Comet bursts into the library waving page two of the SH-1 packet. "Section P2 says SH-1 is built on a large model from another company," she reads.
"So SH-1 just lives on a screen," she adds. "No trucks, no smoke. It has no footprint at all!"
Wren taps his clipboard. "What does the evidence say? Something has to make every answer SH-1 writes."
Librarian Joss looks up from a cart of returned books. "The panel wants to know that too. What does SH-1 run on?"
Nova projects a long line from a screen to a box marked with a question mark. "Would you like a hint?" she asks. "Follow one answer back to where it is made."
Comet turns to you. "Reviewer, what does it take to run SH-1?"
SH-1, the builder and Harbor Point High School are a made-up example. The ideas you use to review SH-1 are real.
This week the crew rates two more rows of the risk register: R8, environmental impacts, and R9, value chain and component integration.
They also add an access note to the SH-1 model card.
SH-1 is generative AI. Generative AI makes new content, such as text, that imitates the data it learned from.
Training, maintaining and running generative AI take a lot of computing resources.
These activities can have large energy and environmental footprints. In this course, a footprint is the mark an activity leaves on the planet.
So "no trucks, no smoke" does not mean no footprint. Running SH-1 takes computing, even when all you see is a screen.
| Stage | What it means | Who does it in the SH-1 story (made up) |
|---|---|---|
| Training | Teaching a model from huge amounts of data | The base model maker, then the builder |
| Maintaining | Keeping the system working and up to date | The builder |
| Running (use) | Making an answer each time someone asks | SH-1, every time a student or teacher uses it |
NIST says the energy generative AI uses depends on four things.
They are the stage (training or use), the kind of content, the hardware and the task.
So two requests to SH-1 may not use the same energy. A long quiz is a different task from a one-word label.
NIST draws the AI lifecycle as a picture. At its center is "People & Planet": human rights and the well-being of society and the planet.
The crew reads that center as a reminder. They will ask how SH-1 affects people and the planet, not only how well it answers.
| Statement | True or false? |
|---|---|
| Running generative AI takes computing resources. | ? |
| Only the training stage of generative AI uses energy. | ? |
| Generative AI can have large energy and environmental footprints. | ? |
| Energy use depends on the stage, the content, the hardware and the task. | ? |
| NIST puts People & Planet at the center of its lifecycle picture. | ? |
| SH-1 and Harbor Point are real. | ? |
Good tracing, reviewer. Tomorrow the crew compares the tasks SH-1 does and rates R8.