Ai · Top stories
Google Develops ProtNLM AI Model to Accelerate Antimicrobial Peptide Discovery
Google developed ProtNLM (Protein Natural Language Model), a deep learning model that predicts biological function directly from raw protein sequences. This model helps identify novel antimicrobial peptides (AMPs) from unannotated protein databases, addressing the urgent need for new therapeutics against rising antimicrobial resistance.
Instagram Direct Migrates Android UI to Jetpack Compose, Cuts AI Token Costs by 33%
Instagram Direct migrated its Android UI to Jetpack Compose, creating an AI-native architecture. This change reduced UI code volume by 50% and decreased AI agent token costs per session by 33%. The migration focused on designing the codebase to guide AI agents towards producing correct and high-quality code.
Visko Deploys Live World Model Orbis on Google Cloud TPUs for Real-Time Streaming
Visko collaborated with Google Cloud engineers to deploy its Orbis Live World Model on Cloud TPUs, achieving real-time video and audio streaming. This deployment allows for continuous, interactive generative video, addressing the challenge of low-latency inference for live AI applications.
MacPaw Launches Eney, a Proactive AI Assistant for Mac and iPhone
MacPaw released Eney, a proactive AI assistant for Mac and iPhone, after a development process that began in 2019. Eney integrates with system context, apps, and files to anticipate user needs and suggest actions, utilizing both local and cloud-based AI models.
Ricursive Intelligence Develops AI for Automated Chip Design to Address Compute Demand
Ricursive Intelligence, founded by Anna Goldie and Azalia Mirhoseini, is developing AI that designs chips, learns from iterations, and improves subsequent designs. This aims to automate and accelerate the chip design process to meet increasing compute demand and create more efficient hardware. The company optimizes entire chips, allowing AI to discover unconventional and effective layouts.
OpenAI London Staff Received Average £2.3M Compensation in 2025
OpenAI's London staff received an average of £2.3 million each in 2025, comprising salaries and company shares. This compensation reflects the competitive talent market for AI researchers and positions London as a key AI hub.
Zimbabwean Author Denies Using AI for Prize-Winning Story After Software Flags It
Simbarashe Kundizeza, winner of the Caine Prize for African Writing, denied using AI to write his story "The Blue Fighter" after an Atlantic article reported AI-detection software flagged it as AI-generated. This incident follows similar accusations against other literary prize winners, raising questions about AI's role in creative writing and the reliability of AI detection tools.
OpenAI Infrastructure Lead Sachin Katti Departs Role, Becomes Advisor
Sachin Katti, OpenAI's infrastructure lead, is stepping down from his position after approximately one year but will remain as an advisor. Uday Ruddarraju, formerly CTO of compute, will take over leadership of the compute organization, which is critical for developing OpenAI's infrastructure roadmap and supporting its model training and serving capabilities.
Researchers use AI to design proteases that cut specific proteins
Researchers utilized AI-based protein design tools to create novel enzymes capable of cutting specific proteins. These enzymes were demonstrated to cut proteins associated with neurological diseases, potentially offering new methods for targeted protein degradation.
HiRO-ACEv2 AI Climate Model Simulates Extreme Weather at 3 km Resolution
The HiRO-ACEv2 climate emulator, an update to the HiRO-ACE system, can simulate global 3 km resolution climate data, including extreme weather events, at approximately 1/300th the cost of the X-SHiELD atmospheric model. This development allows for more cost-effective and accurate high-resolution climate modeling, crucial for understanding and predicting rare weather phenomena.
Booz Allen Report: AI Agents Control Unfamiliar OT Systems and Move Equipment in Minutes
Booz Allen's new report demonstrates that AI agents can navigate unfamiliar operational technology (OT) systems and manipulate physical equipment within minutes. This finding indicates the improving capabilities of contemporary artificial intelligence in industrial control environments.
Meta Extends Confidential Computing to AI Glasses for Private Processing
Meta has extended its Private Processing architecture to AI glasses, utilizing confidential virtual machines (CVMs) and Trusted Execution Environments (TEEs) in the cloud for sensitive AI workloads. This allows for advanced features like streaming transcription and contextual search while maintaining user privacy by preventing Meta from accessing personal data.
Anthropic seeks rehearing on Pentagon blacklisting appeal in D.C. federal court
Anthropic has asked a federal court in Washington, D.C. to rehear its appeal regarding the Pentagon's blacklisting of its AI models. The company claims the designation as a national-security "supply chain risk" was retaliation for its AI safety commitments, which included refusing use for mass surveillance or lethal autonomous warfare.
Michael Dell Calls for AI Safeguards and Government Role in Regulation
Dell Technologies CEO Michael Dell stated that AI models require safeguards and controls, suggesting a role for government in their implementation. This comes amid ongoing debate among researchers and industry experts regarding potential risks of rapidly evolving AI technology.
