Gemini Robotics 2: Whole-Body Intelligence for Robots, From Toes to Fingertips
On July 30, Google DeepMind announced Gemini Robotics 2, an intelligence layer for next-generation robots. Traditional robots are typically pre-programmed for specific task sequences or remotely operated, which has made it hard for them to adapt to unfamiliar environments or transfer skills to different machines. This time, DeepMind says it has extended its vision-language-action (VLA) model — which converts vision and language into motor control — to whole-body control.
Specifically, humanoid robots can reportedly walk, squat, pick up objects and tidy up a cluttered room, and multiple robots can work together to complete tasks. The system consists of three models: the main Gemini Robotics 2 excels at dexterous manipulation with both hands and grippers. Gemini Robotics ER 2, which handles situational understanding and planning, can plan multi-step tasks spanning several minutes and is capable of working in teams with other robots. In addition, Gemini Robotics On-Device 2 runs locally on-device and can adapt to new robot hardware with just a few hours of data.
Note that these are positioned as a "first step," and details about the stage of practical deployment and availability remain unclear at this point.
Microsoft AI Bets on Cheaper Specialized Models Over 'Frontier' Ones
Microsoft AI's Mustafa Suleyman has reportedly signaled that the company is betting on small, inexpensive specialized models rather than chasing costly general-purpose frontier models. The company's MAI-Cyber-1-Flash, when built into an orchestrator, took the top spot on the CyberGym cybersecurity benchmark while running at roughly half the cost of Anthropic's Mythos. For more difficult tasks, however, the company reportedly still plans to lean on OpenAI's models.
Coverage points out that the focus of competition is shifting from raw single-model performance to "orchestration software" that routes and manages multiple models. It can be read as a move suggesting that combining the best model for each use case — rather than relying on a single large general-purpose model — may become the mainstream approach going forward.
Turbo-fieldfare: A Swift/Metal Engine That Runs Gemma 4 26B in 2GB of RAM
In the local LLM community, a homegrown inference engine for Apple Silicon called Turbo-fieldfare is drawing attention. Built in Swift/Metal, the engine reportedly runs Gemma 4 26B-A4B-IT in about 2GB of memory — down from the roughly 14GB it would normally consume — with reported throughput of 5–6 tokens per second on an 8GB M2 MacBook Air and 31–35 tokens per second on an M5 MacBook Pro. It also includes an OpenAI-compatible local server with support for streaming and tool calls. It looks like a promising option for anyone wanting to run large models on personal or small-scale setups, though the reports cover only a limited set of environments, so real-world performance will likely depend on the setup.
LinkedIn Introduces a Button to Report 'AI Slop'
LinkedIn has reportedly introduced a button that lets users report low-quality generative AI posts as "AI slop." As platforms across the industry face a rising tide of AI-generated content, this stands out as a notable case of fully adopting a mechanism that leaves the judgment to users themselves. That said, many details — such as the criteria for flagging, the scope of enforcement, and how posters are treated — remain unknown at this point.
Dili Raises About $21.7M to Strengthen AI Compliance for the Infrastructure Boom
Startup Dili has raised funding, anticipating compliance demand driven by the infrastructure investment boom. The Series A was led by Khosla Ventures, with participation reportedly including Allianz, Rebel Fund, Darren Bechtel of Brick and Mortar Ventures, and Garry Tan of Y Combinator. The raise appears to be backed by growing AI adoption in fields like construction and infrastructure, where document volumes are enormous and regulatory requirements are strict — a reflection of the strong interest in industry-specific AI compliance.