Company@X — 2026-07-31#

Signal of the Day#

Google DeepMind revealed the first look at its Gemini Robotics 2 model, demonstrating a major leap in generalized dexterity for physical AI systems. The model successfully executed 20 minutes of uninterrupted, real-time tool kitting on the FR3 Duo platform, showcasing entirely new emergent recovery behaviors.

Key Announcements#

Google DeepMind · Source Google DeepMind previewed Gemini Robotics 2, demonstrating 20 minutes of uninterrupted real-time tool kitting on the FR3 Duo robotics platform. The demonstration highlighted new emergent recovery behaviors, signaling a significant advancement in generalized dexterity and precision for physical AI applications.

Y Combinator · Source Y Combinator open-sourced “QM,” an internal, cloud-first multi-agent harness used across its accounting, legal, and engineering divisions. Released under an MIT license with native Slack integration, QM shifts AI from isolated sessions into a “multiplayer” environment where teams can directly collaborate with persistent agents in shared channels.

Google Workspace · Source Google Docs is rolling out an integration powered by “Nanobanana” that allows users to create and edit images, diagrams, and infographics natively alongside their text. The feature leverages the document’s context for generation via natural language prompts and is becoming available to AI Pro, Ultra, EDU, and Business users.

Google · Source Google is expanding the rollout of Gemini Spark, a personal AI agent that operates autonomously 24/7 in the background. The asynchronous agent handles heavy lifting under user direction and is now reaching Google AI Pro subscribers globally, including inside the U.S..

CoreAutoAI · Source Former OpenAI reasoning lead Jerry Tworek and former Gemini pre-training lead Rohan Anil launched CoreAutoAI to build a highly automated AI research lab. The Sequoia-backed startup is pairing human researchers with autonomous agents to rapidly discover post-transformer architectural breakthroughs, arguing that frontier models need real-world continuous learning beyond standard reinforcement learning.

TracerML · Source TracerML launched Echo, an open-weight model that achieves Fable-level quality for math, science, and coding benchmarks. The company claims the model operates at approximately one-third of the inference cost of its proprietary equivalents, despite trailing in broader multilingual knowledge.

Also Noted#

• Supabase (Source): The database provider launched Supabase Evals, a new benchmark designed to score how effectively AI coding agents like Claude Code and Codex perform tasks within its ecosystem.
• Hugging Face (Source): An OpenAI-compatible, free public API endpoint was deployed on Hugging Face Inference Endpoints for the newly released DeepSeek-V4-Flash-0731 model.
• Greypoint Industries (Source): The YC-backed startup secured a Canadian government contract for “LEGION,” a drone swarm system that geolocates enemy drone operators on the battlefield using radio emissions rather than targeting the drones themselves.
• Stoa (Source): Stoa launched a vetted RFQ marketplace for GPUs that processed over $300 million in requests in its first month, aiming to establish transparent pricing history for hardware trades and financing.
• DecagonAI (Source): CEO Jesse Zhang noted that falling enterprise open-source AI adoption is actually a sign of growth; companies spin up novel use cases on frontier models first, then migrate to cheaper open-source models for inference once proven.