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
VLA-RFT: Vision-Language-Action Reinforcement Fine-tuning with Verified Rewards in World Simulators
The VLA-RFT framework improves the efficiency and robustness of Vision-Language-Action models, which are crucial for advancing embodied AI decision-making capabilities.
- VLA-RFT fine-tunes Vision-Language-Action models using a data-driven world model, enhancing performance while reducing sample requirements.
- The framework utilizes reinforcement learning to provide a more stable learning signal through predicted visual observations and dense, trajectory-level rewards.
- With fewer than 400 fine-tuning steps, VLA-RFT outperforms traditional supervised methods and shows improved robustness under disturbed conditions.
Vision-Language-Action models integrate visual, linguistic, and action data to facilitate decision-making. Previous methods have struggled with errors and inefficiencies in learning, particularly when transitioning from simulations to real environments. VLA-RFT's introduction could signify a shift towards more practical and effective post-training procedures for these models.
Read the original ↗
Code2Video: A Code-centric Paradigm for Educational Video Generation
Code2Video addresses the challenge of generating high-quality educational videos, enhancing the effectiveness of online learning through improved coherence and discipline-specific content creation.
- Code2Video is a code-centric framework that generates educational videos using executable Python code.
- The framework includes three collaborative agents: Planner, Coder, and Critic, which work together to create coherent and visually structured content.
- Code2Video achieved a 40% improvement over direct code generation, producing videos comparable to human-crafted tutorials.
Code2Video represents an advanced method for educational video generation, addressing limitations found in existing generative models. Traditional video synthesis techniques struggle to produce the necessary clarity and structure required for educational material. Future developments may focus on further enhancing video quality and scalability in educational applications.
Read the original ↗
Beyond Log Likelihood: Probability-Based Objectives for Supervised Fine-Tuning across the Model Capability Continuum
This research could improve supervised fine-tuning techniques for large language models, leading to better performance across diverse tasks and models with varying capabilities.
- New probability-based objectives are proposed for fine-tuning large language models (LLMs), outperforming the traditional negative log likelihood (NLL) in specific contexts.
- The performance of training objectives varies across the model capability continuum, with prior-leaning objectives excelling in stronger models while NLL is more effective for weaker ones.
- The study conducted experiments with seven model backbones, 14 benchmarks, and three domains to examine the effectiveness of these probability-based objectives.
Supervised fine-tuning (SFT) is a common method to refine large language models after initial training, but it can struggle with generalization. This research aims to enhance SFT methods by identifying better training objectives based on a model's capabilities. Following this study, further exploration of different training objectives may lead to more robust performance in LLM applications.
Read the original ↗
ServiceNow AI Releases Apriel-1.5-15B-Thinker: An Open-Weights Multimodal Reasoning Model that Hits Frontier-Level Performance on a Single-GPU Budget
The release of Apriel-1.5-15B-Thinker signifies advancements in AI that could make high-performance multimodal reasoning more accessible through efficient use of resources.
- ServiceNow AI has launched Apriel-1.5-15B-Thinker, a new AI model designed for multimodal reasoning.
- The model features open weights, allowing for greater flexibility and accessibility in AI applications.
ServiceNow AI is focused on developing innovative AI models to enhance business processes and decision-making. The introduction of Apriel-1.5-15B-Thinker represents a significant milestone in making advanced AI technology more resource-efficient and widely usable. Future developments may further push the boundaries of AI performance and efficiency.
Based on the headline only — the full article could not be fetched.
Read the original ↗
Disney Forces Character.AI To Remove Iconic Characters
The removal of Disney characters from Character.AI underscores the ongoing tension between user-generated content and copyright law, impacting both user experience and brand integrity.
- Character.AI has removed several Disney characters after receiving a cease-and-desist letter from Disney regarding copyright infringement.
- Disney accused Character.AI of exploiting its intellectual property and causing reputational harm through unregulated AI interactions.
- The removal of characters like Mickey Mouse and Captain America reflects Disney's efforts to protect its brand from potential misuse in harmful contexts.
Character.AI allows users to create AI companions, mimicking both real and fictional figures. The platform has faced scrutiny due to past incidents involving harmful interactions, raising concerns about safety for younger users. Disney's actions may influence other companies to demand similar compliance regarding their characters.
