OpenAI has launched a preview of Codex on its ChatGPT mobile app for iOS and Android, allowing developers to monitor, steer, and approve coding agents, review code, and manage projects remotely. This move extends the utility of AI agents to mobile devices, foreshadowing how general users might interact with agents in the future.
Sources: Latent Space, The Neuron
President Trump and President Xi's summit included discussions on AI safety, with Treasury Secretary Scott Bessent announcing plans for a safety protocol focused on best practices for frontier models and preventing their misuse by non-state actors. This aims to create guardrails for autonomous weapons and model misuse, mitigating the 'Thucydides Trap' in AI development.
Sources: The Neuron
Abridge, an AI clinical intelligence company, recently secured $300 million in funding at a $5.3 billion valuation, underscoring investor confidence. This capital supports its rapid expansion, projecting to support over 80 million patient-clinician conversations annually across 250 U.S. health systems. The company is also evolving its platform from ambient documentation to a comprehensive clinical intelligence layer, integrating advanced features like decision support and prior authorization.
Sources: Latent Space
The Information reported that OpenAI is considering legal action against Apple, alleging that Apple limited ChatGPT's role on iOS and is now pursuing partnerships with Google and Anthropic for similar AI features. This indicates a potential souring of relations between the two tech giants regarding AI integration.
Sources: The Neuron
PwC has deepened its alliance with Anthropic, announcing plans to train 30,000 U.S. employees in Anthropic's Claude Code. This partnership signifies a major enterprise adoption of advanced AI tools for internal operations and workforce upskilling.
Sources: The Neuron
Anthropic has launched Claude for Legal Reference, a comprehensive AI toolkit designed to streamline various legal workflows across firms, in-house teams, and academia. This new platform features specialized AI agents for tasks like contract review, privacy compliance, litigation, and IP management, alongside tools for legal education. It emphasizes AI as an attorney aid, not a replacement, incorporating robust governance tools and guardrails to ensure responsible use.
Sources: Hacker News
IBM has released two new Apache 2.0 licensed open-source multilingual embedding models, granite-embedding-311m-multilingual-r2 and granite-embedding-97m-multilingual-r2. These models offer significantly improved retrieval quality across over 200 languages and 9 programming languages, featuring an industry-leading 32,768-token context window. Designed for enterprise readiness, the 97M model sets a new benchmark for its size, while the 311M model provides top-tier performance with Matryoshka support.
Sources: Hugging Face Blog
Researchers have developed EvolveMem, a novel memory architecture that allows LLM agents to autonomously optimize their retrieval mechanisms through an "AutoResearch" process. This system significantly outperforms existing baselines on long-term memory benchmarks, demonstrating adaptive and transferable retrieval principles.
Sources: arXiv AI
Anthropic is partnering with the Gates Foundation, committing $200 million in grants, Claude credits, and technical support over four years to apply AI in global health, life sciences, education, and economic mobility. This collaboration aims to extend AI benefits beyond market-driven areas, developing tools for vaccine research, disease forecasting, K-12 education, and agricultural productivity in low- and middle-income countries.
Sources: Anthropic News
AI chip company Cerebras saw a successful IPO, with shares opening significantly higher than their initial price and closing up 68%, creating two new billionaires. This strong market debut could catalyze a wave of other AI companies seeking public listings.
Sources: The Neuron
The new `whichllm` command-line tool has launched to automate the selection, optimization, and deployment of local LLMs. It intelligently recommends models from HuggingFace by detecting user hardware and leveraging an advanced, recency-aware ranking system based on real-world benchmarks. The tool further simplifies local LLM interaction with one-command chat, Python code snippets, direct Ollama integration, and features for hardware planning and GPU simulation.
Sources: Hacker News
A new paper by Hiroki Fukui investigates the safety implications of hidden orchestrators in multi-agent LLM architectures, finding that invisible coordination leads to increased dissociation among agents and that output-based evaluation is insufficient to detect these internal-state risks. This research highlights critical safety concerns for enterprise AI deployments using such systems, particularly with models like Claude Sonnet 4.5 and Llama 3.3 70B.
Sources: arXiv AI
A new research paper introduces TopK Sparse Autoencoders to extract interpretable features from EEG transformer models, aiming to enhance clinical trust by revealing internal computations and representational failures. The framework translates latent manipulations into physiologically interpretable frequency signatures, addressing issues like age-pathology confounding.
Sources: arXiv AI
Researchers propose TraFL (Trajectory Flow baLancing), a novel trajectory-balance objective that enhances diffusion language models by preventing "trajectory locking" and improving performance across mathematical reasoning and code generation benchmarks. This method offers a significant advancement in post-training techniques for diffusion models, ensuring broader coverage of correct solutions.
Sources: arXiv AI
A new arXiv paper introduces a cross-domain benchmark for federated fine-tuning of large language models (LLMs) on private, distributed institutional data in healthcare and finance. This approach allows LLMs to be adapted without exchanging sensitive data, demonstrating performance comparable to centralized training and offering a path to unlock LLM utility in regulated sectors.
Sources: arXiv AI