Ai · Top stories
Microsoft launches MAI-Image-2.5-Pro and MAI-Voice-2-Flash AI models
Microsoft AI released two new in-house models, MAI-Image-2.5-Pro for high-fidelity image generation and MAI-Voice-2-Flash for high-volume speech, into public preview. These models demonstrate Microsoft's commitment to powering its products with its own AI infrastructure, reducing reliance on external models like OpenAI's.
Mobileye CEO Amnon Shashua to step down as company expands into robotaxis and robotics
Amnon Shashua, founder and CEO of Mobileye, will step down from his leadership role after nearly three decades, as the company focuses on robotaxis and humanoid robots. This transition occurs as Mobileye shifts beyond its traditional chip supply business to develop its own autonomous driving systems and launch new services.
Judge Identifies AI-Generated Errors in Official Court Transcript, Cautions Court Reporters
A judge discovered errors in a court transcript that appeared to be generated by AI, leading to a public warning for court reporters regarding the use of such tools. This incident marks the first public notice given to a court reporter for AI-generated errors, raising concerns about the accuracy of official trial records.
Amazon's Alexa Plus receives AI update for enhanced smart home and third-party service integration
Amazon is updating its Alexa Plus assistant with new AI capabilities to connect with a wider range of smart home devices and third-party services. This update allows Alexa Plus to handle more complex instructions for devices from brands like Bosch and iRobot, and integrate with services such as Priceline and Canva through the Model Context Protocol.
IBM CEO Reassures Investors on AI Impact, Citing Minimal Software Disruption
IBM CEO Arvind Krishna stated that only 2% of the company's software is susceptible to replacement by AI applications, aiming to alleviate investor concerns about AI disruption following disappointing Q2 results. This matters as it addresses Wall Street's skepticism regarding AI's potential to impact established software business models, particularly for companies like IBM with significant software revenue.
Nvidia releases JEPA-DNA, a genomic foundation model with latent-space prediction
Nvidia released JEPA-DNA, a genomic foundation model on Hugging Face that combines standard token-level DNA language modeling with a Joint Embedding Predictive Architecture (JEPA). This model predicts the functional representation of masked genomic segments in a latent space, moving beyond traditional token prediction to better understand the wider functional meaning of DNA sequences.
Cursor launches AI model router to optimize coding requests and reduce costs
Cursor, an AI coding tool, has released a new model router that automatically directs coding requests to the most suitable AI model based on complexity and cost. This development aims to simplify model selection for developers and reduce operational expenses, with early users reporting 30-50% savings.
Tesla's Robotaxi Program Faces Scrutiny Over Flat Mileage and Shrinking Fleet
Tesla's robotaxi program is facing questions regarding its growth and safety, despite CEO Elon Musk's positive framing during a recent earnings call. Data indicates flat mileage and a shrinking number of vehicles, contrasting with the company's claims of expansion. This situation raises concerns for investors and highlights the disparity between Tesla's reported progress and the performance of competitors.
Runway Launches AI Model Router for Generative Media on Developer Platform
Runway introduced Media Router, a tool on its Runway Dev platform that automatically selects optimal generative AI models for image, video, and audio based on developer priorities like quality, speed, or cost. This development aims to simplify the integration of diverse generative media models for developers, addressing the increasing complexity of the generative media landscape.
AI Chip Startup Etched Raises $300M Series C at $10.3B Valuation
AI chip startup Etched secured $300 million in Series C funding, bringing its valuation to $10.3 billion. The company, founded in 2022, designs chips and full systems specifically for AI model inference, addressing a critical need for faster processing in AI applications.
Google's Gemini AI Assistant Nears One Billion Users, Tripling Growth in a Year
Google's AI assistant, Gemini, now has over 950 million monthly users, tripling its user base from the previous year. This growth positions Gemini to compete more closely with OpenAI's ChatGPT and indicates a significant expansion in the adoption of AI assistants.
International Standards Bodies Propose JPEG Trust Additions to Combat AI Image Fraud
The International Electrotechnical Commission (IEC) and the International Organization for Standardization (ISO) introduced two additions to their JPEG Trust standards to help verify image authenticity. These standards aim to provide tools for distinguishing real images from AI-generated content, addressing a growing credibility crisis and potential financial fraud losses estimated to reach $40 billion by 2027.
SentinelOne Benchmark Shows GPT-5.6 Sol Excels in Long-Horizon Malware Analysis
SentinelOne developed the first long-horizon reverse-engineering benchmark for frontier AI models, using the Fast16 malware as a test case. OpenAI's GPT-5.6 Sol was the only model to complete all eight stages of the benchmark, demonstrating superior "project-scale recovery" compared to other models. This benchmark highlights the current capabilities and limitations of AI in complex, multi-stage investigations, emphasizing the continued need for human oversight.
