Exploring the complex relationship between AI and environmental sustainability
From carbon footprints to climate solutions
Strategies for a sustainable AI ecosystem
Advanced models optimize energy grids but consume vast resources23, 24
The industry faces a critical pivot point between "Red AI" and "Green AI"29, 30
How can we reconcile these opposing forces in the era of Generative AI?
According to the lecture, how does AI relate to environmental sustainability?
Choose an answer to continue
Advanced models optimize energy grids but consume vast resources. Think about what that means for AI's overall role.
The lecture says advanced models optimize energy grids but also consume vast resources, so AI is not only a climate savior.
Correct. AI acts as both a climate savior and a significant environmental burden.
Week 08 — The AI Sustainability Paradox
Correct. AI acts as both a climate savior and a significant environmental burden.
Week 08 — The AI Sustainability Paradox
Advanced models consume vast resources, which is why the lecture calls AI a significant environmental burden as well as a climate savior.
Correct. AI acts as both a climate savior and a significant environmental burden.
Week 08 — The AI Sustainability Paradox
The lecture says the industry faces a critical pivot point between "Red AI" and "Green AI", not that it has moved past it.
Correct. AI acts as both a climate savior and a significant environmental burden.
Week 08 — The AI Sustainability Paradox
Analyzing the hidden costs of computational power
Beyond carbon: Water, hardware, and waste
The shift from training to inference impact
Operational carbon splits into training and inference emissions
Training a model like GPT-3 emitted over 550 metric tons of CO2e2, 3
Newer models likely exceed these figures by orders of magnitude
Why is transparency regarding training data becoming a rarity?
How does the aggregate impact of billions of daily queries change our sustainability strategy?
According to the lecture, how much of a deployed model's lifecycle emissions comes from inference?
Choose an answer to continue
Think about why the lecture calls inference "the sleeping giant."
The lecture says inference constitutes the majority of lifecycle emissions for deployed models, not a small fraction.
Correct. Inference constitutes the majority of lifecycle emissions for deployed models.
Week 08 — Inference: The Sleeping Giant
Operational carbon splits into training and inference emissions, so inference does produce emissions.
Correct. Inference constitutes the majority of lifecycle emissions for deployed models.
Week 08 — Inference: The Sleeping Giant
Correct. Inference constitutes the majority of lifecycle emissions for deployed models.
Week 08 — Inference: The Sleeping Giant
The lecture does not describe an even split; it says inference constitutes the majority of lifecycle emissions.
Correct. Inference constitutes the majority of lifecycle emissions for deployed models.
Week 08 — Inference: The Sleeping Giant
Median Gemini App text prompt consumes 0.24 Wh of energy4, 5
Emissions per prompt are roughly 0.03 grams of CO2e4, 5
Comparable to running a 60W light bulb for about 14 seconds
Is per-query efficiency enough to offset the explosive growth in total usage?
Data centers require massive cooling systems to maintain operation
What are the ethical implications of data centers in water-scarce regions?
According to the lecture, why do data centers consume so much water?
Choose an answer to continue
Think about what a data center needs in order to keep running.
Correct. Data centers require massive cooling systems to maintain operation.
Week 08 — Water Consumption: A Critical Metric
The water use the lecture describes comes from the massive cooling systems data centers require to maintain operation.
Correct. Data centers require massive cooling systems to maintain operation.
Week 08 — Water Consumption: A Critical Metric
Inference consumes energy, measured in Wh per prompt; the water use comes from the cooling systems data centers need.
Correct. Data centers require massive cooling systems to maintain operation.
Week 08 — Water Consumption: A Critical Metric
The lecture raises water-scarce regions as an ethical question; the consumption itself comes from the cooling systems data centers require.
Correct. Data centers require massive cooling systems to maintain operation.
Week 08 — Water Consumption: A Critical Metric
Studies debate the metric of "bottles of water per conversation"7, 8
Google reports 0.26 mL (five drops) per median query due to efficiency4, 5
Efficiency gains fight against the sheer scale of deployment
How do we effectively communicate water impact to end-users?
"Embodied footprint" includes extraction, manufacturing, and disposal
Manufacturing AI accelerators is chemically intensive and energy-demanding11, 12
2025 LCA shows efficiency gains can improve Compute Carbon Intensity by 3x11
Should hardware longevity be prioritized over peak performance?
