Environmental impact of AI in Prompt Engineering / GenAI - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding AI Energy Consumption

Which factor most directly increases the energy consumption of training a large AI model?

A Increasing the number of training data samples

B Using a simpler model architecture

C Reducing the number of training epochs

D Switching from GPU to CPU for training

💡 Hint

Think about what makes the computer work harder during training.

More training data means the model processes more information, which takes more energy. Simpler models or fewer epochs reduce energy use. CPUs are usually less efficient than GPUs for AI training, so switching to CPU often increases energy use.

❓ Metrics

intermediate

Measuring Carbon Footprint of AI Training

If training a model uses 500 kWh of electricity and the local grid emits 0.4 kg CO2 per kWh, what is the total CO2 emission from training?

A 0.8 kg CO2

B 1250 kg CO2

C 200 kg CO2

D 500 kg CO2

💡 Hint

Multiply energy used by emission per unit energy.

Total CO2 = Energy (kWh) × Emission factor (kg CO2/kWh) = 500 × 0.4 = 200 kg CO2.

❓ Model Choice

advanced

Choosing Models to Reduce Environmental Impact

Which model choice best reduces environmental impact while maintaining reasonable accuracy?

A A very large transformer model trained from scratch on massive data

B Using a large model with many training epochs without early stopping

C Training multiple large models and averaging their predictions

D A smaller pre-trained model fine-tuned on specific data

💡 Hint

Fine-tuning smaller models uses less energy than training large models from zero.

Fine-tuning a smaller pre-trained model requires less energy and time than training a large model from scratch, reducing environmental impact while keeping good accuracy.

🔧 Debug

advanced

Identifying Energy Waste in AI Training Code

What is the main issue causing unnecessary energy use in this training loop? ```python for epoch in range(100): train(model, data) if epoch % 10 == 0: evaluate(model, val_data) # No early stopping or checkpoint saving ```

A Lack of early stopping causes training to continue even if performance stops improving

B The data is not shuffled before training, causing poor model performance

C The training loop uses too few epochs, causing inefficient learning

D The model is evaluated too frequently, wasting energy

💡 Hint

Stopping training early when the model stops improving saves energy.

Without early stopping, the model trains for all 100 epochs even if it stops improving early, wasting energy.

🧠 Conceptual

expert

Evaluating Environmental Impact Trade-offs

Which strategy best balances AI model performance and environmental impact for a company deploying AI services globally?

A Train multiple large models for each region to maximize accuracy locally

B Use smaller, optimized models deployed closer to users with edge computing

C Train large models centrally and serve predictions globally without optimization

D Continuously retrain large models daily regardless of performance changes

💡 Hint

Deploying smaller models near users reduces data transfer and energy use.

Deploying optimized smaller models near users (edge computing) reduces energy from data transfer and large centralized servers, balancing performance and environmental impact.