Modelos, agentes, investigación y negocio de la IA.
Samsung and Google have led the ratification of the NVMe TP4193 standard, which brings SSD virtualization to hardware. This reduces latency and hypervisor overhead, facilitating live migration of virtual machines in AI data centers. However, requiring new hardware will push up enterprise SSD prices.
Google implements a label informing users if an ad was created or edited with AI. The measure aims to increase transparency in digital advertising, though labeling responsibility falls on advertisers for external tools.
SpaceXAI has unveiled Grok 4.5, its most advanced model to date, designed specifically for coding tasks and autonomous agents. This version, the first after the acquisition of Cursor and the company's IPO, aims to directly compete with the flagship models of OpenAI and Anthropic in the enterprise market.
Hugging Face has integrated vLLM as a native modeling backend, delivering inference speeds up to 24 times faster. This breakthrough reduces costs and democratizes access to large language models, with profound implications for startups, enterprises, and the future of work.
Colibrì, a proof of concept by engineer Vincenzo (JustVugg), allows running the 744-billion-parameter GLM-5.2 model on just 25 GB of RAM and an NVMe SSD. Although extremely slow (0.05-0.1 tokens/second), it demonstrates that frontier AI can be brought to consumer hardware, with implications for privacy and democratization of AI.
The growing energy demand from AI data centers is testing the US power grid. Experts and companies are seeking solutions that combine efficiency, renewable energy, and new technologies to prevent collapses and ensure sustainable growth.
Researchers at Noma Security discovered GitLost, an attack that tricks the GitHub AI agent into posting content from private repositories in public issues. The vulnerability requires no credentials or technical exploits, but exploits the agent's trust in unverified content.
OpenAI has received US government approval for a broad launch of GPT-5.6, previously available only to 20 partners. This move signals a new chapter in frontier model oversight.
Researchers have discovered 'HalluSquatting,' a technique that exploits AI hallucinations to execute large-scale botnet attacks. By leveraging nine popular AI tools, attackers can inject malicious commands that models blindly execute, forming botnets without traditional malware.
SK hynix completed the largest IPO by a foreign company in US history, raising $26.5 billion to fund expansion of its HBM production, a key memory for AI accelerators. Demand exceeded supply by seven times, reflecting investor appetite for the artificial intelligence boom.
Democratic Senator Ed Markey introduced the 'AI Accountability Agenda,' a set of bills addressing everything from data centers to children's chatbots. The initiative aims to federalize regulation that has so far been fragmented state by state.
A study on AI sovereignty reveals that 91% of executives lack complete visibility into their systems' dependencies. Systemic drift, an emerging phenomenon, jeopardizes organizational resilience in increasingly interconnected ecosystems.