AI Mornings with Andreas Vig
Your daily AI news briefing in under 10 minutes. New models, product launches, research breakthroughs, and industry shifts, explained clearly, no hype.
AI Mornings with Andreas Vig
TML's Interaction Models & DeepMind's AI Mathematician
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Hey! Welcome to AI Mornings with Andreas Vig. It's the 13th of May 2026. Mira Murati's Thinking Machines Lab just broke its silence with something genuinely different. Instead of racing toward autonomous agents that run TAS on their own, TML introduced what they're calling interaction models, AI built for real-time collaboration. The system takes in voice, video, and text in two zero zero millisecond chunks, letting you talk, show things, interrupt, and steer while it keeps working. A background model handles the slower reasoning and tool use while the live model stays in conversation with you. Morati said the way we work with AI matters as much as how smart it is. It's a contrarian bet against the agentic direction most labs are chasing, and it's one of the first clear differentiators from her lab since she left OpenAI. Google DeepMind is applying the Agentic Coding Playbook to mathematics. They just published research on an AI co-mathematician based on Gemini 3.1, and it set a new high on Frontier Math Tier 4 at 48% more than double Gemini 3.1 Pro's baseline of 19%. The system runs parallel work streams with sub-agents that write code, search literature, and attempt proofs. But here's the remarkable part. An Oxford professor named Mark Lackenby actually solved an open problem from the Kuroka notebook by spotting a clever proof strategy buried inside an output the system's own reviewers had rejected. That's the promise of AI math tools, not replacing mathematicians, but helping them see things they might have missed. Now a story that's flown under the radar but could be a big deal for AI in healthcare. Medicare is launching a 1-0-year program called Access that goes live July 5th, and it's the first federal payment mechanism designed specifically for AI-driven care. Instead of paying for clinician time, the model rewards health outcomes, things like lower blood pressure or reduced pain. One participant, a company called Pair Team, is using a voice AI agent named Flora that handles patient intake and check-ins around the clock. Some patients, like a six, seven-year old woman living out of her car, have spoken to Flora for over an hour. That companionship piece is turning out to be a genuine intervention. The economics only work for AI-first operations, which is exactly the point the government is creating a competitive swim lane for AI innovation in a heavily regulated industry. In startup news, AI voice platform Vapi just hit a 500 million US dollar valuation after raising a 50 million Series B. What's interesting here is how they won Amazon Ring's Business Ring, evaluated more than 40 AI voice vendors before choosing Vapi. Now 100% of Ring's inbound calls go through Vapi's platform. The startup processes between 1 million and 5 million calls a day and has handled over a billion calls total. They're focused on the infrastructure layer, not pre-built applications, which gives enterprises more control over reliability and compliance. Quick research highlight.1's tool calling ability into just 26 million parameters. That's tiny. You can fine-tune it locally on a Mac, and it still beats larger models like Function Gemma 270M and Quen06B on single-shot function calling. It's designed for consumer devices, phones, watches, glasses, where running frontier models isn't practical. Google DeepMind also shared a concept for reimagining the mouse pointer. The idea is an AI-enabled pointer that understands what you're pointing at, not just where. You can say things like, show me directions while pointing at a building in an image, or compare these while selecting products on a web page. The approach turns pixels into actionable entities. It's coming to Chrome and a new Google Book laptop experience, and it's part of a broader push toward interfaces that meet users where they work instead of forcing them into AI-specific windows. A few more things worth knowing about today. University of Warwick astronomers used an AI system called Raven to find over 100 confirmed exoplanets from NASA test data, including 31 never-before spotted planets. Anthropic expanded Claude for Legal with new plugins and connectors for DocuSign, Box, and Westlaw. Exaforce, a cybersecurity AI startup, raised 125 million US dollars at a 725 million valuation for real-time threat detection. China's Quai-Show is planning to spin off its Kling AI video branch at a projected $20 billion valuation with an IPO target of 2027. And Greece is proposing to add AI protections to its constitution, citing threats to democracy one of the first constitutional responses to AI risks. That's it for today. See you tomorrow.