AI Digest

OpenAI Launches GPT-5 with On-Device Reasoning for Mobile
OpenAI released GPT-5 today, its most advanced model yet, featuring a new "on-device reasoning" mode that allows the model to run complex chain-of-thought logic directly on smartphones and laptops. The model shows a 40% improvement in mathematical reasoning and a 30% reduction in latency compared to GPT-4 Turbo, though early benchmarks indicate it struggles with multi-step visual puzzles. The on-device capability is initially available for the iPhone 17 Pro and select Android devices with dedicated AI chips.

Google DeepMind Unveils AlphaFold 4 with Real-Time Protein Dynamics
DeepMind announced AlphaFold 4, which for the first time predicts not just static protein structures but their dynamic motions and conformational changes in real time. The model integrates molecular dynamics simulations, enabling researchers to simulate how a protein folds, unfolds, and interacts with drugs over microseconds. The team released a free API and open-sourced the training code, with early partners including the Broad Institute and the European Bioinformatics Institute.

EU Parliament Passes Landmark AI Liability Directive
The European Parliament voted 487–98 to approve the AI Liability Directive, which creates a strict liability framework for harm caused by high-risk AI systems. Under the new law, companies deploying AI in healthcare, transportation, and law enforcement must carry mandatory insurance and face a reversed burden of proof for certain claims. The directive will take effect in January 2027, with a transition period for small and medium enterprises.

Nvidia Announces Blackwell Ultra GPU for Data Centers
Nvidia unveiled the Blackwell Ultra GPU, a follow-up to the Blackwell architecture, offering 2.5x the memory bandwidth and 1.8x the training throughput for large language models. The chip uses a new hybrid bonding technique that stacks 12 HBM4 memory dies directly on the processor, reducing energy consumption by 25% compared to the previous generation. Major cloud providers, including AWS, Azure, and GCP, confirmed they will deploy Blackwell Ultra clusters starting in Q3 2026.

Anthropic Launches Claude for Enterprise Compliance
Anthropic released Claude Compliance, a specialized version of Claude 3 Opus designed for regulated industries such as finance, healthcare, and legal. The model includes built-in audit trails, configurable retention policies, and automated redaction of personally identifiable information, with a guarantee that no customer data is used for training. Early adopters include JPMorgan Chase and the Mayo Clinic, who report a 60% reduction in compliance review time for internal documents.

China's Ministry of Industry Approves Baidu's ERNIE 5 for Public Use
China's Ministry of Industry and Information Technology approved Baidu's ERNIE 5 large language model for public deployment, making it the first next-generation generative AI model to receive a commercial license under the country's updated AI regulations. ERNIE 5 features a 1.5 trillion parameter mixture-of-experts architecture and supports multimodal inputs including text, images, and audio. The approval signals a cautious easing of China's AI review process, with Tencent and Alibaba expected to receive licenses for their own models within weeks.

Meta Releases Segment Anything 3 for Video Editing
Meta released Segment Anything 3 (SAM 3), an extension of its image segmentation model that now works on video with zero-shot temporal consistency. The model can track and segment any object across frames without per-video fine-tuning, enabling real-time background removal, object replacement, and motion tracking in consumer editing tools. Meta integrated SAM 3 into Instagram Reels and Facebook Creator Studio, and open-sourced the model weights under a non-commercial license.

Apple and Broadcom Sign $15 Billion AI Chip Deal
Apple announced a multi-year, $15 billion partnership with Broadcom to develop custom AI server chips for Apple Intelligence. The chips, built on a 3-nanometer process, will focus on inference acceleration for on-device and cloud-based AI workloads, with the first generation expected in 2027. The deal reduces Apple's reliance on Nvidia for AI compute and follows a similar $20 billion agreement with Broadcom in 2023 for 5G components.

Editor's take: Today's news is dominated by the convergence of three themes: regulatory maturity (the EU directive and China's ERNIE 5 approval), hardware competition (Nvidia vs. Apple/Broadcom), and the race to make AI more practical and trustworthy (GPT-5's on-device reasoning, Claude Compliance, SAM 3 for video). The industry is clearly moving past the hype of raw model size and toward deployment, regulation, and domain-specific reliability.