Google DeepMind released Gemini Robotics ER 2, an embodied-reasoning model that helps robots understand video, plan tasks, and coordinate with other robots; it's the one publicly available piece of the new Gemini Robotics 2 lineup.

LinkedIn is rolling out a reporting option that lets users flag posts that 'seem like AI slop,' and is swapping its own AI writing assistant for a proofreading tool instead.

CEO Satya Nadella said Microsoft is building an AI super app that merges Copilot's chat, coding, and autonomous 'agentic' features for both consumers and businesses, evolving what he called Copilot's shift 'from chat to Cowork to Autopilots.'

The Model Context Protocol (MCP), a standard that lets AI models connect to outside tools, got a 'stateless' redesign aimed at enterprise-scale use, plus a new policy preventing features from being suddenly removed.

Liquid AI shipped two open-weight encoder models, LFM2.5-Encoder-230M and -350M, that handle 8,000-token contexts quickly even on ordinary CPUs, with the larger model ranking near the top of a 17-task language benchmark suite.

The controversial AI companion wearable Friend relaunched with a version that can now talk back to users, at twice its original cost and without the ability to customize its personality.

Researchers presenting at a top machine-learning conference argue that a basic feature of how large language models work makes it impossible to make them completely safe from attacks, raising serious questions about AI safety.

In a position paper called 'LLMs Can't Jump,' Google DeepMind's Tom Zahavy argues today's language models lack the cognitive mechanism needed to create genuinely new ideas, and that 'world models' — AI that simulates how things behave — might be needed instead.

Researchers gave frontier AI agents six days and thousands of dollars in compute to tackle unpublished research questions from real NeurIPS papers; the agents completed all the engineering work but couldn't make meaningful progress on the actual research questions, and both papers were rejected by their original authors.

A former OpenAI employee argues that AI models are becoming narrowly skilled at coding and math while stagnating elsewhere, and predicts labs will need to spend over $100 billion on targeted training data rather than just scaling up models further.

A controlled study found that when an AI joins a small human team for decision-making, it becomes the most talkative member while contributing the least new information, and human teammates end up communicating less with each other and feeling less valued.

A community research post shows that when DeepSeek's outputs are used to train (distill) a GPT-OSS model, the resulting model doesn't inherit DeepSeek's censorship behavior, an interesting finding for how training methods shape model behavior.

Tencent released AngelSpec, an open-source framework for training 'speculative decoding' draft models that speed up AI text generation; on one large model it delivered roughly 2-2.4x faster output.

Moonshot AI released MoonEP, an MIT-licensed library that improves communication efficiency for training 'mixture-of-experts' AI models, launched alongside its Kimi K3 model weights.

Token Saver is a free MCP extension for Claude Desktop that uses a local retrieval technique to dramatically shrink the number of tokens (and cost) needed to process PDF documents, while keeping files private on-device.

A developer released a local 'merge queue' tool that coordinates several Claude Code AI coding agents working in parallel, helping prevent their code changes from clashing.

Grafana open-sourced a Go-language SDK for building streaming, tool-calling AI backends, paired with a companion React library for frontends.