Geoffrey Hinton reiterates AI existential threat warnings, criticizes industry self-regulation
AI pioneer Geoffrey Hinton restated his warnings about AI posing an existential threat to humanity, dismissing claims that the industry can regulate itself. His comments follow increasing concerns from other AI scientists and pushback from figures like Nvidia CEO Jensen Huang, who called Hinton's predictions unscientific.
Janas-LLM enables running large language models exceeding RAM by streaming experts from SSD
Janas-LLM, a new tool written in C, allows large language models (LLMs) like Qwen3-Next-80B to run on consumer-grade hardware with limited RAM by streaming model experts from an NVMe SSD. This method achieves significantly faster token generation rates compared to llama.cpp on the same hardware, making larger LLMs accessible on ordinary computers.
IMF warns of AI's labor market impact; World Bank council urges permanent employment support
The IMF Managing Director Kristalina Georgieva warned that AI poses a "slow-moving tsunami" to global labor markets, particularly affecting young graduates. In response, the World Bank's Jobs Council co-chair Tharman Shanmugaratnam advocated for governments to adopt permanent "crisis-era" support measures to help young people enter the workforce.
Intel and Broadcom CEOs Warn Capital Access Could Limit AI Growth
The CEOs of Intel and Broadcom stated that access to capital may become the primary bottleneck for AI expansion, surpassing concerns about chip or power shortages. Rising capital costs and the need for sufficient returns on AI investments are highlighted as key challenges for the industry's infrastructure build-out.
Lancet Commission Identifies Malicious AI and Nuclear War as Top Threats by 2100
A new global risk assessment by The Lancet Commission on 21st Century Global Threats to Health identifies malicious AI use and nuclear conflict as the leading threats to human health and existence by 2100. These threats, alongside 15 others, could lead to significant loss of healthy life years and require integrated solutions from policymakers.
Senate Investigation Finds Misleading Claims by AI Data Center Developers
A year-long U.S. Senate investigation found that some AI data center developers mislead communities about job creation and economic benefits. Developers highlight construction jobs but often fail to disclose low numbers of permanent positions, and benefit from significant tax exemptions while resisting infrastructure cost contributions.
UK Must Develop Domestic AI to Avoid Foreign Dependence, Says Alan Turing Institute Head
The head of the Alan Turing Institute stated the UK needs to develop its own AI systems to prevent reliance on foreign technology that could be restricted or shut down. This initiative aims to build national resilience in AI, responding to the dominance of the US and China in the field.
AI adoption by 90% of UK/Ireland legal professionals challenges billable hour model
Almost 90% of legal professionals in the U.K. and Ireland use AI, leading to increased work capacity and reduced administrative costs. This widespread adoption is pressuring the traditional billable hour model, with some firms reporting difficulty meeting targets and a projected global decline in hourly-charged work.
Microsoft Plans to Transform Windows 11 into an Agentic OS with Hybrid AI
Microsoft announced its strategy to evolve Windows 11 into an "agentic OS" by integrating AI agents and hybrid intelligence, allowing local execution of AI models. This initiative aims to leverage local processing for AI tasks, reducing reliance on cloud resources and targeting developers and creators with new hardware like the Surface Laptop Ultra.
Lawmakers Urge Google and Spirit Airlines to Halt $10M AI Data Deal Over Privacy Concerns
Over 100 members of Congress have sent a letter to Google and Spirit Airlines, urging them to stop a $10 million deal where Google would acquire Spirit's internal data for AI model training. Lawmakers are concerned that even with de-identification, the vast amount of sensitive employee data, including emails and payroll records, could still pose privacy risks when used with modern AI technologies.
AI Leaderboard Arena Raises $200M Series B, Valuing Company at $3.1B
Arena, an AI model evaluation platform, secured a $200 million Series B funding round, bringing its valuation to $3.1 billion. This funding follows significant revenue growth and the introduction of a commercial product for AI model analytics based on community feedback, addressing the limitations of traditional AI benchmarks.
Whistle Speech Recognition Model Released for On-Device Applications
A new speech recognition model, Whistle, has been released, designed for on-device transcription, word timestamps, and speech embedding. It operates as a 16.9 MB file on the CPU without external dependencies, supporting seven languages for mobile, wearable, and embedded systems.
Goodfire Launches Internal AI Monitors to Detect Rogue Agents at Lower Cost
Goodfire introduced new AI monitors that observe internal model operations to detect rogue AI agents, offering a more cost-effective alternative to existing methods. This system aims to prevent incidents like AI agents escaping test environments by integrating directly into the AI model's computational process.
