Posts

Mellea: Programación Generativa

Image
  Mellea: Programación Generativa Un programa generativo es cualquier programa que contiene llamadas a un LLM (Large Language Model, o modelo de lenguaje de gran escala) . Esto abarca desde un simple envoltorio ( wrapper ) de un prompt hasta un complejo sistema de razonamiento de múltiples pasos. El término es deliberadamente amplio: lo importante no es cuántas llamadas a un LLM realiza un programa, sino los desafíos estructurales que surgen cuando se combinan operaciones estocásticas realizadas por LLM con código determinista. Mellea es una biblioteca diseñada para escribir correctamente programas generativos. El desafío fundamental Los programas clásicos son deterministas. Dada la misma entrada, producen la misma salida. Es posible razonar sobre ellos, probarlos y confiar en que los resultados obtenidos durante las pruebas pueden generalizarse. Las llamadas a los LLM no son deterministas. El mismo prompt , enviado al mismo modelo y utilizando la misma temperatura, puede produci...

Mellea - Generative Programming

Image
  Mellea: Generative Programming A generative program is any program that contains calls to an LLM. This covers everything from a simple prompt wrapper to a complex multi-step reasoning system. The term is deliberately broad: what matters is not how many LLM calls a program makes, but the structural challenges that arise when you combine stochastic LLM operations with deterministic code. Mellea is a library for writing generative programs well. The fundamental challenge Classical programs are deterministic. Given the same input, they produce the same output. You can reason about them, test them, and trust that the test results generalise. LLM calls are not deterministic. The same prompt, sent to the same model, with the same temperature, may produce different outputs. These outputs may each be valid responses to the prompt in a natural-language sense, but one may satisfy the downstream requirements of your program and another may not. Generative programs interleave these two modes....

Who controls the machine? Anthropic extends the MCP vision into the physical world

Image
  Who controls the machine? Anthropic extends the MCP vision into the physical world Today’s top story, from reporter Aili McConnon:   As AI systems gain the ability to operate laboratory equipment, factory machinery and robotics platforms, the biggest question becomes who remains in control?   That was the central concern raised by panelists during a discussion about Anthropic’s new Model Hardware Standard (MHS) on the Mixture of Experts podcast . Anthropic recently introduced MHS as a research preview for scientific labs and advanced manufacturers. The standard provides a common way for AI agents to discover, understand and operate physical devices, including microscopes, liquid handlers, robotic arms and manufacturing equipment. Anthropic says it could reduce integration projects that traditionally take weeks or months to hours or even minutes.   In many ways, MHS is the physical-world counterpart to Anthropic’s Model Context Protocol (MCP). MCP established a c...

DeepSeek publicó un modelo multimodal experimental de 305 mil millones (305B) de parámetros.

Image
  DeepSeek publicó un modelo multimodal experimental de 305 mil millones (305B) de parámetros, DeepSeek-V4-Flash-Vision-Exp, en Hugging Face bajo una licencia MIT.  Es el primer modelo de visión de la familia V4-Flash, el cual añade un codificador de visión (vision encoder) y un alineador sobre la ruta directa del modelo de texto que incluye atención DFlash, MoE (Mezcla de Expertos), Hyper-Connections y DSpark. La tarjeta del modelo (model card) es escueta. Su única frase evaluativa dice: "En comparación con DeepSeek-V4-Flash-0731, DeepSeek-V4-Flash-Vision-Exp logra mejoras sustanciales en sus capacidades de agente multimodal, mientras mantiene un rendimiento comparable en tareas de agente que solo requieren texto". La tabla de referencia (benchmarks) posiciona a Vision-Exp con un 36.5 en ApexBench Pass@1, subiendo desde el 26.2 del modelo V4-Flash-0731 que solo admite texto, y quedando un poco por detrás del 39.4 de Opus-4.8. En Agents' Last Exam obtiene una puntuación d...

DeepSeek releases 305B V4-Flash-Vision-Exp under MIT license

Image
  DeepSeek published a 305B-parameter experimental multimodal model, DeepSeek-V4-Flash-Vision-Exp , to Hugging Face under an MIT license. It is the first vision model in the V4-Flash family, adding a vision encoder and aligner on top of the text model's DFlash attention, MoE, Hyper-Connections and DSpark forward path. The model card is spare. Its lone evaluative sentence reads: "Compared to DeepSeek-V4-Flash-0731, DeepSeek-V4-Flash-Vision-Exp achieves substantial improvements on its multimodal agent capabilities, while maintaining comparable performance on text-only agent tasks." The benchmark table puts Vision-Exp at 36.5 on ApexBench Pass@1, up from 26.2 for the text-only V4-Flash-0731 and shy of Opus-4.8's 39.4. Agents' Last Exam scores 27.3, ahead of Opus's 25.7. ZeroBench Pass@5 lands at 35.0 against Opus's 34.0, and Chartography at 64.3 against 65.0. On text-only agent tasks the gap is narrower: Terminal Bench 2.1 at 83.9 versus Opus's 85.0, Dee...