Rearranges video embeds for better readability

Moves margin video embeds below the corresponding introductory paragraph in the AI for Good chapter.

This improves readability by ensuring the introductory text is read before the video is presented, providing context upfront.

book/contents/core/ai_for_good/ai_for_good.qmd

@@ -73,10 +73,10 @@ AI technologies, such as Cloud ML, Mobile ML, Edge ML, and Tiny ML, are unlockin
 
 ### Agriculture {#sec-ai-good-agriculture-7412}
 
-{{< margin-video "https://youtu.be/MD61bddZtbg?si=Ake2uP8vC_lsvYhd" "Plant Village Nuru" "PlantVillage" >}}
-
 ![**Mobile Disease Detection**: Example of edge machine learning, where a smartphone app uses a trained model to classify plant diseases directly on the device, enabling real-time feedback in resource-constrained environments. this deployment reduces reliance on network connectivity and allows for localized, accessible agricultural support.](images/png/plantvillage.png){#fig-plantvillage}
 
+{{< margin-video "https://youtu.be/MD61bddZtbg?si=Ake2uP8vC_lsvYhd" "Plant Village Nuru" "PlantVillage" >}}
+
 In Sub-Saharan Africa, cassava farmers have long battled diseases that devastate crops and livelihoods. Now, with the help of mobile ML-powered smartphone apps, as shown in @fig-plantvillage, they can snap a photo of a leaf and receive instant feedback on potential diseases. This early detection system has reduced cassava losses from 40% to just 5%, offering hope to farmers in disconnected regions where access to agricultural advisors is limited [@ramcharan2017deep].
 
 Across Southeast Asia, rice farmers are confronting increasingly unpredictable weather patterns. In Indonesia, Tiny ML sensors are transforming their ability to adapt by monitoring microclimates across paddies. These low-power devices process data locally to optimize water usage, enabling precision irrigation even in areas with minimal infrastructure [@tirtalistyani2022indonesia].
@@ -95,24 +95,24 @@ In parallel, Cloud ML is advancing healthcare research and diagnostics on a broa
 
 ### Disaster Response {#sec-ai-good-disaster-response-4034}
 
-In disaster zones, where every second counts, AI technologies are providing tools to accelerate response efforts and enhance safety. Tiny, autonomous drones equipped with Tiny ML algorithms are making their way into collapsed buildings, navigating obstacles to detect signs of life. By analyzing thermal imaging and acoustic signals locally, these drones can identify survivors and hazards without relying on cloud connectivity [@duisterhof2021sniffy]. These drones can autonomously seek light sources (which often indicate survivors) and detect dangerous gas leaks, making search and rescue operations both faster and safer for human responders.
-
 {{< margin-video "https://www.youtube.com/watch?v=wmVKbX7MOnU" "Light Seeking" "TU Delft" >}}
 
-{{< margin-video "https://www.youtube.com/watch?v=hj_SBSpK5qg" "Gas Seeking" "TU Delft" >}}
+In disaster zones, where every second counts, AI technologies are providing tools to accelerate response efforts and enhance safety. Tiny, autonomous drones equipped with Tiny ML algorithms are making their way into collapsed buildings, navigating obstacles to detect signs of life. By analyzing thermal imaging and acoustic signals locally, these drones can identify survivors and hazards without relying on cloud connectivity [@duisterhof2021sniffy]. These drones can autonomously seek light sources (which often indicate survivors) and detect dangerous gas leaks, making search and rescue operations both faster and safer for human responders.
 
 At a broader level, platforms like Google's [AI for Disaster Response](https://crisisresponse.google/) are leveraging Cloud ML to process satellite imagery and predict flood zones. These systems provide real-time insights to help governments allocate resources more effectively and save lives during emergencies.
 
 Mobile ML applications are also playing a critical role by delivering real-time disaster alerts directly to smartphones. Tsunami warnings and wildfire updates tailored to users' locations enable faster evacuations and better preparedness. Whether scaling globally with Cloud ML or enabling localized insights with Edge and Mobile ML, these technologies are redefining disaster response capabilities.
 
+{{< margin-video "https://www.youtube.com/watch?v=hj_SBSpK5qg" "Gas Seeking" "TU Delft" >}}
+
 ### Environmental Conservation {#sec-ai-good-environmental-conservation-d41f}
 
 Conservationists face immense challenges in monitoring and protecting biodiversity across vast and often remote landscapes. AI technologies are offering scalable solutions to these problems, combining local autonomy with global coordination.
 
-{{< margin-video "https://youtu.be/ci95eyvTyXo?si=iD8TZiVAfuci4QeN" "Elephant Edge" "ElephantEdge" >}}
-
 EdgeML-powered collars are being used to unobtrusively track animal behavior, such as elephant movements and vocalizations, helping researchers understand migration patterns and social behaviors. By processing data on the collar itself, these devices minimize power consumption and reduce the need for frequent battery changes [@verma2022elephant]. Meanwhile, Tiny ML systems are enabling anti-poaching efforts by detecting threats like gunshots or human activity and relaying alerts to rangers in real time [@bamoumen2022tinyml].
 
