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
Baidu Releases ERNIE-4.5-VL-28B-A3B-Thinking: An Open-Source and Compact Multimodal Reasoning Model Under the ERNIE-4.5 Family
The introduction of Baidu's ERNIE-4.5-VL-28B-A3B-Thinking model highlights advancements in multimodal reasoning capabilities, which could impact various AI applications.
- Baidu has released a new multimodal reasoning model named ERNIE-4.5-VL-28B-A3B-Thinking.
- This model is open-source and designed to be compact, while still offering advanced reasoning capabilities.
Baidu's ERNIE series focuses on enhancing language and reasoning models. The release builds on previous iterations in the ERNIE family and aims to improve accessibility for developers and researchers in AI.
Based on the headline only — the full article could not be fetched.
Read the original ↗
The global race for the AI app layer is still on
Understanding the competitive dynamics in AI application development highlights opportunities for growth and investment outside traditional model-building, influencing future funding strategies and market trends.
- The U.S. is leading in funding for large AI models, but Europe is emerging as a significant player in the AI application layer.
- European and Israeli AI applications have raised 66 cents for every dollar raised by American counterparts in 2025, according to Accel's report.
- A new generation of AI-native applications is achieving $100 million in annual recurring revenue significantly faster than before, changing growth expectations for software companies.
The report from VC firm Accel evaluates the current state of AI and cloud markets, suggesting that while the U.S. dominates large models, Europe is developing a robust ecosystem of application startups. The emergence of effective AI-native applications marks a shift in speed and efficiency not seen in previous tech waves, indicating a maturing landscape.
Read the original ↗
How to Reduce Cost and Latency of Your RAG Application Using Semantic LLM Caching
Reducing costs and latency in RAG applications can enhance efficiency and improve user experience, making advanced technology more accessible and practical for developers.
- The article discusses methods to reduce cost and latency in RAG applications.
- It emphasizes using semantic LLM (Large Language Model) caching for optimizing performance.
RAG applications involve retrieval-augmented generation, combining information retrieval with language generation. Semantic LLM caching refers to storing and reusing information to speed up response times and lower costs. Developers and organizations interested in optimizing their AI applications should consider these techniques.
Based on the headline only — the full article could not be fetched.
Read the original ↗
Ace Your Exams with AI: Step-by-step Guide for Google Gemini as you...
The ability of students to effectively use AI tools like Google Gemini for exam preparation could transform traditional study methods, making learning more efficient and accessible.
- Google Gemini is an AI chatbot designed to assist students in midterm exam preparation, offering various features for personalized learning.
- Students can upload study materials to Gemini and generate customized study plans based on their syllabus and time frame.
- Gemini enables real-time question answering, allowing students to clarify doubts and understand complex topics during their study sessions.
Google Gemini is a multimodal AI model developed by Google that can process various types of information. This guide emphasizes its application for academic success, particularly for exam preparation. Users can explore its capabilities for tailored learning experiences.
Read the original ↗
RLVE: Scaling Up Reinforcement Learning for Language Models with Adaptive Verifiable Environments
Improving language model reasoning through dynamic problem difficulty adjustments could enhance AI's problem-solving capabilities and efficiency during training.
- Reinforcement Learning with Adaptive Verifiable Environments (RLVE) dynamically adjusts problem difficulty for language models, improving performance over traditional methods.
- RLVE-Gym, a large suite of 400 verifiable environments, facilitates environment scaling that enhances reasoning capabilities during training.
- Training with RLVE results in a 3.37% average improvement across six reasoning benchmarks, compared to a 0.49% gain from standard RL training under higher computational demands.
Reinforcement learning (RL) is a type of machine learning where agents learn to make decisions by receiving rewards. Previous methods often relied on static problem distributions, which can hinder learning. The introduction of RLVE aims to optimize the training process for language models by adapting challenges as learning progresses.
Read the original ↗
Artificial Intelligence (372)
The development of AI tools for managing email can significantly enhance productivity for users inundated with communications.