Read the original ↗
OpenAI staff grapples with the company’s social media push
The Sora app's launch raises critical questions about OpenAI's balance between profitability and its mission to develop AI that benefits humanity, reflecting broader industry challenges.
- OpenAI launched its Sora app, a social media platform featuring AI-generated videos, causing internal concerns about alignment with its nonprofit mission.
- Current and former researchers at OpenAI express mixed feelings about Sora, with some fearing it may replicate the issues seen in traditional social media apps.
- CEO Sam Altman stated that the consumer business, including Sora, is intended to fund AI research while ensuring that their mission remains a priority.
OpenAI is known for developing advanced AI technologies, with previous projects like ChatGPT focused on utility rather than entertainment. The introduction of Sora marks a significant shift as the company explores consumer-facing applications, which may impact its foundational goals of creating beneficial AI.
Read the original ↗
OpenAI mocks Musk’s math in suit over iPhone/ChatGPT integration
This legal dispute highlights tensions in the tech industry over access to AI technologies, potentially influencing antitrust enforcement and market competition.
- xAI, founded by Elon Musk, claims Apple’s integration of ChatGPT into iPhone features gives OpenAI a monopoly on prompts, violating antitrust laws.
- OpenAI and Apple seek to dismiss the lawsuit, arguing that Musk's claims are based on flawed calculations and speculative assumptions.
- OpenAI asserts that integration affects only the latest iPhone models and requires users to opt in for data usage, limiting the potential prompt pool.
- Apple emphasizes that Musk's grievances stem from complaints about his own app Grok not being featured as a 'Must Have' app, suggesting dissatisfaction with Apple's editorial choices rather than actual competition harm.
OpenAI developed ChatGPT, a widely used AI chatbot, while Musk's xAI is a newer entrant in the AI space. The lawsuit reflects ongoing rivalry and concerns about market control in the rapidly evolving AI industry. Observers should watch for outcomes that could set precedents for antitrust cases involving tech giants and emerging AI firms.
Read the original ↗
Regression Language Models for Code
The development of a unified Regression Language Model (RLM) enhances the ability to predict code execution outcomes, potentially improving efficiency in software development and performance assessment.
- A unified Regression Language Model predicts numeric outcomes like memory usage and latency from code across various programming languages and hardware platforms.
- The RLM can predict metrics from code in languages such as Python and C++, achieving a Spearman-rank correlation exceeding 0.9 on specific programming submissions.
- With over 300 million parameters, the model performs well against traditional benchmarks, including surpassing previous methods used for graph neural networks.
- The RLM consistently predicts architecture latencies across different hardware, showing its versatility in multiple coding and execution environments.
Regression Language Models aim to predict numeric outcomes of software executions, addressing challenges in programming languages' complexity. Prior methods often required extensive feature engineering, while the RLM integrates multiple predictors into a single model. Future advancements could see RLMs integrated into development environments for real-time performance analysis.
Read the original ↗
TTT3R: 3D Reconstruction as Test-Time Training
TTT3R improves 3D reconstruction performance and efficiency, making it a significant advancement in the ability to process complex visual data, with applications across various industries.
- TTT3R enhances length generalization in 3D reconstruction by adjusting memory updates based on alignment confidence.
- This method results in a twofold increase in global pose estimation compared to baseline models.
- TTT3R operates at 20 frames per second while requiring only 6 GB of GPU memory to handle thousands of images.
Test-Time Training refers to a machine learning approach where models are updated at inference time. The development of proactive techniques like TTT3R is crucial for improving the adaptability and efficiency of AI in real-time applications. Continued advancements in 3D reconstruction technology may lead to better performance in fields such as autonomous driving and robotics.
Read the original ↗
OpenAI’s Trillion-Dollar Bet, Generating Viruses, Modeling Planet Earth, Paying for Training Data
The advancements in AI-driven technologies present significant implications for industries like climate science, music copyright, and healthcare through enhanced data processing and modeling capabilities.
- OpenAI is planning to collaborate with Oracle, Nvidia, and Softbank to develop 20 gigawatts of data center capacity, potentially costing trillions.
- Researchers have developed AI models that can generate custom viruses from genetic data, highlighting advanced capabilities in genomic engineering.
- Google's AlphaEarth Foundations model integrates satellite imagery to monitor climate and land use on a global scale.
- Sweden's STIM has created a framework for training AI on music while ensuring original artists receive compensation.
- LandingAI's Agentic Document Extraction enables efficient conversion of PDFs into formats compatible with Large Language Models (LLMs).
