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
With new in-house models, Microsoft lays the groundwork for independence from OpenAI
Microsoft's introduction of in-house AI models indicates a strategic shift towards reducing reliance on OpenAI, which could reshape the competitive landscape in AI development.
- Microsoft has trained new AI models, MAI-Voice-1 and MAI-1-preview, for internal use, signaling a move towards independence from OpenAI.
- MAI-Voice-1 is designed for high-fidelity speech generation, while MAI-1-preview is a large language model aimed at enhancing Copilot, Microsoft's AI chatbot.
- These models were developed to cater to specific use cases that OpenAI may not prioritize, highlighting a trend towards specialized AI solutions.
Microsoft has historically relied on OpenAI for its AI tools, utilizing their models in products like Copilot. The introduction of in-house models follows reports of Microsoft's interest in developing foundational models over the past year. The shift toward internal AI development reflects broader trends in the industry towards tailored solutions rather than general-purpose models.
Read the original ↗
How Intuit killed the chatbot crutch – and built an agentic AI playbook you can copy
Intuit's strategic pivot toward AI aims to enhance product efficiency, showcasing a transformative approach that could set standards in enterprise software development.
- Intuit launched Intuit Assist, a generative AI assistant, aiming for a major AI-driven update by September 2023, but its initial chat feature failed to meet expectations.
- After recognizing issues with the chat feature in QuickBooks, Intuit underwent a nine-month transformation to reinvent its AI strategy, eliminating ineffective workflows.
- The company has expanded its AI team significantly and restructured into small cross-functional teams focused on customer-centric rapid prototyping.
- Intuit's internal AI platform, GenOS, supports continuous development of AI agents, leveraging extensive domain-specific data for improved functionality in financial tools.
Intuit, known for financial products like QuickBooks and TurboTax, pivoted its strategy after initial AI ventures fell short. This transformation emphasizes building user-friendly AI solutions that integrate seamlessly into existing workflows.
Read the original ↗
Zuckerberg’s AI hires disrupt Meta with swift exits and threats to leave
The leadership turmoil at Meta threatens to undermine its ambitious AI goals and attract prominent talent from competing firms, impacting the company's competitive stance in the AI market.
- Shengjia Zhao, a key AI hire, threatened to quit Meta shortly after being named chief AI scientist, reflecting leadership instability.
- Several new AI staff, including former OpenAI researchers, have left or threatened to leave Meta within weeks of joining.
- Zuckerberg's restructuring of Meta's AI group aims to accelerate development, but has faced criticism for fostering internal competition and dissatisfaction among recruits.
Meta is undergoing its most significant leadership overhaul in 20 years, focusing efforts on developing artificial intelligence. The company has recently transitioned to a new team structure while also facing challenges in retaining key talent. Stakeholders are watching how these changes affect Meta's ability to innovate in AI.
Read the original ↗
Building and Optimizing Intelligent Machine Learning Pipelines with TPOT for Complete Automation and Performance Enhancement
The development of automated machine learning pipelines, as demonstrated with TPOT, enhances efficiency and effectiveness in deploying machine learning solutions across various applications.
- TPOT automates and optimizes machine learning pipelines using evolutionary algorithms and a pipeline design approach.
- The tutorial utilizes Google Colab for streamlined setup and reproducibility in building data models.
- Key aspects include defining custom scorers and using advanced models like XGBoost with stratified cross-validation strategies.
TPOT, or Tree-based Pipeline Optimization Tool, is an open-source library for automating machine learning processes. This tutorial provides a practical guide to using TPOT, allowing users to build and optimize pipelines effectively. Future developments in automated machine learning are likely to focus on efficiency and scalability across diverse datasets.
Read the original ↗
Step-Audio 2 Technical Report
The development of Step-Audio 2 signifies a breakthrough in audio processing, improving capabilities in speech recognition and understanding for real-world applications.
- Step-Audio 2 is an end-to-end multi-modal large language model designed for audio understanding and speech conversation.
- It integrates latent audio encoding and reinforcement learning, achieving state-of-the-art performance in automatic speech recognition and audio understanding.
- The model incorporates discrete audio token generation and retrieval-augmented generation to enhance conversation responsiveness.
- Trained on millions of hours of speech data, Step-Audio 2 aims to provide intelligence across diverse conversational scenarios.