- AI is increasingly being utilized for managing emails effectively.
- A new tool for Gmail inbox management is introduced using the n8n platform.
Artificial Intelligence (AI) refers to the simulation of human intelligence in machines. The growth of AI applications continues to transform various sectors, including email management. Advances in platforms like n8n allow developers to create customized automation solutions for users.
Based on the headline only — the full article could not be fetched.
Read the original ↗
Teaching Pretrained Language Models to Think Deeper with Retrofitted Recurrence
This research highlights advancements in language model efficiency, potentially leading to more effective AI systems that require less computational power while maintaining high performance.
- This study explores converting pretrained non-recurrent language models to depth-recurrent models to enhance performance.
- Using a curriculum of recurrences during training helps preserve performance while cutting down on computational costs.
- Experiments indicate that retrofitting existing models yields better results compared to simply post-training non-recurrent models.
Language models are algorithms designed to understand and generate human language. Recent innovations have focused on enhancing model depth and performance while managing computational resources effectively. Future developments will likely aim to improve efficiency in other areas of AI model training.
Read the original ↗
Google says new cloud-based “Private AI Compute” is just as secure as local processing
The development of Google's Private AI Compute system could shift data processing towards the cloud while maintaining security, impacting how users interact with AI technologies.
- Google's new Private AI Compute offers cloud-based AI processing claimed to be as secure as local processing.
- The system connects devices to Google's secure servers via an encrypted link, leveraging its Tensor Processing Units (TPUs).
- This service allows for more powerful AI models like Gemini to be utilized, enhancing user experiences without compromising data privacy.
Google's Private AI Compute aims to integrate generative AI into more products while ensuring user data remains secure. This approach is similar to Apple's Private Cloud Compute and represents a shift towards cloud-dependent AI functionality. The effectiveness and adoption of this system could redefine user interaction with AI tools.
Read the original ↗
Grounding Computer Use Agents on Human Demonstrations
The introduction of the GroundCUA dataset enhances the development of computer-use agents, enabling more efficient and effective mapping of instructions to UI elements, which could improve user interaction with software.
- GroundCUA is a large-scale desktop grounding dataset created from expert human demonstrations, covering 87 applications across 12 categories.
- The dataset includes 56,000 screenshots with over 3.56 million human-verified annotations, providing a comprehensive resource for training models.
- GroundNext models, built on the GroundCUA dataset, achieve state-of-the-art results with less than one-tenth the training data compared to previous models.
GroundCUA addresses a gap in high-quality datasets for desktop environments, which are less common than those for web and mobile. This enables the creation of more reliable AI agents that can understand and navigate user interfaces effectively. Future developments may focus on enhancing these models' capabilities and their real-world applications.
Read the original ↗
SoftBank’s Nvidia sale rattles market, raises questions
SoftBank's divestment from Nvidia signifies a notable shift in investment strategy, raising concerns about market stability and the future trajectory of AI funding.
- SoftBank has sold its entire $5.8 billion stake in Nvidia to reinvest in AI ventures.
- The sale impacted Nvidia's stock, which fell nearly 3% following the announcement.
- Masayoshi Son plans a $30 billion commitment to OpenAI and aims to participate in a $1 trillion AI manufacturing project in Arizona.
SoftBank, founded by Masayoshi Son, is a major player in technology investments, previously known for its Vision Fund and high-profile bets like Alibaba. This sale marks SoftBank's second exit from Nvidia, previously selling a stake at a loss in 2019, and signals a shift towards aggressive AI investments.
Read the original ↗
How AI startups should be thinking about product-market fit
Understanding product-market fit is crucial for AI startups as it impacts long-term sustainability and successful integration of their products in evolving markets.
- AI startups must evaluate their product-market fit differently due to rapid technology changes, moving beyond traditional strategies.
- Founders should monitor 'durability of spend' as companies shift from experimental budgets to core operational budgets.
- Classic engagement metrics, like daily and monthly active users, remain vital for assessing customer interaction with AI tools.
