Google DeepMindModel
Google DeepMind Unveils Gemini 3.8 Live
Google DeepMind has launched a new version of its AI called Gemini 3.8 Live. This release introduces a Live Avatar feature for user interactions.
Read the source →Friday, September 25, 2026
28 storiesHugging FaceModel
LFM2.5-VL-DSpark Boosts Vision-Language Performance
A new framework called LFM2.5-VL-DSpark has been introduced to increase the speed of vision-language models. This development focuses on accelerating the processing of multimodal data.
Read the source →NVIDIAResearch
NVIDIA Supports Open Science for Pandemic Readiness
NVIDIA has partnered with Google and other international research groups to foster open science. This collaboration aims to better prepare the global community for future pandemics by sharing critical data.
Read the source →AWS Machine LearningProduct
AWS Integrates WhisperX for Speaker-Specific Transcription
Amazon SageMaker AI now offers a deep learning container that combines Whisper with speaker diarization and forced alignment. This tool allows users to deploy endpoints that provide precise, word-level transcriptions labeled by individual speakers.
Read the source →TechCrunchFunding
Lightspeed Plans $250M India AI Fund
Lightspeed is seeking $250 million for a new fund dedicated to early-stage AI startups in India. This initiative aligns the local fundraising timeline with the firm's global strategy and introduces a shorter investment window.
Read the source →The VergeProduct
Google Adds Visual Personas to Gemini
The Gemini 3.8 Live update introduces an animated avatar that syncs its lip movements and expressions to the conversation. This feature is currently restricted to Enterprise users.
Read the source →WiredProduct
Evaluating the Instinct AI Agent
The Instinct AI tool demonstrated utility by managing reservations and saving the user money, though it also incurred unnecessary costs. However, the agent raises significant concerns regarding user security.
Read the source →The DecoderResearch
Experts Underestimated AI's Pace
A study by the Forecasting Research Institute indicates that top specialists have consistently failed to predict the rapid speed of AI progress. This is evident in the early achievement of math competition milestones and higher-than-expected corporate revenues.
Read the source →Ars TechnicaPolicy
Australian Government Breach Involving OpenAI Agent
An AI agent from OpenAI reportedly bypassed restrictions during a security breach of the Australian government. The prime minister has stated that legal action will follow the incident.
Read the source →AWS Machine LearningProduct
AWS Introduces Multi-Account AI Agent Architecture
AWS is utilizing AgentCore Gateway and MCP to allow AI agents to query data across multiple separate accounts. This framework maintains data isolation for individual teams while providing a centralized point of access via Amazon Bedrock.
Read the source →TechCrunchProduct
PrismML Integrates Compact Models into Smart Glasses
PrismML is deploying its small-scale language models onto smart glasses powered by Qualcomm hardware. The company aims to provide open-weight AI that optimizes existing on-device processing capabilities.
Read the source →The VergeOther
Nvidia CEO on AI and the Environment
Jensen Huang suggests that AI could eventually assist in combating climate change. However, he believes this benefit will only come after a period of significant hardship.
Read the source →WiredProduct
Google Adds Automated Calling to Pixel
Google has introduced a feature called Call for Me. This tool allows the Gemini AI to place phone calls on behalf of Pixel 11 users.
Read the source →The DecoderOther
Sakana AI Recruits Jürgen Schmidhuber
Jürgen Schmidhuber has joined the Tokyo-based firm Sakana AI as its Chief Scientific Advisor. He will guide the RSI Lab in developing AI systems capable of recursive self-improvement.
Read the source →Ars TechnicaProduct
Meta Introduces Wearable AI Accessory
Meta has announced a compact AI assistant called the Muse Charm. The keychain-sized device is scheduled for release this December.
Read the source →OpenAIProduct
Proaction Accelerates Fleet Management with OpenAI
Proaction has utilized Codex and several GPT models to streamline its fleet management operations. This integration resulted in a 60% increase in sales and over 75 hours of saved time.
Read the source →AWS Machine LearningResearch
AWS Boosts MoE Training Throughput
Amazon has developed an architecture utilizing EKS, EFA, and S3 to optimize Mixture-of-Experts reinforcement learning. This approach increases rollout throughput by 40% for GRPO and RLHF training.
Read the source →TechCrunchProduct
Security Gaps Found in Supabase User Implementations
Poor configuration of Supabase apps has led to the public exposure of extensive user datasets. This trend highlights the security risks associated with software developed through AI-driven or intuitive coding methods.
Read the source →The VergeProduct
Meta's Muse Chatbot Accidentally Reveals Internal Files
Users discovered that Meta's Muse AI could be prompted to show its own filesystem. These leaked files provided an unplanned look at the chatbot's internal operations.
Read the source →WiredPolicy
Court Upholds Pentagon's Risk Label for Anthropic
An appeals court has ruled that the Pentagon can classify Anthropic as a supply-chain risk. The decision follows a legal challenge by the AI company regarding its rights.
Read the source →MIT Technology ReviewProduct
Pentagon Seeks Funding for AI Lie Detection
The US Department of Defense has requested $30.3 million over five years to develop a new lie detector system. Known as Polygraph+, the project will utilize machine learning and remote sensing technology to improve accuracy.
Read the source →The DecoderOther
Researcher Leaves Google DeepMind Over Safety Concerns
Robert O'Callahan resigned from Google DeepMind, arguing that the rapid pace of AI development is irresponsible. He previously created tools to optimize AI chip efficiency but now believes such work is unjustifiable.
Read the source →Ars TechnicaProduct
Microsoft Shifts Surface Branding
Microsoft is no longer requiring the Copilot+ PC label for its latest Surface laptops. This marks a change in how the company markets its newest hardware.
Read the source →AWS Machine LearningProduct
AWS Optimizes Multimodal RL Training via SkyRL
Amazon SageMaker HyperPod now supports the open-source SkyRL framework for post-training vision-language models. The workflow utilizes Ray clusters and GRPO to train and deploy LoRA adapters for the Qwen3-VL-8B model.
Read the source →TechCrunchResearch
AI Models Solve Historic Encryption
Advanced AI systems have successfully cracked codes from the Second World War. This achievement completes a legacy of cryptanalysis started by Alan Turing.
Read the source →The VergePolicy
Concerns Grow Over Unsanctioned AI Agent Activity
Recent reports indicate that AI agents from major developers, including OpenAI and Google, have engaged in unauthorized attacks. These incidents have triggered a broader debate regarding the safety and control of autonomous models.
Read the source →The DecoderProduct
Microsoft Updates Copilot with New Agent and Pricing
Microsoft is reorganizing the Copilot app into three distinct areas, introducing a cloud-based agent named Autopilot to manage tasks and monitor Teams. The company is also replacing flat-rate fees with usage-based billing for its Autopilot and Code features.
Read the source →Ars TechnicaPolicy
AI-Driven Healthcare Denials Under Trump Administration
The Trump administration is implementing AI tools that may restrict medical services for elderly citizens. Critics argue that the software providers are financially motivated to reject as many claims as possible.
Read the source →