Choosing a model is also a responsibility decision. Bigger models use more energy and hardware every time they run, so picking a model that is larger than the job needs wastes resources and adds to carbon emissions. AWS's guidance is to ask whether AI is needed at all, choose the smallest model that does the job well, reuse existing models instead of training new ones, and run them on infrastructure that only uses capacity when there's work to do. The exam frames this as environmental considerations and sustainability.
Ask Whether Generative AI Is Needed
AWS's first sustainability principle for generative AI is to check whether it's the right tool. If a keyword search, a rule or a small traditional model can do the job, it will usually use far less computing power than a large generative model. Generative AI is for creating new content; it's an expensive way to look something up.
Right-Size the Model
Larger models aren't automatically better for a given task. For a narrow job, such as sorting support tickets into a few categories, a smaller model often reaches the accuracy target with much less computing per request. AWS recommends comparing candidate models on the task, then choosing the smallest one that meets the quality bar. Bedrock model evaluations help with that comparison.
Several techniques make a capable model cheaper to run:
- Distillation trains a smaller "student" model to copy a larger "teacher" model's answers on your task. Amazon Bedrock offers this as Model Distillation.
- Quantization stores the model's numbers with less precision, which shrinks it and speeds it up with little loss of quality.
- Fine-tuning a smaller model for a specific task can match a large general model on that task.
Reuse Instead of Retrain
Training a foundation model from scratch takes huge amounts of data, time and energy. Starting from an existing pre-trained model, and adapting it only as much as needed with prompts, retrieval or light fine-tuning, avoids repeating that cost. Reuse is usually both the cheaper and the more sustainable choice.
When customization is needed, the lightest method that works is the responsible one. Better prompts or retrieval add almost no extra computing. Fine-tuning a smaller model costs some training energy once. Training a new model from scratch should be a last resort, kept for cases where no existing model can do the job.
Run It Efficiently
How a model is hosted matters as much as which model it is:
- Serverless and auto scaling inference uses capacity only when requests arrive, instead of keeping servers running while idle.
- Managed services, such as Amazon Bedrock, share well-utilized infrastructure across many customers.
- Purpose-built chips, such as AWS Inferentia for inference, are designed to deliver more work per watt than comparable general instances.
On the Exam
When a question mentions sustainability, environmental impact or carbon footprint, the answer usually involves doing less computing: a smaller or distilled model, reusing a pre-trained model, serverless or auto scaling hosting, or efficient chips. Options that train a new model from scratch, pick the largest model "to be safe", or keep dedicated instances running around the clock are the traps.