- Qualitative feedback from users can provide deeper insights into customer loyalty and product effectiveness.
- Product-market fit should be seen as a continuum, requiring ongoing adaptation and strength over time.
Product-market fit refers to the alignment of a product with market demand, a crucial aspect for startups. AI startups are facing unique challenges due to the fast-paced evolution of technology, which influences traditional approaches to achieving product-market fit. Observing changes in customer spending habits and engaging in qualitative assessments are key strategies for success in this landscape.
Read the original ↗
Immortality startup Eternos pivots to a personal AI that sounds like you
The shift in focus to personal AIs reflects growing trends in AI personalization, allowing individuals to create tailored models that enhance personal and professional engagement.
- Eternos has rebranded as Uare.ai, shifting its focus from a legacy preservation service to developing personal AIs that reflect users' unique life stories and traits.
- The company raised $10.3 million in seed funding, led by Mayfield and Boldstart Ventures, to support its new mission.
- Uare.ai will offer users the ability to train their personal AIs using personal data through text, voice, and video inputs.
Eternos originally aimed to help users create digital replicas to preserve their stories after death. The rebranding to Uare.ai signifies a transition towards leveraging AI for ongoing personal and professional applications.
Read the original ↗
Google announces even more AI in Photos app, powered by Nano Banana
The integration of the Nano Banana AI model into Google Photos enhances the app's capabilities, allowing users to perform complex image edits more easily and intuitively.
- Google Photos is adding the Nano Banana AI model to improve its image-editing capabilities on both Android and iOS platforms.
- The updated Help Me Edit feature allows users to make specific edits using names from their private face groups, which streamlines the editing process.
- A new 'Create with AI' section will provide pre-formed prompts for various edits, aiding less-experienced users.
- An 'Ask' button is introduced for contextual assistance in photo editing and information retrieval, initially available only in the US.
Google Photos is an application by Google that allows users to store and edit images. The company has been progressively integrating advanced AI features into its products to enhance user experience. Users should anticipate further improvements and additional features as AI technology evolves.
Read the original ↗
Wonderful raised $100M Series A to put AI agents on the front lines of customer service
Wonderful's significant funding indicates a strong belief in AI's role in transforming customer service, potentially reshaping how enterprises interact with customers across global markets.
- Israeli startup Wonderful has raised $100 million in a Series A funding round led by Index Ventures, increasing its total funding to $134 million.
- Wonderful aims to help enterprises deploy AI agents across voice, chat, and email, adapting to various languages and cultural norms.
- The startup plans to expand its services to Germany, Austria, the Nordics, and Portugal in 2025, with further growth in the Asia-Pacific region expected in 2026.
Wonderful, which recently emerged from stealth mode, focuses on creating customer-facing AI agents tailored for diverse markets. As AI agents become increasingly crucial for cost-effective customer service, Wonderful's technology aims to enhance traditional support infrastructures.
Read the original ↗
Lawyers keep giving weak-sauce excuses for fake AI citations in court docs
The prevalence of inaccurate AI-generated citations in legal documents undermines court integrity and raises questions about accountability in legal practices involving AI technologies.
- Judges are facing an increase in fake AI-generated case citations, leading to a rise in sanctions against lawyers.
- One judge noted the simplest way for lawyers to avoid severe penalties is to admit their use of AI and self-report errors.
- Common excuses for using faulty AI citations include ignorance of AI capabilities and technology issues, which judges largely reject as credible.
AI technologies are increasingly used in legal writing, but this has led to numerous cases of lawyers submitting documents with fabricated citations. Courts have begun imposing sanctions as awareness of the issue grows, and ongoing guidance is being established regarding AI usage in legal contexts.
Read the original ↗
The Best Ergonomic Mouse to Keep Wrist Strain at Bay
Choosing an ergonomic mouse can significantly reduce the risk of wrist strain and enhance overall comfort during prolonged computer use.
- The Logitech MX Master 4 is recommended as a top ergonomic mouse due to its improved design and usability features.
