Ai · Top stories
Claude AI designed protein binders against scorpion neurotoxins
An internal research model of Claude AI designed 211 protein minibinders targeting scorpion neurotoxins, with 17 designs experimentally confirmed to bind. This project demonstrates AI's potential to accelerate drug discovery for neglected diseases, specifically antivenom development.
Meta Ordered to Name Executives Who Authorized Copyrighted AI Training Data Downloads
A magistrate judge ordered Meta to identify executives who approved the use of pirated books from shadow libraries for AI training data. This order follows Meta's attempts to attribute the downloads to individual staff members, despite internal communications indicating executive-level authorization.
Spotify removes 50 million AI-generated tracks from recommendations, but not platform
Spotify announced it removed 50 million AI-generated tracks from its recommendation system in 2026, following policy changes implemented a year prior. The company has not updated the total number of "spammy tracks" removed from the platform since September 2025, which remains at 75 million, indicating these tracks may still be accessible. This approach suggests Spotify is attempting to manage AI-generated content in recommendations while also integrating AI features into its platform.
FairStress Benchmark Evaluates LLM Demographic Fairness Beyond Final Decisions
FairStress is a new benchmark designed to measure demographic fairness in large language models (LLMs) by assessing how readily models change decisions when challenged, rather than just their final output. This benchmark aims to uncover hidden asymmetries in LLM decision-making across various demographic groups and is being integrated into Olmo-eval.
AI Model Darmok Solves Cryptic Crossword Faster Than Human Champion
The AI model Darmok solved the final puzzle of the Times's annual cryptic crossword competition three minutes faster than the human winner. This marks a significant advancement in AI's ability to handle complex wordplay and allegorical reasoning, a challenge previously thought difficult for machines.
Figma Expands MCP Support Despite Co-founder's Concerns About Protocol Imperfections
Figma announced 20 new partners for the Model Context Protocol (MCP), an open-source standard for AI agents to connect to data sources, despite co-founder Dylan Field's public concerns about the protocol's limitations and potential for abuse. The company is expanding its ecosystem while acknowledging MCP is not a perfect solution, particularly regarding asymmetric context sharing and data security.
Eka Robotics Raises $150M Series A, Valuing Company at $1.45B for Dexterous AI Model
Eka Robotics, founded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, secured a $150 million Series A funding round, bringing its valuation to $1.45 billion. The funding will advance its physical AI model, which demonstrates advanced dexterous manipulation capabilities, exceeding human performance in certain tasks.
AI Tool Predicts Cancer Survival from Facial Photographs
Mass General Brigham researchers developed an AI tool that predicts cancer patient survival rates from a single facial photograph. The system, detailed in The Lancet Digital Health, estimates biological age and directly forecasts survival, offering a new method for assessing patient health.
doubleAI Introduces Maxwell, a New Category of AI Models for Code Execution
doubleAI has developed "code execution models" (CEMs) named Maxwell, which integrate AI between software code and hardware. These models observe data flow within running programs and dynamically modify algorithms and instructions to optimize execution, without human intervention. This approach aims to improve program performance and resource utilization by allowing AI to adapt code during runtime.
AI Code Reviewer Reduces Inference Costs by 68% and Review Time by 63%
An AI code reviewer reduced its average review cost by 68% to $0.77 and median review time by 63% to 1 minute 32 seconds. This optimization was achieved through code changes, model selection, and managing token usage, addressing a previous cost increase from $1.19 to $2.46 per review.
Tenstorrent raises $1.5 billion at $6.2 billion valuation for open-weight AI chips
Tenstorrent, an AI chip startup, secured $1.5 billion in funding, increasing its valuation to $6.2 billion. This capital will be used to scale chips designed for open-weight AI models and video generation, positioning the company to compete with Nvidia.
Mercury 3 Diffusion LLM Released, Outperforms GPT-6 Luna and Claude Haiku 5.5
The Mercury 3 large language model has been released, demonstrating improved performance over GPT-6 Luna and Claude Haiku 5.5 in benchmarks while running 4-8x faster. This release indicates that diffusion-based LLMs can scale effectively for complex commercial applications, addressing a key question in the AI community.
AI coding tools increase game cloning, causing indie developers to delay public demos
AI coding tools are making game cloning more efficient, leading indie developers to become cautious about sharing early work. Alliance Arts delayed its public demo for "Drunken Goddess Reflex" due to concerns about AI-driven imitation, opting for closed sessions instead. This trend indicates a shift in how developers approach publicizing their projects, impacting game development cycles and intellectual property protection.
