Como cliente Amazon Prime obtén 3 meses de Audible gratis
Everything Green, Everything Broken
LLMs in Production, the Failures That Raise No Error and the Five Numbers Nobody Counts
No se ha podido añadir a la cesta
Solo puedes tener 50 títulos en tu cesta para poder pagar.
Vuelve a intentarlo más tarde
Vuelve a intentarlo más tarde
Error al eliminar la lista de deseos.
Vuelve a intentarlo más tarde
Se ha producido un error al añadirlo a la biblioteca
Inténtalo de nuevo
Se ha producido un error al seguir el podcast
Inténtalo de nuevo
Error al dejar de seguir el podcast
Escúchalo ahora gratis con tu suscripción a Audible
Después de los 30 días, 9,99 €/mes. Cancela tu siguiente plan mensual cuando quieras.
Disfruta de forma ilimitada de este título y de una colección con 90.000 más.
Escucha cuando y donde quieras, incluso sin conexión.
Sin compromiso. Cancela tu siguiente plan mensual cuando quieras.
Everything Green, Everything Broken
Narrado por:
AI Voice A synthetic voice
Compra ahora por 9,99 €
-
Narrado por:
-
AI Voice A synthetic voice
-
De:
-
Nikolaos Broikos
Este título utiliza una narración voz virtual
Voz virtual es una narración para audiolibros generada por ordenador.
A model refused to answer while the answer sat in front of it, complete and correct, and then invented a data quality problem to justify the refusal, in the register of a careful bug report. A filter identified sensitive material perfectly, labelled it, wrote a log line proving it had seen it, and let it through. An upgrade rewrote the format of more than a quarter of responses while every accuracy test reported a clean migration. A cost model was wrong in two directions at once.
Six failures. Zero errors. Several of them producing positive evidence of health.
None of it was hidden. All of it sat in plain text, in fields nobody aggregated, phrased in a way nothing matched, or summarised by a statistic that absorbs exactly that signal. This book is about the class of failure that leaves no trace your monitoring was built to catch, and the five cheap numbers that catch it instead. None of them needs a labelled dataset, a vendor, a budget or anybody's permission.
The good news is the argument. The unpredictable part of your system was always the small part. The provable part was always most of it, and almost nobody tests it, because the presence of a model made an entire subsystem feel unmeasurable when it is not.
For engineers and technical leaders running something real, who suspect the green dashboard is not the whole story and want to know exactly where to look.
adbl_web_anon_alc_button_suppression_t1
No hay reseñas aún
