DeepSeek Acaba de Resolver Uno de los Mayores Problemas de la IA

por Two Minute Papers

Resumen

Two Minute Papers explica la técnica EnGram de DeepSeek, que le da a los modelos de IA componentes pre-hechos (como una despensa de ingredientes) en lugar de generar todo desde cero. El video cubre cómo este enfoque hace la IA más eficiente y discute las limitaciones actuales de la técnica y posibles mejoras.

IntermedioLanzamiento de ModeloBenchmarksCódigo Abierto

Herramientas Discutidas

DeepSeek EnGram

Presented as solving major AI efficiency problem

Desglose del Puntaje

Puntaje bruto: 46= 46/100

Verificación Automatizada

11 / 40

Testing de Prompts0

Ejecución de Código0

Validación de Links—

Verificación de Claims8

Precisión de Versiones—

Análisis de Calidad (IA)

23 / 40

Originalidad3

Especificidad4

Completitud5

Densidad de Valor5

Honestidad / Limitaciones6

Modelo: anthropic/claude-sonnet-4

Señales de Contexto

12 / 20

Frescura3

Historial del Autor2

Engagement Real7

Prompts Testeados

Ejecutamos cada prompt del video contra LLMs reales y verificamos que el resultado coincida con lo que el creador afirmó.

FALLO3635ms

Prompt

What is the capital of France?

El creador afirmó: The AI will recall the basics, France is a country in Western Europe, and Paris is a major city there. But wait, am I confusing it with any other city? Let me think. There's Lyon, Marseille, Toulouse... but none of those are the capital. Paris is definitely the most well-known. I think the Eiffel Tower is there, which is a big landmark. Also, the French government is based in Paris. The president's residence, the Élysée Palace, is there. The Louvre Museum, which is in Paris, adds to its

Respuesta del LLM

The capital of France is **Paris**. Known for its art, fashion, and culture, Paris is home to iconic landmarks such as the Eiffel Tower, the Louvre Museum, and Notre-Dame Cathedral.

Resultado de verificación: The creator's claim describes a complex reasoning process with self-questioning and deliberation about multiple French cities, but the actual output is a simple, direct factual response with no evidence of the claimed internal reasoning or consideration of alternative cities.

FALLO6137ms

Prompt

Use only emoji to explain how a transformer neural network works and its advantage. Be creative!

El creador afirmó: The AI will explain transformers with emoji

Respuesta del LLM

Here’s my emoji-only explanation of a transformer neural network and its advantages: 🧠💡 (smart idea) 📖➡️🔢 (text to numbers) 🔄✨ (magic self-attention) 👀📊 (looks at all words) 🔗🤝 (connects relationships) 🚀⚡ (fast & powerful) 🌐🗣️ (great for language) 🎯✅ (accurate results) **Advantage:** ⚡🆚🐢 (faster than older models) 📈🔍 (scales well with data) 🧩🔄 (handles long sequences) Hope you like it! 😄🤖

Resultado de verificación: The output includes explanatory text in parentheses like '(smart idea)' and '(text to numbers)' plus additional commentary, violating the 'only emoji' requirement despite the creator's claim that the AI will explain transformers with emoji.

Tests de Verificación

FALLOEjecución de Código533ms

El creador afirmó: This is a Pygame script using Pygame to create a bouncing yellow ball within a rotating square, with proper collision detection.

Resultado de verificación: La prueba no produjo el resultado esperado

APROBADOVerificación de Herramientas7157ms