source. Meanwhile, the Pentagon is scaling AI in defense, and San Francisco’s AI boom is sparking culture clash. LLMs? Still not great at long conversations.

See you tomorrow,
David