AI hardware becomes obsolete faster than general-purpose servers13, 15
Global e-waste reached 62 million tonnes in 202216
Specialized chips contain hazardous materials and rare earth elements16, 17
What policy mechanisms could enforce better recycling rates for AI hardware?
Shift toward liquid cooling: Direct-to-chip or immersion cooling5
DeepMind-style optimization reduces cooling energy by up to 40%18
Balancing Water Usage Effectiveness (WUE) against Power Usage Effectiveness (PUE)
Can data centers ever become truly "net-positive" for their local environments?
According to the lecture, which two measures must data centers balance against each other?
Choose an answer to continue
Think about the two resources that a data center's cooling draws on.
Compute Carbon Intensity and the embodied footprint came up under the hardware lifecycle; the balance named for data centers is WUE against PUE.
Correct. Data centers balance Water Usage Effectiveness (WUE) against Power Usage Effectiveness (PUE).
Week 08 — Data Center Innovation
Those are the two forms of liquid cooling data centers are shifting toward, not two measures to balance.
Correct. Data centers balance Water Usage Effectiveness (WUE) against Power Usage Effectiveness (PUE).
Week 08 — Data Center Innovation
Those are the two parts operational carbon splits into, not the balance named for data centers.
Correct. Data centers balance Water Usage Effectiveness (WUE) against Power Usage Effectiveness (PUE).
Week 08 — Data Center Innovation
Correct. Data centers balance Water Usage Effectiveness (WUE) against Power Usage Effectiveness (PUE).
Week 08 — Data Center Innovation
Leveraging pattern recognition for the planet
Weather forecasting and disaster warning
Optimizing renewable energy systems
AI outperforms traditional Numerical Weather Prediction (NWP) in speed and efficiency
GraphCast: Predicts weather 10 days in advance with high accuracy19
Operates at 0.25-degree resolution using Graph Neural Networks19
How does faster disaster prediction translate to tangible lives saved?
According to the lecture, what does GraphCast do?
Choose an answer to continue
GraphCast is the AI model the lecture pairs with Graph Neural Networks. Think about how far ahead it forecasts.
GraphCast predicts weather 10 days in advance with high accuracy.
Correct. GraphCast predicts weather 10 days in advance with high accuracy.
Week 08 — Revolutionizing Weather Prediction
GraphCast is an AI model; the lecture says AI outperforms traditional Numerical Weather Prediction (NWP) in speed and efficiency.
Correct. GraphCast predicts weather 10 days in advance with high accuracy.
Week 08 — Revolutionizing Weather Prediction
Correct. GraphCast predicts weather 10 days in advance with high accuracy.
Week 08 — Revolutionizing Weather Prediction
GraphCast operates at 0.25-degree resolution using Graph Neural Networks.
Correct. GraphCast predicts weather 10 days in advance with high accuracy.
Week 08 — Revolutionizing Weather Prediction
NVIDIA's FourCastNet v3 uses Spherical Fourier Neural Operators20, 21
Enables rapid ensemble forecasting to predict extreme weather probabilities21
Democratizes forecasting: Runs on a few GPUs instead of supercomputers22
What new applications emerge when weather forecasting becomes accessible to smaller organizations?
Integrating variable sources like wind and solar into the grid
AI predicts solar irradiance and wind speeds with high precision23
Reinforcement Learning manages Hybrid Energy Storage Systems (HESS)23, 24
Can AI be the key factor that allows grids to run on 100% renewable energy?
According to the lecture, what manages Hybrid Energy Storage Systems (HESS)?
Choose an answer to continue
Think of an approach that learns from the results of its own decisions, such as when to store or release energy.
NVIDIA's FourCastNet v3 uses Spherical Fourier Neural Operators for weather forecasting, not for managing energy storage.
Correct. Reinforcement Learning manages Hybrid Energy Storage Systems (HESS).
Week 08 — Optimizing Renewable Energy
Correct. Reinforcement Learning manages Hybrid Energy Storage Systems (HESS).
Week 08 — Optimizing Renewable Energy
GraphCast uses Graph Neural Networks for weather prediction, not for managing energy storage.
Correct. Reinforcement Learning manages Hybrid Energy Storage Systems (HESS).
Week 08 — Optimizing Renewable Energy
Numerical Weather Prediction is the traditional forecasting approach AI outperforms, not a way of managing energy storage.
