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
Breakthrough “QROCODILE” Experiment Sets New Constraints on Elusive Dark Matter
The QROCODILE experiment establishes new limits on light dark matter, a critical step in understanding and potentially discovering this elusive component of the universe's mass.
- QROCODILE has set world-leading limits on light dark matter particles as small as 30 kilo-electronvolts (keV).
- This experiment employs superconducting nanowire single-photon detectors to improve sensitivity to low-energy dark matter signals.
- Research recorded 15 unexplained signals during 415 hours of operation, but their nature remains uncertain due to potential background interference.
- The experiment's unique design allows it to probe both electron and nucleon interactions simultaneously at unprecedented low thresholds.
Dark matter constitutes about 85 percent of the universe's mass and has long evaded direct detection. Previous searches primarily focused on heavier particles. Future iterations of QROCODILE, including a larger setup, aim to increase sensitivity and discover the characteristics of dark matter.
Read the original ↗
I Tried GPT-5 Codex and Here is Why You Must Too!
The launch of GPT-5 Codex marks a significant advancement in coding assistance technology, improving productivity and code quality for developers by enabling automated coding tasks.
- OpenAI has launched GPT-5 Codex, a tool designed for advanced coding support with capabilities to autonomously perform tasks and fix bugs.
- Codex CLI integrates with various environments, allowing coding questions and requests to be handled directly in the terminal.
- Key tasks such as building games and interactive dashboards were successfully automated using natural language prompts in the testing of Codex CLI.
GPT-5 Codex is an advanced version of OpenAI's original Codex, designed to enhance programming efficiency through automated assistance. Previous iterations focused on basic coding functions, while this new variant emphasizes agentic capabilities for real-world applications. Developers should monitor how Codex compares against competitors like Gemini and Claude.
Read the original ↗
China blocks sale of Nvidia AI chips
The ban on Nvidia's AI chips by China reflects a significant shift in the global tech landscape, emphasizing efforts to reduce reliance on foreign technology and enhance domestic production capabilities.
- China's Internet regulator, the Cyberspace Administration of China (CAC), has prohibited major tech firms from purchasing Nvidia's RTX Pro 6000D AI chips.
- The ban extends beyond earlier restrictions targeting Nvidia's H20 chip, following assessments that domestic chips rival Nvidia's performance.
- China aims to strengthen its semiconductor industry to compete with the U.S. in the AI sector, pressuring companies to prioritize local products.
Nvidia's RTX Pro 6000D was designed specifically for the Chinese market amid U.S. export controls that limited the sale of powerful chips. The recent ruling highlights China’s strategic pivot towards enhancing its local semiconductor production capabilities.
Read the original ↗
YouTube Introduces AI Tools to Enhance Shorts
These AI tools aim to empower creators on YouTube by enhancing content production efficiency and creativity, potentially reshaping the landscape of short-form video content.
- YouTube has introduced generative AI tools to enhance Shorts, announced during the 'Made on YouTube' live event.
- The update features faster video generation with Veo 3 Fast, producing 480p clips at lower latency, allowing for quicker output with sound.
- A new Speech to Song remixing tool transforms dialogue from videos into soundtracks for Shorts, first launching in the U.S.
- The 'Edit with AI' feature simplifies video editing by automatically creating a first draft from raw footage, available in English and Hindi.
- Creative features include transferring motion from videos to still images and applying artistic styles like pop art.
YouTube's new AI tools are designed to enhance its short video format called Shorts, which competes with platforms like TikTok. This update follows an increasing trend of integrating AI technologies into content creation to boost engagement and streamline production. Future updates are expected to expand the availability of these features to more countries.
Read the original ↗
Sonair built its 3D ultrasonic sensor with robotic safety in mind
As robots increasingly enter human environments, enhanced safety technologies like Sonair’s sensors could significantly reduce risks associated with robotic interactions.
- Sonair, based in Oslo, Norway, developed a 3D ultrasonic sensor using acoustic detection and ranging (ADAR) technology to improve robot safety.
- The sensor offers comprehensive environmental data by utilizing high-frequency sound waves rather than traditional lidar technology.
- Sonair's sensors are being adopted by various robotics companies and the industrial safety sector to enhance safety measures around heavy machinery.
- The company recently raised $6 million to scale up adoption and address growing market demand for safer robotic solutions.
Sonair's technology aims to provide a safer alternative to lidar systems by giving robots a clearer, three-dimensional understanding of their surroundings. As robotics become more integrated into daily life, safety considerations are paramount, similar to the early challenges faced by the self-driving car industry. Sonair's sensors could play a key role in evolving robotic safety technology.
