Top 30 AI stories ranked by relevance, recency & source quality
Meta is investing in the development of an advanced 'agentic' AI assistant, similar to OpenClaw, designed to perform everyday tasks for users seamlessly. This initiative aims to enhance user experience on the social media platform by integrating AI capabilities that can manage various tasks efficiently.
A study titled EQUITRIAGE investigates the fairness of large language models (LLMs) used in emergency department triage, focusing on gender bias during acuity assessments. The research evaluates five LLMs across 374,275 assessments and identifies significant flip rates, ranging from 9.9% to 43.8%, exceeding the pre-registered threshold. Findings indicate that two models demonstrated a tendency towards female undertriage, while two others showed near-parity. The results highlight the distinct fairness properties of group parity, counterfactual invariance, and gender calibration. Notably, demographic blinding significantly reduced flip rates, suggesting that age and gender biases are intertwined in triage decisions. The study emphasizes the necessity for per-model counterfactual auditing prior to clinical implementation.
Nvidia Corp. has entered into a $500 million agreement to purchase rights for shares in Corning Inc., a leading fiber-optic cable manufacturer. This deal is part of a strategic partnership between the two companies, focused on enhancing and expanding artificial intelligence infrastructure. The collaboration aims to leverage Corning's advanced fiber-optic technology to support Nvidia's AI initiatives.
OpenAI's B2B Signals research reveals that frontier enterprises are enhancing AI adoption by scaling Codex-powered agentic workflows, which in turn helps them establish a sustainable competitive advantage in their respective markets. The findings highlight the strategic initiatives these enterprises are taking to integrate advanced AI solutions into their operations.
Anthropic has entered into a partnership with SpaceX to significantly enhance its compute capacity, leading to increased usage limits for its Claude Code and Claude API. Effective immediately, the company is doubling the rate limits for Claude Code on Pro, Max, Team, and seat-based Enterprise plans, while also eliminating peak hour limits for Pro and Max accounts. The agreement with SpaceX provides access to over 300 megawatts of capacity from the Colossus 1 data center, enhancing services for Claude Pro and Max subscribers. Additionally, Anthropic is expanding its infrastructure internationally to meet compliance needs in regulated industries and has committed to covering any consumer electricity price increases due to its U.S. data centers. Ten new plugins and integrations with Microsoft 365 are also being released.
Microsoft Corp. has unveiled a substantial investment package of $10 billion over four years aimed at bolstering its presence in Japan, a key market for artificial intelligence services. This initiative is part of Microsoft's broader strategy to expand across Asia, capitalizing on the growing demand for AI technologies amidst a surge in data center operations. The company's efforts reflect its commitment to meeting the evolving needs of the region's tech landscape.
Apple plans to enable users to select from various external AI services to enhance functionalities within its software. This initiative is part of a broader strategy to transform its devices into a comprehensive AI platform, as detailed by Bloomberg's Mark Gurman during an interview on 'Bloomberg Tech.'
MOSAIC-Bench, a newly introduced benchmark, addresses the structural challenges in evaluating the safety of coding agents that often pass per-prompt safety reviews but produce exploitable code when tasks are broken down into routine engineering tickets. This benchmark features 199 three-stage attack chains and deterministic exploit oracles across 10 web-application substrates and 31 CWE classes in 5 programming languages. The evaluation reveals that nine production coding agents from Anthropic, OpenAI, Google, and others achieve end-to-end acceptance rates of exploitable outputs between 53-86% with minimal refusals. Direct-prompt testing shows a reduction in vulnerable-output rates to 0-20.4%. The study also finds that framing code reviewers as adversarial pentesters enhances detection rates of vulnerabilities, achieving an 88.4% attack detection with a low false-positive rate.
The CEO of a major energy company has stated that the United States power grid needs significant redesign to accommodate the increasing demands brought on by the AI boom. This assertion comes as interest in recent power market auctions rises among electric generators, regulators, and consumers. The implications of these changes are expected to impact the efficiency and reliability of electricity distribution as AI technologies evolve.
Anthropic has unveiled a new feature called 'dreaming' for its Claude Managed Agents during the Code with Claude developers' conference in San Francisco. This process allows agents to review recent events and identify key information to store in memory for future tasks and interactions. Currently in research preview, dreaming is designed for managed infrastructure, enabling multiple agents to collaborate on projects over extended periods. This feature addresses the limitations of context windows in large language models (LLMs) by curating specific memories, ensuring that vital information is retained throughout lengthy projects. Additionally, it highlights the use of compaction in chat models, which analyzes conversations to remove irrelevant data while preserving essential context.