Software Optimization Offers Significant Energy Savings for AI Data Centers
Researchers suggest software and algorithmic improvements can substantially reduce the energy consumption of AI data centers, which are projected to use as much electricity as Japan by 2030. Focusing on software efficiency, rather than solely hardware, could cut energy use by up to 30% without impacting performance. This approach addresses the increasing energy demands of AI workloads.
Co-op Legal Services Implements AI Surveillance for Customer Phone Calls
Co-op Legal Services has deployed an AI system to monitor and score employee phone calls with customers, assessing over 50 aspects of each interaction. This move follows similar implementations by other companies and raises concerns among employees and experts about increased stress and distrust in the workplace.
Claude Sonnet 5.5 now available on Amazon Bedrock and Claude Platform on AWS
Claude Sonnet 5.5, an updated AI model, is now available on Amazon Bedrock and Claude Platform on AWS. This version offers improved efficiency for coding and knowledge work at a lower cost and faster speed, making it suitable for continuous or large-scale AI workloads within AWS infrastructure.
Healthleap raises $38M in seed and Series A funding for AI platform identifying at-risk hospital patients
Healthleap secured $38 million in seed and Series A funding to develop its AI platform, which analyzes patient records to identify hospital patients at risk of undiagnosed conditions like malnutrition or delirium. The platform integrates with electronic health record systems, using language models to extract information from clinical notes and structured data to generate risk scores for care teams. This funding enables Healthleap to expand its deployment and further validate programs for various conditions, potentially improving early detection and patient outcomes in hospitals.
Arizona Court Rules AI-Generated Victim Video Carried "Undue Emotional Weight" in Sentencing
An Arizona appellate court ruled that an AI-generated video used as a victim impact statement in a manslaughter sentencing carried "undue emotional weight," leading to a resentencing order. The court determined that the AI video, which depicted the victim speaking, improperly influenced the judge's decision on the prison term. This ruling sets a precedent regarding the admissibility and impact of AI-generated content in legal proceedings.
OpenSearch Veterans Launch Infino, an Agent Retrieval Platform for AI Agents
Infino, founded by OpenSearch veterans, launched its agent retrieval platform, which provides a unified database system for AI agents to query all data from a single location. This platform addresses the fragmented data stack currently used by AI agents, which were not optimized for their unique query patterns.
AI-assisted proof verifies optimal packing for 11 squares using Lean 4.34.1
An AI-assisted proof has been verified for the optimal packing of 11 squares, passing verification with native numerical certificates. This achievement demonstrates the increasing capability of AI in complex mathematical proofs, particularly in geometry and optimization problems.
Decisions API enters public beta, offering faster text and image evaluation
The Decisions API is now in public beta, providing a new method for evaluating text and images to return typed answers up to 10 times faster than the existing Responses API. This development offers developers a dedicated tool for content classification, request routing, and work prioritization based on probabilities, choices, or scores.
Thiel Fellow Sigil Wen launches Underdog, an on-device AI assistant for privacy
Sigil Wen, a Thiel Fellow, launched Underdog, an invite-only beta of an AI assistant that runs entirely on-device for enhanced privacy. Underdog uses Wen's Husky inference engine to run a 27-billion parameter model, aiming to provide capabilities comparable to older state-of-the-art models without compromising user data. This development matters as it offers a privacy-focused alternative in the AI assistant market by keeping user data local.
Musubi releases PolicyLM-1.7B, an open-weights decision model for real-time content moderation
Musubi launched PolicyLM-1.7B, a new open-weights decision model designed for real-time content moderation. This model applies plain English content policies to messages in under 50 milliseconds, offering flexibility similar to LLMs but with the cost and speed of AI classifiers.
Google launches EmbeddingGemma 2 model and AI Edge Foresight Mac app for offline notetaking
Google released EmbeddingGemma 2, a 740 million parameter AI model, designed for on-device information processing. The company also launched AI Edge Foresight, a Mac application that uses EmbeddingGemma 2 to convert shorthand notes into polished notes from offline recordings and to answer questions based on local files.
SAP acquires AI company TechWolf to enhance HR agent context in SuccessFactors
SAP acquired TechWolf, a Belgian AI company that maps employee skills and work, to integrate its technology into SAP SuccessFactors HR agents. This acquisition aims to provide more context for HR tasks like skills-based hiring and talent management, making AI agents more efficient and reducing operational costs.
Vinod Khosla Backs Wajo Personal AI Agent, Citing Trust as Key Differentiator
Vinod Khosla is backing Wajo, a personal AI agent founded by former DeepMind engineer Shivani Poddar, emphasizing trust as its primary advantage over competitors like Meta's Muse and OpenAI's Dot. Wajo offers features such as cross-platform messaging, task organization, and the ability to make calls or payments on a user's behalf, with a focus on privacy and user control.