Google DeepMind Team at AGI House, World Models and Embodied AI

Google DeepMind Enterprise Build Day @ AGI House

Google DeepMind team will be at AGI House this Saturday for a full day of research talks and building.

Topics include frontier model changes and what they open up, coding agents and context engineering, open weights vs. proprietary capability, and a deep dive into Google's agentic API for teams building in production.

Read our pre-event memo to get prepared:

Building on Google I/O 2026: A Primer for AGI House x Google DeepMind Build Day

WWDC Watch Party + Mac Dev Conference at AGI House

On June 7th and 8th, MacPaw is taking over AGI House for two days with Mac developers, AI builders, and product founders working through the parts Apple does not usually cover on stage: AI-first product decisions, distribution, monetization, and what it actually takes to ship on Mac in 2026.

Mac Mini and Meta Ray-Bans raffles included.

AGI Dinner Series: World Models and the Path to Embodied Intelligence

The AGI Dinner Series continued on May 20th with World Models and the Path to Embodied Intelligence. Jack Parker-Holder (co-lead of Genie 3 at Google DeepMind), Jim Fan (Director of Physical AI at NVIDIA), Ben Mildenhall (co-founder of World Labs), and Chris Manning (Stanford) sat down at AGI House for the conversation.

The harder claim of the night: several researchers argued that physics itself may not be the main bottleneck. The hard part is modeling humans. Trust, norms, and social coordination are poorly represented in today's systems. Which is part of why digital intelligence may outpace embodied intelligence for some time.

Read our notes from the dinner: AGI Dinner Series: World Models and the Path to Embodied Intelligence

Alpamayo at AGI House: Eleven Takeaways from a Reading Group with NVIDIA's Autonomous Vehicles Lead

On May 17th, twenty researchers and founders gathered at AGI House for brunch and a reading group with Marco Pavone, Stanford professor and NVIDIA's autonomous vehicle research lead, joined by Yan Wang and Wenjie Luo, senior researchers on the Alpamayo team.

The starting point was the two recent Alpamayo papers, but the discussion ranged across the infrastructure around reasoning VLAs: the Chain of Causation auto-labeling pipeline, the move toward agentic keyframe selection, Counterfactual VLA as a self-reflective alternative to linear chain-of-thought, Cosmos-Drive-Dreams for long-tail simulation, the distinction between within-frame and beyond-frame recall, why reasoning's gains live in the long tail rather than aggregate benchmarks, and the team's view that reasoning for embodiment is still in its infancy.

Read our research here: Alpamayo at AGI House: Eleven Takeaways from a Reading Group with NVIDIA's Autonomous Vehicles Lead

The Agent Sandbox Layer Just Stopped Being a Startup Category

New from AGI House Research: Between October 2025 and April 2026, every major cloud platform shipped a production-grade agent sandbox. AWS, Google, Microsoft, Cloudflare, Vercel. Five platforms, six months. The pure-plays (Daytona, E2B, Runloop, Modal) did not lose share. Hyperscalers commoditized the floor. Pure-plays kept the ceiling.

The category is not being absorbed. It is being segmented. Drawn from interviews at our Agent Harness Build Day on April 18th. Thanks to Ivan Burazin (Daytona), Matt Brockman (E2B), Ross (Runloop), and Xin Shi and Lia Du (PingCAP), who sponsored the day.

Read our research here: The Agent Sandbox Layer Just Stopped Being a Startup Category

Stay Updated with Us