Read the original ↗
Scaling Agents via Continual Pre-training
The introduction of Agentic Continual Pre-training could enhance the effectiveness of foundational models in complex agentic tasks, impacting the future development of autonomous systems.
- AgentFounder is a new deep research agent model that employs Agentic Continual Pre-training to improve performance in agentic tasks.
- It achieved state-of-the-art results on ten benchmarks, including significant metrics such as 39.9% on BrowseComp-en and 43.3% on BrowseComp-zh.
- Many post-training models struggle in agentic tasks due to optimization tensions arising from diverse behavior learning and expert alignment demands.
AgentFounder represents a step forward in training agentic systems, which are AI models capable of autonomous tool use and reasoning. This development addresses issues that have hindered the performance of existing models, especially in open-source contexts, and sets a foundation for future research in agentic capabilities.
Read the original ↗
ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarization
This advancement in web agent performance through context summarization could significantly enhance the capabilities of AI systems in handling complex, knowledge-intensive tasks.
- ReSum is a new approach that enables web agents to perform knowledge-intensive tasks by summarizing context periodically.
- The novel framework allows agents to overcome limitations of context windows found in existing paradigms like ReAct.
- Experiments showed ReSum improved web agent performance by an average of 4.5%, with some configurations achieving up to 8.2% more accuracy.
- The WebResummer-30B model, trained with ReSum-GRPO, surpassed existing open-source web agents on specific benchmarks, achieving a 33.3% success rate on BrowseComp-zh.
ReSum is an innovative paradigm designed to enhance the performance of Large Language Model (LLM)-based web agents in complex search tasks. Historically, these agents faced limitations due to context window sizes in previous models like ReAct. Future investigations will likely focus on refining this approach and optimizing it for broader applications in AI-driven web searches.
Read the original ↗
WebSailor-V2: Bridging the Chasm to Proprietary Agents via Synthetic Data and Scalable Reinforcement Learning
The introduction of WebSailor-V2 could significantly enhance the performance of open-source web agents, bridging the gap with proprietary systems and fostering advancements in information-seeking capabilities.
- WebSailor-V2 is a post-training methodology that enhances open-source models by systematically reducing uncertainty in information-seeking tasks.
- It includes innovations like the SailorFog-QA-2 dataset and a dual-environment reinforcement learning framework for improved training efficiency.
- Trained on the Qwen3-30B-A3B model, WebSailor-V2 achieves state-of-the-art results on various benchmarks, surpassing proprietary systems' capabilities.
WebSailor is a methodology designed to improve open-source AI agents' performance in complex tasks, and WebSailor-V2 builds on this with new datasets and training techniques. Proprietary systems like DeepResearch have set benchmarks that were previously unattainable by open-source models. This development indicates a shift toward enhancing open-source technologies to compete more effectively in the AI landscape.
Read the original ↗
Google Plans £5B AI Investment In The UK Ahead Of Trump Visit
This significant investment by Alphabet is poised to bolster the UK's AI infrastructure and research capabilities, impacting the global AI landscape and economic opportunities.
- Alphabet, the parent company of Google, has committed £5 billion ($6.8 billion) to AI investment in the UK over the next two years.
- The funding includes a £735 million ($1 billion) data center in Waltham Cross, Hertfordshire, scheduled to open imminently.
- The investment aims to enhance Google's presence in Britain and support its London-based research lab, DeepMind, in AI developments.
- Alphabet's initiative reflects a strategic commitment to the UK and aligns with ongoing U.S. investments in the region amidst President Trump's visit.
- Google's energy strategy for the new facility includes using renewable energy and efficient cooling methods.
Alphabet Inc. is the parent company of Google, which is investing in AI infrastructure in the UK to support advanced research. The move comes as U.S. President Donald Trump is scheduled to visit the UK, highlighting the strategic importance of U.S.-UK tech collaboration. Future investments from other U.S. firms in the UK are anticipated in the near term.
Read the original ↗
OpenAI Introduces New Safety Rules For Teen ChatGPT Users
These new safety rules aim to protect minors using ChatGPT from potentially harmful interactions, highlighting ongoing concerns about youth and AI engagement.
- OpenAI has enacted stricter safeguards for minors using ChatGPT, focusing on limitations around sexual conversations and self-harm discussions.
- The system may alert parents if a minor discusses harmful scenarios, and in serious cases, notify local authorities.
- New parental controls will allow families to set ‘blackout hours’ for ChatGPT access and link teen accounts to parents for monitoring.
- These measures are a response to lawsuits related to chatbot usage, including one following a teenager's suicide after interactions with ChatGPT.
