#DeepMind

Five years after AlphaFold 2, scientists have woven protein structure prediction into everyday research and drug discovery workflows while startups push accuracy further to meet drug-design needs.

'SIMA 2 upgrades DeepMind's embodied agent with a Gemini reasoning core, doubling task performance and enabling multimodal, self-improving behavior across commercial and generated 3D worlds.'

'DeepMind's new SIMA 2, powered by Gemini, learns to navigate and solve tasks across games like Goat Simulator 3 and procedurally generated worlds, improving through generated tips and retries.'

'CodeMender is an AI agent from DeepMind that localizes root causes, synthesizes validated patches, and upstreams fixes using Gemini Deep Think and program-analysis tooling, contributing 72 security patches in six months.'

Gemini Robotics 1.5 splits high-level embodied reasoning and low-level visuomotor control to enable agentic, multi-step robot behavior and zero-shot motion transfer across heterogeneous platforms

'DeepMind's Gemini Robotics 1.5 enables robots to perform everyday chores by combining reasoning, physical adaptability and web-sourced knowledge, while sparking corporate adoption and policy debates.'

DeepMind demonstrates a mathematical limit on fixed-size dense embeddings that causes retrieval failures in RAG systems at scale, and the LIMIT benchmark exposes this ceiling even on small toy tasks.

'DeepMind CEO Demis Hassabis warns AI may deliver change 10 times greater and 10 times faster than the Industrial Revolution, urging urgent cooperation and guardrails amid job, legal and ethical risks.'

Google DeepMind introduced LSM-2 with AIM, a novel framework allowing AI models to learn effectively from incomplete wearable sensor data, enhancing health-related predictions and robustness.

Google DeepMind released Gemini Robotics On-Device, a local AI model that enables robots to perform complex tasks in real time without cloud connectivity, opening new possibilities for edge robotics.