29 comments

[ 2.7 ms ] story [ 236 ms ] thread
Does this account for the fact that AI cannot write or draw without relying on human written or drawn samples?
Could an “untrained” human being write or draw, alone, without samples?
(comment deleted)
You could say that drawing and writing occurred spontaneously several times.
Those samples already existed and thus don't cost anything.
(comment deleted)
Did you learn to write or draw without studying the samples of artists and authors before you?
Kids one hundred percent learn to draw without sampling from other authors.
Do kids draw in a private room, with no other kids around, and are blindfolded from birth until one day, they're given crayons and paper? Do they also not see pictures in books read to them by adults? They may not have taken an art history class and ingested an Internet's worth of Artstation and Greg Rutkowski, but they have plenty of outside influence.

Of course, their limited motor skills get in the way, but they're working on it.

The most stereotypical kid drawing is of themselves in a house with their parents. They do not need to sample other drawings to get to that.

Regardless, if we go back a few centuries, humans definitely demonstrated their ability to physically draw on paper, without sampling from sources other than their eyes.

But even if we were to consider that human eyes are effectively a way to “collect” samples, that’s another thing not taken into account in the aforementioned article.

(comment deleted)
With this study's hack back-of-the-envelop math counting "human being alive while creating" as part of their calculation, AI should have to include the entire carbon output of any artist or writer that exists in it's training data for the time they were creating that art.
Better study how much CO2 does a 'functional' LLM training cycle emits..
Well, I would say most humans will keep breathing even if you give the tasks to an AI. From that perspective AI will emit more CO2 not less.
Only those who are still alive. Sarcasm? Or business optimization?
yes i wonder how much energy is used each time gpt4 answers something.

anyone know?

Just a quick glance shows that they consider the overall emissions of a persons day to day life to be involved in the act of writing:

"The emission footprint of a US resident is approximately 15 metric tons CO2e per year (14), or approximately 1.7kg CO2e per hour. Therefore, assuming that a person’s emissions while writing are in line with their overall annual impact, we propose that the carbon footprint for a US resident generating a page of text (250 words) is approximately 1400 grams CO2e."

That's a pretty big assumption, which seems hardly scientific. You would think that they would isolate the act of writing and measure that against the CO2 emissions of equivalent AI.

They should also measure the CO2 emissions of a human at resting state vs. a human engaged in the act of writing. This would reveal the exact 'cost' of human writing.
They should also measure the CO2 emissions of all the humans who wrote all the words that the AI was trained on.
> That's a pretty big assumption, which seems hardly scientific.

It's hard not to point out that the authors aren't scholars in this field, but essentially AI apologists.

(comment deleted)
This feels very much like crypto. Same sort of absurd "justifications".
I only skimmed the article, so may have missed something, but it doesn't seem that they are taking into account how much CO2 emission humans emit just for living.

If I spend an hour writing my emissions will be about the same as if I had instead spent that hour reading, or watching TV, or playing my guitar, or doing most other things I do.

Computers can be turned off when not working so only need to cause CO2 emission when actually working. Humans cannot be turned off between jobs.

Like the cyanobacteria before us which killed off most previous life by overloading this planet with oxygen, maybe we're just prepping the world for the next forms of life which will live on CO2.