Correct. Reinforcement Learning manages Hybrid Energy Storage Systems (HESS).
Week 08 — Optimizing Renewable Energy
Maximizing yield while minimizing chemical inputs and water usage
Analysis of satellite imagery and soil sensors for Variable Rate Technology26, 28
Smart irrigation systems optimize water delivery in real-time27
How can we ensure smallholder farmers have access to these advanced tools?
Prioritizing efficiency alongside accuracy
Architectural innovations and model compression
The rise of Small Language Models
Red AI: Buying performance with massive computational cost29, 30
Green AI: Treating carbon efficiency as a primary evaluation metric
Decoupling AI progress from exponential resource consumption
What cultural shifts in research are needed to value efficiency as much as accuracy?
According to the lecture, what characterizes Green AI?
Choose an answer to continue
Red AI buys performance with massive computational cost. Green AI changes what a model is judged on.
That describes Red AI, not Green AI.
Correct. Green AI treats carbon efficiency as a primary evaluation metric.
Week 08 — Red AI vs. Green AI
The lecture's direction is the opposite: decoupling AI progress from exponential resource consumption.
Correct. Green AI treats carbon efficiency as a primary evaluation metric.
Week 08 — Red AI vs. Green AI
Green AI treats carbon efficiency as a primary evaluation metric, and the lecture asks how research can value efficiency as much as accuracy.
Correct. Green AI treats carbon efficiency as a primary evaluation metric.
Week 08 — Red AI vs. Green AI
Correct. Green AI treats carbon efficiency as a primary evaluation metric.
Week 08 — Red AI vs. Green AI
Quantization: Reducing precision (e.g., 32-bit to 4-bit) to save energy
Pruning: Removing redundant weights to create sparse models
Knowledge Distillation: Training small student models from large teachers31, 32
Is there a "minimum viable precision" for most business applications?
According to the lecture, what is knowledge distillation?
Choose an answer to continue
The lecture describes it with a teacher and a student. Think about which one is small.
That is quantization.
Correct. Knowledge distillation trains small student models from large teachers.
Week 08 — Model Compression Techniques
That is pruning.
Correct. Knowledge distillation trains small student models from large teachers.
Week 08 — Model Compression Techniques
Correct. Knowledge distillation trains small student models from large teachers.
Week 08 — Model Compression Techniques
This reverses the lecture: small student models are trained from large teachers.
Correct. Knowledge distillation trains small student models from large teachers.
Week 08 — Model Compression Techniques
The "Small is Sufficient" trend: Using task-specific models34
Adoption could save roughly 28% of global AI electricity by 202534
Avoiding the use of massive generalist models for simple queries
Why do organizations still default to the largest available models?
Hardware inspired by the human brain's architecture
Spiking Neural Networks (SNNs): Neurons only consume energy when active35, 36
Eliminates the energy cost of moving data between memory and processors
How close are we to seeing neuromorphic chips in consumer devices?
According to the lecture, when do neurons in Spiking Neural Networks (SNNs) consume energy?
Choose an answer to continue
Neuromorphic hardware is inspired by the human brain's architecture. Think about what a "spike" is.
In Spiking Neural Networks, neurons only consume energy when active.
Correct. In Spiking Neural Networks (SNNs), neurons only consume energy when active.
Week 08 — Neuromorphic Computing
Correct. In Spiking Neural Networks (SNNs), neurons only consume energy when active.
Week 08 — Neuromorphic Computing
Neuromorphic computing eliminates the energy cost of moving data between memory and processors.
Correct. In Spiking Neural Networks (SNNs), neurons only consume energy when active.
Week 08 — Neuromorphic Computing
The lecture ties SNN energy use to neuron activity: neurons only consume energy when active.
Correct. In Spiking Neural Networks (SNNs), neurons only consume energy when active.
Week 08 — Neuromorphic Computing
Intel Loihi 2 and BrainChip Akida showing commercial viability36, 37
Potential for 100x to 1000x efficiency gains in specific tasks36, 38
Ideal for edge AI and sensory processing applications
Will neuromorphic computing replace or augment traditional GPU architectures?
Closing the loop on waste
Biodiversity monitoring
The emerging regulatory landscape
Treating waste as a resource through better sorting and logistics
AI robotics sort waste streams with high speed and accuracy40, 41
Digital Waste Passports track recovery of valuable materials42
How can AI incentivize consumers to participate more effectively in recycling?
