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
Evaluating Deep Learning Models with Custom Loss Functions and Cali...
Implementing custom loss functions enhances model performance by aligning training with specific business goals, particularly in critical applications like fraud detection and medical diagnosis.
- Custom loss functions can address the limitations of traditional evaluation metrics such as accuracy and cross-entropy in deep learning models.
- These functions can optimize for specific goals, such as penalizing false negatives more than false positives, which is crucial in scenarios like fraud detection.
- Calibration techniques are needed to ensure that models provide reliable probability estimates, which is vital for decision-making processes.
- Overconfidence in model predictions can lead to significant misclassifications and risks, particularly in high-stakes environments.
Deep learning models are often evaluated on conventional metrics, which do not capture the complexities of real-world applications. Custom loss functions can tailor training processes to better reflect actual utility and business objectives. As machine learning continues to advance, improvements in model calibration and evaluation methods will be critical for achieving reliable outcomes.
Read the original ↗
Is GPT-5 really worse than GPT-4o? Ars puts them to the test.
The comparison of GPT-5 and GPT-4o illustrates the challenges faced by AI models in meeting user expectations and highlights the complexities of evaluating AI performance.
- OpenAI's GPT-5 has received user complaints about its tone, creativity, and accuracy, prompting the company to reintroduce the earlier GPT-4o model as an option.
- Ars Technica conducted tests on both models using various prompts to compare their performance and response quality.
- While GPT-5 often provided more detailed and nuanced answers, GPT-4o demonstrated strengths in clarity and directness in certain contexts.
OpenAI's GPT models are widely used for natural language generation, with GPT-5 being the latest iteration following GPT-4o. The user feedback and testing results indicate evolving standards for AI capabilities and user satisfaction. Users should monitor updates and refinements in future AI model releases.
Read the original ↗
ChatGPT’s mobile app has generated $2B to date, earns $2.91 per install
ChatGPT's mobile app revenue demonstrates its strong market dominance in consumer AI applications, illustrating the significant financial potential in this rapidly growing sector.
- ChatGPT's mobile app has generated $2 billion in global consumer spending since its launch in May 2023.
- The app has made $1.35 billion in 2025, a 673% increase from $174 million in the same period in 2024.
- ChatGPT's app has been installed approximately 690 million times globally, significantly surpassing competitors like Grok, which has 39.5 million installs.
- In the U.S., consumers have spent an average of $10 per ChatGPT app download, accounting for 38% of its total revenue.
- The app is currently seeing about 45 million downloads per month, an increase of 180% compared to earlier in 2024.
ChatGPT is a conversational AI application developed by OpenAI, available on iOS and Android platforms. Its rapid revenue growth outpaces competitors in mobile consumer spending, highlighting its strong position in the AI market. Future developments to watch include further enhancements in AI capabilities and additional monetization strategies.
Read the original ↗
STream3R: Scalable Sequential 3D Reconstruction with Causal Transformer
STream3R addresses limitations in current 3D reconstruction methods, offering a more efficient and effective approach that could enhance real-time 3D perception applications.
- STream3R reformulates 3D reconstruction as a decoder-only Transformer problem, utilizing causal attention to process image sequences more efficiently.
- The method outperforms state-of-the-art techniques in static and dynamic scene benchmarks, demonstrating better generalization in challenging scenarios.
- STream3R's design allows for compatibility with large language model training methods, facilitating large-scale pretraining and fine-tuning for 3D tasks.
3D reconstruction involves creating a three-dimensional model from images; traditional methods often require complex optimization. STream3R introduces a novel approach based on Transformer architecture, signaling a significant shift in how 3D tasks can be performed. Future developments may focus on further enhancing real-time capabilities in various streaming environments.
Read the original ↗
ToonComposer: Streamlining Cartoon Production with Generative Post-Keyframing
ToonComposer enhances cartoon production efficiency and quality, potentially transforming the creative workflows of artists in the animation industry.
- ToonComposer is a generative model that merges inbetweening and colorization into a single post-keyframing stage.
