Modelos, agentes, investigación y negocio de la IA.
Frontier AI models are trained at enormous cost, and a significant fraction is recouped in the few months after launch. Each week of delay shrinks the profitability window, risking the economic viability of labs and the massive infrastructure that supports them.
Alibaba launches Qwen-AgentWorld, two models trained to predict the state of agent environments across seven domains. This inverse approach enables training more robust agents and overcomes limitations of production data. Results show improvements on benchmarks never seen during training.
IBM has announced a 0.7 nm (7 Angstroms) manufacturing process using stacked 3D transistors, achieving nearly 100 billion transistors on a fingernail-sized die. The company projects reaching 1 Angstrom within the next decade, marking a milestone in semiconductor miniaturization.
After months of unchecked usage, companies are implementing AI token rationing systems to prevent employees from exhausting budgets with small tasks. This trend marks a shift toward more efficient management of AI resources.
ByteDance, the parent company of TikTok, is in preliminary talks with banks to obtain an offshore loan of approximately $20 billion, the largest it has ever sought. The funds would be used to boost its development in artificial intelligence amid growing global competition.
Blackstone, the world's largest alternative asset manager, plans to invest about $30 billion in artificial intelligence data centers in Japan over the next three to five years, according to its president Jonathan Gray in an interview with Nikkei. The investment is a boost for Japan as a destination for critical AI infrastructure.
Adobe has acquired Topaz Labs, a company specializing in AI-powered image and video enhancement. The integration of its upscaling, restoration, and local processing tools promises to transform creative workflows, blending generated and captured content.
The US government is pressuring Meta to allow federal security review of its most advanced AI models, according to a New York Times report. Meta is the only major developer that has not agreed, raising questions about security risks and transparency in the industry.
Vercel has launched an observability feature for AI agent sessions on its platform, allowing inspection of each turn, model call, and tool. The tool, integrated into the Vercel dashboard, requires no OpenTelemetry setup and offers two viewing modes: Developer and Business.
MRAgent, developed by the National University of Singapore, introduces an active and reconstructive memory approach that reduces token consumption to 118K per query, compared to 3.26M for LangMem. This breakthrough promises more efficient and accurate AI agents in complex tasks.
Pro-data center candidates lose primaries in Utah, while New Jersey bans new construction. Discontent over energy consumption, noise, and real estate speculation becomes a political force.
A high-severity vulnerability in Amazon Q Developer, discovered by Wiz Research, allowed a malicious repository to silently execute commands on the developer's machine and steal their AWS credentials. The flaw, identified as CVE-2026-12957, has been patched by Amazon, but the case highlights security risks in AI assistants integrated into development environments.