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
OneReward: Unified Mask-Guided Image Generation via Multi-Task Human Preference Learning
The introduction of OneReward represents a significant advancement in generative AI capabilities, streamlining processes for diverse image manipulation tasks without compromising efficiency or generalization.
- OneReward is a unified reinforcement learning framework that enhances generative capabilities across multiple tasks using a single vision-language model.
- The framework allows for mask-guided image generation, applicable to tasks such as image fill, object removal, and text rendering.
- Unlike existing methods that require task-specific supervised fine-tuning, OneReward eliminates this need, improving overall efficiency.
OneReward improves generative AI by utilizing a single model for various image manipulation tasks, overcoming limitations of traditional methods that depend on specific model training. The framework addresses the inefficiencies seen in previous approaches, paving the way for more adaptable AI solutions in image generation.
Read the original ↗
A Survey of Reinforcement Learning for Large Reasoning Models
The survey examines the challenges faced in enhancing reasoning capabilities in Large Language Models through Reinforcement Learning, impacting future developments in AI and computational intelligence.
- The paper surveys advancements in Reinforcement Learning (RL) for enhancing Large Language Models (LLMs) in complex reasoning tasks.
- Challenges identified include scalability, computational resources, algorithm design, and training data for RL applications in LLMs.
- The review highlights the need for scalable approaches to improve RL's contribution toward Artificial Superintelligence (ASI).
Reinforcement Learning is a methodology used to improve the performance of AI models by allowing them to learn from interactions with their environment. This paper revisits the state of RL applied to LLMs and discusses opportunities for future research and development. Researchers aim to overcome foundational challenges to better harness RL for reasoning tasks in LLMs.
Read the original ↗
3D and 4D World Modeling: A Survey
This survey offers a foundational reference for advancing research in 3D and 4D world modeling, critical for AI's understanding of dynamic environments.
- The survey reviews 3D and 4D world modeling, defining terminology and establishing a taxonomy for various approaches.
- It highlights the importance of RGB-D imagery, occupancy grids, and LiDAR point clouds for large-scale scene modeling.
- The survey identifies gaps in existing literature and discusses applications, challenges, and research directions for future work.
3D and 4D world modeling are essential areas in AI that help agents understand their environments. Previous research primarily focused on 2D generative methods, overlooking the growing field of 3D representation. This survey aims to unify and clarify the existing studies in this area.
Read the original ↗
AgentGym-RL: Training LLM Agents for Long-Horizon Decision Making through Multi-Turn Reinforcement Learning
The development of the AgentGym-RL framework enables more effective training of large language model agents for complex decision-making, potentially advancing automation technologies significantly.
- AgentGym-RL is a modular reinforcement learning framework for training large language model agents in diverse environments without supervised fine-tuning.
- The framework includes ScalingInter-RL, a method that balances exploration and exploitation during training to enhance decision-making capabilities.
- Extensive experiments demonstrate that agents trained using AgentGym-RL perform at or above the level of commercial models on 27 different tasks.
AgentGym-RL addresses the lack of a unified reinforcement learning framework for training intelligent agents without relying on supervised fine-tuning. As the field evolves, there is a growing need for agents capable of independent decision-making in real-world scenarios. Future developments will focus on enhancing agent capabilities and expanding the framework's applications.
Read the original ↗
Meet mmBERT: An Encoder-only Language Model Pretrained on 3T Tokens of Multilingual Text in over 1800 Languages and 2–4× Faster than Previous Models
The development of mmBERT is significant as it advances multilingual natural language processing, addressing limitations of previous models while improving efficiency and inclusivity for low-resource languages.
- mmBERT is a modern encoder-only language model pre-trained on 3 trillion tokens, covering 1,833 languages.
- It features a performance boost of 2–4 times over earlier models, processing sequences up to 8192 tokens.
- The model achieves competitive scores on English and multilingual benchmarks, surpassing existing models like XLM-R.
- mmBERT's innovations include an annealed learning schedule and a merging strategy that enhances performance, especially for low-resource languages.
mmBERT, developed by researchers at Johns Hopkins University, addresses a gap in multilingual encoder capabilities dominated by XLM-R for over five years. With a unique training methodology, it is designed to be more efficient and adaptable, especially for underrepresented languages. Future developments may focus on further expanding its applications in natural language processing.
