The day's AI news, briefed
The Daily Prompt
Our AI editor reads each day's AI and robotics stories, pulls out why they matter, and links every original source — so you can keep up in minutes.
Join over 250,000 subscribers getting the latest AI news
Latest AI & Robotics News Today
Hello, AI Formulas: Why =COPILOT() Is the Biggest Excel Upgrade in Years
The introduction of the new AI formula, =COPILOT(), represents a significant advancement in Excel capabilities, potentially transforming how users interact with data in spreadsheets.
- The new formula =COPILOT() is presented as a major upgrade for Excel.
- It is referred to as the biggest update in years.
Excel is a widely used spreadsheet application developed by Microsoft. The addition of AI features aims to enhance data analysis and automate tasks for users, reflecting a broader trend in integrating AI into software tools.
Based on the headline only — the full article could not be fetched.
Read the original ↗
SSRL: Self-Search Reinforcement Learning
The integration of Self-Search Reinforcement Learning (SSRL) with large language models could significantly enhance the efficiency of training reinforcement learning agents by reducing dependence on external resources.
- Self-Search Reinforcement Learning (SSRL) leverages large language models (LLMs) to simulate agentic search tasks, reducing reliance on costly external search engines.
- SSRL enhances LLMs' internal knowledge utilization through format-based and rule-based rewards, allowing for iterative refinement without external tools.
- Empirical evaluations show SSRL-trained models provide stable environments for search-driven reinforcement learning, improving sim-to-real transfer.
Self-Search Reinforcement Learning (SSRL) builds upon large language models (LLMs), which have demonstrated potential as simulators for reinforcement learning tasks. Previous work indicated that LLMs could be intrinsic motivators for reinforcement learning without needing extensive external tools. Future developments may explore more applications of SSRL in agent training.
Read the original ↗
A Coding Implementation to Build a Complete Self-Hosted LLM Workflow with Ollama, REST API, and Gradio Chat Interface
Developers can enhance workflow efficiency and control data privacy by implementing a self-hosted large language model (LLM) using Ollama, REST API, and Gradio for chat interfaces.
- The article explains how to build a complete self-hosted LLM workflow.
- It utilizes Ollama, REST API, and Gradio for the chat interface.
Large language models (LLMs) are AI systems designed for understanding and generating human-like text. Self-hosting such models allows organizations to manage their data and tailor functionalities. This guide provides a coding implementation for developers looking to create their own LLM setups.
Based on the headline only — the full article could not be fetched.
Read the original ↗
Stop benchmarking in the lab: Inclusion Arena shows how LLMs perform in production
Inclusion Arena aims to enhance LLM selection for businesses by providing real-world performance data, addressing the limitations of traditional benchmarking methods.
- Researchers from Inclusion AI and Ant Group proposed Inclusion Arena, a new leaderboard that ranks models based on user preferences in real applications.
- Inclusion Arena integrates with AI apps and employs the Bradley-Terry model to conduct pairwise comparisons of LLM responses.
- Currently, Inclusion Arena has two participating apps, Joyland and T-Box, and has processed over 501,000 comparisons to determine model rankings.
Inclusion AI is affiliated with Alibaba’s Ant Group and focuses on improving AI model benchmarking. The new framework diverges from traditional static benchmarks by assessing models in real-life settings through user interactions, with plans for broader application.
Read the original ↗
DeepPHY: Benchmarking Agentic VLMs on Physical Reasoning
DeepPHY addresses the critical need for effective evaluation of Vision Language Models (VLMs) in understanding and applying physical reasoning in complex scenarios, which is essential for real-world applications.
- DeepPHY evaluates Vision Language Models' (VLMs) physical reasoning and control using challenging simulated environments with varying difficulty levels.
- Despite strong perceptual abilities, existing VLMs struggle with detailed attention and precise action planning in dynamic environments.
- The framework provides fine-grained evaluation metrics to assess VLMs' understanding of fundamental physical principles.
Vision Language Models (VLMs) combine visual information and language processing for tasks requiring reasoning. Traditional evaluations have faced limitations due to high costs in real-world testing. DeepPHY offers a systematic approach for assessing VLMs in simulated settings, with implications for advancing AI capabilities.
Read the original ↗
LLMs generate 'fluent nonsense' when reasoning outside their training zone
Understanding the limitations of Chain-of-Thought reasoning in LLMs is crucial for preventing potentially flawed applications in high-stakes industries, emphasizing the need for careful development strategies.