TwelveLabs Pegasus 1.5 Video Language Model Now Available on Amazon Bedrock
TwelveLabs' Pegasus 1.5, a video language model, is now available on Amazon Bedrock. This update allows for the analysis of long-form video to generate structured, timestamped metadata, enabling better searchability and utilization of video content.
Deepgram optimizes Nova-3 voice AI model for AWS Trainium2, improving efficiency
Deepgram collaborated with AWS to port its Nova-3 speech-to-text model to AWS Trainium2, achieving 11-70% more audio processing per dollar for batch workloads. This optimization allows Deepgram to scale its voice AI services more efficiently and potentially offer lower prices to customers.
Meta's GenRecs Tokenization Improves Ad Recommendations with LLM-based Semantic Vocabulary
Meta introduced GenRecs Tokenization, a new data representation method for its ad recommendation systems. This approach uses a stable semantic vocabulary from large language models (LLMs) to represent ad content and user preferences, moving beyond traditional item IDs. The system has delivered a +0.35% lift in conversions for advertisers by enabling more intelligent and precise ad matching.
Stack Overflow Details LLM System Maturity Model Levels for Safety and Operations
Stack Overflow has outlined two additional levels in its six-level maturity model for Large Language Model (LLM) systems. Level 4, "Safety and Governance," focuses on implementing layered guardrails, handling Personally Identifiable Information (PII) at boundaries, maintaining immutable audit trails, and using scoped memory. Level 5, "Operating an LLM System," addresses operability through observability signals, cost controls, routing, failover mechanisms, and kill switches to manage LLM systems in production environments.
CERN and Mistral Partner to Develop AI Foundation Models for Scientific Discovery
CERN and Mistral signed a statement of intent to form a strategic partnership focused on developing AI foundation models for scientific discovery. This collaboration aims to create a Large Physics Model trained on LHC data to improve high-energy physics analysis and potentially benefit the broader scientific community.
Australian Institute Calls for Stronger Workplace Privacy Laws Amid AI Monitoring Rise
The Human Technology Institute in Australia has warned that employers are increasingly using AI and other technologies for workplace surveillance, often without employee awareness. The institute is advocating for the Australian federal government to integrate privacy protections into workplace laws to address this growing trend.
Meta Completes First Phase of Prometheus Gigawatt-Scale AI Supercluster
Meta has completed the initial phase of its Prometheus AI supercluster in New Albany, Ohio, which is designed to train next-generation foundational models like the Muse family. This infrastructure project aims to scale to approximately one million GPUs, supporting larger models and longer context windows for AI development.
Laplace Intelligence develops AI red-teaming for robotics and self-driving cars
Laplace Intelligence, co-founded by Howie Sun and Henry Liu, is developing automated red-teaming solutions for physical AI systems like robots and autonomous vehicles. The company aims to identify and fix potential failures in simulation before real-world deployment, addressing the unpredictability of AI agents in physical applications.
Swedish Researchers Develop AI System to Detect Concealed Handguns Using Wi-Fi Signals
Researchers at the Swedish Defense Research Agency (FOI) developed WiFireNet, an AI system that identifies concealed handguns by analyzing how objects distort ambient Wi-Fi signals. This system offers a non-intrusive method for detection, utilizing channel state information from Wi-Fi interactions. While experimental, it presents a new approach to security screening without visible cameras or metal detectors.
World Bank Advocates for "Small AI" to Drive Development in Emerging Economies
The World Bank is promoting the development of "small AI" models for developing countries, which can operate on mobile phones with limited connectivity and localized data. This initiative aims to accelerate economic development by adapting AI to local needs, such as providing agricultural advice to farmers in their native languages.
OpenAI Confirms Development of Humanoid Robot, Focuses on AI 'Brain'
OpenAI CEO Sam Altman confirmed the company is developing a humanoid robot, prioritizing the AI 'brain' over the physical form. The project aims to create a robot capable of navigating human-designed environments, with a demo anticipated by 2027.
Replit CEO States Company No Longer Partners with Anthropic, Citing Cost and Performance
Replit CEO Amjad Masad announced that the AI coding platform has ended its partnership with Anthropic, citing cost, partnership factors, and performance benchmarks favoring OpenAI's models. This decision follows Replit's August partnership announcement with OpenAI, though previous indications suggested continued use of Anthropic's Claude models.
Microsoft and CMU Launch PRAISE Program to Train Early Career Software Engineers in AI
Microsoft and Carnegie Mellon University (CMU) launched the Preceptorship for AI-accelerated Software Engineering (PRAISE) program. This initiative offers early-career Microsoft engineers online CMU courses and hands-on mentorship to develop AI-accelerated software engineering skills, addressing the gap in AI education for junior engineers.