Read the original ↗
A California bill that would regulate AI companion chatbots is close to becoming law
If enacted, this bill would establish California as a leader in regulating AI, particularly in protecting vulnerable users from potential harms associated with AI companion chatbots.
- California's SB 243 bill aims to regulate AI companion chatbots and protect minors and vulnerable users.
- The bill has gained bipartisan support in the State Assembly and Senate and now awaits Governor Newsom's approval.
- It requires chatbot operators to implement safety protocols, including alerts that users are interacting with an AI and not a real person.
AI companion chatbots are systems designed to provide human-like responses and meet users' social needs. This bill was introduced in January following the death of a teenager linked to AI chatbot interactions. Regulatory scrutiny on AI platforms has increased, particularly concerning their impact on minors' mental health.
Read the original ↗
Satellites Have Been Mysteriously Failing in Orbit—Now Scientists Know the Reason Why
Understanding the causes of spacecraft failures can lead to better protective measures, ultimately reducing communication disruptions and improving the reliability of satellite operations.
- Scientists at Los Alamos National Laboratory identified the cause of sudden spacecraft failures in orbit, known as spacecraft environment discharges (SEDs).
- SEDs correlate with electron activity in the surrounding space environment, specifically during solar storms.
- New data from the Department of Defense’s STP-Sat6 satellite revealed hundreds of discharge events linked to peaks in electron activity.
- Around 75% of the time, electron peaks occur 30 to 45 minutes before SED events, indicating a direct link.
Spacecraft environment discharges (SEDs) are electrical discharges that can cause significant damage to satellite electronics and communications systems. Previous unexplained satellite failures raised concerns about the impact of space weather. This new research may lead to future monitoring systems to predict and mitigate these failures.
Read the original ↗
Perplexity reportedly raised $200M at $20B valuation
Perplexity's significant funding increase reflects the growing investment interest in AI-powered search solutions, highlighting competitive pressures in the digital search landscape.
- Perplexity raised $200 million, increasing its valuation to $20 billion.
- This funding follows a $100 million raise two months prior at an $18 billion valuation.
- The startup's total funding to date reaches $1.5 billion since its founding three years ago.
- Perplexity's annual recurring revenue (ARR) is nearing $200 million, up from over $150 million last month.
Perplexity is an AI-powered search startup that competes with Google by providing conversational responses to queries. The company has rapidly grown in market presence and funding, pushing back against Google's search dominance. Upcoming developments to watch include potential adjustments in response to regulatory actions affecting Google's services.
Read the original ↗
DCReg: Decoupled Characterization for Efficient Degenerate LiDAR Registration
Improving LiDAR point cloud registration enhances robotic perception and navigation, which is crucial for automated systems in challenging environments.
- DCReg addresses ill-conditioned LiDAR point cloud registration issues by detecting and mitigating degeneracies using Schur complement decomposition.
- The framework separates registration into clean subspaces for accurate characterization of motion directions, allowing for better identification of constraints.
- A novel preconditioner stabilizes ill-conditioned directions while preserving well-constrained observable information, leading to significant performance improvements.
LiDAR point cloud registration is essential for enabling robots to perceive and navigate within their environments. Traditional methods often struggle with accuracy in geometrically challenging settings, necessitating innovative solutions like DCReg to overcome these limitations.
Read the original ↗
Sources: AI training startup Mercor eyes $10B+ valuation on $450M run rate
Mercor's potential $10 billion valuation highlights the increasing demand for specialized AI training services, reflecting the industry's growth and investment trends.
- Mercor is in discussions for a Series C funding round aiming for a valuation of over $10 billion, up from a previous target of $8 billion.
- The company is approaching $450 million in annualized run-rate revenue and has already generated $6 million in profit in the first half of the year.
- Mercor was co-founded in 2022 by three individuals in their early twenties and now serves top AI labs like OpenAI and Meta.
Founded in 2022, Mercor connects AI companies with domain experts for model training. The startup is generating significant revenue from its contracts and is competing with established firms like Scale AI while expanding its service offerings.
Read the original ↗
Thinking Machines Lab wants to make AI models more consistent
Improving the consistency of AI model responses could enhance reliability in AI applications and streamline reinforcement learning, impacting how AI systems are developed and implemented across industries.
