← Back to course
AI and You 9-12 / Week 10 / Tuesday
2/6
Week 10 · Footprint and Supply Chain

Tuesday

Making versus labeling
// What SH-1 depends on, and who can use it
⏱ about 20 min

Tuesday: Making Versus Labeling

Comet spreads the three SH-1 features across the library table: Ask, Quiz and Check.

"They all live inside SH-1," she says, "so they must all use the same energy. Easy. R8 is Low."

Wren reads the builder's printouts again. "What does the evidence say? Quiz writes whole quizzes. Check only marks an answer right or wrong."

"So are those the same job?" Comet asks, frowning at the cards.

Nova sorts the cards into two glowing columns, then leaves them unlabeled. "Would you like a hint?" she asks. "Look at what each feature hands back."

Wren slides his pencil across to you. "Reviewer, which features make something new, and which only put a label on something?"

Making and labeling

Some AI tasks make something new, such as a summary or a quiz. Other tasks sort or label things, such as marking a sentence as one type or another.

In one study, making things such as summaries used more energy and produced more carbon than sorting things such as labeling text.

One study is not a rule for every system. It does not tell us SH-1's numbers.

SH-1 featureWhat it hands backMaking or labeling?
AskA new written answer to a questionMaking new text
QuizA new practice quiz written from notesMaking new text
CheckA mark: right or wrongLabeling
MAKING OR LABELING?
  • Read the question.
  • Tap your answer.
SH-1 Quiz writes a new practice quiz from Biology notes. Making or labeling?
SH-1 Check marks a student's answer as wrong. Making or labeling?
SH-1 Ask writes a paragraph answer to a history question. Making or labeling?
In the one study, which kind of task used more energy?

Smaller models, and no agreed method

Smaller versions of trained models can lower the impact of using them. Making those smaller versions still has an impact.

There is no agreed way yet to estimate generative AI's environmental impacts.

The SH-1 packet gives no energy information at all. So the crew has nothing to measure SH-1 with.

Rating R8

Remember: risk combines how likely an event is with how big its consequences would be.

The crew cannot rate either one for R8. They have no SH-1 energy information, and there is no agreed method to estimate it.

So they rate R8 "Unknown" for likelihood and "Unknown" for size of harm. Then they write to the builder to ask for information.

An honest "Unknown" is better than a guess. Comet's "Low" was a guess.

RowRiskSH-1 exampleLikelihoodSize of harm
R8Environmental impactsEvery request runs on computing, and most make new textUnknownUnknown
The crew's note to the builder (made up)
Please share what you know about the energy SH-1 uses for training and for each kind of request.
Please tell us whether SH-1 uses a smaller version of its base model.
If you do not know, please say so. That answer helps us too.
StatementTrue or false?
In one study, making things used more energy than labeling things.?
That one study gives the exact energy SH-1 uses.?
Smaller versions of trained models can lower the impact of using them.?
Making a smaller model has no impact at all.?
There is an agreed method to estimate generative AI's environmental impacts.?
WHY THIS EXERCISEKnowing what a study does and does not show keeps the review honest.
WHAT THE CREW DOES WHEN IT CANNOT MEASURE
  • ?Rate R8 again if the builder answers.
  • ?Find that the packet gives no energy information.
  • ?Write to the builder to ask for information.
  • ?Rate R8 Unknown for likelihood and size of harm.
  • ?Look for evidence about SH-1's energy use.
WHY THIS EXERCISESaying "Unknown" out loud turns a gap into a question someone can answer.
With no information and no agreed method, the crew rates R8 ____.
Check marks answers right or wrong. That kind of task is called ____.

Careful thinking, reviewer. Tomorrow in the Explorer Lab you tally a whole week of SH-1 requests.

← Monday