Google DeepMind
·
8 months ago
An AI company published the third version of its Frontier Safety Framework, expanding risk assessment categories and refining protocols for identifying severe hazards from advanced models. The update introduces a new Critical Capability Level focused on harmful manipulation and adds Tracked Capability Levels as of April 17, 2026 to catch less extreme risks earlier. The framework now requires safety case reviews before external launches and expands such reviews to large-scale internal deployments of models that could accelerate AI development.
Google DeepMind
·
8 months ago
Google DeepMind is partnering with Commonwealth Fusion Systems to use artificial intelligence for controlling plasma in fusion reactors, focusing on optimizing the SPARC tokamak machine designed to achieve net energy gain. The collaboration centers on three areas: developing TORAX, a fast plasma simulator built in JAX; using reinforcement learning to find optimal operating parameters; and creating AI-based real-time control systems to manage heat distribution and plasma stability. The partnership aims to accelerate the timeline toward delivering fusion energy to the grid by running millions of virtual experiments before SPARC operation begins.
Google DeepMind
·
8 months ago
Google DeepMind and Kaggle launched Game Arena, an open-source benchmarking platform where AI models compete in strategic games to evaluate their capabilities. The platform will host its inaugural chess tournament on August 5 with eight frontier models competing, with final rankings determined by over one hundred matches between each model pair. Game Arena addresses limitations of current benchmarks by using games with clear winning conditions to measure reasoning and planning rather than memorization, with plans to expand to Go, poker, and video games.
Google DeepMind
·
8 months ago
Google integrated a new image editing model called Nano Banana into the Gemini app, enabling users to edit photos while maintaining consistent appearance of people and pets across multiple versions. The model is described as the top-rated image editing model available, with capabilities including multi-turn editing, photo blending, and style transfer. Users can now apply costume changes, background modifications, and complex edits while preserving specific elements of their photos, with all generated images marked by visible and invisible watermarks.
Google DeepMind
·
8 months ago
Google DeepMind released VaultGemma, a 1-billion-parameter language model trained with differential privacy, accompanied by new scaling laws that describe how privacy, compute, and data budgets interact during training. The model was trained with a sequence-level privacy guarantee of (ε ≤ 2.0, δ ≤ 1.1e-10) and showed no detectable memorization of training data in empirical tests. VaultGemma performs comparably to non-private models from approximately five years ago, establishing a baseline for measuring progress as privacy-preserving training methods improve.