- It utilizes sparse sketches for fine control, requiring only one sketch and a colored reference frame.
- The model has been evaluated using the PKBench benchmark, showing improvements in visual quality, motion consistency, and production efficiency over existing methods.
ToonComposer addresses inefficiencies in traditional cartoon production, which typically involves fragmented stages like keyframing and colorization. By integrating these stages, ToonComposer aims to facilitate artists' workflows and reduce manual labor. This innovation is significant as it leverages recent AI advancements to enhance creative processes in animation.
Read the original ↗
Artificial Intelligence (346)
The ongoing exploration of AI agents in data science could significantly reshape workflows and decision-making processes in the field.
- The article discusses the use of AI agents by a data scientist in 2025.
- It highlights the evolving role of AI in data analysis and decision-making.
AI agents are software programs designed to automate tasks and assist users in various domains, including data science. The concept has gained traction as AI technology has advanced, enhancing productivity and efficiency. Observers should monitor how these technologies will influence data workflows and outcomes.
Based on the headline only — the full article could not be fetched.
Read the original ↗
NVIDIA aims to solve AI’s issues with many languages
Resolving AI's challenges with multiple languages could enhance communication and accessibility in global applications, impacting various industries.
- NVIDIA is addressing issues related to AI's capability to process multiple languages.
- The company's focus is aimed at improving AI's performance across diverse linguistic contexts.
NVIDIA is a technology company known for its work in graphics processing units (GPUs) and AI systems. The company has previously developed technologies to enhance AI capabilities. Observers should monitor NVIDIA's advancements in language processing technology going forward.
Based on the headline only — the full article could not be fetched.
Read the original ↗
Pass@k Training for Adaptively Balancing Exploration and Exploitation of Large Reasoning Models
This research highlights a novel approach in reinforcement learning that could enhance the capabilities of large reasoning models and improve their exploration efficiency.
- Using Pass@k as a reward in reinforcement learning with verifiable rewards can improve the balance between exploration and exploitation.
- Pass@k Training enhances the exploration abilities of large reasoning models, outperforming existing models like GPT-4o and Claude-3.7.
- The study derives an analytical solution for the advantages of using Pass@k, suggesting it can lead to more efficient training processes.
Reinforcement learning with verifiable rewards (RLVR) often struggles with balancing exploration, which involves trying new strategies, and exploitation, which focuses on optimizing known strategies. This research introduces Pass@k as a reward mechanism, emphasizing its potential to improve the exploration capabilities of large reasoning models. Future developments will likely focus on refining advantage designs within RLVR.
Read the original ↗
NextStep-1: Toward Autoregressive Image Generation with Continuous Tokens at Scale
NextStep-1's capabilities advance the field of image generation, promising improved accuracy and efficiency in creative applications across industries.
- NextStep-1 is a 14 billion parameter autoregressive model designed for text-to-image generation and image editing.
- It utilizes discrete text tokens and continuous image tokens to achieve high-fidelity results without the quantization loss seen in other models.
- The model demonstrates state-of-the-art performance in text-to-image tasks and will be made available to the research community.
NextStep-1 represents a new approach in autoregressive models for image generation, allowing for improved processing methods. Previous models typically relied on computationally heavy techniques or vector quantization, which could degrade image quality. Future developments may focus on further enhancing its capabilities and accessibility in creative fields.
Read the original ↗
How AI will Change the iGaming Industry in 2026
Understanding how AI will transform the iGaming industry is crucial for businesses and players, as it will shape future gaming experiences and operational efficiencies.
- AI technology is projected to significantly impact the iGaming industry by 2026.
- The transformation will include enhancements in player experience and operational processes.
The iGaming industry encompasses online gaming and betting, where technology plays a key role in shaping user experiences. AI's integration is expected to revolutionize various aspects of this sector, improving personalization and security.
Based on the headline only — the full article could not be fetched.
