Archives
All the articles I've archived.
El tigre y la jaula: por qué la alineación debe ir más allá del modelo
La cadena de razonamiento no basta para verificar la alineación de un sistema de IA. Una propuesta de ingeniería que combina representaciones, evaluación, autoridad humana y contención.
The Tiger and the Cage: Why Alignment Must Go Beyond the Model
A chain of thought is not enough to verify the alignment of an AI system. An engineering approach that brings together representations, evaluation, human authority, and containment.
La métrica del token está muerta
Tokens are still useful for billing and I/O. But as reasoning moves behind opaque summaries — and increasingly into latent computation — they may be becoming a terrible proxy for how much work a model actually did.
The Token Metric Is Dead
Tokens are still useful for billing and I/O. But as reasoning moves behind opaque summaries — and increasingly into latent computation — they may be becoming a terrible proxy for how much work a model actually did.