Step-Audio 2 improves audio recognition technology, building on previous advancements in multi-modal language models. It utilizes millions of hours of data to enhance its performance. Future developments might focus on integrating more languages and improving interaction quality.
Read the original ↗
Host an event beyond the main event: Apply to host a Side Event at TechCrunch Disrupt 2025
Hosting a Side Event at TechCrunch Disrupt 2025 allows brands to engage directly with key industry players, enhancing visibility and networking opportunities.
- TechCrunch Disrupt 2025 will take place from October 27-29 at Moscone West in San Francisco, expecting over 10,000 attendees.
- Events can be held from October 25-31, encompassing various formats like dinners, workshops, or pitch-offs.
- There is no cost to apply for hosting a Side Event; selected hosts will receive promotional support to attract attendees.
TechCrunch Disrupt is an annual technology conference featuring founders, venture capitalists, and industry innovators. Side Events provide opportunities for smaller gatherings that complement the main event, fostering deeper connections. Interested parties should consider applying soon, as event participation is expected to be competitive.
Read the original ↗
Multi-View 3D Point Tracking
This research introduces a method that improves 3D point tracking efficiency, which could enhance applications in robotics, AR, and computer vision by reducing equipment costs and complexity.
- The new multi-view 3D point tracker uses a feed-forward model with transformers and k-nearest-neighbors for robust tracking.
- It performs well with as few as four cameras, unlike previous methods requiring over 20.
- The tracker achieves median trajectory errors of 3.1 cm on Panoptic Studio and 2.0 cm on DexYCB benchmarks.
- The model has been trained on 5,000 synthetic multi-view sequences, showcasing strong generalization to diverse camera setups.
- Provided datasets and models aim to establish a new standard in multi-view 3D tracking research.
MVTracker is the first data-driven system designed to track arbitrary 3D points across multiple cameras. Previous methods often struggle with depth and require significant optimization resources. Future developments in multi-view tracking might build on this approach by integrating more robust features or expanding the model's usability in real-world scenarios.
Read the original ↗
Billionaire Ambani taps Google, Meta to build India’s AI backbone
The development of a national AI infrastructure in India, supported by major tech firms, could accelerate the country's global competitiveness in artificial intelligence.
- Mukesh Ambani launched the new subsidiary Reliance Intelligence to develop India's AI backbone.
- Reliance has formed partnerships with Google Cloud and Meta to enhance AI infrastructure in India.
- Meta and Reliance's joint venture involves a combined investment of ₹8.55 billion (approximately $100 million) to create enterprise AI solutions.
Reliance Industries, led by billionaire Mukesh Ambani, is aiming to establish a comprehensive AI infrastructure with global partners. This initiative is part of a broader strategy to strengthen India's position in the AI sector amidst competition from the U.S. and China.
Read the original ↗
Navy Tests Hydrogen Powered Balloon and Drone Systems to Enhance Over-the-Horizon Surveillance
These tests could enhance the U.S. Navy's surveillance capabilities, reduce reliance on traditional fuel supplies, and improve logistical efficiency in contested maritime environments.
- The U.S. Navy is testing hydrogen-powered balloons and drones for over-the-horizon surveillance.
- These systems aim to improve logistics and extend operational range in remote areas.
- Demonstrations were conducted by the Naval Research Laboratory in California this past spring.
The U.S. Navy is exploring hydrogen as a power source for advanced surveillance systems due to the logistical challenges of traditional fuels. The combination of hydrogen-filled balloons and drones is aimed at enhancing maritime monitoring and data collection capabilities in contested environments.
Read the original ↗
Marketing AI boom faces crisis of consumer trust
The erosion of consumer trust in marketing AI could hamper its growth and acceptance across industries, impacting how businesses engage with audiences.
- The marketing AI sector is experiencing a crisis of consumer trust.
- Concerns about data privacy and ethical use of AI are contributing to this decline.
Marketing AI refers to artificial intelligence technologies used to analyze consumer data and optimize marketing strategies. The industry's growth has been rapid, but consumer skepticism is rising. Monitoring how companies address these trust issues will be crucial moving forward.
Based on the headline only — the full article could not be fetched.
Read the original ↗
Google warns that mass data theft hitting Salesloft AI agent has grown bigger
The broader scope of the data breach affecting the Salesloft Drift AI agent raises significant security concerns for companies relying on third-party integrations for sensitive data management.
- Google warns that all Salesloft Drift security tokens may be compromised due to unauthorized access to Google Workspace email accounts.