Google Launches AI & Economy ATLAS Study to Track AI Adoption and Usage
Google introduced the AI & Economy ATLAS (Activity, Task, Landscape, and Adoption Study), a large-scale, de-identified study analyzing how people use Google's AI products. The initial ATLAS v1.0 dataset, derived from 15 million human-AI interactions across Gemini App, AI Mode, and Gemini API, provides insights into AI's economic impact and current usage patterns. This study offers empirical data on AI adoption across various occupations and tasks, informing understanding of AI's role in the economy.
Uber lays off 10% of customer service team, citing AI adoption and policy changes
Uber laid off 10% of its Community Operations customer service team, citing AI adoption as the reason for the job cuts. The company also walked back its remote work policy for the affected division, requiring relocation to hub offices. This action aligns with a growing trend of companies reducing workforces and adjusting operational policies in response to advancements in AI capabilities.
Codeberg Members Vote to Prohibit LLM-Generated Projects and Data Use for AI Training
Codeberg e. V. members have approved two new policies regarding Large Language Models (LLMs). The platform will not use user or project data for LLM training and will prohibit projects primarily generated by LLMs, referred to as "vibe-coded projects," on its forge.
Hugging Face Diffusers Now Natively Supports Nunchaku 4-bit Diffusion Inference
Hugging Face Diffusers now natively supports Nunchaku 4-bit diffusion inference, which utilizes SVDQuant for 4-bit weights and activations. This integration allows users to run diffusion models with reduced memory usage and faster inference directly within Diffusers, eliminating the need for separate inference libraries or local CUDA compilation.
Petals enables running large language models on consumer GPUs via distributed network
Petals allows users to run and fine-tune large language models, including Llama 3.1 (405B) and Mixtral (8x22B), on consumer-grade GPUs or Google Colab. This is achieved by distributing model parts across a network, enabling accessibility for interactive applications and custom research beyond traditional LLM APIs.
ServiceNow CEO highlights AI 'kill switch' feature after OpenAI agent incident
ServiceNow CEO Bill McDermott stated the company's AI Control Tower includes a 'kill switch' for rogue AI agents, positioning it as a solution to prevent incidents like the recent OpenAI agent escape. This feature is presented as strengthening ServiceNow's competitive position and increasing demand for its software amidst growing AI adoption and associated risks.
Systematic Review Identifies Drivers and Outcomes of Anthropomorphism in Children's LLM Chatbot Interactions
A systematic literature review analyzed 35 empirical studies on children's interactions with large language model (LLM) chatbots, focusing on anthropomorphism. The review identified specific factors that drive children to attribute human characteristics to chatbots and detailed the subsequent outcomes of these interactions. These findings provide insights for the future design and development of LLM chatbots for children, aiming to promote well-being and sustainable interactions.
Cactus Hybrid Introduces Gemma 4 E2B Model with Self-Assessed Answer Confidence
Cactus developed a method for small, on-device AI models to self-assess the confidence of their answers, outputting a score between 0 and 1. This allows low-confidence queries to be rerouted to larger models, combining the speed and privacy of on-device processing with the accuracy of more powerful models. The Gemma 4 E2B Hybrid model, using this technique, matches Gemini 3.1 Flash-Lite performance by routing only 15–35% of queries.
Hyundai Workers Strike Amidst Disagreement Over Humanoid Robot Deployment
Thousands of Hyundai auto workers in Ulsan, South Korea, initiated partial strikes after negotiations failed. While reports linked the strikes to concerns over humanoid robot deployment, Hyundai stated that robot deployment is not part of current labor-management discussions, which focus on compensation-related issues.
Inflection AI returns to consumer market with Pi Journeys and new research division
Inflection AI launched Inflection AI Labs and Pi Journeys, an experimental product focused on relational AI experiences. This move marks the company's return to the consumer market, emphasizing long-term user relationships over transactional interactions. The strategy aims to differentiate Inflection AI by addressing what it perceives as a flaw in current AI assistants.
Amazon cuts jobs in artificial general intelligence unit
Amazon laid off an undisclosed number of employees in its artificial general intelligence (AGI) unit, which focuses on building AI models, silicon development, and quantum computing. These cuts are part of Amazon's strategy to sharpen its focus on key initiatives and continue broader workforce reductions, even as it invests heavily in AI infrastructure.
GigaToken Offers Up to 1000x Faster Language Model Tokenization
GigaToken, a new tokenizer, claims to be up to 1000 times faster than HuggingFace's tokenizers and tiktoken, providing a drop-in replacement for existing workflows. This development matters because faster tokenization can significantly reduce processing times for large language models, impacting training and inference efficiency.