Passive Acoustic Monitoring (PAM) tracks species via soundscapes44, 45
Computer vision identifies species in real-time via camera traps44, 46
Non-invasive methods for tracking elusive or endangered species
Can global biodiversity data be standardized to drive international policy?
Mandates transparency regarding "AI systems resource performance"48, 49
General-Purpose AI providers must publish energy consumption data50, 51
Applies to models trained with more than 10^23 FLOPs50
Will the EU's regulations become the de facto global standard for AI sustainability?
Under the EU AI Act, what must General-Purpose AI providers publish, according to the lecture?
Choose an answer to continue
Article 40 mandates transparency regarding "AI systems resource performance."
Correct. General-Purpose AI providers must publish energy consumption data.
Week 08 — The EU AI Act: Article 40
The obligation the lecture names is publishing energy consumption data.
Correct. General-Purpose AI providers must publish energy consumption data.
Week 08 — The EU AI Act: Article 40
10^23 FLOPs is the threshold above which the requirement applies, not a limit models must stay below.
Correct. General-Purpose AI providers must publish energy consumption data.
Week 08 — The EU AI Act: Article 40
The obligation the lecture names is publishing energy consumption data.
Correct. General-Purpose AI providers must publish energy consumption data.
Week 08 — The EU AI Act: Article 40
ISO/IEC 42001: Management system standard including sustainability53
ISO/IEC TR 20226: Focuses on energy, water, and e-waste aspects54, 55
Moving from voluntary "greenwashing" to rigorous compliance
How can organizations prepare for these upcoming reporting requirements?
Software tools like CodeCarbon and Green Algorithms estimate emissions3, 56
Compute Carbon Intensity (CCI) quantifies carbon per unit of work11
Tools may underestimate true usage by missing overheads like cooling56
What is the margin of error we should accept in carbon accounting tools?
According to the lecture, why may software tools like CodeCarbon underestimate true usage?
Choose an answer to continue
Think about the energy a data center uses beyond the computation itself.
Correct. Tools may underestimate true usage by missing overheads like cooling.
Week 08 — Measurement Tools
These tools estimate emissions; the lecture says they may underestimate by missing overheads like cooling.
Correct. Tools may underestimate true usage by missing overheads like cooling.
Week 08 — Measurement Tools
Compute Carbon Intensity quantifies carbon per unit of work; the underestimate comes from missing overheads like cooling.
Correct. Tools may underestimate true usage by missing overheads like cooling.
Week 08 — Measurement Tools
The lecture gives a different reason: the tools may miss overheads like cooling.
Correct. Tools may underestimate true usage by missing overheads like cooling.
Week 08 — Measurement Tools
Integrate carbon tracking into CI/CD pipelines
Default to Small Language Models (SLMs) where possible
Schedule training runs for times with low grid carbon intensity
How can individual developers influence organizational choice of models?
Require Life Cycle Assessments (LCA) from hardware vendors
Prioritize data centers with water-free or recycled water cooling
Align reporting with EU AI Act and ISO standards immediately
What are the risks of ignoring the environmental component of ESG goals?
Standardizing water reporting to prevent "water washing"
Hardware-software co-design for neuromorphic architectures
Mitigating Jevons Paradox: Efficiency leading to increased total consumption
Can we innovate fast enough to outpace the environmental damage of scaling?
According to the lecture, what does Jevons Paradox describe?
Choose an answer to continue
Think about what can happen to total use when each use becomes more efficient.
This reverses the lecture: Jevons Paradox is efficiency leading to increased total consumption.
Correct. Jevons Paradox is efficiency leading to increased total consumption.
Week 08 — Future Directions
That is closer to "water washing", which standardizing water reporting aims to prevent.
Correct. Jevons Paradox is efficiency leading to increased total consumption.
Week 08 — Future Directions
That is hardware-software co-design, a separate future direction.
Correct. Jevons Paradox is efficiency leading to increased total consumption.
Week 08 — Future Directions
Correct. Jevons Paradox is efficiency leading to increased total consumption.
Week 08 — Future Directions
AI is a powerful tool with a heavy environmental price tag
The shift to inference dominance demands new efficiency strategies
Regulation and standards are transforming the industry
Sustainable AI is not just an option, it is a license to operate.