AI companies are pushing assistants into more autonomous roles while increasing the infrastructure and business models around them. Anthropic opened a preview of a proactive Claude app, and Meta paired early agent adoption figures with a higher spending floor for data centers. OpenAI expanded ChatGPT advertising internationally. Elsewhere, Hugging Face released a model that can reason over DNA, employee resignations renewed scrutiny of military AI governance, and a robotics startup raised an unusually large pre-seed round to develop touch-based machine intelligence.
Anthropic previews a Claude agent that initiates work
CNBC reported that Anthropic opened a US research preview of Claude Duet for Pro and Max subscribers. The app is designed to identify items that need attention, continue requested work in the background and ask before taking consequential actions. Unlike a conventional chatbot that waits for a prompt, Duet can surface a morning briefing and monitor connected work tools through the day.
The launch makes initiative, not just response quality, an AI feature. That increases usefulness for scheduling and follow-up tasks, but also raises the cost of mistaken assumptions. A proactive agent needs narrow permissions, visible activity logs and clear confirmation points. The research-preview label matters because the product will be judged as much by what it declines to do as by how much work it completes.
Meta lifts its AI infrastructure spending floor
CNBC reported that Meta raised the low end of its 2026 capital spending forecast to $135 billion, leaving the top end at $145 billion. Meta also reported quarterly revenue of $64.6 billion, up 26% from a year earlier. Mark Zuckerberg said more than 10 million people in the US and Canada had set up Muse personal agents, with about one quarter using them on a typical day. Shopping was the largest reported use case.
Those figures connect compute investment to a product that may generate subscriptions, developer fees and commerce. They do not yet prove that agent revenue will justify the infrastructure bill, since Meta's advertising systems consume the same capacity. The next signal will be whether Muse use persists after launch and whether agent activity creates measurable value without weakening privacy or user control.
OpenAI takes ChatGPT advertising to 89 countries
CNBC reported that OpenAI expanded ChatGPT Ads to 26 additional countries, bringing availability to 89 markets. OpenAI said more than 100,000 advertisers have run campaigns since the program launched eight months ago. Ads remain limited to Free and Go accounts, while Plus, Pro and Enterprise subscriptions remain ad-free. The company says ads are labeled, separated from answers and do not give advertisers access to private conversations.
Advertising gives OpenAI a way to subsidize low-cost access and diversify revenue beyond subscriptions and APIs. It also introduces incentives that users will watch closely when recommendations appear near commercial decisions. Separation from model answers needs to be technically enforceable and independently testable. Clear reporting on targeting, measurement, advertiser review and harmful-content controls will matter as the system adds formats and buying options.
Hugging Face releases an open model for DNA
Hugging Face released Carbon-X, an open 14.8-billion-parameter model for understanding and generating DNA sequences. The model combines text and DNA tokens, supports a context window of 40,960 tokens and was trained for 302 billion tokens after an initial vocabulary-extension stage. Its model card reports tasks including mutation classification, regulatory-region identification, sequence generation and plain-language explanations, with weights and deployment examples available to researchers.
Carbon-X makes genomic analysis more accessible to teams that want to inspect or run a specialized model themselves. Benchmark results are a starting point, not clinical validation. Predictions about pathogenic variants can be wrong, and biological meaning depends on experimental and population context. Useful adoption will require careful evaluation, privacy protections for genetic data and expert review before outputs affect research or patient decisions.
DeepMind resignations sharpen the military AI debate
The BBC reported that several Google DeepMind employees have resigned or are preparing to leave over Google's military AI contracts. Research engineer Sofia Liguori said she resigned this week because of the company's expanding work with the US and Israeli militaries. Google said it reviews public-sector contracts against its acceptable-use policies and AI principles, and that employees have protected channels for raising concerns.
The dispute exposes a gap between written principles and the visibility employees have into real deployments. Military users can compress decision times, combine outputs from several vendors and operate in classified settings, making outside scrutiny difficult. Contracts that require human oversight still need evidence about how much time and authority people retain. Independent testing, incident reporting and credible internal escalation channels are essential when errors may have irreversible consequences.
Enact raises $30 million to teach robots through touch
Tech.eu reported that Swiss robotics startup Enact Intelligence emerged from stealth with a $30 million pre-seed round. Investors include Redalpine, Notion Capital and Firstminute Capital, with Mistral CEO Arthur Mensch, Yann LeCun and Hugging Face co-founder Thomas Wolf participating as angels. Enact is developing models, data pipelines, actuators and sensors that let robots learn from force and touch rather than relying mainly on cameras.
Tactile feedback could make manipulation more reliable when objects are partly hidden, reflective or delicate. Building the full stack may also help Enact collect the interaction data that general-purpose robot models lack. The large early round reflects investor appetite for physical AI, but industrial deployment will test durability, safety and data efficiency in ways that laboratory demonstrations cannot.
What matters today
Today's developments move AI from conversation toward action. Claude Duet and Muse are competing to manage ongoing tasks, while OpenAI is commercializing the attention those assistants attract. Carbon-X and Enact extend model interfaces into genomes and physical contact. The DeepMind resignations show that expanded capability also raises harder questions about authority and accountability. The durable advantage will belong to systems that can act usefully while keeping permissions, incentives and human oversight legible.