The Sky Sorter team reads the board's feedback out loud.
"Add dim comets to training. Post a transparency sign. Name who is responsible," Rocket reads. "That is a lot."
Raven is already writing. "Feedback is a gift. It tells us exactly where to look."
Nova adds one more line to the list. "And think about energy," she says.
"Big generative models take a lot of computing to train and run. Even our design choices have a footprint."
Rocket taps his pencil. "So the brief needs a job, a plan, the fixes and the footprint. Let's go."
Training, maintaining and running generative AI takes a lot of computing.
That can mean large energy and environmental footprints.
In one study, generating text, like writing summaries, used more energy than sorting text into labels.
Sky Sorter is a sorting tool, not a generator. It also runs on paper, so its footprint is tiny.
Designers can ask: does this job really need a big generative model?
| Statement | True or false? |
|---|---|
| Training and running generative AI can use a lot of energy. | ? |
| In one study, generating summaries used less energy than labeling text. | ? |
| Designers can think about energy when they choose what to build. | ? |
Good designers seek feedback, then change their design to meet users' needs.
Then they test again to check that the change helped.
| Section | What to write |
|---|---|
| Job | What the system is for, in one sentence. |
| Users | Who will use it and who it affects. |
| Data | What training data it needs and who supplies it. |
| Test | How you will test it on new, held-back cards. |
| Bias fix | What bias you found and how you will manage it. |
| Privacy plan | What personal information you remove or protect. |
| Transparency | What users can find out, and who is responsible. |
| Human check | When a person reviews the output. |
| Footprint | Why this design does not use more computing than it needs. |
A strong brief. Tomorrow you look back across the whole course before you present it.