Read the original ↗
Top 10 Free AI Essay Writers
AI essay writing tools are becoming significant resources for students, enabling them to efficiently produce quality work despite increasing academic pressures.
- A list of the top 10 free AI essay writers has been compiled, highlighting their features and suitability for academic use.
- Litero is identified as a top choice, specifically designed for academic integrity and structured writing assistance.
- The article discusses various tools, outlining their strengths, such as grammar correction, idea generation, and citation support.
As academic demands increase, AI tools designed for writing help students manage their workloads more effectively. Previous writing aids often lacked the academic focus found in today's specialized platforms, like Litero. Future considerations include how these tools evolve to align with educational standards while maintaining academic integrity.
Read the original ↗
Human-in-the-loop work drives AI powering Alibaba’s smart glasses
The development of AI for Alibaba's smart glasses showcases the integration of human feedback in enhancing AI capabilities, which may influence future technology advancements.
- Alibaba's smart glasses utilize AI powered by human-in-the-loop work.
- This approach enhances the functionality and user experience of the smart glasses.
Alibaba is a major technology company known for its e-commerce platform and innovations in AI. Human-in-the-loop refers to systems where human input is used to improve machine learning models. Future developments may focus on expanding the capabilities of such technologies in consumer devices.
Based on the headline only — the full article could not be fetched.
Read the original ↗
Dynamic Fine-Tuning (DFT): Bridging the Generalization Gap in Supervised Fine-Tuning (SFT) for LLMs
Dynamic Fine-Tuning (DFT) addresses the generalization gap in Supervised Fine-Tuning (SFT) for large language models (LLMs), impacting their performance and application in various tasks.
- Dynamic Fine-Tuning (DFT) is introduced as a method to improve generalization during the training of large language models (LLMs).
- The method aims to bridge the performance gap that may occur in Supervised Fine-Tuning (SFT) for LLMs.
Supervised Fine-Tuning (SFT) is widely used to refine the capabilities of large language models (LLMs) after their initial training. Previous methods have struggled with generalization, leading to decreased model effectiveness across new tasks. Observers will look for how DFT may enhance LLM applications in real-world scenarios.
Based on the headline only — the full article could not be fetched.
Read the original ↗
Europe’s Top AI Models of 2025: Multilingual, Open, and Enterprise-Ready
The development of multilingual, open, and enterprise-ready AI models in Europe will enhance technological accessibility and capabilities across various industries.
- Europe is advancing towards its top AI models expected in 2025.
- These AI models will feature multilingual capabilities, making them usable in multiple languages.
Artificial intelligence is rapidly evolving, with significant investments and innovations occurring globally. The focus on multilingual and open models reflects a shift towards inclusivity and enterprise integration in AI technology. Monitoring the progression of these models will be essential as they may drive future AI applications.
Based on the headline only — the full article could not be fetched.
Read the original ↗
Transformer Explainer: Interactive Learning of Text-Generative Models
The Transformer Explainer tool enhances understanding of complex AI models, making advanced generative technologies more accessible to non-experts and promoting broader educational outreach.
- Transformer Explainer is an interactive visualization tool that demonstrates the workings of the GPT-2 model in a web browser.
- Users can experiment with their own inputs and observe real-time predictions from the model's internal components.
- The tool requires no installation and thus is accessible to a wider audience interested in generative AI techniques.
Transformers, particularly models like GPT-2, have significantly impacted machine learning and AI. However, these models are often challenging for non-experts to grasp due to their complexity. The Transformer Explainer tool aims to democratize education on these technologies by providing a user-friendly interface for learning.
Read the original ↗
Efficient AI Agents Don’t Have to Be Expensive: Here’s Proof
The potential for more affordable efficient AI agents could broaden access to advanced technologies, impacting various industries and applications.
- Efficient AI agents can be developed without high costs.
- The article presents evidence supporting the affordability of these AI solutions.
AI agents are software programs that use artificial intelligence to perform tasks and make decisions. The trend toward cost-effective AI solutions may influence market dynamics and innovation. Observers should watch for developments in affordable AI technologies and their applications.