- Thinking Machines Lab, backed by $2 billion in seed funding, is focused on creating AI models that generate more consistent responses.
- The team's blog post discusses tackling randomness in AI responses, which is a recognized issue in non-deterministic systems like current AI models.
- Researchers suggest that controlling GPU kernel orchestration may lead to more deterministic AI responses, facilitating smoother reinforcement learning training.
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, aims to advance AI technologies with a strong research focus. The startup offers insights into its work through public blog posts, a shift from the more closed approach of larger AI companies. Observers will monitor whether the lab can deliver practical solutions aligning with its significant valuation.
Read the original ↗
Stronger Chatbot Guardrails, Weaker Google Monopoly, AI-Assisted Education, 10 Million Tokens of Context
The shift towards skills-based education and new regulations for chatbots could reshape the landscape of AI in education and online interactions, impacting millions of users.
- Coursera's annual conference highlighted a shift from knowledge-based to skills-based education, benefiting individuals and institutions.
- Google is required to share its search index with competing AI firms while retaining ownership of Chrome and Android.
- Meta and OpenAI announced new regulations to enhance the safety of chatbot interactions with minors.
- A Texas-based school program uses AI to replace a traditional 6-hour school day with just 2 hours of personalized learning.
- A transformer-like architecture called ATLAS can handle context with up to ten million tokens, enhancing large language model capabilities.
Coursera is an online learning platform that offers courses from various institutions. The company is responding to the evolving demands of the job market by focusing on practical skills. Recent regulatory changes are pushing technology companies to make their data more accessible to enhance competition in the AI field.
Read the original ↗
NVIDIA AI Releases Universal Deep Research (UDR): A Prototype Framework for Scalable and Auditable Deep Research Agents
NVIDIA's Universal Deep Research framework enhances flexibility and efficiency for AI agents, allowing enterprises to customize research processes without the need for retraining models, addressing critical limitations of existing tools.
- NVIDIA introduced Universal Deep Research (UDR), a prototype open-source framework for customizable deep research agents that operates without retraining LLMs (large language models).
- UDR improves on existing deep research tools by decoupling research strategies from models, enabling users to design workflows with no fixed procedures.
- The framework processes user-defined strategies through a structured system, compiling them into Python code and executing them efficiently.
Universal Deep Research (UDR) aims to overcome the limitations of Deep Research Tools (DRTs), which often bind users to single models and rigid workflows. By providing a model-agnostic architecture, UDR facilitates innovation across various sectors like science and enterprise. Observers should watch for its integration and reception in high-stakes environments.
Read the original ↗
I want to love Apple’s new iPhone Air, but the iPhone 17 is a better deal
Consumer choices in smartphone technology influence market trends and pricing strategies within the tech industry.
- The new iPhone Air features a 6.5-inch screen and the A19 Pro chip but is only marginally more expensive than the iPhone 17.
- The iPhone 17 offers longer battery life, promising 30 hours of video playtime compared to the Air's 27 hours.
- While the Air has better storage options, the highest tier costs $1,400, approaching the price of a Pro model with similar storage.
Apple's latest smartphones, the iPhone Air and iPhone 17, showcase advancements in design and technology. Analysts and consumers are comparing features like battery life and camera quality to assess value. Observers should watch for upcoming comparisons and consumer preferences leading up to the holiday season.
Read the original ↗
One of Google’s new Pixel 10 AI features has already been removed
The removal of the Daily Hub feature from Pixel 10 phones highlights challenges in delivering effective AI-driven user experiences, impacting customer satisfaction and trust in new technologies.
- Google has discontinued the Daily Hub feature just weeks after the Pixel 10 series launch due to its limited effectiveness.
- Daily Hub appeared prominently on devices, but it primarily offered basic information like weather and calendar reminders.
- A Google spokesperson stated the feature is on hold for enhancements and will return once improved.
Google has integrated advanced AI capabilities in its Pixel phone lineup, leveraging its Tensor processor for on-device processing. Past similar features, such as Samsung's Now Brief, faced issues in providing valuable insights, indicating a broader challenge in mobile AI functionality.
Read the original ↗
OpenAI and Oracle reportedly ink historic cloud computing deal
The collaboration between OpenAI and Oracle could significantly impact the cloud computing market, enabling enhanced AI capabilities through unprecedented compute power access.