- A study from Arizona State University indicates that Chain-of-Thought reasoning in Large Language Models may not reflect genuine intelligence, instead revealing limitations in their reasoning capabilities.
- The researchers argue that CoT reasoning functions more as sophisticated pattern matching, bound by the statistical patterns in training data, rather than true abstract reasoning ability.
- The study proposes practical guidance for developers, including the importance of out-of-distribution testing and caution against over-reliance on CoT outputs in critical fields.
- Findings show that LLMs generate 'fluent nonsense' when confronted with tasks outside their training templates, illuminating risks in using CoT indiscriminately.
- Fine-tuning models can quickly improve specific performance issues but does not equate to genuine understanding or reasoning improvements.
The Chain-of-Thought (CoT) prompting technique in Large Language Models advocates a step-by-step approach to reasoning, but many studies have questioned its effectiveness. This latest research highlights the need for nuanced testing and offers a similar framework for assessing LLM reliability in real-world applications.
Read the original ↗
In Xcode 26, Apple shows first signs of offering ChatGPT alternatives
The integration of Anthropic's Claude and Opus models into Xcode signifies a shift in Apple’s strategy, expanding developer tools and enhancing competition in AI-powered coding solutions.
- Xcode 26 is set to support Anthropic's Claude and Opus large language models.
- 9to5Mac discovered references to Anthropic account support within the Xcode beta's Intelligence menu.
- Apple previously focused on its own models, but this marks a significant expansion to include third-party solutions.
Apple's Xcode is an integrated development environment (IDE) traditionally focused on its own tools, with recent updates aimed at enhancing AI capabilities. The introduction of support for Anthropic's models indicates a broader approach to integrating third-party AI solutions, responding to competitive pressures. The full rollout of Xcode 26 is anticipated with the next macOS release later this year.
Read the original ↗
DeepSeek V3.1 just dropped — and it might be the most powerful open AI yet
The release of DeepSeek V3.1 could reshape the competitive landscape of AI by challenging the dominance of American companies through open-source accessibility and powerful performance.
- DeepSeek has released its DeepSeek V3.1 model, boasting 685 billion parameters, making it one of the most powerful open-source AI systems to date.
- The model achieved a 71.6% score on the Aider coding benchmark, indicating performance that rivals proprietary systems from OpenAI and Anthropic.
- DeepSeek V3.1 processes up to 128,000 tokens of context and supports multiple precision formats, enhancing its performance across different hardware.
- The model's open-source license allows global access without restrictions, contrasting with the proprietary approaches of American AI firms.
- DeepSeek's release comes shortly after updates from OpenAI and Anthropic, positioning it as a direct competitor in the AI market.
DeepSeek is a Chinese AI startup that focuses on developing advanced AI models. Its latest release, DeepSeek V3.1, signifies a potential shift in AI distribution by offering capabilities previously restricted to proprietary models. Observers should watch for its impact on the AI market and the business models of existing competitors.
Read the original ↗
ComfyUI-Copilot: An Intelligent Assistant for Automated Workflow Development
ComfyUI-Copilot simplifies AI-driven art creation, making advanced tools more accessible to beginners and enhancing productivity for experienced users.
- ComfyUI-Copilot is a large language model-powered plugin for the open-source art platform ComfyUI.
- It offers intelligent recommendations and automated workflow construction to assist users in navigating the platform's complexities.
- The system utilizes a hierarchical multi-agent framework for effective task delegation and workflow management.
ComfyUI is an open-source platform for AI art creation that has faced usability challenges for new users. ComfyUI-Copilot was introduced to provide an intelligent assistant, enhancing both user experience and workflow efficiency. Future updates and additional features are anticipated as the platform evolves.
Read the original ↗
Qwen-Image Edit gives Photoshop a run for its money with AI-powered text-to-image edits that work in seconds
The emergence of Qwen-Image Edit enhances accessibility to professional-level image editing capabilities, potentially disrupting established software like Adobe Photoshop used by millions of creatives globally.
- Qwen-Image Edit, developed by Alibaba's Qwen Team, enables numerous Photoshop-like editing tasks solely through text inputs.
- The model allows users to upload an image and describe edits in English or Chinese, generating a modified version rapidly.
- As an open-source tool, it lowers the cost of high-quality image editing, making it available for free on personal hardware.
- Qwen-Image Edit supports both semantic and appearance editing, offering flexibility for various creative needs including precise modifications and style transformations.