+{{< margin-video "https://youtu.be/ci95eyvTyXo?si=iD8TZiVAfuci4QeN" "Elephant Edge" "ElephantEdge" >}}
+
 At a global scale, Cloud ML is being used to monitor illegal fishing activities. Platforms like [Global Fishing Watch](https://globalfishingwatch.org/) analyze satellite data to detect anomalies, helping governments enforce regulations and protect marine ecosystems. These examples highlight how AI technologies are enabling real-time monitoring and decision-making, advancing conservation efforts in profound ways.
 
 ### AI's Holistic Impact {#sec-ai-good-ais-holistic-impact-de43}
@@ -321,16 +321,14 @@ In machine learning applications, this pattern requires careful consideration of
 
 #### Google's Flood Forecasting {#sec-ai-good-googles-flood-forecasting-7eca}
 
-{{< margin-video "https://youtu.be/ET04pDj-RvM?si=l7P0nBv1h2rXOzIE" "AI for Flood Forecasting" "Google" >}}
-
-<!-- VJ: For PDF, maybe we could do something like this... or improve on it.-->
-
 Google's [Flood Forecasting Initiative](https://blog.google/technology/ai/google-ai-global-flood-forecasting/) demonstrates how the Hierarchical Processing Pattern supports large-scale environmental monitoring. Edge devices along river networks monitor water levels, performing basic anomaly detection even without cloud connectivity. Regional centers aggregate this data and ensure localized decision-making, while the cloud tier integrates inputs from multiple regions for advanced flood prediction and system-wide updates. This tiered approach balances local autonomy with centralized intelligence, ensuring functionality across diverse infrastructure conditions[^fn-Google-Flood].
 
 [^fn-Google-Flood]: Google's Flood Forecasting Initiative has been instrumental in mitigating flood risks in vulnerable regions, including parts of India and Bangladesh. By combining real-time sensor data with machine learning models, the initiative generates precise flood predictions and timely alerts, reducing disaster-related losses and enhancing community preparedness.
 
 At the edge tier, the system likely employs water-level sensors and local processing units distributed along river networks. These devices perform two critical functions: continuous monitoring of water levels at regular intervals (e.g., every 15 minutes) and preliminary time-series analysis to detect significant changes. Constrained by the tight power envelope (a few watts of power), edge devices utilize quantized models for anomaly detection, enabling low-power operation and minimizing the volume of data transmitted to higher tiers. This localized processing ensures that key monitoring tasks can continue independently of network connectivity.
 
+{{< margin-video "https://youtu.be/ET04pDj-RvM?si=l7P0nBv1h2rXOzIE" "AI for Flood Forecasting" "Google" >}}
+
 The regional tier operates at district-level processing centers, each responsible for managing data from hundreds of sensors across its jurisdiction. At this tier, more sophisticated neural network models are employed to combine sensor data with additional contextual information, such as local terrain features and historical flood patterns. This tier reduces the data volume transmitted to the cloud by aggregating and extracting meaningful features while maintaining critical decision-making capabilities during network disruptions. By operating independently when required, the regional tier enhances system resilience and ensures localized monitoring and alerts remain functional.
 
 At the cloud tier, the system integrates data from regional centers with external sources such as satellite imagery and weather data to implement the full machine learning pipeline. This includes training and running advanced flood prediction models, generating inundation maps, and distributing predictions to stakeholders. The cloud tier provides the computational resources needed for large-scale analysis and system-wide updates. However, the hierarchical structure ensures that essential monitoring and alerting functions can continue autonomously at the edge and regional tiers, even when cloud connectivity is unavailable.

book/contents/core/sustainable_ai/sustainable_ai.qmd

@@ -891,7 +891,6 @@ The carbon footprint of AI model design varies significantly depending on the co
 
 : **Model Carbon Footprint**: Training large AI models generates substantial carbon emissions, directly correlating with computational demands measured in flops; for example, training GPT-3 requires energy equivalent to the lifetime emissions of hundreds of cars. Understanding these emissions is crucial for developing sustainable AI practices and selecting energy-efficient hardware like tpus to minimize environmental impact. Source: {#tbl-training-emissions}
 
-
 Addressing the sustainability challenges of the design phase requires innovations in training efficiency and computational resource management. Researchers have explored techniques such as sparse training, low-precision arithmetic, and weight-sharing methods to reduce the number of required computations without sacrificing model performance. The use of pre-trained models has also gained traction as a means of minimizing resource consumption. Instead of training models from scratch, researchers can fine-tune smaller versions of pre-trained networks, leveraging existing knowledge to achieve similar results with lower computational costs.
 
 Optimizing model search algorithms further contributes to sustainability. Traditional neural architecture search methods require evaluating a large number of candidate architectures, but recent advances in energy-aware NAS approaches prioritize efficiency by reducing the number of training iterations needed to identify optimal configurations. Companies have also begun implementing carbon-aware computing strategies by scheduling training jobs during periods of lower grid carbon intensity or shifting workloads to data centers with cleaner energy sources [@gupta2022].