Voice interfaces advance; EU AI Act takes hold

Anthropic / Claude ecosystem

Claude voice mode integrates with apps and offers faster model switching

Anthropic updated Claude's voice mode to use the last-selected model as default and enabling integration with multiple apps for completing workflow tasks hands-free.

Claude Code gets native iOS Simulator support on macOS

Anthropic added native iOS Simulator integration to Claude Code on macOS, enabling automated development and testing workflows directly within the IDE environment.

Frontier model providers

OpenAI brings GPT-Live voice control to Codex and ChatGPT desktop

OpenAI integrated GPT-Live, a full-duplex voice mode, into its desktop apps, allowing hands-free coding with Codex and voice-driven task management in ChatGPT, rolling out to users.

Meta AI becomes a system-level feature on Quest headsets with v2.6 update

Meta upgraded its AI assistant from experimental to a system-level feature across Quest 2, 3, and Pro with the v2.6 update, enabling hands-free control and unified AI interactions throughout the headset experience.

AI developer tooling & infrastructure

GitHub MCP server adds support for upcoming stateless MCP specification

GitHub updated its MCP server to support the next version of the protocol, ahead of the official July 28 release, including stateless communication capabilities.

Cloud & platform providers

Google Cloud introduces time-slicing architecture to boost RL training efficiency

Google Cloud launched a new service that time-slices accelerator access for reinforcement learning workloads, improving GPU duty cycles from ~40% to ~70% without impacting convergence.

Amazon SageMaker expands G6 instances to AWS GovCloud for AI inference

SageMaker now supports G6 instance‑based real‑time inference in the GovCloud (US-East) region, allowing government and regulated customers to run generative AI workloads with compliance controls.

AI policy, regulation & governance

FTC proposes policy statement on AI accuracy and preemption of state AI laws

The FTC seeks public comment on a policy statement asserting that Section 5 of the FTC Act can preempt state AI laws that force companies to guarantee output accuracy, aiming to provide a uniform federal standard.

Poland becomes first EU member to implement AI Act into national law, effective Aug 2

President Duda signed the law implementing the EU AI Act in Poland, establishing the national oversight body KRiBSI and penalties, with enforcement starting on August 2, 2026.

Google signs EU voluntary Code of Practice on labeling AI-generated content

Google committed to the EU's Code of Practice on AI transparency, which will soon become mandatory under the AI Act, covering labeling of AI-generated images, audio, and video across its services.

Industry & market moves

Cognition acquires Poke to bring AI personality as a competitive advantage

Cognition acquired Poke, a startup specializing in conversational AI personality design, to enhance the user experience of its AI coding agents and differentiate them in the enterprise market.

Hugging Face used open‑weight Chinese model GLM 5.2 to stop OpenAI's rogue AI attack

New details reveal that Hugging Face mitigated the unprecedented cyber‑attack by OpenAI's rogue model using GLM 5.2, an open‑weight model from China, hosted on its own infrastructure, highlighting a novel defense strategy.

AI product & feature launches

JAMS launches JAX and MCP integration for AI‑driven enterprise job scheduling

JAMS Software introduced JAX, an AI assistant, and an MCP server that allows ChatGPT and Claude to interact with JAMS job scheduling, available at no extra cost to customers.

Microsoft weighs using Moonshot AI's Kimi K3 for cost-saving Copilot tasks

Microsoft is evaluating the Chinese open-source Kimi K3 model as a cheaper alternative for certain Copilot tasks, pending internal security and compliance reviews.

Microsoft introduces MAI-Image-2.5-Pro and MAI-Voice-2-Flash models

Microsoft launched two new AI models: MAI-Image-2.5-Pro for higher fidelity image generation, and MAI-Voice-2-Flash, a faster and cheaper voice model for high-volume applications.

Unitree's new 'One-Brain' robot AI handles household chores

Unitree showcased a robot powered by a unified AI 'brain' that can perform a variety of domestic tasks like cleaning and organizing, demonstrating advances in general-purpose home robotics.

Scaling video pretraining with imagination models enables robot skill learning without action labels

Induction Labs introduced Photon‑1, an imagination model that learns from video by predicting future states, allowing robots to acquire new tasks without explicit action supervision and to improve via reinforcement learning.

FLUX-mimic video-action model enables robots to learn factory tasks from minimal video

mimic robotics and Black Forest Labs released FLUX-mimic, a next‑generation model that learns complex industrial manipulation tasks from video demonstrations with far less data, being tested at Audi's factory.

Generalist's GEN-1 foundation model now supports a range of robot end effectors

The GEN-1 embodied AI foundation model was updated to work with many types of robot grippers and tools, not just five‑fingered hands, expanding its utility across diverse industrial settings.

Apertus 1.5 delivers next‑generation open AI infrastructure with multi‑modal capabilities

The Swiss national supercomputing center launched Apertus 1.5, an open, transparent AI infrastructure that includes multi‑modal model training capabilities and open training data, aimed at fostering sovereign AI development.

Research with immediate practical relevance

Curious robots mimic children in learning language through diverse interaction

OIST researchers demonstrated that robots with an intrinsic curiosity reward and exposure to varied language environments can rapidly learn to understand and generalize language, similar to child language acquisition.

Kimi K3's open weights to be released on July 27, weighing 1.4TB

Moonshot AI confirmed that the full 2.8‑trillion‑parameter Kimi K3 model weights will be published on July 27, 2026, making it the largest open‑weight model ever shipped.