- OpenAI signed a five-year deal with Oracle to purchase $300 billion worth of cloud computing power, starting in 2027.
- This potential agreement is considered one of the largest cloud contracts ever, bolstering Oracle's position in the cloud market.
- Previously, OpenAI had shifted some of its computing needs from Microsoft Azure to Oracle, in addition to a deal with Google earlier this year.
OpenAI is a leading artificial intelligence company that requires vast computational resources for its operations. The deal with Oracle follows prior engagements, indicating OpenAI's strategy to diversify its cloud resources amid growing AI competition. Future developments may reveal more about how this partnership will enhance OpenAI's capabilities.
Read the original ↗
Reverse-Engineered Reasoning for Open-Ended Generation
The introduction of the REER paradigm could significantly enhance the performance of AI in open-ended generation tasks, where conventional methods have limitations.
- REER is a new paradigm for deep reasoning that reverse-engineers reasoning processes to enable models to perform well on open-ended tasks.
- Traditional methods like reinforcement learning and instruction distillation face challenges in open-ended generation, including unclear reward signals and high costs.
- The researchers curated and open-sourced the DeepWriting-20K dataset containing 20,000 deep reasoning trajectories for open-ended tasks to support their model training.
- Their model, DeepWriter-8B, outperforms strong open-source baselines and competes effectively with major proprietary models like GPT-4o and Claude 3.5.
Deep reasoning refers to complex problem-solving methods that have achieved progress in areas like mathematics but struggled in creative tasks. The development of REER marks a significant shift in tackling open-ended generation challenges in AI, making it a key focus area for future research.
Read the original ↗
Developers joke about “coding like cavemen” as AI service suffers major outage
The outage underscores the risks associated with reliance on AI tools in software development, as even brief service interruptions can disrupt workflow for many developers.
- Anthropic experienced a complete service outage for approximately 30 minutes, impacting its AI infrastructure and leaving developers unable to access key services.
- The outage affected Claude.ai, the API, Claude Code, and the management console, prompting immediate reactions from users on Hacker News.
- During the downtime, some developers turned to alternative AI solutions, while others joked about reverting to traditional coding methods.
Anthropic is a company that provides AI tools for software development, such as Claude.ai and Claude Code. This outage reflects ongoing reliability issues users have noted, particularly during peak US working hours. As AI coding assistants become integral to development, monitoring their reliability remains a key concern.
Read the original ↗
Better Than GPT-5? We Try ERNIE X1.1, Baidu’s Latest AI Model
The development of ERNIE X1.1 marks Baidu's transition towards more reliable AI models, aiming to reduce errors and improve adherence to user instructions, affecting future applications in AI-assisted technologies.
- Baidu announced ERNIE X1.1 at the Wave Summit in Beijing, presenting it as a reasoning-first model with a focus on practical reliability.
- The new model shows a 34.8% improvement in factuality, a 12.5% increase in instruction following, and a 9.6% enhancement in agentic capabilities over its predecessor, ERNIE 4.5.
- ERNIE X1.1 is available via the ERNIE Bot website, the Wenxiaoyan mobile app, and through Baidu's Qianfan API, which supports enterprise deployments.
ERNIE X1.1 is an AI model developed by Baidu, built upon the ERNIE 4.5 base and designed to improve reasoning and reliability. This model's introduction reflects a broader shift in AI development towards practical applications and user-centered performance. Future evaluations will be necessary to verify Baidu's claims of parity with other leading models like GPT-5 and Gemini 2.5 Pro.
Read the original ↗
Anthropic reports outages, Claude and Console impacted
The outages of Anthropic's services, including Claude and its API, highlight the ongoing reliability challenges faced by AI platforms, impacting users' productivity and trust in technology.
- Anthropic reported a service outage affecting its APIs, Console, and Claude earlier today.
- Users began experiencing issues around 12:20 p.m. ET, with the company acknowledging the outage shortly afterward.
- The company released a status update stating that it implemented several fixes and was monitoring the situation.
Anthropic is an AI company known for its language model, Claude. The platform has experienced previous outages and bugs, causing concerns among users about reliability. Continued monitoring and fixes will be essential for restoring user confidence.
Read the original ↗