Modelos, agentes, investigación y negocio de la IA.
A new study shows that platforms like YouTube and X are functioning as 'gateways' to apps that create non-consensual sexual deepfakes. For less than a dollar, anyone can undress a person in an image. The phenomenon poses serious privacy risks and demands urgent action from platforms.
DeepMind CEO Demis Hassabis has proposed creating an independent standards body, inspired by FINRA, to evaluate and certify frontier AI models before deployment. The initiative aims to mitigate national security risks such as cyberattacks or biological threats, and could be voluntary initially and mandatory later.
Yann LeCun has left Meta to found AMI Labs, backed by one billion dollars, aiming to build artificial intelligence that learns from sensory experience and common sense, like a young child, rather than relying on vast amounts of text.
Mark Zuckerberg has confirmed that Meta is exploring an AI cloud business to rent out its computing power. The initiative, internally called Meta Compute, aims to monetize massive AI infrastructure investments and position Meta as a new player in the cloud market.
A study from Towards Data Science quantifies the real electricity cost of running large language models (LLMs) on a local GPU. The results show that the cost per million tokens varies significantly across models, and efficiency does not always correlate with model size.
OpenAI has restructured its organization by merging safety and research teams, leading to the departure of Johannes Heidecke, head of safety systems. The move aims to streamline AI development but raises questions about the independence of safety oversight.
SRE teams are exploring AI agents to manage complex incidents, but trust is only earned through safe failures and transparency. This article analyzes the five pillars for building reliable agents in production.
Two research teams have proposed HBM memory designs that rotate DRAM chips 90 degrees, enabling liquid cooling and higher density. The technique promises 540 tokens/s on GPT-3 versus 296 for HBM4, and could redefine the future of AI data centers.
Intel is allocating €5 billion (about $5.7 billion) to expand its campus in Leixlip, Ireland, to manufacture AI and high-performance processors. The investment represents 30% of the company's capital budget for 2026 and underscores its commitment to European technological sovereignty.
Researchers at Anthropic have revealed that Claude, their artificial intelligence model, shows different values when asked in different languages. The finding raises questions about the universality of AI ethical alignment and its implications for global users.
A Databricks analysis shows that measuring AI cost only by token is misleading: the price per task reveals that seemingly expensive models can be more cost-effective. Additionally, the intermediary software (harness) impacts the final cost by up to 3 times.
Envision CEO Lei Zhang unveiled the 'Mission Gobi' plan at VivaTech, aiming to develop 5 GW of green AI capacity in deserts by 2030. The initiative reverses traditional logic: instead of bringing energy to data centers, it brings data centers to where abundant renewable energy is, thus avoiding competition with homes and industries.