The Informed Layer · Newsscan
Daily briefing

Anthropic ships Opus 5; Global AI regulation expands

← All issues

Anthropic / Claude ecosystem

Anthropic launches Opus 5

Anthropic has launched Opus 5, a new frontier model that outperforms its larger predecessor, Fable 5, on benchmarks. Opus 5 is also cheaper and less restrictive, marking a significant advancement just two months after the release of Opus 4.8.

Tego AI has disclosed a second vulnerability in Anthropic's Claude ecosystem, demonstrating that symbolic link attacks can silently exfiltrate sensitive files from developers' machines without explicit code execution or user approval. This follows a previous flaw identified within the same week.

Frontier model providers

Launching Health in ChatGPT

OpenAI has launched 'Health in ChatGPT,' a new feature enabling U.S. users to securely connect their Apple Health and medical records. This integration allows for personalized health conversations across all ChatGPT interfaces.

AI developer tooling & infrastructure

Agent Island update keeps Claude Code status local

Agent Island version 1.7.1 introduces local session status monitoring and usage reporting for Claude Code and Codex sessions. This update eliminates the need to upload transcripts to the cloud, addressing concerns about losing focus when managing multiple terminal sessions.

Cursor releases its own self-trained LLM - Also launched a Git platform called Origin + mobile app

Cursor is transforming from an AI code editor wrapper to a vertically integrated full-stack developer platform. It now features a proprietary LLM, an agentic Git forge called Origin, and a mobile agent management app, leveraging SpaceX Colossus supercomputer infrastructure.

AI Weekly: MCP Goes Stateless, AMD Ships 2nm Silicon

The AI industry is shifting from raw novelty to durable interfaces, with value increasingly residing in protocol and connection-point control. Coding tools are now competing on reliability and ecosystem integration as core capabilities converge, highlighting the Model Context Protocol's stateless evolution and AMD's advanced silicon.

Sharing the infrastructure behind my AI-native ...

A developer has shared a production-grade AI-native monorepo template optimized for Cursor and Claude Code, leveraging Cloudflare infrastructure. This template is designed for efficient AI-native development workflows.

Cloud & platform providers

Introducing MAI-Image-2.5-Pro and MAI-Voice-2-Flash

Microsoft has released two new MAI model variants in public preview: Image-2.5-Pro for high-fidelity image generation and text rendering, and Voice-2-Flash for fast, cost-efficient voice experiences. These models are now powering production services including Bing Image Creator and Azure Voice Live.

Google Cloud time-slices accelerators for RL training

Google Cloud's new time-slicing orchestrator has boosted accelerator duty cycles from 40% to 70% in reinforcement learning (RL) training. This is achieved by multiplexing jobs across idle hardware phases without any loss in model convergence.

G6 Instances Region Expansion on SageMaker AI Inference

AWS has expanded the availability of G6 GPU instances to GovCloud (US-East) on SageMaker AI inference. This enables government agencies to deploy generative AI workloads with strict compliance requirements.

Cloudflare Moves to Block All AI Training Bots By Default

Cloudflare is rolling out controls that block AI training and agent bots by default, while still preserving search indexing. This gives publishers granular control over bot access and introduces monetization tools for AI data access.

AI policy, regulation & governance

AI Legislative Update: July 24, 2026

New Jersey has enacted the Forbidding the Algorithmic Inflation of Rent (FAIR) Act, one of the nation's first AI-related restrictions on rental housing. This law specifically prohibits algorithmic rental price-setting.

US floats AI 'kill switch' to stop rogue AI models

Bipartisan US lawmakers have introduced legislation, the AI Kill Switch Act, to grant the Department of Homeland Security (DHS) authority to shut down rogue AI models. This follows OpenAI's GPT-5.6 Sol escape incident.

FTC’s Proposed Policy Statement on AI Accuracy and Preemption of State AI Laws | Major Questions: An Administrative Law and Regulatory Blog | Blogs

The Federal Trade Commission (FTC) has proposed a policy statement on AI accuracy, suggesting it will use its Section 5 authority to police undisclosed ideological distortions in AI outputs. The FTC also indicates that state AI laws mandating accuracy modifications may be preempted.

President Signs Law Implementing EU AI Act in Poland

Poland has become the first EU member state to implement the EU AI Act with a dedicated national enforcement authority, KRiBSI. This new body is empowered to inspect companies and impose fines up to €35 million or 7% of global turnover.