- Ergonomic mice are designed to fit better in the hand, potentially reducing risks of injuries like carpal tunnel syndrome.
- Studies indicate that angled and vertical mice alleviate discomfort by aligning with the natural wrist position.
Ergonomic mice differ from standard mice as they prioritize comfort and hand ergonomics to prevent strain injuries. The market largely focuses on gaming mice, often neglecting ergonomic design. Possible future developments may include improved ergonomic features and designs tailored to various uses.
Read the original ↗
Meta’s chief AI scientist Yann LeCun reportedly plans to leave to build his own startup
LeCun's departure could impact Meta's AI research direction, especially as the company seeks to strengthen its competitiveness against rivals like OpenAI and Google.
- Yann LeCun is reportedly planning to leave Meta to establish his own startup focused on world models, according to anonymous sources.
- LeCun is already in discussions to secure funding for his new venture, which aims to continue his research work.
- His departure comes as Meta undergoes significant changes in its AI development strategies to keep up with competitors.
- LeCun has expressed skepticism about the current marketing of AI technologies, particularly large language models (LLMs).
Yann LeCun is a leading figure in AI research, recognized for his work as Meta's chief AI scientist and as a professor at New York University. His exit comes at a time when Meta is reshaping its AI division, partly due to competition from companies such as OpenAI and Google. Future developments to watch include LeCun's startup efforts and Meta's ongoing AI restructuring.
Read the original ↗
New Study Finds Half of Americans Don’t Know Alcohol Raises Cancer Risk
Limited public awareness about the cancer risks associated with alcohol consumption can lead to higher rates of alcohol-related cancers and deaths, affecting overall public health.
- A study from the University of Texas MD Anderson Cancer Center reveals that over half of Americans are unaware that alcohol increases cancer risk.
- The research, based on nearly 7,000 survey responses, indicates that regular drinkers are often the least informed about this risk.
- Only 37.1% of survey participants recognized the connection between alcohol and cancer, while about 1% mistakenly believed alcohol lowers cancer risk.
The World Health Organization classifies alcohol as a Group 1 carcinogen, linking it to various cancers. Previous studies have highlighted similar gaps in health awareness, particularly around smoking and diet. Future health campaigns may need to focus more on alcohol-related risks to improve public understanding.
Read the original ↗
The Former Staffer Calling Out OpenAI’s Erotica Claims
Concerns over OpenAI's planned allowance for erotic interactions with chatbots highlight broader mental health impacts of AI, raising challenges for responsible development and use of technology.
- Steven Adler, a former OpenAI safety lead, voiced concerns about the mental health risks of chatbots engaging in erotic conversations.
- OpenAI CEO Sam Altman announced that the company would permit erotica for verified adult users, prompting Adler's op-ed warning against trusting those claims.
- Adler claims OpenAI lacks adequate measures to manage risks associated with erotic AI interactions, which he argues should be addressed before proceeding.
OpenAI is a leading artificial intelligence research organization known for developing AI technologies like chatbots and language models. As AI applications expand, discussions about safety and potential societal impacts intensify. Adler's insights reflect ongoing debates in the AI community regarding ethical standards and user safety.
Read the original ↗
Booze Without the Burn? An Enzyme-Tinkering Startup Aims to Make Spirits Smoother
The development of a smoother alcoholic beverage through enzyme modification could challenge traditional distillation practices and alter consumer drinking habits, raising questions about safety and consumption limits.
- Voodoo Scientific has discovered methods to reduce the harsh burning sensation of alcohol using enzymes during fermentation.
- The company identified nine compounds responsible for the irritation caused by spirits, which are by-products of the fermentation process.
- Their product, Viriato, aims to transform the drinking experience by creating a smoother tipple without compromising quality.
Voodoo Scientific is a startup focused on improving alcoholic beverages by addressing the pain response caused by drinking. Their innovative approach involves identifying and neutralizing specific compounds during the fermentation process. Future developments will be closely monitored as the company tests its processes in various distilleries worldwide.
Read the original ↗