OpenAI is a leading AI research organization known for developing advanced AI models. Recent developments illustrate the growing partnerships aimed at expanding computational resources and leveraging AI across various sectors. Key areas to follow include advancements in AI for climate modeling and healthcare applications.
Read the original ↗
Character.AI removes Disney characters after receiving cease-and-desist letter
The removal of Disney characters from Character.AI indicates the increasing tension between tech companies and copyright holders in the AI space, impacting how users interact with branded content.
- Character.AI ceased offerings of Disney characters after receiving a cease-and-desist letter from Disney.
- Disney accused Character.AI of infringing its copyrights by allowing users to create chatbots featuring its iconic characters.
- The letter highlighted concerns over harmful content generated by the chatbots, potentially damaging Disney's reputation.
Character.AI is a platform that allows users to create AI chatbots based on real or fictional characters. The company has faced criticism for unfiltered chatbot content in the past, leading to legal repercussions. The removal of Disney characters showcases the ongoing disputes between AI developers and content owners over copyright enforcement.
Read the original ↗
California’s new AI safety law shows regulation and innovation don’t have to clash
The passage of California's SB 53 law illustrates a model for balancing innovation in AI technologies with necessary safety regulations, impacting how future legislation may approach the tech industry.
- California Governor Gavin Newsom recently signed SB 53, requiring large AI labs to disclose their safety protocols.
- The law aims to prevent catastrophic risks associated with AI technologies, such as cyberattacks and bio-weapons.
- The Office of Emergency Services will enforce compliance with the safety protocols mandated by SB 53.
SB 53 is a new AI safety law that addresses the need for transparency in AI safety measures. This law is unique because it aims to ensure safety without stifling technological innovation, contrary to prior regulations that faced strong opposition from tech companies. Future developments may include federal legislation that could affect state-level regulations on AI.
Read the original ↗
OpenAI’s new social app is filled with terrifying Sam Altman deepfakes
The launch of OpenAI's Sora app highlights the potential for misuse of realistic deepfake technology, raising concerns about disinformation and copyright violations in user-generated content.
- OpenAI's social app Sora allows users to create deepfakes, including realistic videos of CEO Sam Altman asking humorous questions.
- Sora uses an opt-out system for copyright content, reversing typical practices that require user consent for content use.
- The app features advanced video generation that enhances realism, making it easier to spread potentially misleading content across the internet.
OpenAI's Sora app is a new social media platform focused on video content creation through deepfakes, which involves using AI to create realistic representations of individuals. Concerns have emerged regarding its implications for copyright law and potential misuse for spreading disinformation. The platform gives users control over who can generate videos of their likeness, but raises ethical questions around privacy and safety.
Read the original ↗
OpenAI’s Sora 2 lets users insert themselves into AI videos with sound
The launch of Sora 2 enables users to create realistic AI-generated videos, raising potential issues regarding misuse and the ethics of deepfake technology in social media settings.
- OpenAI introduced Sora 2, a video-synthesis AI model that generates videos with synchronized sound effects, marking a first for the company.
- The new iOS social app allows users to insert themselves into AI-generated videos through a feature called 'cameos.'
- Sora 2 boasts improved visual realism and can simulate complex physical movements, enhancing video generation accuracy over the previous model.
- OpenAI has implemented safety measures and user controls to address potential misuse of the app's capabilities, particularly for teen users.
OpenAI is a leading AI research company known for developing advanced language and video models. Sora 2 follows the original Sora model and other recent competitors like Google's Veo 3 and Alibaba's Wan 2. As the technology evolves, concerns about ethical usage and content generation continue to grow.
Read the original ↗
Waymo can keep testing robotaxis in NYC until end of 2025
Waymo's extended permit for robotaxi testing in NYC highlights the regulatory challenges facing autonomous vehicle deployment in urban environments, impacting future mobility options.
- New York City regulators have extended Waymo's permit to test autonomous vehicle technology until the end of 2025.
- Under the permit, Waymo can operate up to eight Jaguar I-Pace vehicles with a human safety operator in Manhattan and Downtown Brooklyn.
- Waymo's ability to launch a commercial robotaxi service in NYC remains contingent on obtaining additional licenses from local authorities.
Waymo is an Alphabet-owned company focused on self-driving technology and currently operates commercial services in several U.S. cities. The company's presence in NYC has faced regulatory hurdles since 2021, including obtaining necessary permits for driverless operation.