Based on the headline only — the full article could not be fetched.
Read the original ↗
We-Math 2.0: A Versatile MathBook System for Incentivizing Visual Mathematical Reasoning
We-Math 2.0's enhancements for Multimodal Large Language Models could significantly improve mathematical reasoning capabilities, impacting fields that rely on advanced computational methods and education.
- We-Math 2.0 integrates a structured mathematical knowledge system and model-centric data space modeling to enhance MLLMs' mathematical reasoning abilities.
- The system features a five-level hierarchical knowledge structure comprising 491 points and 1,819 principles.
- It includes datasets MathBook-Standard and MathBook-Pro, which offer diverse challenges for training MLLMs across different difficulty levels.
- A reinforcement learning framework, MathBook-RL, utilizes two-stage fine-tuning to improve reasoning alignment and performance on benchmarks.
- MathBookEval serves as a comprehensive benchmark assessing reasoning capabilities across all 491 knowledge points.
We-Math 2.0 is designed to address the inadequacies of MLLMs in performing complex mathematical reasoning tasks. Previous methods have emphasized dataset construction without adequately considering structured knowledge systems. This new approach aims to provide a foundational framework for future advancements in AI-driven education.
Read the original ↗
That 'cheap' open-source AI model is actually burning through your compute budget
The research indicates that the cost advantages of open-source AI models may be overstated, as they can consume far more resources than closed alternatives, impacting enterprise budgeting.
- Open-source AI models use 1.5 to 4 times more tokens than closed models, with some models using up to 10 times more for basic knowledge questions.
- The inefficiency is especially acute in Large Reasoning Models, which require extensive computation for simple tasks.
- Closed source models tend to optimize for efficiency, which can offset higher API costs associated with their use.
Open-source AI models are often perceived as cost-effective alternatives to proprietary models. However, a new study reveals significant inefficiencies in resource usage that challenge this belief, particularly for common tasks.
Read the original ↗
ASTRA: Autonomous Spatial-Temporal Red-teaming for AI Software Assistants
ASTRA's capability to discover safety flaws in AI coding assistants is crucial for enhancing the security of software development, especially in high-risk areas like cybersecurity.
- ASTRA is an automated agent system designed to uncover safety flaws in AI-driven code generation.
- It builds knowledge graphs to model software tasks and vulnerabilities, enhancing its vulnerability exploration process.
- In real-world evaluations, ASTRA identified 11-66% more issues than existing safety techniques.
- ASTRA won the Amazon Nova AI Challenge by outperforming elite teams in an adversarial evaluation.
ASTRA addresses the safety concerns surrounding AI coding assistants, which are rapidly being adopted in software development. Traditional red-teaming methods often fail to simulate realistic scenarios, making ASTRA's results significant. Future developments in AI safety tools will likely be informed by ASTRA's findings.
Read the original ↗
Qwen2.5-Omni Technical Report
The development of Qwen2.5-Omni signifies advancements in multimodal processing, enabling improved performance across multiple modalities such as text, audio, and video, which could enhance various applications in AI.
- Qwen2.5-Omni is a multimodal model capable of processing text, images, audio, and video in real-time, achieving state-of-the-art results in multimodal benchmarks.
- It uses a dual-track architecture called Thinker-Talker for simultaneous text and audio generation, reducing interference between the two outputs.
- The model introduces innovative techniques like block-wise processing for encoders and TMRoPE for aligning audio and video timestamps.
- Qwen2.5-Omni demonstrates comparable performance to the Qwen2.5-VL model and outperforms the Qwen2-Audio model, especially in speech instruction following.
- Its streaming Talker component surpasses most existing models in the naturalness of audio generation.
Qwen2.5-Omni is developed as an end-to-end model designed to handle diverse input types and generate outputs effectively. This development follows previous breakthroughs in multimodal AI models, focusing on integrating various sensory inputs. Future trends may involve further enhancements in model universality and application in real-time scenarios.
Read the original ↗