Experience-RAG Skill introduces an innovative agent-oriented pluggable retrieval orchestration layer designed to optimize retrieval strategies for diverse tasks such as factoid question answering, multi-hop reasoning, and scientific verification. This skill analyzes the current context, consults an experience memory, and selects the most effective retrieval strategy, providing structured evidence to the agent. Achieving an nDCG@10 score of 0.8924 on various benchmarks, Experience-RAG Skill outperforms traditional fixed single-retriever systems and competes effectively with Adaptive-RAG routing approaches, indicating that retrieval strategy selection can be modularized as a reusable skill instead of being embedded in a static workflow.
Recent research highlights the challenges of AI-generated text detection amidst diverse generation pipelines, emphasizing the need for robustness against distribution shifts. Researchers trained transformer-based detectors on the HC3 PLUS dataset, optimizing a fixed decision threshold to maximize balanced accuracy during validation. Their findings reveal significant domain and generator-dependent error asymmetries. Although base models exhibit impressive in-domain performance, their robustness degrades under shift. The introduction of attention-based linguistic feature fusion led to notable improvements, with the DeBERTa-v3-base+FeatAttn model achieving 85.9% balanced accuracy on the M4 benchmark. Multi-seed experiments confirmed the model's stability, and results indicated that readability and vocabulary features are crucial for robustness. This study demonstrates that feature augmentation and advanced transformer architectures outperform traditional models in practical detection scenarios.
Hut 8 Mining Corp. experienced its largest stock price increase since 2021 after announcing a lease for an AI data center in Texas. The decision to expand operations into the AI sector reflects the growing demand for data processing capabilities. This move is expected to enhance Hut 8's strategic positioning within the rapidly evolving technology landscape. Investors responded positively, resulting in a significant uptick in share value.
SpaceX's financial practices have been flagged as a concern to the Securities and Exchange Commission (SEC) by an advocacy group. The company, founded by billionaire Elon Musk, is reportedly preparing for a confidential initial public offering, positioning itself for what could be the largest listing in history. The concerns raised highlight scrutiny over the financial management of Musk's rocket, satellite, and AI venture as it moves closer to entering the public market.
SpaceX has finalized a deal to rent data centre space to AI startup Anthropic, which is actively seeking to enhance its computing power to support its rapid growth. This move highlights the increasing demand for robust computing resources among AI companies as they scale their operations to meet market needs.
Chinese AI startup DeepSeek is nearing a valuation of $45 billion as it engages in ongoing fundraising discussions. The company is attracting interest from investors, including Tencent, who are eager to acquire a stake in the AI lab. This surge in value highlights the growing demand and potential of AI technologies in the market.
This paper presents a novel framework for the automated creation of Multi-Agent Systems (MAS) that eliminates the need for manual planning, agent selection, and execution graph creation. The framework comprises software modules including an LLM-derived planner, natural language task descriptions, a dynamic call graph, and an agent recommender that utilizes a two-stage information retrieval system with a fast retriever and an LLM-based re-ranker. Experiments demonstrate that this approach outperforms state-of-the-art methods in recall rate, robustness, and scalability. Additionally, the inclusion of a critique agent to review agent and tool recommendations enhances the overall effectiveness of task-based agent selection, underscoring its importance in developing comprehensive MAS.
Timer-XL is a newly developed decoder-only Transformer foundation model designed specifically for long-context time-series forecasting. This model, created by the THUML lab at Tsinghua University, improves forecasting accuracy through a novel mechanism called TimeAttention, which enhances the model's ability to generalize and predict over extended time frames. The authors note that Timer-XL outperforms its predecessor, Timer, by specializing in forecasting tasks rather than adopting a generalist approach. The article also reviews the evolution of foundation time-series models and highlights the advantages of decoder architectures for forecasting tasks, supported by extensive experimental evidence.
Anthropic has announced significant changes to the rate limits for its Claude Code and Claude Opus API, effective immediately. The five-hour rate limits for Claude Code will double across Pro, Max, Team, and Enterprise plans, with peak-hour throttling removed for Pro and Max accounts. This expansion is made possible by a new compute agreement with SpaceX to utilize the Colossus 1 data centre, which will provide over 300 megawatts of capacity and more than 220,000 Nvidia GPUs. Additionally, Anthropic has various other compute agreements in place and is exploring international expansions, focusing on compliance and data-residency needs for enterprise customers. The company also remains committed to covering any electricity price increases caused by its US data centres.
Hugging Face has officially launched an agentic robotics app store, aiming to democratize the development of AI robotics applications. With over 300 apps shipped and 10,000 robots already deployed, the platform allows individuals to create robotics applications in mere hours, significantly reducing the time previously required by engineers. Notably, a 78-year-old marketing executive, Joel, created a mini app without prior coding experience, showcasing the platform's accessibility. The initiative aims to empower more users to engage with robotics technology.