Read the original ↗
Liquid AI Released LFM2-Audio-1.5B: An End-to-End Audio Foundation Model with Sub-100 ms Response Latency
The release of Liquid AI's LFM2-Audio-1.5B model could enhance applications in audio processing and responsiveness for AI systems.
- Liquid AI has released LFM2-Audio-1.5B, an end-to-end audio foundation model.
- The model features a response latency of under 100 milliseconds.
Liquid AI develops advanced AI models for various applications, including audio processing. The LFM2-Audio-1.5B model aims to improve the efficiency and speed of audio responses in AI technology.
Based on the headline only — the full article could not be fetched.
Read the original ↗
The Dragon Hatchling: The Missing Link between the Transformer and Models of the Brain
The development of BDH could significantly enhance the interpretability and performance of AI language models by simulating brain-like processes, impacting future research in artificial intelligence.
- BDH, a biologically inspired Large Language Model, merges scale-free network architecture with Hebbian learning to emulate brain processes.
- This model is designed to maintain the performance of Transformers while providing better interpretability and flexibility.
- BDH demonstrates performance comparable to GPT-2 in language and translation tasks, utilizing the same range of parameters and training data.
- The architecture uses spiking neurons and synaptic plasticity, reinforcing learning connections with specific language inputs.
- BDH's sparse activation vectors offer insights into its interpretability, distinguishing it from conventional neural network models.
The BDH, or Baby Dragon Hatchling, represents a new approach in AI that integrates biological principles from brain function with machine learning frameworks. Traditional Transformers, widely used in AI, rely on complex matrix operations that obscure their interpretability. Future developments may focus on refining BDH's performance and exploring its implications for machine learning.
Read the original ↗
Can today’s AI video models accurately model how the real world works?
Inconsistent performance in AI video models reveals significant challenges for their practical deployment in real-world applications, highlighting the need for further advancements in technology.
- Google's DeepMind tested its Veo 3 model on physical reasoning tasks, revealing both successes and notable inconsistencies.
- Veo 3 performed well on generating plausible videos of certain actions but failed in various other tasks, with significant error rates in trials.
- Only 16 out of 62 tasks tested showed failures across all trials, indicating a mix of capabilities even among tasks with some success rates.
Generative video models like Veo 3 aim to understand and represent real-world physical properties. Earlier models, such as Veo 2, showed more limitations in these capabilities. Ongoing research will focus on improving consistency and reliability in performance.
Read the original ↗
Sam Altman Announces Sora 2, a “ChatGPT for Creativity”
The introduction of Sora 2 could significantly impact digital content creation by streamlining the process and enhancing user creativity, potentially reshaping online interaction and storytelling.
- Sam Altman announced Sora 2, described as a 'ChatGPT for Creativity' designed to transform ideas into videos and art.
- Sora 2 features a cameo tool for users to insert themselves or friends into videos, enhancing content personalization.
- Safeguards are implemented to prevent bullying and misuse of likenesses in deepfakes, addressing potential risks associated with creative AI tools.
- Early users report Sora as intuitive and open, promoting collaboration and playful content creation.
Sora 2 is a new AI model from OpenAI aimed at enhancing creative projects through rapid video creation. This launch follows a trend in the evolution of AI-powered tools, which have seen greater acceptance among users. It will be important to monitor user engagement and how Sora 2 balances creativity with safety.
Read the original ↗
Creative machines and where AI meets imagination at TechCrunch Disrupt 2025
AI's role in creativity is poised to revolutionize the creative industries, impacting how stories are told, experiences are designed, and audiences are engaged.
- TechCrunch Disrupt 2025 will host discussions on the intersection of AI and creativity featuring leaders from TwelveLabs, Wonder Dynamics, and Pocket Entertainment.
- The event highlights how AI is reshaping storytelling and the creative process, demonstrating machines as active collaborators in arts and media.
- Founders Prateek Dixit, Nikola Todorovic, and Soyoung Lee will address innovations in audio storytelling, 3D character animation, and video search technologies.
TechCrunch Disrupt is an annual event that showcases emerging technologies and innovation in the startup world. This year's focus on AI and creativity underscores the growing trend of integrating AI technologies across creative sectors. Attendees can learn about advancements that are transforming the way stories and experiences are created.
Read the original ↗