Modelos, agentes, investigación y negocio de la IA.
Direct Preference Optimization (DPO) has proven to be an effective technique for aligning language models with human preferences without the need for complex reinforcement learning. But its application goes far beyond chatbots: from image generation to robotics, DPO is transforming how we train AI systems.
Hugging Face and AWS have introduced a set of building blocks that simplify training and inference of foundation models. This initiative aims to democratize access to high-performance AI infrastructure, reducing costs and technical barriers for businesses and developers.
AWS has launched the general availability of its AWS MCP Server, a managed remote server that enables AI agents and coding assistants to securely and authentically access all AWS services. This release includes a reduced set of tools, support for IAM context keys, and the transition from Agent SOPs to Skills, reducing hallucinations and improving efficiency.
Vercel's AI Gateway data reveals a seismic shift: DeepSeek V4 has achieved what no low-cost model had before—production-grade quality. Meanwhile, Anthropic continues to dominate spending with 65% of the total, and cost per token rises despite the low-cost disruption.
0Din is a Mozilla bug bounty program focused exclusively on generative AI vulnerabilities. It aims to engage the global security community to identify flaws such as jailbreaks, prompt injection, and training data leakage, with rewards varying by severity.
Mozilla has released version 0.8 of Llamafile, an open-source project that allows running large language models on local hardware simply and efficiently. The new tinyBLAS library democratizes access to GPUs, including AMD, removing barriers like the need to install CUDA.
LLM-based chatbots improve on standard tests, but user experience doesn't advance at the same pace. The missing key is purposeful dialogue: multi-turn conversations oriented toward a goal, allowing iteration, negotiation, and mutual understanding.
LLMs, successful in natural language, are now applied to financial series. However, the noisy and adversarial nature of markets limits their accuracy. This analysis explores the state of the art, real opportunities, and challenges quants face when trying to predict prices with these tools.
Mamba, a model based on state-space models (SSMs), challenges the dominance of Transformers by offering comparable performance with much higher efficiency on long sequences. Its linear architecture eliminates the quadratic bottleneck of attention, enabling contexts of up to a million tokens and speeds up to 5 times faster.
At the upcoming VivaTech 2026, enterprise artificial intelligence will take center stage, reflecting a European trend: integrating AI into existing complex systems rather than pursuing massive consumer language models. This approach promises efficiency and sustainability but poses regulatory and scalability challenges.
Dario Amodei, CEO of Anthropic, has published an essay calling for government regulation of frontier AI models, inspired by the FAA. The proposal includes mandatory third-party testing, compute thresholds, and the ability to block deployments that pose a risk to public safety. This would have direct implications for companies integrating AI into their operations.
For the first time, a Western government applies real-time export control on a commercial AI already deployed. The order forced Anthropic to globally disconnect its most advanced models, Fable 5 and Mythos 5, after an alleged jailbreak that would have allowed access to offensive cybersecurity capabilities. The precedent marks a before and after in AI regulation.