- Following the breach, Google revoked the affected tokens and disabled integration between Salesloft Drift and all Workspace accounts.
- Initially, the breach was thought to be limited to Salesforce integrations, but further investigation revealed a wider impact.
- Salesloft has engaged Mandiant, a Google-owned incident response service, to assist in investigating the breach.
Salesloft Drift is an AI-powered chat agent that allows websites to interact with customers in real-time. The platform was acquired by Salesloft 18 months ago and can integrate with various services, including Salesforce and Google Workspace. Users are urged to review and secure their third-party integrations to mitigate risks from the breach.
Read the original ↗
Finetune LLM Agents without Finetuning LLMs! Thanks to Memento!
The Memento framework addresses the challenge of ongoing AI learning, potentially transforming AI development by allowing agents to adapt without the high costs of fine-tuning.
- Memento is a new framework allowing LLM agents to learn continuously using an external Case Bank, avoiding costly fine-tuning.
- The framework enables agents to store past experiences, retrieve similar cases for new tasks, and adapt strategies accordingly.
- Memento achieved the top score of 79.4% on the GAIA benchmark, demonstrating its effectiveness over traditional training-heavy systems.
Memento provides a memory-driven solution for large language model (LLM) agents, allowing them to learn from experiences rather than requiring constant fine-tuning. This approach represents a significant shift in AI agent development, emphasizing adaptability and efficiency. The framework promises to enhance the capabilities of AI agents across various tasks.
Read the original ↗
The State of Voice AI in 2025: Trends, Breakthroughs, and Market Leaders
The rapid advancement and widespread adoption of voice AI technology are transforming business processes and consumer interactions, highlighting the critical role this technology plays in shaping the future of commerce and healthcare.
- Voice AI is projected to grow from $3.14 billion in 2024 to $47.5 billion by 2034, with a 34.8% compound annual growth rate (CAGR).
- The intelligent virtual assistant segment is expected to reach $27.9 billion in 2025, up from $20.7 billion in 2024.
- Healthcare organizations report that 70% credit voice AI with improved operational outcomes, highlighting its impact across multiple sectors.
- Modern voice AI can detect emotional cues from speech, allowing agents to adapt their responses and enhance user satisfaction.
- Smartphones dominate voice AI interactions, with 91% of users preferring mobile apps for voice engagement.
Voice AI technology encompasses systems that enable interaction through spoken language. The market has expanded significantly due to advances in speech recognition and natural language understanding, shifting from basic command systems to more complex applications in various fields. Continued developments in this sector will likely focus on privacy and regulatory issues as well as improving emotional intelligence in AI systems.
Read the original ↗
How to Cut Your AI Training Bill by 80%? Oxford’s New Optimizer Delivers 7.5x Faster Training by Optimizing How a Model Learns
By significantly reducing the cost and time of AI model training, Oxford's Fisher-Orthogonal Projection optimizer has the potential to transform AI research and development economics.
- Oxford researchers developed the Fisher-Orthogonal Projection (FOP) optimizer, which can speed up AI training by 7.5 times and potentially reduce costs by up to 87%.
- FOP maintains intra-batch gradient variance, enabling faster and more stable convergence even at large batch sizes, unlike existing optimizers that struggle.
- The optimizer has demonstrated improved accuracy on datasets like ImageNet-1K and CIFAR-10, outperforming traditional methods such as SGD and AdamW in both speed and stability.
- Practitioners can easily implement FOP in existing workflows using PyTorch with minimal adjustments, facilitating faster experimentation and model development.
AI model training often requires significant computational resources, leading to high costs that can hinder innovation. The introduction of FOP offers a new approach that leverages previously discarded data signals to enhance training efficiency. Future developments may focus on real-world applications and further improvements in optimizer technology.
Read the original ↗
How AI and Automation Are Transforming Energy Billing Systems for Businesses
Businesses are increasingly adopting AI and automation in their energy billing systems, which may lead to more efficient operations and cost savings.
- AI and automation are transforming energy billing systems for businesses.
- These technologies are expected to streamline processes and improve accuracy in billing.
Energy billing systems are crucial for businesses to manage their energy consumption and costs. The integration of AI and automation into these systems signifies a shift towards more innovative and efficient solutions. Future developments should focus on the impact on overall energy management strategies.
Based on the headline only — the full article could not be fetched.