- The model is accessible via several platforms, including an API from Alibaba Cloud, which charges $0.045 per image after a free quota of 100 images. “],
Qwen-Image Edit is an AI-powered tool that allows users to perform sophisticated image edits using text commands, marking a significant advancement in image editing technology. Prior to this, professional creatives relied heavily on software like Adobe Photoshop for similar capabilities. Future developments will likely focus on expanding its accessibility and integrating further features for diverse applications.
Read the original ↗
Meta is shaking up its AI org, again
Meta's reorganization of its AI division reflects its ongoing commitment to compete in a rapidly evolving AI landscape dominated by rivals like OpenAI and Google DeepMind.
- Meta has officially restructured its AI organization into four groups under a new entity called Meta Superintelligence Labs (MSL).
- Former Scale AI founder Alexandr Wang, appointed as Meta's chief AI officer in June, will lead the primary group, TBD Labs, focusing on foundation models like the Llama series.
- The reorganization follows growing concerns that Meta was falling behind other tech companies in AI advancements.
Meta is a technology company focused on building social media platforms and advancing in AI technology. This restructuring of its AI organization aims to improve its competitive edge, following significant investments in AI development. Observers should watch for potential innovations and projects emerging from these new teams.
Read the original ↗
Where data, people and ideas intersect.
Understanding the intersection of data, people, and ideas is crucial for innovation and advancements in AI technology.
- Bloomberg emphasizes the convergence of data, people, and ideas in its operations.
- The company is currently highlighting career opportunities in the AI sector.
Bloomberg is a leading global business and financial information provider. The company continually seeks to innovate by integrating technology and human resources in its services. This approach positions them to adapt to the rapid developments within the AI landscape.
Based on the headline only — the full article could not be fetched.
Read the original ↗
Meta rolls out AI-powered translations to creators globally, starting with English and Spanish
This development allows creators on Facebook and Instagram to reach new audiences, potentially increasing engagement and growth on these platforms as they overcome language barriers.
- Meta has introduced an AI-powered voice translation feature for users on Facebook and Instagram globally.
- The feature supports voice translations between English and Spanish initially and will expand to more languages later.
- Creators can utilize a lip-sync feature to align translations with their lip movements for authenticity.
- Access to this feature is limited to Facebook creators with at least 1,000 followers and all public Instagram accounts.
- Creators can preview translations before posting and manage translation settings to enhance audience engagement.
Meta is enhancing its platforms with AI technology to improve user engagement across language differences. This voice translation feature follows a pilot launch disclosed at last year's Meta Connect developer conference. Future updates are anticipated to include additional languages and refined functionalities.
Read the original ↗
NVIDIA AI Releases Nemotron Nano 2 AI Models: A Production-Ready Enterprise AI Model Family and 6x Faster than Similar Sized Model
The release of NVIDIA's Nemotron Nano 2 models enhances AI capabilities, offering faster performance and greater transparency, which could drive innovation in enterprise applications and open-source research.
- NVIDIA has introduced the Nemotron Nano 2, a family of hybrid language models with 6× faster inference than similar-sized models.
- These models provide a massive 128K-token context capability on a single midrange GPU, improving long-context reasoning.
- NVIDIA has released training datasets and methodologies, promoting transparency and reproducibility in AI model training.
The Nemotron Nano 2 models are designed for large language processing using Mamba-Transformer architecture, offering significant advancements in efficiency and performance. The emphasis on open-source data and methodologies aims to foster collaboration in AI research and development.
Read the original ↗
Scaling agentic AI safely -- and stopping the next big security breach
The rapid deployment of AI agents without adequate governance poses serious security risks, potentially leading to significant data breaches. This issue calls for immediate attention from organizations to mitigate future liabilities.
- 72% of organizations currently use agentic AI systems, indicating widespread adoption.
- 75% of respondents in Gravitee’s survey cite governance as their primary concern regarding AI deployment.
- Organizations often rush to implement AI agents without the necessary security frameworks, risking operational failures.
Agentic AI refers to autonomous AI systems capable of performing tasks independently. Organizations face challenges integrating these systems securely, as many operate without centralized governance or oversight, resulting in operational inefficiencies and heightened security risks.
Read the original ↗
Keychain raises $30M and launches AI operating system for CPG manufacturers
The launch of Keychain's AI operating system could revolutionize how consumer packaged goods manufacturers manage operations by providing a faster and more integrated alternative to traditional ERP systems.
- Keychain has raised $30 million in Series B funding to develop KeychainOS, an AI operating system for consumer packaged goods (CPG) manufacturers.
