Modelos, agentes, investigación y negocio de la IA.
The UK plans to use artificial intelligence to estimate the age of asylum seekers who lack documents. However, an internal report reveals that the systems frequently misclassify children as adults, with serious legal consequences. The measure raises concerns about bias and unreliability.
An AI company is offering free door-to-door cleaning services to collect data from real environments. The goal is to train robots that will one day replace the very workers doing the cleaning today. This model raises ethical questions about consent and the valuation of human labor.
The Pentagon has deployed GenAI.mil, a generative AI platform, for 1.5 million employees to draft reports, including those for Congress. CTO Emil Michael promotes its use with a pragmatic approach: 'Use GenAI.mil, do the best you can.'
Google and the University of California, San Diego have transformed retired Pixel smartphones into functional computing clusters, exploring a low-cost data center alternative that also combats electronic waste. This experiment could redefine sustainability in cloud infrastructure.
An article from Towards Data Science proposes a pipeline to extract only relevant images from PDFs in RAG systems, avoiding the cost of converting all images to text. The key is a filtering step with lightweight models that evaluates the importance of each image before sending it to a paid OCR service.
OpenAI introduced LifeSciBench, a set of 750 tasks to measure AI capability in life sciences. Its flagship model, GPT-Rosalind, failed 63.9% of the tests, demonstrating that AI is still far from replacing scientists.
Automated AI agents are becoming the new attack vector. Jay Chaudhry, CEO of Zscaler, proposes zero trust as the necessary security architecture to protect them, marking a paradigm shift in enterprise cybersecurity.
AI has made code generation nearly free and instant, but technical debt is skyrocketing. Charity Majors and Simon Willison warn that more discipline, not less, is needed to avoid chaos in production systems.
China has announced the creation of the Space Computing Industry Innovation Center, a forced alliance between rocket, satellite, chip manufacturers and AI labs to develop orbital data centers without connection to the terrestrial power grid. This move, coming a week before Musk's AI1 event, aims to position China as a leader in space AI infrastructure.
Amazon openly criticizes the AI governance model with human oversight ('human-in-the-loop'), arguing that humans are inconsistent and prone to 'normalization of deviance.' The company proposes a machine-based approach with strategic, not operational, human oversight.
Miami-based startup Subquadratic has released independent results supporting its claim of breaking quadratic attention, the mathematical bottleneck limiting language models. Its SubQ model achieves a 12 million token context window with sublinear scaling, a breakthrough that could redefine long-text processing in AI.
Norway becomes the first country to ban generative AI in primary schools, citing risks to critical thinking and minors' privacy. The decision, affecting hundreds of thousands of students, could set a precedent for other nations.