OpenAI readies superapp; AI policy retreats & advances

Anthropic / Claude ecosystem

Anthropic opens Claude safety testing to external researchers

Anthropic has launched the Claude Researcher Access Program, opening early-access safety testing to external researchers. The program provides API credits and pre-release model access, with a primary focus on adversarial red-teaming rather than general capability evaluation.

Frontier model providers

OpenAI plans biggest ChatGPT overhaul yet to build a superapp ahead of potential IPO: FT

OpenAI is planning a major overhaul of ChatGPT to transform it into an integrated 'super app.' This redesign will bundle AI agents, coding tools, and third-party integrations, aiming to compete in the expanding generative AI superapp market and drive paid conversions ahead of a potential IPO.

Google Shrank Gemma 4 by 72% and Unsloth Fixed the 4-Bit Bug Nobody Else Caught on One 4090, and 4-Bit Shouldn't Be This Good | by Chew Loong Nian - AI ENGINEER | Jun, 2026 | Medium

Google DeepMind has released Quantization-Aware Training (QAT) checkpoints for Gemma 4, achieving a 72% size reduction with minimal quality loss. Additionally, Unsloth's dynamic quantization method recovered 15 points of accuracy over naive conversion by fixing a scaling-factor mismatch.

AI developer tooling & infrastructure

No significant new developments.

Cloud & platform providers

Google Cloud release notes  |  Google Cloud Documentation

Google Cloud has released multiple new features, including the general availability of Cloud Location Finder and Cloud Data Fusion 6.11.1.3. Additionally, public preview integrations for Gemini Enterprise and trace data dashboards have been launched.

AI policy, regulation & governance

Greece Prepares To Shape A National Artificial Intelligence Policy

Greece is actively moving to transpose the EU AI Act into domestic legislation, aiming to be among the first four to five EU member states to do so. Simultaneously, Greece is expanding its AI partnerships with major players like OpenAI, Mistral, and ElevenLabs.

Federal government accused of AI policy retreat as US tech giants plan Australian investments - Australia

Australia's government has reportedly abandoned standalone AI safety legislation, opting instead for a hands-off regulatory approach. This decision has raised concerns about consumer protection, especially as US tech giants continue to invest heavily in Australian data centers.

Industry & market moves

Google DeepMind, Wellcome Sanger Institute Launch AI Genomics Consortium | GenomeWeb

Google DeepMind and the Wellcome Sanger Institute have established a joint consortium. The goal is to build large genomic datasets to support AI-driven biological research, accelerating discoveries in genomics and health.

South Korean tech giant Naver has become the first Korean company to join Nvidia's Nemotron Alliance and agreed to build gigawatt-scale AI infrastructure with Nvidia. This strategic partnership aims to expand sovereign AI capabilities across Asia-Pacific, Europe, and the Middle East, with an initial 55MW AI factory deployment in 2027.

S&P Global Expands AI Ecosystem Reach with Cohere to Power Trusted, Agentic Workflows for Financial Institutions

S&P Global's financial data now integrates directly into Cohere's secure enterprise AI platform North. This integration enables financial institutions to build trusted agentic workflows with cited and verifiable financial intelligence.

OKI partners U.S. Lazarus AI for defense-focused reliable AI | IBTimes JP

Japanese electronics firm OKI has partnered with U.S. Lazarus AI to deploy source-cited, auditable AI in defense and critical infrastructure. This collaboration aims to address hallucination risks that have previously prevented AI adoption in these high-stakes sectors.

How NVIDIA and LG Group's AI Factory Spans Six Companies

LG Group's physical AI data factory will generate synthetic robotics training data using NVIDIA Cosmos world models. This data will be commercially sold to Korean startups and global robotics companies pursuing autonomous systems, leveraging a partnership across six LG subsidiaries and NVIDIA.

AI product & feature launches

Unisound Releases U2: A Native Agentic Large Model Built for Execution, Capable of Autonomously Decomposing and Completing 100+ Steps in Complex Real-World Workflows

Unisound has released U2, a native agentic large language model designed for multi-step task execution, capable of autonomously decomposing and completing over 100 steps in complex real-world workflows. U2 features hybrid thinking that dynamically switches between implicit and explicit reasoning modes.

Research with immediate practical relevance

Today's AI Talks Like “Nobody.” New Research Gives It Real Personality. | Stanford HAI

New research from Stanford HAI, published in npj Artificial Intelligence, introduces PsychAdapter. This system enables fine-grained control over AI-generated language personality traits, facilitating realistic simulations for clinical training and personalized educational content.

Xiaomi MiMo and TileRT Push a 1-Trillion-Parameter Model Past 1000 Tokens Per Second on Commodity GPUs - MarkTechPost

Xiaomi and TileRT have achieved over 1000 tokens per second inference on a 1-trillion-parameter model using commodity GPUs. This was accomplished through an integrated approach combining FP4 quantization, DFlash speculative decoding, and microsecond-scale runtime optimization.

UNIST Develops Real-Time 3D Spatial AI That Understands Human Language - Seoul Economic Daily

The Ulsan National Institute of Science and Technology (UNIST) has developed LightSplat, an open-vocabulary 3D spatial recognition technology. LightSplat reduces memory consumption to 1/64th of existing methods and achieves 50–400x faster semantic injection while maintaining superior recognition accuracy for real-time 3D spatial understanding.