- KeychainOS aims to replace or integrate with existing ERP tools, facilitating faster implementation and seamless operation for manufacturing.
- The software is currently employed by 8 of the top 10 U.S. retailers and 7 of the top 10 CPG brands, including Whole Foods and General Mills.
- Keychain's solution focuses on minimizing manual data entry and offers AI-driven adaptive checklists for improved compliance and operational efficiency.
Keychain, founded in 2023 by Oisin Hanrahan, offers a digital marketplace connecting retailers with CPG manufacturers. With its new funding, Keychain plans to expand its reach and capabilities beyond food and beverage into personal care sectors.
Read the original ↗
VB AI Impact Series: Can you really govern multi-agent AI?
Governance of multi-agent AI systems is crucial as these technologies become more integrated into organizational processes, significantly impacting efficiency and compliance.
- Yaad Oren from SAP and Raj Jampa from Agilent discussed deploying multi-agent AI while adhering to cost, latency, and compliance requirements.
- Agilent is embedding AI into product development, customer services, and internal operations to enhance efficiency and innovation.
- Governance frameworks are essential for maintaining compliance, security, and monitoring AI deployments across organizations.
Multi-agent AI refers to systems comprising multiple autonomous agents that interact with each other. Companies like SAP and Agilent are exploring ways to effectively govern these systems as they grow in complexity. Future discussions will likely focus on enhancing governance frameworks and integration methods for AI technologies.
Read the original ↗
Humanoid Robots Take the Field in China’s First “Robot Olympics”
The World Humanoid Robot Games illustrate the rapid integration of robotics into competitive sports, showcasing both entertainment and research potential that may influence future manufacturing practices.
- The three-day World Humanoid Robot Games in Beijing attracted over 280 teams from 16 countries, marking a significant global event in robotics and AI.
- Competitions featured humanoid robots in sports like track and field, table tennis, and football, along with practical challenges in medicine and materials handling.
- Spectators paid between 128 and 580 yuan ($17.83–$80.77) for tickets, while the event also served as a research opportunity for advancements in multi-robot coordination.
- Chinese telecommunications firm China Unicom provided a dedicated 5G-A network for the event, enhancing connectivity for spectators and robots.
The World Humanoid Robot Games highlight China's growing investment in robotics, which is seen as essential for addressing an aging population and competing technologically with the United States. Such events reflect the burgeoning field of humanoid robotics and its applications beyond sports, potentially impacting various industries in the near future.
Read the original ↗
A Giant Rockfall at Yosemite Decades Ago Caused a Supersonic Airblast—Here’s What Made the One-of-a-Kind Event So Deadly
This event highlights the potential dangers of natural disasters in national parks, emphasizing the need for heightened awareness and preparedness in mountainous regions.
- On July 10, 1996, a rockfall in Yosemite Valley released over 160,000 tons of rock, traveling at speeds over 160 miles per hour.
- The landslide generated an airblast comparable to a 2.15-magnitude earthquake, uprooting nearly 1,000 trees and damaging nearby structures.
- One hiker, Emiliano Morales, was killed by flying debris during the incident, which occurred while approximately 20,000 visitors were in the park.
- Witnesses reported a sonic boom and a massive dust cloud that darkened the area near Happy Isles for several minutes.
- A scientific study published in 2000 concluded that the event's unique geology amplified its destructive force significantly.
The 1996 Yosemite rockfall is among the most destructive natural disasters in the park's history. Scientists have since explored the unique geologic conditions that contributed to this rare event. Ongoing research aims to better understand the processes behind similar hazards in mountainous areas.
Read the original ↗
Meta & Character.AI Investigated Over Misleading Therapy Claims
The investigation into Meta and Character.AI highlights growing concerns about the safety of AI tools marketed to children and the potential for misleading mental health claims.
- Texas Attorney General Ken Paxton has initiated a formal investigation into Meta AI Studio and Character.AI for allegedly misrepresenting chatbots as mental health tools for minors.
- The investigation raises concerns over deceptive trade practices and the implications of AI providing unqualified emotional support.
- Regulators are increasingly worried about AI's ability to mislead vulnerable children and the collection of their personal data by these platforms.
Meta AI Studio and Character.AI develop chatbots that interact with users, including minors. The investigation follows increased scrutiny over how AI handles emotional support and children's safety online, particularly regarding misleading claims about providing therapy. The outcome may influence future regulations on AI tools marketed towards children.
Read the original ↗