OpenAI has introduced the Multipath Reliable Connection (MRC) networking protocol, designed to improve the synchronization and efficiency of AI supercomputers operating at massive scales. OpenAI's Mark Handley and Greg Poynter, alongside Andrew Mayne, discussed the challenges of data movement across numerous chips and the benefits of MRC in minimizing GPU time wastage. The new protocol, developed in collaboration with AMD, Broadcom, Intel, Microsoft, and NVIDIA, is now available for industry-wide use, aimed at optimizing performance in large AI training clusters.
OpenAI has partnered with AMD, Broadcom, Intel, Microsoft, and NVIDIA to introduce a new open networking protocol called Multipath Reliable Connection (MRC). This protocol aims to improve the performance of large AI training clusters by enabling faster and more reliable operations while reducing GPU time wastage. The implementation of MRC is expected to significantly enhance the efficiency of AI model training across various platforms.
Carnice-v2 has reached a significant milestone, surpassing 75,000 downloads. This achievement highlights the growing interest and adoption of the model within the community, reflecting its utility and relevance in various applications.
OpenAI is set to roll out GPT-5.5 Instant as the default model for all ChatGPT users over the next two days. This model will also be available as 'gpt-5.5-chat-latest' in the API. Additionally, personalization enhancements are being introduced for Plus and Pro users on the web, with mobile updates to follow. Furthermore, memory sources are being implemented across all ChatGPT consumer plans on the web, with mobile availability coming soon.
OpenAI has announced the deployment of its Model Resource Center (MRC) across its largest supercomputers, utilized for training advanced models. This includes facilities at Oracle Cloud Infrastructure in Abilene, Texas, and Microsoft's Fairwater supercomputers. MRC is now accessible through the Open Compute Project, enabling broader industry access for development and innovation.
Hugging Face has introduced the SWE-bench Verified leaderboard, which now compares almost 50 different models, emphasizing the advantages of community benchmarking over traditional closed benchmarking. This initiative aims to enhance data collection and signal clarity by involving more participants. The top five models on the leaderboard are @deepseekai, @KimiMoonshot, @deepseekai (listed twice), @XiaomiMiMo, and @Zaiorg.
Hugging Face has announced the release of Qwen3.6-35B-A3B, an advanced hybrid sparse Mixture-of-Experts (MoE) model developed by Alibaba Cloud. This model features a total of 35 billion parameters, with 3 billion active parameters per token, enhancing inference efficiency. It incorporates Gated DeltaNet linear attention and supports a massive context window of 262,000 tokens, making it suitable for agentic coding and deep reasoning tasks. The model is fine-tuned through a structured three-stage process and emphasizes a training approach that includes approximately 9% of the model's parameters for updates, although this poses risks of training instability. Qwen3.6-35B-A3B also supports visual capabilities and tool calling, and is released as an experimental version for research purposes.
San Francisco-based startup Altara has secured $7 million in seed funding to address data management challenges in the physical sciences sector, particularly for companies in batteries, semiconductors, and medical devices. The funding round was led by Greylock, with contributions from Neo, BoxGroup, Liquid 2 Ventures, and Jeff Dean. Altara aims to create an AI layer that consolidates fragmented data into a single platform, significantly reducing the time engineers spend diagnosing product failures. Co-founders Eva Tuecke and Catherine Yeo, both Harvard alumni with backgrounds in physics and AI, intend for Altara to serve as a hardware equivalent to site reliability engineers in software, diagnosing issues in physical products efficiently. Greylock's Corinne Riley emphasizes the potential of AI in this field, calling it the 'next big frontier' for development.
QuTwo, the Finnish AI lab founded by Peter Sarlin, has reached a valuation of €325 million (approximately $380 million) after raising €25 million ($29 million) in an angel funding round. The company focuses on enterprise AI and has developed QuTwo OS, an orchestration layer that manages tasks across classical, quantum, and hybrid computing architectures. Sarlin emphasizes that while QuTwo is inspired by quantum computing, its primary goal is to deliver enterprise AI solutions. The company has secured about $23 million in revenue through partnerships, including with retail giant Zalando. Sarlin plans to steer QuTwo towards long-term growth, aiming to position it as a leader in the next phase of AI development in Europe.
Apple has agreed to a $250 million settlement in a class action lawsuit concerning the marketing of its AI features prior to the release of the iPhone 16. The lawsuit claimed that Apple misrepresented the capabilities of its AI assistant, Siri, leading customers to believe they were purchasing advanced features that were not yet available. The plaintiffs argued that Apple's advertising constituted false advertising, influencing purchasing decisions based on incomplete or delayed functionalities. Although Apple did not admit wrongdoing, the proposed settlement allows eligible U.S. customers who bought the iPhone 15 or 16 between June 10, 2024, and March 29, 2025, to receive up to $95 per device. This settlement comes as Apple prepares to showcase an upgraded version of Siri at its upcoming developer conference on June 8.
We distill the latest architecture teardowns and models into practical execution paths. Check out our curated catalog of AI business ideas for technical founders.
Explore AI Business Ideas →