Robots read instruments; Cloudflare builds AI infra

Anthropic / Claude ecosystem

No significant new developments.

Frontier model providers

Gemini Robotics-ER 1.6 Can Read a Lab Instrument. That Changes Everything About the Robots We Are Building.

Google DeepMind's Gemini Robotics-ER 1.6 introduces a groundbreaking instrument-reading capability, allowing robots to autonomously interpret analog and digital displays and make decisions in complex, unstructured environments. This innovation could revolutionize automation in sectors like pharmaceuticals and healthcare.

OpenAI adds AI pets to its Codex coding tool

OpenAI has introduced an animated companion feature, 'Codex Pets,' to its Codex coding tool. These AI pets provide real-time project status updates directly within the coding environment, eliminating the need for users to switch tabs.

AI developer tooling & infrastructure

What changed in Iris v0.4.0 - DEV Community

Iris v0.4.0 has been released, bridging deterministic rule-based and semantic LLM-as-Judge evaluation in a single MCP-native runtime. This update closes a competitive gap, adding citation verification and OpenTelemetry observability to the platform.

GitHub - cloudflare/vinext: Vite plugin that reimplements the Next.js ...

Cloudflare has open-sourced 'vinext,' a Vite plugin that reimplements the Next.js API surface. This plugin enables developers to deploy Next.js applications to any platform, including Cloudflare Workers, AWS, and Vercel, promoting vendor-neutral compatibility.

Cloud & platform providers

Cloudflare Builds High-Performance Infrastructure for Running LLMs - InfoQ

Cloudflare has introduced its custom Infire inference engine and a disaggregated prefill architecture designed to optimize GPU utilization and reduce memory overhead. This new infrastructure aims to enable faster and more scalable inference for trillion-parameter language models.

Google Cloud launches AI Protection: Security for the AI era

Google Cloud has launched AI Protection, a comprehensive suite designed to discover AI inventory, secure AI assets with the now generally available Model Armor, and manage AI threats across multi-cloud environments. This suite provides end-to-end security for AI systems.

Cloudflare Launches Cloudforce One Threat Events Platform

Cloudflare has launched Cloudforce One threat events platform, designed to deliver real-time intelligence on cyberattacks. The platform provides actionable context and Indicators of Compromise (IoCs) to help security teams respond faster to evolving threats.

Cloudflare integrates Content Credentials preservation into its Images service

Cloudflare is integrating Content Credentials preservation into its Images service. This feature allows creators to maintain image provenance and authenticity metadata at scale using C2PA standards, addressing concerns about deepfakes and manipulated content.

AI policy, regulation & governance

OpenAI claims DeepSeek using distillation to replicate US models

OpenAI has formally informed US lawmakers that DeepSeek is employing distillation techniques and circumventing access controls to replicate US AI models for its own training. This raises significant intellectual property and national security concerns in the AI industry.

Microsoft, Amazon Hand Pentagon More Control Over AI Systems

The Pentagon has secured expanded agreements with multiple major AI companies, including Microsoft, Amazon (AWS), Nvidia, Oracle, Reflection AI, OpenAI, Google, and SpaceX, for the use of advanced AI tools on classified military networks. This move replaces an earlier reliance on Anthropic's Claude after a contractual dispute over autonomous weapons restrictions.

Access to major illegal adult content websites in South Korea blocked overnight with cooperation from Cloudflare

South Korea's Media and Communications Commission, with cooperation from Cloudflare, executed an overnight blocking of major illegal adult content and copyright infringement websites. This action addresses the widespread distribution of non-consensual intimate imagery.

US lawmakers move to mandate first comprehensive review of China’s AI capabilities

For the first time, US lawmakers are mandating a comprehensive State Department assessment of China's AI capabilities and leaders. This legislation aims to establish verification frameworks for advanced AI development oversight.

Industry & market moves

Anthropic in talks to buy AI inference chips from UK startup Fractile

Anthropic is reportedly in negotiations to purchase AI inference chips from Fractile, a UK-based startup. This potential deal aims to support Anthropic's growing inference workloads, indicating a strategic move to secure specialized hardware for its AI models.

Mistral AI acquires Koyeb and accelerates cloud expansion

Mistral AI has made its first acquisition, purchasing Paris-based Koyeb, an AI infrastructure company. This strategic move aims to consolidate European AI infrastructure and vertically integrate Mistral AI's operations from model development to deployment.

Starcloud Secures $170 Million Funding to Pioneer Orbital Data Centers for AI Compute | Aerospace & Defense News

Starcloud has secured $170 million in Series A funding, achieving a $1.1 billion valuation in just 17 months, making it the fastest Y Combinator company to unicorn status. The funding will be used to scale orbital data centers for AI compute.

Musk testimony dominated first week Musk v. Altman trial in Oakland

The first week of the high-stakes federal civil trial between Elon Musk and OpenAI leadership, including Sam Altman and Greg Brockman, was dominated by Musk's testimony. The lawsuit alleges that OpenAI illegally converted from a nonprofit to a for-profit entity, with potential damages claims of $134 billion.

AI chipmaker Cerebras targets up to $4bn IPO at $40bn valuation

AI chipmaker Cerebras Systems is reportedly targeting an IPO of up to $4 billion at a $40 billion valuation. This follows a transformative $10 billion-plus compute agreement with OpenAI and a CFIUS-cleared refinancing after its withdrawal from an earlier IPO attempt in 2024.

AI product & feature launches

Xiaomi's open-weight MiMo-V2.5-Pro takes aim at Claude Opus with hours-long autonomous coding

Xiaomi has released its open-weight MiMo-V2.5-Pro model, demonstrating a significant leap in autonomous coding capabilities. The model successfully completed a compiler project in 4.3 hours and operates with 40–60% fewer tokens than Anthropic's Claude Opus 4.6 for similar coding tasks.

Alibaba’s Metis Agent Cuts Redundant AI Tool Calls by 96% While Setting New Accuracy Benchmarks

Alibaba's Metis Agent, a multimodal reasoning agent built on Qwen3-VL-8B-Instruct, has demonstrated a 96% reduction in redundant AI tool calls through Hierarchical Decoupled Policy Optimization. It also achieved state-of-the-art benchmarks on visual and reasoning tasks with only 8 billion parameters.

Moreh's LLM Inference Breakthrough on Tenstorrent Galaxy: DGX A100 Performance at One-Third the Cost

Moreh has achieved a breakthrough in LLM inference systems, delivering DGX A100-level performance at one-third the cost using Tenstorrent Galaxy Blackhole with Moreh vLLM. This was accomplished through chip-level, cluster-level, and infrastructure optimization multipliers.

Research with immediate practical relevance

Even the latest AI models make three systematic reasoning errors, ARC-AGI-3 analysis shows

An analysis of 160 frontier model runs on the ARC-AGI-3 benchmark, including GPT-5.5 and Opus 4.7, reveals three persistent systematic reasoning failures: inability to build world models from local observations, false analogies to training data, and failure to validate success.

In Harvard study, AI offered more accurate emergency room diagnoses than two human doctors | TechCrunch

A Harvard Medical School study, published in Science, found that OpenAI's o1 model provided more accurate emergency room diagnoses than human internal medicine physicians. In 76 real patient cases, the AI achieved 67% accuracy, surpassing the 55% and 50% accuracy rates of two human doctors.