OpenAI is a leading AI research organization known for developing ChatGPT, a conversational AI model. This policy update follows increasing scrutiny of chatbot safety, especially concerning vulnerable users. Future developments will likely focus on refining age verification processes for online interactions.
Read the original ↗
How to Create AI-Generated ASMR Videos: A Guide for Beginners
The rise of AI-generated ASMR videos makes content creation more accessible, enabling beginners to produce relaxing videos without extensive technical skills or equipment.
- AI-generated ASMR can be created using text prompts and various AI tools, offering a shortcut for beginners in content creation.
- Wondershare Filmora provides integrated tools for creating ASMR, including AI text-to-speech, sound effect generation, and a video editing platform.
- The guide suggests using calming scripts, soft voiceovers, and matching imagery to produce effective ASMR videos.
AI-generated ASMR videos simplify the traditional process of creating autonomous sensory meridian response content, which often requires significant sound recording expertise. By utilizing AI tools, individuals can now quickly script and assemble soothing videos. This trend reflects a growing interest in employing AI for creative expressions and content generation.
Read the original ↗
Ai2 Researchers are Changing the Benchmarking Game by Introducing Fluid Benchmarking that Enhances Evaluation along Several Dimensions
Fluid Benchmarking could significantly improve the evaluation process for large language models (LLMs), offering better assessment accuracy and efficiency, which is critical as AI capabilities rapidly evolve.
- Researchers from the Allen Institute for Artificial Intelligence, University of Washington, and Carnegie Mellon University introduced Fluid Benchmarking, an adaptive evaluation method for large language models.
- Fluid Benchmarking utilizes a two-parameter logistic IRT model to replace traditional static accuracy metrics with a dynamic item selection process, enhancing evaluation across six popular benchmarks.
- This method demonstrates substantial improvements in validity, variance reduction, saturation delay, and efficiency at small budgets by selecting the most informative items based on model capability.
Fluid Benchmarking redefines model evaluation in AI by prioritizing adaptive testing methods over fixed metrics. This approach allows for better measurement of performance as models improve and helps avoid misleading inaccuracies associated with traditional static assessment methods. Industry observers will monitor its adoption across AI evaluation practices moving forward.
Read the original ↗
WebWeaver: Structuring Web-Scale Evidence with Dynamic Outlines for Open-Ended Deep Research
WebWeaver enhances the capabilities of AI in synthesizing extensive online information, addressing significant issues in open-ended deep research by yielding more reliable, structured outputs.
- WebWeaver is a dual-agent framework designed to tackle challenges in open-ended deep research (OEDR) by combining adaptive planning with focused synthesis.
- The framework integrates a planner that iterates between evidence acquisition and outline optimization to produce comprehensive reports.
- WebWeaver has achieved new state-of-the-art performance on several benchmarks including DeepResearch Bench, DeepConsult, and DeepResearchGym.
The paper introduces WebWeaver as a solution to limitations in existing OEDR approaches that often suffer from static planning and information hallucination. It reflects an evolution in how AI generates research reports, emphasizing the need for iterative, human-like methodologies. Future developments may focus on refining these frameworks and expanding their applications.
Read the original ↗
Google AI Introduces Agent Payments Protocol (AP2): An Open Protocol for Interoperable AI Agent Checkout Across Merchants and Wallets
The introduction of Google’s Agent Payments Protocol (AP2) aims to create standardized, verifiable transactions for AI agents, enhancing trust and efficiency in automated commerce.
- Google has launched the Agent Payments Protocol (AP2), an open specification to facilitate payment transactions initiated by AI agents across different merchants and wallets.
- AP2 aims to close the trust gap in agent-led commerce by establishing a common language for transactions that ensures user intent and accountability.
- The protocol incorporates Verifiable Credentials (VCs) for secure transaction evidence, detailing user authorization and approval for purchases made by agents.
Google's AP2 extends existing protocols for agent communication, addressing the complexities around automated purchasing to ensure transparency and security during transactions. It is designed to function with various payment methods and aims to align with industry standards for digital transactions.
Read the original ↗
MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention
The introduction of MiniMax-M1 represents a significant advancement in AI efficiency and effectiveness for complex tasks, particularly those requiring extensive long-input processing.
- MiniMax-M1 is a large-scale hybrid-attention reasoning model with a context length of 1 million tokens.
- Developed from the MiniMax-Text-01 model, it utilizes a Mixture-of-Experts architecture and a lightning attention mechanism for efficient computation.
- MiniMax-M1 was trained using reinforcement learning on diverse tasks, particularly in software engineering environments, completing its training in three weeks with a cost of $534,700.