HUMXN offers free home repairs in three cities to gather AI training data
AI lab HUMXN is providing free plumbing, HVAC, and other home services in Minneapolis, Chicago, and Miami. This initiative aims to collect real-world data from skilled technicians to train AI robots on physical world interactions.
Websites Block AI Agents from Performing Tasks, Citing Security and Business Concerns
Major websites like Yelp, Delta Air Lines, JPMorgan Chase, and Charles Schwab are blocking AI agents from accessing their services. This prevents agents from performing tasks such as booking reservations or managing accounts, raising questions about the future functionality of AI agents on the internet.
Senior UK Labour Figures Express Growing Concerns Over AI Risks and Regulation
Senior Labour Party figures in the UK are increasingly concerned about the potential dangers of AI, including its misuse by malicious actors and the government's perceived slow response to AI safety. This concern has led to calls for stronger regulation and cybersecurity measures to address the evolving threats posed by advanced AI.
Model Context Protocol (MCP) adopted by AI assistants and developer tools for decentralized search
The Model Context Protocol (MCP), launched by Anthropic in November 2024 and transferred to the Agentic AI Foundation in December 2025, enables AI assistants to fetch information directly from publishers via server tools. This protocol allows for decentralized search, providing up-to-date, verifiable, and licensed information directly from source servers rather than relying on frozen model memories.
Google Open-Sources EnvHarness for Dynamic AI Agent Training Environments
Google researchers open-sourced EnvHarness, a framework that dynamically adapts agent-training environments to an agent's weaknesses without altering the agent or environment. This framework aims to improve learning effectiveness and efficiency by allowing agents to continuously learn from evolving environments.
Anthropic's Claude AI designs more efficient quantum circuits, reducing resource needs
Anthropic's Claude AI designed quantum circuits that improve upon existing published designs for various problems in chemistry, optimization, physics simulation, and number theory. These AI-generated designs reduce the quantum resources required, potentially accelerating the development of practical quantum computers.
Elorian releases Glimpse, an AI model designed for iterative visual reasoning
Elorian, an AI startup, launched its first model, Glimpse, which processes images by repeatedly referencing the source material rather than converting it to tokens once. This approach allows Glimpse to perform actions like cropping regions or extracting video frames during reasoning, addressing limitations in current multimodal models' visual understanding.
Graylark Technologies' Raven AI Tool Identifies Homes from Interior Photos for Law Enforcement
Graylark Technologies has updated its AI tool, Raven (formerly GeoSpy), to allow law enforcement to identify specific homes from a single interior photograph. This new capability cross-references interior details with millions of short-term rental listings and real estate sites like Zillow and Airbnb, enabling geolocation without relying on image metadata. The Los Angeles Police Department and Miami-Dade Sheriff's Office have acquired access to Raven.
AWS releases Strands Decider 2B and Strands Box for AI agent development
AWS introduced Strands Decider 2B, a 2-billion-parameter model for decision-making within AI agent workflows, and Strands Box, a sandbox for controlling agent execution. These open-source projects provide tools for building AI agents that interact with external systems and execute code in a controlled environment.
SpaceX agrees to testify to NYC Council on AI safety after subpoena
SpaceX will testify before the New York City Council regarding AI safety risks after receiving a subpoena. The council is investigating potential risks from advanced AI models to public safety, cybersecurity, and privacy, and sought testimony from major AI companies.
Benchmark Partners Acknowledge AI Boom Straining VC Rule Against Backing Rivals
Venture capital firm Benchmark's partners stated that the traditional rule against backing competing companies is eroding due to the rapid evolution of AI. This shift is driven by the fast-paced development in AI, leading to increased overlap and competition among portfolio companies.
Krea 3 Image Editing Model Released, Focuses on Production-Scale Edits
Krea 3 is a new image editing model designed for production environments, capable of making specific changes within an image while preserving other areas. It offers faster editing times and a per-image cost structure, aiming to integrate with leading organizations for industry advancement.
Amazon's Panos Panay envisions AI ambient devices replacing app-centric interactions
Amazon's SVP of Devices, Alexa & Leo, Panos Panay, stated that ambient computing devices, rather than apps, represent the future of personal AI. These devices operate in the background using AI and sensors to anticipate user needs, aiming to reduce direct user interaction with tools. This shift is exemplified by Amazon's new Bee devices, which passively record audio and generate insights while prioritizing user privacy through real-time processing and encrypted transcripts.
Accenture Caps Employee Claude AI Spending at $29 Per Month
Accenture has implemented a $29 per month spending cap for employees using the Claude AI tool, effective October 1. This change requires employees exceeding the limit to secure internal project funding for additional usage, reflecting a broader industry trend of companies managing AI token consumption costs.