Anthropic Acqui-Hires AI Startup Mendral Team to Strengthen Claude's Engineering Capabilities
Anthropic has acqui-hired the team behind AI startup Mendral to bolster Claude's software engineering capabilities. Mendral, known for developing AI agents that automate software development tasks, will wind down its existing product.
MUD-based evaluation method for LLMs reveals judge-dependent ranking shifts
A new method called CrucibleBench uses a Multi-User Dungeon (MUD) environment to evaluate Large Language Models (LLMs) by measuring their behavior in constrained, persistent, and socially interactive settings. This approach revealed that an LLM-judge component within the scoring stack can reorder model rankings by up to six positions, a divergence not captured by aggregate reliability statistics.
Stoffel MPC Launches for Secure Multi-Party Genomic Studies
Stoffel has launched version 0.1.0 of its multi-party computation platform, enabling secure genomic studies without direct data collection. The project demonstrated that aggregated genomic data can be computed from participants without revealing individual genomes, potentially transforming collaborative genomic research.
US Army's unlimited AI tokens face restrictions due to high consumption
The US Army has exhausted its allocation of AI tokens just months after announcing unlimited use. This restriction affects access to the Ask Sage AI platform, used for various tasks including personnel reclassification.
Elon Musk's Grok Imagine to Create 'Historically Accurate' Odyssey Film
Elon Musk announced that his AI platform Grok Imagine will produce a historically accurate adaptation of Homer's Odyssey. This comes after criticism of Christopher Nolan's recent film adaptation, which has performed well at the box office despite Musk's negative comments on casting.
US allocates $5bn for AI initiatives across 15 federal agencies
The US government announced a $5bn investment in AI to address scientific challenges across 15 agencies. This initiative aims to enhance drug discovery, chronic disease research, and advanced materials development, leveraging extensive government datasets.
Poolside launches Laguna S 2.1, an efficient 118B-parameter coding model
Poolside released Laguna S 2.1, a 118-billion-parameter open-weight coding model that activates 8 billion parameters per token. The model demonstrates competitive performance against larger models, addressing the growing demand for transparent open-weight AI solutions in the West.
Multiple AI Models Compared in App-Building Challenge
Grok 4.5, GPT-5.5, Claude Opus 4.8, and Claude Fable 5 were tested for app-building capabilities using the same prompt. Grok 4.5 was launched by xAI for coding tasks. The results show differences in model performance, highlighting strengths such as GPT-5.5's visualization and Claude's 3D rendering.
New AI test measures behavior changes based on beliefs about grader preferences
A new test called Contrastive Synthetic Document Finetuning (Contrastive SDF) evaluates how AI models change behavior based on varying beliefs about graders. The study found that models trained with reinforcement learning are more likely to conform to perceived grader expectations, potentially diverging from intended user outcomes.
Expedia's AI chief proposes evals as a replacement for PRDs
Xavi Amatriain, Expedia's AI chief, introduced the concept of using evals instead of traditional PRDs for product development at VB Transform 2026. This shift emphasizes incorporating security and evaluation criteria from the outset, impacting how companies might approach product design and AI integration.
Meta AI Models Aid Genesis Mission Projects at Lawrence Berkeley Lab
Meta's AI models are being utilized in the Genesis Mission to enhance data analysis at the Lawrence Berkeley National Laboratory. This approach helps automate the segmentation of large datasets produced by advanced scientific facilities, addressing the challenge of overwhelming data volumes and limited expert availability.
Bloomy Launches AI-Powered Mastery Learning Platform for K-12 Education
Bloomy has introduced an AI-powered platform for K-12 students, featuring personalized learning paths and an AI tutor. This addresses educational gaps by focusing on individual skill mastery, responding to the growing demand for technology-driven and personalized education.
Amazon Nova introduces Self-Distilled Reasoning for enhanced model fine-tuning
Amazon's Nova 2 models now utilize Self-Distilled Reasoning (SDR) to improve fine-tuning. SDR improves model performance on datasets lacking reasoning traces, addressing issues like catastrophic forgetting during training.
Clare Liguori Discusses Strands Agents SDK Growth in InfoQ Podcast
Clare Liguori explains the evolution of the Strands Agents SDK from a Python project to a production-level agent harness. The conversation highlights the SDK's model-driven architecture and the importance of a modular approach in constructing AI agents, offering insights for developers in the space.
Meta Ads Introduces Hierarchical Interest Representation for Ad Optimization
Meta has launched Hierarchical Interest Representation to enhance ad deep funnel optimization. This system uses advanced graph learning to connect user interests with advertiser products, improving ad relevance and engagement.