Read the original ↗
Casino Mastery in the Age of AI: What’s Coming Next
As AI transforms gaming strategies, understanding its impact on casino operations is essential for stakeholders to adapt and remain competitive in the industry.
- The article discusses the evolving role of AI in casinos.
- It focuses on upcoming changes and innovations in gaming technology prompted by AI advancements.
AI is increasingly being integrated into various sectors, including gaming and casinos, to enhance strategies and user experiences. Stakeholders should monitor these changes to adapt effectively to the evolving market.
Based on the headline only — the full article could not be fetched.
Read the original ↗
Vocal Image is using AI to help people communicate better
This story highlights how AI is being utilized to enhance communication skills, providing accessible coaching resources to a broad audience and supporting the personal development of individuals, including marginalized groups.
- Vocal Image, an Estonia-based startup, has reached 4 million app downloads and aims to improve voice and communication skills using AI-powered coaching.
- The subscription-based app was founded by Nick Lahoika, who overcame personal speaking challenges, and offers an interactive library of exercises and feedback.
- Vocal Image secured a $3.6 million seed round and has increased its annual recurring revenue to $12 million with 50,000 paid users.
- The startup utilizes community-contributed voice samples to enhance its AI accuracy, accumulating over 1 million real-voice samples through user engagement.
Vocal Image, originally a YouTube channel, has evolved into a voice coaching app that combines personal experience with AI technology for communication improvement. Founded by Belarusian exiles, it has quickly grown and garnered investment, positioning itself in a competitive edtech market. Future developments may include expanded features and localization for diverse user needs.
Read the original ↗
CogVLA: Cognition-Aligned Vision-Language-Action Model via Instruction-Driven Routing & Sparsification
CogVLA significantly reduces computational costs in Vision-Language-Action models, enhancing their efficiency and scalability, which is crucial for real-world applications in robotics and AI.
- CogVLA is a new framework designed to improve efficiency and performance in Vision-Language-Action tasks through instruction-driven routing and sparsification.
- The model achieves state-of-the-art results, with success rates of 97.4% on the LIBERO benchmark and 70.0% in real-world robotic tasks.
- CogVLA reduces training costs by 2.5-fold and decreases inference latency by 2.8-fold compared to OpenVLA.
Vision-Language-Action models integrate visual and language processing to enable machines to understand and perform tasks. Traditional models require extensive training, leading to high computational demands. With CogVLA's advancements, future models may see broader deployment and better performance outcomes.
Read the original ↗
OpenAI Releases an Advanced Speech-to-Speech Model and New Realtime API Capabilities including MCP Server Support, Image Input, and SIP Phone Calling Support
OpenAI’s latest Realtime API and speech-to-speech model could significantly enhance enterprise communication by integrating advanced voice AI capabilities with traditional telephone infrastructure.
- OpenAI has launched the Realtime API and GPT-Realtime, marking advancements in speech-to-speech technology and exiting beta with new enterprise features.
- GPT-Realtime processes audio directly through a single system, improving performance metrics by up to 26% over previous models in accuracy tests.
- The API now supports SIP phone connections and asynchronous function calling, enabling smoother interactions in business applications without interrupting conversation flow.
- Pricing has been set at $32 per million audio input tokens and $64 per million output tokens, representing a 20% reduction from the prior model's costs.
OpenAI is a leading AI research organization known for developing advanced models like GPT for various applications. The new Realtime API showcases a shift towards unified voice processing, crucial for enterprise-level deployment. Future developments will focus on addressing ongoing challenges in accuracy and contextual understanding in diverse environments.
Read the original ↗
Performance Prediction for Large Systems via Text-to-Text Regression
This advancement in text-to-text regression technology could significantly improve predictive modeling for large systems, leading to better resource efficiency and performance across various industries.
- A text-to-text regression model predicts resource efficiency for Google's Borg system with high accuracy, achieving near perfect rank correlation and significantly lower mean squared error than traditional tabular methods.
- The model, an encoder-decoder with 60 million parameters, adapts to new tasks effectively, needing only 500 examples for training.
- Ablation studies emphasize the importance of encoders, increased sequence length, and the model's ability to quantify uncertainty in predictions.
Text-to-text regression models convert complex inputs into a textual format for predictive tasks, making them versatile for various data types, including system logs and configuration files. Traditional methods often struggle with such complex data, highlighting the significance of this new approach. Future developments may further explore the application of this technology in diverse large-scale systems.
Read the original ↗