DeepMind / Google: robotics, Gemini agents and AGI commentary

DeepMind/Google introduced Gemini Deep Research Agent and a unified Interactions API for developers. - Gemini Deep Research Agent is built on Gemini 3 Pro with multi-step RL fine-tuning - Features: autonomous planning, targeted iterative search, deep integration of private docs and web, verifiable citations - New Interactions API provides a single endpoint for models and agents with structured outputs and high developer control - Published internal benchmarks: DeepSearchQA 66.1%, Humanity's Last Exam 46.4%, BrowseComp 59.2% - Google expanded Google AI Plus to 36 more countries with Nano Banana, Veo access and Gemini integrated into apps - DeepMind/Google robotics commentary: VLA + LLM hybrids, Motion Transfer for cross-body learning, emphasis on simulation data - DeepMind scientist Shane Legg estimated 50% chance of minimal AGI by 2028 under his definition - Robotics research stresses data bottlenecks, world models and tactile sensing for dexterous hand tasks
Key insights
Gemini Deep Research Agent released for rigorous developer research
DeepMind announced the Gemini Deep Research Agent, built on Gemini 3 Pro and trained with multi-step reinforcement learning to improve accuracy and lower hallucination. The agent autonomously plans research, iteratively searches, integrates private documents with public web data, and outputs verifiable citations. DeepMind published new internal benchmarks—DeepSearchQA 66.1%, Humanity's Last Exam 46.4%, BrowseComp 59.2%—to demonstrate the agent’s performance on complex, causal-chain research tasks. The agent is available via the Interactions API for developer integration and structured outputs.
Interactions API unifies model and agent calls for developers
Google released the Interactions API, a unified endpoint for interacting with models and agents that supports high control, structured outputs (JSON), and developer-defined prompt constraints. The API is the integration surface for the Gemini Deep Research Agent and is available via Google AI Studio and Gemini API keys, enabling applications like finance due diligence and biotech literature aggregation.
Google expands Google AI Plus and Nano Banana access globally
Google announced Google AI Plus expansion to 36 additional countries, increasing availability to 77 countries total. The plan offers higher limits for Nano Banana image generation, access to Veo 3 Fast, FlowbyGoogle, Whisk, NotebookLM expansions, Gemini integration in Gmail/Docs/Sheets, and enhanced storage; promotional pricing includes 50% off for six months for new subscribers.
DeepMind and Google robotics research emphasizes simulation, motion transfer and world models
DeepMind robotics leads highlighted integration of LLMs and reinforcement learning: Vision-Language-Action (VLA) plus world models, Motion Transfer enabling cross-body data reuse, and the need for tactile sensing for dexterous manipulation. Researchers stressed simulation and synthetic data as scalable sources and noted remaining gaps between simulated and real-world robotics performance. Claims included that robotic locomotion and gait are largely solved, while hand-level dexterity remains at toddler levels.
DeepMind-associated scientists offer AGI timeline and capability caveats
DeepMind figures articulated an AGI estimate and capability gaps: Shane Legg estimated a 50% chance of attaining a minimal AGI by 2028 under a definition requiring completion of typical human cognitive tasks. Commentary noted language skills are superhuman across many languages, but persistent learning, visual reasoning subtleties and long-term continual learning remain bottlenecks.