Martínez-Plumed et al. (2021) — pasajes subrayados

Esta es la nota de detalle de Martínez-Plumed et al. (2021): aquí están los pasajes que subrayé en el PDF, uno por uno, con mi comentario. La nota principal es el resumen concreto; esta sirve cuando quieras citar o volver al texto.

Pasajes subrayados por color

Resumen: 16 anotaciones aplicadas (azul 2, amarillo 3, morado 2, verde 3, rojo 4, naranja 2). Las páginas son las del artículo, que coinciden con las físicas del PDF.

🔵 Definiciones / conceptos operativos

• «The current TRL scale consists of 9 levels. Each level characterises the maturity of the development of a technology, from the mere idea (level 1) to its full deployment on the market (level 9).» (p. 3) Mi nota: Definición operativa de los nueve TRL. Te da un vocabulario preciso para describir la madurez del componente de IA.
• «Consequently, we need to assign readiness levels according to different layers of generality: a technology that is specialised for a very particular, controlled, domain may reach higher TRL than a technology that has to be more general-purpose and open-ended.» (p. 3) Mi nota: Concepto de generalidad. Te ayuda a definir el alcance de la IA local: cuanto más acotada la tarea, más confiable.

🟡 Ideas principales

• «Many so-called breakthroughs in AI are associated with highly-cited research papers or good performance in some particular benchmarks. However, research breakthroughs do not directly translate into a technology that is ready to use in real-world environments.» (p. 1) Mi nota: Tesis central del artículo. Te sirve para no prometer más de lo que la IA local puede dar en tus 12 sedes.
• «The key question is not whether AI is envisaged to work or already working in restricted situations, but whether an AI technology is sufficiently developed to be applicable in the real world, as a viable product leading to business value and real transformation.» (p. 2) Mi nota: Cambia la pregunta: no basta con que la IA funcione en una demostración; debe funcionar en el aula multigrado real de El Tablazo.
• «For example, there is no such a thing as TRL 3 or TRL 7 for machine translation, unless we also specify the layer of generality (quality, autonomy, etc.) that is expected for the technology.» (p. 22) Mi nota: Primera lección del método: decir qué tan madura es una IA exige decir para qué tarea. Aplícalo al describir tu ecosistema.

🟣 Metodología

• «In this work, we have assessed the TRL of each technology (at a particular layer) by asking experts (including ourselves) to follow the rubric in A to estimate the particular level in the scale» (p. 25) Mi nota: Método: juicio de expertos con una rúbrica. Podrías adaptar la idea para que tus docentes valoren el componente de IA.
• «What we have covered in this paper is an exemplar-based methodology where (1) we identify the technology, its category and its scope, (2) we recognise and define the layers of generality that are most meaningful for the technology and appropriate to estimate the TRLs accurately, (3) we find evidence in the scientific literature and industry to identify the points on the readiness-vs-generality chart, and (4) we use the chart to understand the state of the art of the technology and, if drawn historically, extrapolate its future trends.» (p. 28) Mi nota: Los cuatro pasos del método en una frase. Sirve si quieres ubicar cada herramienta del ecosistema digital en un gráfico de madurez.

🟢 Confirmaciones / respaldo empírico

• «These two dimensions lead to the new readiness-vs-generality charts, which show that higher TRLs are achievable for low-generality technologies, focusing on narrow or specific abilities, while high TRLs are still out of reach for more general capabilities.» (p. 1) Mi nota: Hallazgo principal: lo estrecho está maduro, lo general no. Por eso tu componente de IA debe tener tareas acotadas.
• «In this regard, current MT software and applications17 are best suited when we need quick, one-off translations and accuracy is not of importance.» (p. 11) Mi nota: La traducción automática sirve para lo rápido, no para lo preciso. Cuida su uso con materiales para niños.
• «Note that although all the above speech recognition-powered products and software are market-ready products (TRL 9) with high levels of robustness and accuracy, the capabilities achieved by these systems are still limited to restricted domains, also having problems with noisy environments, different accents, disorganised conversations, echoes, speaker distance from the microphone, etc.» (p. 12) Mi nota: El reconocimiento de voz falla con ruido y acentos: justo las condiciones de un aula multigrado rural. Tenlo en cuenta si usas voz con los niños.

🔴 Contradicciones / tensiones / límites

• «The latter includes some usually neglected dimensions, in addition to performance, such as data, expert knowledge, human oversight, software resources, computing cycles, hardware and network facilities, development time, etc., apart from monetary costs (Martinez-Plumed et al., 2018).» (p. 2) Mi nota: Costos que se olvidan: equipos, energía, supervisión humana, red. Úsalos como lista de chequeo en la fase de observación de tus sedes.
• «Readiness does not necessarily fit appropriateness or feasibility: a mature technology (e.g., an automated or self-driving train) may possess a greater or lesser degree of readiness to be used in a particular context (e.g., underground,2 airports,3 etc.), but the technology may not be ready to be applied to other contexts (e.g., general railways).» (p. 3) Mi nota: Límite clave: madurez técnica no es pertinencia. Que una IA esté lista en la ciudad no significa que lo esté en una vereda sin conexión.
• «Layer 3 is still largely in research and evaluation phases; it is limited in that current approaches (e.g., language models and acoustic models) cannot handle the complexities of a free speech recognition application in unrestricted contexts with multiple speakers for a myriad of languages and different regional accents for the same languages.» (p. 13) Mi nota: Varias voces a la vez y acentos regionales siguen sin resolverse. Es un límite técnico que tu modelo debe asumir, no ocultar.
• «However, the estimates would still be based on expert evidence but not quantitative evidence.» (p. 25) Mi nota: Límite que los autores reconocen: sus niveles son estimaciones de expertos. Cítalo con esa cautela.

🟠 Citas textuales «para usar»

• «In this regard, a given technology may be ready but the costs of deployment may not be affordable for the mass.» (p. 28) Mi nota: Frase citable para el contexto rural: lo técnicamente listo puede ser inalcanzable para una sede veredal.
• «In other words, technological readiness does not mean technological success.» (p. 28) Mi nota: Cita corta para tu marco: la madurez de la IA no garantiza que funcione en tu comunidad; eso lo decide la mediación.