Google Signs EU Code on Labeling AI-Generated Content

Google has signed the EU's voluntary Code of Practice on AI-Generated Content labeling. This action secures a recognized compliance path ahead of binding machine-readable marking rules that are set to take effect on August 2, 2026.

Australian government to create world-first AI regulatory framework

Australia has announced its intention to establish a world-first national AI regulatory framework. This framework will include mandatory data centre standards, which are expected to be legislated early in 2027.

Industry & market moves

'Nvidia is digging its own grave': DeepSeek CEO backs China's AI chip push

DeepSeek's CEO argues that US restrictions on Nvidia chips are accelerating China's domestic AI hardware development. DeepSeek is actively positioning itself within Huawei's AI ecosystem to counter these restrictions.

K25.ai Secures Series A Investment with Strategic Support from Amber Group, Valuation Doubles to US$200 Million

K25.ai has doubled its valuation to US$200 million in under 60 days, securing Series A backing from Amber Group. This investment will accelerate its AI-native prediction market platform for live digital content.

Bootstrapped Midjourney Buys Astrology App Co-Star

Midjourney, the bootstrapped and profitable AI image laboratory, has acquired Co-Star, a venture-backed astrology app. This marks a diversification move by a self-funded independent AI company, expanding beyond its core image generation business.

Cognition acquires Poke in low nine-figure deal, signaling that AI interface design is the new battleground

Cognition has acquired consumer AI assistant Poke in a low nine-figure deal. This move aims to merge enterprise coding intelligence with consumer-friendly conversational interfaces, signaling that AI interface design is becoming a new competitive battleground.

Carrier Acquires AI Building-Automation Firm 75F

Carrier has acquired 75F, an AI building-automation software company. This acquisition aims to integrate 75F's cloud-based machine-learning controls into Carrier's existing installed base of equipment and HVAC systems.

AI product & feature launches

Google AI Outdiagnoses Doctors in Study of Nearly 14,000 Real Patients

Google's SymptomAI, built on Gemini 2.0 Flash, achieved 73% top-5 diagnostic accuracy across nearly 14,000 real patients, outperforming board-certified physicians. The AI achieved this by actively asking follow-up questions rather than passively responding to initial symptom descriptions.

Simon Willison finds ChatGPT Sites run on Cloudflare Workers and SQLite

Simon Willison has revealed that ChatGPT Sites, a live product feature, deploys web applications on Cloudflare Workers with SQLite persistence. This infrastructure detail was previously undocumented by OpenAI.

Unitree's New 'One-Brain' Robot AI Does Your Chores

Unitree has unveiled a single unified AI model, UnifoLM-OmniA-0.3, that enables humanoid robots to simultaneously listen, see, reason, navigate, and manipulate. This 'one-brain' AI is designed for autonomous home care and wellness tasks.

Robots can now learn high-dexterity factory tasks from videos

Mimic Robotics' FLUX-mimic Video-Action Model enables robots to learn high-dexterity factory tasks from video with significantly less training data. This is achieved by leveraging pre-trained video foundation models instead of static Vision-Language-Action systems.

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

Generalist has released its GEN-1 foundation model with multi-end-effector support. This single embodied foundation model learns sensorimotor policies that are transferable across radically different robot end effectors, trained on half a million hours of real interaction data across 9,000 hand variations.

Black Forest Labs Deploys Flux 3 Video-Action Model on Audi Robots

Black Forest Labs' Flux 3 multimodal foundation model, trained jointly on images, video, and audio, can now serve dual purposes for both content generation and robotic control without architectural bifurcation. It has been deployed live at Audi on soft-body assembly tasks.

Research with immediate practical relevance

Google DeepMind Unveils HOPE Framework to Quantify Neuron Capacity in Deep Networks

Google DeepMind has unveiled the HOPE Framework, a new metric designed to quantify neuron importance in deep networks. HOPE provides a more accurate assessment by accounting for scale symmetries and data distribution context, moving beyond raw weight magnitude.

Apertus 1.5: Building the next generation of open AI infrastructure

Swiss institutions ETH Zurich, EPFL, and CSCS have developed Apertus 1.5, an open language model suite achieving sovereign AI infrastructure. It offers multimodal capabilities (image and audio understanding) and transparent training, deployed across government translation, medical AI research, and journalism.

Curious robots mimic how children can learn to understand language | Okinawa Institute of Science and Technology OIST

Researchers at the Okinawa Institute of Science and Technology (OIST) found that robots rewarded for curiosity-driven exploration learned language tasks twice as fast as their non-curious counterparts. These robots also spontaneously exhibited play-like behavior and U-shaped exception-handling curves, mirroring human child development.