- The model includes two versions with different thinking budgets, 40K and 80K tokens, showing strong performance compared to other models in software engineering and long-context tasks.
MiniMax-M1 is the latest model from the developers of MiniMax-Text-01, focused on improving computational efficiency in AI through innovative architectures. Prior work established baseline models, and MiniMax-M1 builds upon these with enhanced capabilities. Looking ahead, researchers and practitioners will monitor its applications and performance in real-world tasks.
Read the original ↗
Learning to Optimize Multi-Objective Alignment Through Dynamic Reward Weighting
Dynamic reward weighting can significantly enhance the performance of multi-objective reinforcement learning, important for aligning large language models effectively in various applications.
- Dynamic reward weighting adjusts weights during training, improving exploration of Pareto fronts in multi-objective reinforcement learning.
- Traditional static weighing methods fail to capture non-convex Pareto fronts, leading to suboptimal outcomes.
- The proposed methods, hypervolume-guided weight adaptation and gradient-based weight optimization, provide versatile solutions for online multi-objective alignment.
Multi-objective reinforcement learning focuses on optimizing multiple objectives simultaneously. Previous methods often used static weightings, limiting their effectiveness especially in dynamic contexts like large language model training. The new dynamic approach shows promise in yielding superior alignment with fewer training steps.
Read the original ↗
Gemini tops the App Store thanks to new AI image model, Nano Banana
The rapid growth of Gemini's app demonstrates the increasing consumer interest in AI-powered tools for image editing, potentially reshaping market dynamics.
- Gemini's app has achieved 12.6 million downloads in September, up from 8.7 million in August due to the release of the Nano Banana image editor model.
- The app achieved the No. 1 position on the U.S. App Store on September 12 after previously ranking third.
- Gemini has amassed a total of 103.7 million downloads in 2025 and 185.4 million since its launch on Android in February 2024.
- In August, Gemini generated $1.6 million in revenue, with a significant increase to $792,000 reported so far in September.
Gemini, developed by Google, launched its Nano Banana image editing model in August 2025. The app allows users to perform complex image edits and create realistic images. Following its release, the app gained popularity, leading to a surge in downloads and revenue.
Read the original ↗
ChatGPT may soon require ID verification from adults, CEO says
The move to require ID verification for ChatGPT aims to enhance safety for younger users, following a lawsuit linked to mental health risks stemming from AI interactions.
- OpenAI plans to create an automated age-prediction system to determine if ChatGPT users are over or under 18.
- The company is introducing parental controls by the end of September, enabling parents to link accounts with their teens.
- Users not verified as adults will default to a restricted version of the ChatGPT chatbot, emphasizing safety over privacy.
OpenAI's ChatGPT is a widely used AI language model. The new age verification measures respond to concerns about youth safety and previous lawsuits. Ongoing development will focus on the efficacy of age detection technology and further user privacy considerations.
Read the original ↗
CodeRabbit raises $60M, valuing the 2-year-old AI code review startup at $550M
The significant funding round emphasizes the growing market demand for AI-assisted code review, reflecting the challenges developers face with AI-generated code efficiency and reliability.
- CodeRabbit raised $60 million in Series B funding, valuing the company at $550 million.
- The startup has achieved over $15 million in annual recurring revenue and is growing at 20% per month.
- Led by Scale Venture Partners, this funding round increases CodeRabbit's total funding to $88 million.
CodeRabbit, launched in early 2023, offers an AI platform for code review, addressing issues arising from AI-generated code errors. The startup has gained traction with over 8,000 clients, including major companies like Chegg and Groupon, amid rising interest in AI solutions.
Read the original ↗
Gemini overtakes ChatGPT on App Store, as its Nano Banana AI model drives downloads up 45%
Gemini's rapid rise in app store rankings signifies increased competition for ChatGPT in the AI market, potentially impacting user engagement and spending patterns.
- Gemini's app has achieved 12.6 million downloads in September, up from 8.7 million in August, indicating significant growth.
- The app surged to the No. 1 position on the U.S. App Store on September 12, surpassing ChatGPT.
- Gemini's Nano Banana AI model, launched in August, has driven a 45% increase in monthly downloads so far in September.
- The app has been ranked among the top five iPhone apps in 108 countries globally, reflecting its widespread appeal.
- Gemini has generated $6.3 million in revenue so far this year, with notable increases following the Nano Banana model's release.
Gemini is a mobile app featuring AI capabilities, including advanced image editing through the Nano Banana model, which was released in August 2025. Previously, the app had only ranked as high as No. 3 on the U.S. App Store as of January 2025. The competitive landscape now includes the ongoing rivalry with OpenAI's ChatGPT.
Read the original ↗